Process anomaly real-time monitoring method and system for quality control

Abstract

The application relates to the technical field of industrial anomaly monitoring, and discloses a process anomaly real-time monitoring method and system for quality management. The method extracts spatial coordination features of data streams synchronously collected by multiple-source isomorphic sensors to obtain a spatial feature vector; determines a working condition state according to the spatial feature vector, extracts time sequence features, calculates a state deviation value of the time sequence features deviating from a historical distribution model, divides the working condition state into stages according to the state deviation value, and determines a process stage; acquires an abnormal boundary corresponding to the process stage, extracts upper and lower limit extreme values to obtain an extreme value interval; when a real-time value exceeds the extreme value interval, calculates a deviation amplitude according to a difference value, carries out abnormal scoring to obtain an abnormal score, further determines a warning level, generates a warning label, carries out rule matching, and outputs a graded warning instruction. The application realizes linkage of spatial coordination features, staged abnormal boundaries and graded warning, and improves the graded response efficiency of process anomalies.

Description

Technical Field

[0001] This invention relates to the field of industrial anomaly monitoring technology, specifically to a method and system for real-time monitoring of process anomalies for quality control. Background Technology

[0002] Quality control in industrial control systems relies on real-time monitoring of process parameters to ensure product qualification rates. Process data collected synchronously by multi-source isomorphic sensors at different spatial locations is an important basis for judging the process status.

[0003] In existing technologies, process anomaly monitoring typically involves independently analyzing the data streams of each sensor and judging whether the real-time values ​​of individual sensors exceed limits based on pre-set fixed thresholds. However, this method of independent analysis and fixed threshold judgment ignores the collaborative characteristics between data streams of similar equipment in different spatial locations. Furthermore, as production conditions continue to evolve, the distribution characteristics of historical data deviate significantly from the current production status. Fixed thresholds are difficult to adapt to the normal fluctuation range that changes with the process stage, leading to misjudgments of normal fluctuations and generating a large number of invalid alarms. This results in a sluggish response to global anomalies, making it impossible for operators in the control center to distinguish the severity of alarms, severely impacting the efficiency and accuracy of anomaly response.

[0004] Therefore, existing technologies have the technical problem of being unable to effectively distinguish and classify the severity of process anomaly signals, making it difficult for operators in the control center to distinguish alarm priorities and resulting in low efficiency in responding to process anomalies. Summary of the Invention

[0005] This invention provides a method and system for real-time monitoring of process anomalies for quality control, in order to solve the technical problem that the existing technology cannot effectively distinguish and classify the severity of process anomaly signals, making it difficult for operators in the control center to distinguish alarm priorities and resulting in low efficiency in responding to process anomalies.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for real-time monitoring of process anomalies for quality control, comprising:

[0007] Acquire the data stream synchronously collected by multiple isomorphic sensors, extract the spatial collaborative features of the data stream, and obtain the spatial feature vector;

[0008] The working condition is determined based on the spatial feature vector, the time series features of the working condition are extracted to obtain the time series features, and the state deviation value of the time series features from the preset historical distribution model is calculated.

[0009] The operating conditions are divided into stages based on the state deviation values ​​to determine the process stages;

[0010] Obtain the abnormal boundary corresponding to the process stage, extract the upper and lower limit extreme values ​​of the abnormal boundary, and obtain the extreme value range;

[0011] When the real-time value in the data stream exceeds the extreme value range, the deviation range is calculated based on the difference between the real-time value and the boundary of the extreme value range, and an anomaly score is obtained.

[0012] The warning level is determined based on the anomaly score, a warning label is generated, and rule matching is performed based on the warning label to output a graded warning instruction.

[0013] In a second aspect, the present invention provides a real-time process anomaly monitoring system for quality control, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the real-time process anomaly monitoring method for quality control as described above.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for real-time monitoring of process anomalies for quality control as described in any of the preceding claims.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) This invention extracts the spatial collaborative features of the data streams synchronously collected by multi-source isomorphic sensors at different spatial locations to obtain feature vectors, and condenses the collaborative change relationship between different measuring points into low-dimensional features, overcoming the shortcomings of existing technologies that analyze the data streams of each sensor independently and ignore the spatial collaborative relationship. Since the off-diagonal elements of the spatial covariance matrix quantify the correlation between different measuring points, after eigenvalue decomposition and principal component extraction, global collaborative information is retained while suppressing local noise, thereby improving the sensitivity of global anomaly identification and reducing invalid alarms caused by small fluctuations in local data.

[0017] (2) This invention obtains time-series features by extracting time-series features of operating conditions, and calculates the state deviation value of the time-series features from the preset historical distribution model. Based on the state deviation value, the operating conditions are divided into stages to obtain the process stages. Since the historical distribution model describes the distribution range of time-series features under normal operating conditions, the stage division triggered by the degree of deviation makes the monitoring benchmark dynamically switch with the process stage, which overcomes the defect that the fixed threshold cannot adapt to the continuous evolution of operating conditions and avoids normal fluctuations being misjudged as abnormal when the process stage is switched.

[0018] (3) This invention obtains the extreme value range by acquiring the abnormal boundary corresponding to the process stage and extracting the upper and lower extreme values ​​of the abnormal boundary. When the real-time value exceeds the extreme value range, the deviation is calculated based on the difference between the real-time value and the boundary of the extreme value range, and an abnormal score is obtained. Based on this, the warning level is determined, a warning label is generated, and a graded warning instruction is output. Since the abnormal boundary is obtained by envelope calculation and projection of the historical operating trajectory of the corresponding process stage, the extreme value range is dynamically adapted with the process stage. Combined with the abnormal score and graded warning, the monitoring system can not only define the abnormal boundary that changes with the process stage, but also respond to the abnormal signal in a graded manner according to the degree of risk, thereby improving the accuracy and response efficiency of abnormal monitoring. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the real-time monitoring method for process anomalies in quality control provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a real-time process anomaly monitoring system for quality control provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some embodiments of the present invention, and other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0022] Reference Figure 1 This invention provides a method for real-time monitoring of process anomalies for quality control, comprising the following steps:

[0023] S11, acquire the data stream synchronously collected by the multi-source isomorphic sensors, extract the spatial collaborative features of the data stream, and obtain the spatial feature vector;

[0024] S12, determine the working condition state based on the spatial feature vector, extract the time series features of the working condition state to obtain the time series features, and calculate the state deviation value of the time series features from the preset historical distribution model.

[0025] S13, Divide the operating conditions into stages according to the state deviation value to determine the process stage;

[0026] S14, obtain the abnormal boundary corresponding to the process stage, extract the upper and lower limit extreme values ​​of the abnormal boundary, and obtain the extreme value range;

[0027] S15, when the real-time value in the data stream exceeds the extreme value range, the deviation range is calculated based on the difference between the real-time value and the boundary of the extreme value range, and an anomaly score is obtained.

[0028] S16. Determine the warning level based on the abnormal score, generate a warning label, perform rule matching based on the warning label, and output a graded warning instruction.

[0029] This invention is applicable to real-time monitoring of process anomalies in various industrial scenarios. For ease of explanation, this embodiment uses the production quality control scenario of a metal smelting heating furnace as an example, and uses multiple identical temperature sensors arranged at different spatial positions along the furnace as multi-source isomorphic sensors to illustrate the method of this invention.

[0030] In step S11, twenty-four temperature sensors of the same type are arranged along the upper, middle and lower parts of the furnace chamber, and the furnace temperature is collected synchronously within the same sampling window as a data stream.

[0031] In one implementation, spatial co-location features of the data stream are extracted to obtain a spatial feature vector, including:

[0032] Extract the amplitude of the data stream at different spatial locations and construct a spatial distribution matrix;

[0033] The spatial distribution matrix is ​​centered to obtain a centered matrix;

[0034] The spatial covariance matrix is ​​obtained by multiplying the centered matrix with its transpose.

[0035] The spatial covariance matrix is ​​decomposed into eigenvalues ​​to obtain an initial set of eigenvectors;

[0036] Based on the initial feature vector set, the main directional features are selected, and the main directional features are vectorized to obtain spatial feature vectors.

[0037] It should be noted that the data stream is a sequence of temperature signals synchronously collected by multiple temperature sensors of the same model within a preset time window. The temperature amplitude of each temperature sensor at the same sampling moment is extracted to construct a spatial distribution matrix with the number of rows equal to the number of sensors and the number of columns equal to the number of sampling points. Each row represents the time-series data of one temperature sensor, and each column represents the temperature amplitude at different spatial locations in the furnace at the same moment, thereby preserving the spatial layout relationship between the multi-source isomorphic sensors in the furnace. In this embodiment, twenty-four temperature sensors of the same model are arranged along the upper, middle, and lower parts of the furnace to synchronously collect temperature data for a continuous ten-minute period, constructing a spatial distribution matrix of twenty-four rows and sixty columns with one sampling point every ten seconds.

[0038] Furthermore, the spatial distribution matrix is ​​centered to obtain a centered matrix. For each row of the spatial distribution matrix, all data values ​​in that row are summed and then divided by the number of sampling points in that row to obtain the average value for that row. The average value of that row is then subtracted from each data value in that row, and the centered matrix is ​​constructed from the results of processing all rows. The centered processing eliminates static offsets caused by different sensor installation positions and calibration references.

[0039] Next, the spatial covariance matrix is ​​obtained based on the centralization matrix. The diagonal elements of the spatial covariance matrix reflect the temperature fluctuation energy at a single measuring point within the sampling window, while the off-diagonal elements reflect the synchronous heating, synchronous cooling, or local imbalance relationships between different measuring points. The formula for calculating the spatial covariance matrix is ​​as follows: ;

[0040] In the formula, C is the spatial covariance matrix. For a centered matrix, Let n be the transpose of the centered matrix, and n be the number of sampling points. The degrees of freedom for the unbiased estimation of covariance. Through the above calculations, the data stream is no longer processed as a series of isolated single-point temperatures, but is transformed into a matrix characterizing the coordinated changes in the furnace space.

[0041] Further, the spatial covariance matrix is ​​decomposed into eigenvalues ​​to obtain an initial set of eigenvectors. The spatial covariance matrix is ​​a real symmetric matrix, and the characteristic equation Cv = λv is solved using the Jacobi iteration method, where C is the spatial covariance matrix, v is the eigenvector to be solved, and λ is the eigenvalue corresponding to that eigenvector. Solving this characteristic equation yields p eigenvalues ​​equal to the number of effective sensors and p corresponding eigenvectors. All p eigenvectors are arranged in descending order of their corresponding eigenvalues, and the set of these p eigenvectors constitutes the initial set of eigenvectors. The magnitude of the eigenvalues ​​maps to the energy weights of the spatial distribution pattern, and the eigenvectors describe the cooperative direction of temperature changes at different spatial locations.

[0042] It should be noted that, for the selection of principal direction features based on the initial feature vector set, the feature values ​​corresponding to the initial feature vector set are arranged in descending order, and the ratio of the sum of the first k largest feature values ​​to the sum of all feature values ​​is taken as the variance contribution rate. The smallest k value that satisfies the variance contribution rate reaching the preset contribution rate threshold is taken, and the feature vector corresponding to the k largest feature values ​​is taken as the principal direction feature.

[0043] The formula for calculating the variance contribution rate is as follows: .

[0044] In the formula, η is the variance contribution rate. Let be the i-th feature value arranged in descending order, k be the number of features selected in the principal direction, and p be the total number of feature values. Through the above processing, the coordinated changes among multiple temperature measurement points are compressed into a low-dimensional feature vector, which is subsequently used for operating condition identification and abnormal boundary adaptation.

[0045] It should be noted that the preset contribution rate threshold is not arbitrarily set, but determined by statistically analyzing the cumulative variance contribution rates corresponding to different numbers of principal components in historical qualified batches. Specifically, during the calibration phase, the industrial control system selects qualified batch data covering the preheating, heating, holding, and unloading preparation phases. It calculates the eigenvalue sequence for each sampling window using the aforementioned centralization, covariance calculation, and eigenvalue decomposition methods. Then, it calculates the lower quantile of the cumulative variance contribution rate of the top k principal components, which is the value at the 5% level after sorting the cumulative variance contribution rates of all qualified batch samples from smallest to largest. When k is between 3 and 5, if the lower quantile is consistently higher than 90%, 90% is set as the preset contribution rate threshold, and the recommended range of k is written into the model parameter table. During the usage phase, the system truncates at the minimum k value that satisfies the preset contribution rate threshold, ensuring that the feature vector retains the main spatial cooperative patterns while preventing irrelevant noise from entering subsequent operating condition identification.

[0046] It should be noted that in this embodiment, k is taken as 3 to 5 because the main spatial coordination modes of the furnace temperature field are concentrated in the first 3 to 5 principal components. Taking this range can ensure that the cumulative variance contribution rate can stably reach more than 90%, and can also avoid introducing high-order principal components that characterize sensor noise.

[0047] Furthermore, the main directional features are vectorized to obtain feature vectors. The first k feature values ​​are divided by the sum of all feature values, and the resulting k ratios are used as weight coefficients for the corresponding main directional features. Each main directional feature is multiplied by its corresponding weight coefficient, and then concatenated end-to-end in a fixed descending order of feature values ​​to form a spatial feature vector. This concatenation method allows the spatial feature vector to simultaneously represent the cooperative direction and cooperative strength. For dimensionality changes caused by different numbers of effective measurement points, the industrial control system aligns according to the effective measurement point mask in the measurement point configuration table and uses the same effective measurement point mask in the baseline feature vector and the historical distribution model to avoid dimensional inconsistencies caused by temporary sensor downtime.

[0048] In step S12, the operating condition is determined based on the spatial feature vector, and the time series features of the operating condition are extracted to obtain time series features, including:

[0049] Calculate the vector distance between the spatial feature vector and the preset reference feature vector, and determine the working condition based on the vector distance;

[0050] Extract the timestamp information of the working condition status, construct a state time axis, and slice the state time axis to obtain multiple working condition status slices.

[0051] Based on the statistical analysis of the state change relationships between adjacent slices of the multiple operating condition slices, a state transition matrix is ​​obtained;

[0052] Based on the state transition matrix, the time series features of the working condition within the current production cycle are extracted to obtain the time series features.

[0053] It should be noted that the preset benchmark feature vectors are established according to the different operating conditions. The industrial control system extracts feature vectors from multiple sampling windows from qualified product batches, and calculates the feature mean vector and feature covariance matrix for each of the four operating conditions: preheating, heating, heat preservation, and furnace unloading preparation. The feature mean vector is then used as the benchmark feature vector for the corresponding stage. Euclidean distance or Mahalanobis distance can be used to calculate the vector distance. In this embodiment, to reduce the influence of the correlation between principal components on the distance value, Mahalanobis distance is used to calculate the distance between the feature vector and the benchmark feature vector for each stage. The formula for calculating the vector distance is:

[0054] ;

[0055] In the formula, Let f be the distance between the feature vector f and the s-th working condition. Let Σ be the baseline feature vector corresponding to the s-th operating condition. s Let be the feature covariance matrix corresponding to the s-th operating condition. is the inverse matrix of the feature covariance matrix. Among the vector distances between the obtained feature vectors and the baseline feature vectors corresponding to each operating condition, the smallest vector distance is selected, and the operating condition category label corresponding to the smallest vector distance is taken as the operating condition, that is, it represents which category the current production is in: preheating, heating, heat preservation, or furnace unloading preparation. After obtaining the operating condition, the system simultaneously records the current time and binds the recorded timestamp to the operating condition.

[0056] It should be noted that the operating condition status is not an isolated category label, but rather an identification result obtained for each sampling window, representing the production period in which that sampling window is located. The industrial control system takes the category label corresponding to the minimum vector distance in each sampling window as the operating condition status of that sampling window, and binds and stores it with the timestamp of that sampling window and the data collected within that window.

[0057] Next, the timestamp information of the operating conditions is extracted to construct a state timeline. The state timeline is then sliced ​​to obtain multiple operating condition slices. The industrial control system records the start and end timestamps of the operating condition records for each sampling window. The operating conditions obtained within consecutive sampling windows are arranged in time-stamp order. In this embodiment, the ten-minute spatial feature window is updated in five-minute increments, and the current production cycle is divided into multiple operating condition slices. Each operating condition slice is essentially a state sequence fragment composed of operating condition labels from multiple consecutive sampling windows arranged in chronological order, used to reflect the continuous change process of the operating conditions within the current production cycle.

[0058] Then, based on the statistical analysis of the state change relationships between adjacent slices of the multiple operating conditions, a state transition matrix is ​​obtained. The multiple operating condition slices are arranged in chronological order and input as an observation sequence into a Hidden Markov Model (HMM). The number of hidden states in the HMM is the same as the number of operating condition categories; in this embodiment, the number of hidden states is 4, corresponding to the four operating conditions: preheating, heating, heat preservation, and furnace unloading preparation. A forward-backward algorithm is used to iteratively estimate the model parameters, calculate the probability of transitions between each operating condition, and output a 4×4 state transition matrix. The sum of the elements in each row of the state transition matrix is ​​1, and the element value represents the probability of transitioning from the operating condition corresponding to the current row to the operating condition corresponding to the column. Next, the main diagonal elements and the maximum transition probability values ​​on the off-diagonal are extracted from the state transition matrix. The main diagonal elements represent the self-holding probability of each operating condition, and the maximum transition probability value on the off-diagonal represents the most significant state migration trend. These two are concatenated in a fixed order to obtain a temporal feature, which simultaneously represents the persistence and migration trend of the operating conditions within the current production cycle.

[0059] The Hidden Markov Model (HMM) was trained during the calibration phase as follows: monitoring data from qualified product batches was collected, and the processing steps S1 and S2 were used to obtain the working condition slice sequences for each batch, which served as the training sample set. The number of hidden states was set to 4, corresponding to the four working conditions: preheating, heating, heat preservation, and unloading preparation. The number of observation symbols was also 4, corresponding one-to-one with the working condition category. When initializing the model parameters, the initial state probability vector was set according to the working condition distribution of the first sampling window in the training samples, the state transition probability matrix was set according to the frequency of adjacent state transitions in the training samples, and the observation probability matrix was fixed as an identity matrix. A forward-backward algorithm was used to iteratively train the observation sequences of each batch. In each iteration, the forward and backward probabilities were calculated, and the probability of being in each hidden state at each time point and the expected number of transitions between states were statistically analyzed. Based on this, the initial state probability vector and the state transition probability matrix were updated, and the observation probability matrix was fixed as an identity matrix. Iterations were performed with three consecutive log-likelihood increments less than 1 × 10⁻⁶. -4Alternatively, the process can terminate upon reaching a maximum of 100 iterations, outputting the trained state transition probability matrix, observation probability matrix, and initial state probability vector. During the usage phase, the current production cycle's operational state slice sequence is input into the trained model, and the transition probabilities between each operational state are output as the state transition matrix.

[0060] In one implementation, calculating the state deviation value of the temporal feature from the preset historical distribution model includes:

[0061] The time-series features are input into the preset historical distribution model for feature mapping to obtain the mapped feature vector;

[0062] Calculate the spatial distance between the mapped feature vector and the preset historical reference vector to obtain the deviation value;

[0063] Calculate the difference between the deviation value and the preset anomaly determination threshold;

[0064] The difference is multiplied by a preset weighting coefficient to obtain a weighted difference, and the weighted difference is mapped to a preset standard interval to obtain a state deviation value.

[0065] It should be noted that the current time-series features are normalized to mean and variance using the normalization parameters fixed during the training phase, mapped to a high-dimensional space via a radial basis function kernel, and then projected onto the principal component directions extracted from the historical distribution model to obtain the mapped feature vector. The normalization parameters, kernel function parameters, and principal component directions used in the mapping process are all model parameters fixed during the training phase and are not adjusted during the usage phase.

[0066] It should be noted that the preset historical distribution model is used to characterize the distribution boundary of temporal features within a normal production cycle. During the model training phase, the industrial control system obtains temporal features for each qualified production batch according to steps S11 and S12, and combines the temporal features of different batches into a training sample set. For batches that experience furnace shutdowns, manual interventions, sensor replacements, or product quality inspection failures, even if some sampling window values ​​are within the normal range, they are not included in the training sample set. During training, the temporal features are normalized for mean and variance, and a radial basis function kernel is used for nonlinear mapping. Principal components whose cumulative variance contribution rate reaches the preset model contribution rate threshold are extracted in the mapping space to form a low-dimensional representation of the historical distribution model. The resulting mapped feature vector has the same dimension as the number of principal components extracted from the historical distribution model and is used to calculate the spatial distance with the historical benchmark vector to quantify the degree to which the current temporal features deviate from the normal distribution.

[0067] The formula for calculating the radial basis function kernel is: .

[0068] In the formula, K(x,y) is the radial basis function kernel value, x and y are two time-series features in the mapping space, ‖xy‖ is the Euclidean distance between them, and γ is the kernel function bandwidth, controlling the radial range of the kernel function. The preset model contribution rate threshold is determined during the calibration stage. Qualified batch data covering each stage of preheating, heating, holding, and unloading preparation are selected. Time-series features are extracted in the same way, and kernel principal component analysis is performed. The cumulative variance contribution rate distribution corresponding to 3 principal components is statistically analyzed, and the cumulative variance contribution rate at the 5th percentile is taken as the lower limit of the threshold. In this embodiment, it is taken as 90%.

[0069] Next, the spatial distance between the mapped feature vector and the preset historical baseline vector is calculated as the deviation value. The deviation value can be calculated using the following formula:

[0070] ;

[0071] In the formula, E represents the degree of deviation, and z is the mapped feature vector obtained by mapping the current time-series features through the historical distribution model. Let W be the preset historical baseline vector and W be the weight matrix. The preset historical baseline vector is determined as follows: for all training samples, the arithmetic mean of the mapped feature vectors obtained through the above feature mapping is calculated dimension by dimension, and the average values ​​of each dimension constitute the historical baseline vector. The weight matrix is ​​determined as follows: for each mapping dimension, the variance of the mapped feature vectors of all training samples in that dimension is calculated, and the reciprocal of this variance is taken as the diagonal element of the weight matrix, while the off-diagonal elements are set to zero, forming a diagonal weight matrix. Mapping dimensions with smaller variances are more sensitive to the normal state, and their deviations are given higher weights in the deviation degree values.

[0072] Then, the deviation value is subtracted from the preset anomaly judgment threshold to obtain the difference. This difference is then multiplied by a preset weighting coefficient to obtain a weighted difference. Finally, the weighted difference is mapped to a preset standard interval to obtain the state deviation value. The calculation formulas for the weighted difference and the state deviation value are as follows:

[0073] ;

[0074] ;

[0075] In the formula, G is the weighted difference, B is the state deviation value, E is the deviation degree value, and α is the preset weight coefficient. To preset the anomaly detection threshold, The upper bound of the weighted difference is used as a reference.

[0076] Among them, the preset anomaly detection threshold The deviation value is determined as follows: for all sampling windows of historical normal operation data, the deviation values ​​are calculated separately, sorted from smallest to largest, and the value at the 95th percentile is taken as the deviation value. That is, 95% of the normal window deviation values ​​do not exceed In this embodiment The value is set to 0.8. The preset weighting coefficient α is determined as follows: the deviation values ​​of normal samples and confirmed abnormal samples are statistically analyzed respectively. The separability index is obtained by dividing the difference between the means of the two types of samples by the sum of the standard deviations of the two types of samples. The coefficient that makes the weighted difference the most distinguishable between the two types of samples is selected from the candidate coefficient set as α. In this embodiment, α is set to 1.5.

[0077] It should be noted that the weighted difference is referenced to the upper bound G. ref The weighted difference is determined as follows: for each historical outlier sample, a weighted difference is calculated, and the resulting weighted differences are sorted from smallest to largest. The 95th percentile value is taken as G. ref In this embodiment, G ref Take 3.0, the weighted difference reaches or exceeds G. ref The state deviation value is set to a maximum of 100. Using the above method, when the time-series feature is still within the historical normal distribution range, the state deviation value is 0; when the time-series feature exceeds the historical normal distribution range, the weighted difference increases with the degree of exceedance, and is mapped to a preset standard interval to obtain the state deviation value, which serves as the triggering basis for the staged analysis in step S13.

[0078] In step S13, the operating condition is divided into stages based on the state deviation value to determine the process stage, including:

[0079] When the state deviation value is greater than zero, multidimensional feature data of the working condition is obtained;

[0080] The multidimensional feature data is normalized and mapped to obtain mapped operating condition data;

[0081] Stage mapping is performed based on the mapped operating condition data to obtain stage labels;

[0082] The process stage is determined based on the stage label.

[0083] It should be noted that the state deviation value is used to determine whether the current operating condition has deviated from the historical normal evolution. When the state deviation value is greater than 0, it indicates that the current time series characteristics have exceeded the normal range characterized by the historical distribution model. Based on this, the industrial control system initiates the stage division of the operating condition state, that is, it acquires the multi-dimensional feature data of the operating condition state corresponding to the last sampling window and executes the subsequent division steps. The multi-dimensional feature data includes four dimensions: furnace space temperature difference, temperature rise rate, deviation of the target heat preservation range, and heating power adjustment, which are jointly generated by the data stream and the process control records of the industrial control system.

[0084] When the state deviation value is equal to 0, it indicates that the current time series characteristics are still within the normal range described by the historical distribution model. The industrial control system does not trigger the stage division, but only records the current operating condition and continues to monitor.

[0085] Specifically, the furnace temperature difference is obtained by subtracting the highest and lowest temperature measurement points within the same sampling window; the temperature rise rate is obtained by dividing the difference in average temperature between adjacent sampling windows at the same measurement point or area by the sampling interval; the deviation from the target heat preservation range is determined by the distance between the current average temperature of the area and the target temperature range at the corresponding time point in the target heating curve; and the heating power adjustment is determined by the change in the power setpoint recorded in the combustion control or electric heating control log. These multi-dimensional characteristic data reflect the furnace space uniformity, heating trend, deviation from the heat preservation target, and changes in control input, respectively, and are used to support the differentiation of process stages.

[0086] During normalization mapping, the minimum, maximum, mean, and standard deviation fixed during the training phase are read for each dimension. For dimensions with clear physical upper and lower bounds for furnace space temperature difference, temperature rise rate, and deviation from the target insulation range, max-min normalization is used; for dimensions where heating power adjustment is significantly affected by batch specifications, standard fraction normalization is used. The max-min normalization calculation formula is as follows:

[0087] ;

[0088] In the formula, Let be the normalized value of the d-th dimension. Let d be the current value of the d-th dimension. Let d be the lower bound of the d-th dimension in the training samples. This is the upper bound of the d-th dimension in the training samples.

[0089] The formula for normalizing standard scores is: .

[0090] In the formula, The value of the d-th dimension after standard score normalization. Let d be the current value of the d-th dimension. Let be the mean of the d-th dimension in the training samples. Let d be the standard deviation of the d-th dimension in the training samples.

[0091] Normalize each dimension one by one as described above, and combine the normalized values ​​of all dimensions in a fixed dimensional order to obtain the mapping condition data. When the current value exceeds the upper and lower bounds of the training phase, retain the out-of-bounds magnitude and write the out-of-bounds flag into the mapping condition data, so that the support vector machine model can use the out-of-bounds information during the recognition phase, rather than simply truncating the degree of anomaly.

[0092] It should be noted that the mapped operating condition data is input into a support vector machine (SVM) for stage mapping. The SVM model is pre-trained during the calibration phase, using the mapped operating condition data within each process stage time period as training samples and the stage category after verification of the process segment records as sample labels. A radial basis function kernel is used to train and obtain the classification decision function, and the support vectors, discriminant function, normalization parameters, and stage label mapping table are stored in a fixed manner. When in use, the current mapped operating condition data is substituted into the classification decision function, and the stage category number is output according to its position relative to each decision boundary. The stage category number is the stage label, and the corresponding process stage is determined according to the stage label mapping table. In this embodiment, stage labels one to four correspond to the abnormal temperature rise rate stage, furnace temperature difference imbalance stage, holding temperature drift stage, and furnace outlet temperature deviation stage, respectively.

[0093] The labels for the support vector machine model are derived from process segment records and quality inspection records in the production batch. For four types of anomalies—abnormal temperature rise rate, furnace temperature imbalance, holding temperature drift, and furnace exit temperature deviation—the corresponding mapped operating condition data within the abnormal time period are extracted as positive samples. For periods where no anomalies occur under the same operating condition, the mapped operating condition data are extracted as negative samples. For buffer periods near stage transitions, if the process segment record does not match the temperature curve change, the buffer sample is removed to avoid ambiguous label boundaries affecting the decision boundary. During training, the penalty coefficient C and kernel function bandwidth are... As candidate hyperparameters, a grid search is performed within a preset value range, and the stage recognition accuracy on the validation set and the confusion rate between adjacent stages are used as evaluation metrics. After training, the support vectors, penalty coefficients, kernel function bandwidth, normalization parameters, and stage label mapping table are fixed.

[0094] The support vector machine model employs a multi-class support vector machine with radial basis function kernels, using a one-to-one strategy to distinguish four stage categories: abnormal temperature rise rate stage, furnace temperature difference imbalance stage, holding temperature drift stage, and furnace exit temperature deviation stage. The input is normalized 4D mapped operating condition data, and the output is the stage category number. During the training phase, mapped operating condition data within each process stage time period are used as samples, and the stage category verified by the process segment records is used as the label. The loss function is a hinge loss with a penalty term. The penalty coefficient C and kernel function bandwidth γ are optimally selected within a preset range through grid search. The stage identification accuracy on the validation set and the confusion rate between adjacent stages are used as evaluation indicators. Training continues until the dual problem converges.

[0095] Therefore, step S13 does not simply divide the process stages based on the planned time, but rather uses multi-dimensional feature data under the current operating conditions, combined with the trained support vector machine decision boundary, to determine the process stages. The output of this step not only determines which set of historical operating trajectories and anomaly boundaries to call in step S14, but also determines the stage encoding used in the anomaly scoring and response rules in steps S15 and S16.

[0096] In step S14, the abnormal boundary corresponding to the process stage is obtained, and the upper and lower limits of the abnormal boundary are extracted to obtain the extreme value interval, including:

[0097] Obtain the historical running trajectory corresponding to the process stage, and perform convex hull calculation on the historical running trajectory to obtain a multi-dimensional envelope region;

[0098] Orthogonally project the multidimensional envelope region to obtain a two-dimensional boundary profile;

[0099] Connect the vertices of the two-dimensional boundary contour to obtain the abnormal boundary corresponding to the process stage;

[0100] The coordinate values ​​of each contour vertex of the abnormal boundary are extracted to generate an extreme value set. The upper limit extreme value and the lower limit extreme value are extracted from the extreme value set to obtain the extreme value interval.

[0101] It should be noted that the historical operating trajectories are stored separately according to the process stage. Each historical operating trajectory consists of multi-dimensional monitoring data points from continuous sampling windows within a qualified product batch, arranged chronologically. These multi-dimensional monitoring data points include at least the furnace space temperature difference, temperature rise rate, deviation from the target insulation range, and heating power adjustment, and are associated with the batch number, billet specification, furnace number, and stage label. When acquiring historical operating trajectories, the industrial control system first retrieves the corresponding trajectory set according to the process stage, and then filters the trajectory data under the same process conditions based on the current product specification and furnace number. Before the filtered historical operating trajectories enter the convex hull calculation, batches that fail quality inspection and batches subject to manual intervention are removed to ensure that the multi-dimensional envelope region represents normal production fluctuations.

[0102] It should be noted that the historical operating trajectories corresponding to the process stages are derived from normal fluctuation data recorded in historical production under the same anomaly category. In other words, a historical operating trajectory categorized as a certain process stage refers to an operating trajectory in historical data that has been determined by the support vector machine to belong to the same anomaly category, and whose multidimensional monitoring data points are within the normal fluctuation range of that category. The historical operating trajectories themselves are stored separately according to their respective process stages.

[0103] During convex hull calculation, each multidimensional monitoring data point is treated as a point in the multidimensional space. An incremental convex hull algorithm or a fast convex hull algorithm is used to find the smallest convex polyhedron containing all normal data points. For cases where the large number of sample points increases the computational burden, stratified sampling is first performed according to time sequence and spatial distribution, retaining extreme points in each dimension, sample points near the boundaries, and trajectory turning points before convex hull calculation. This sampling rule retains key samples forming the boundary and does not destroy the boundary range through random discarding. After the multidimensional envelope region is calculated, it is stored together with the process stage, sample source range, calculation time, and effective dimension set for use in orthogonal projection processing.

[0104] It should be noted that orthogonal projection of the multidimensional envelope region involves projecting the coordinates of the multidimensional envelope region onto the target two-dimensional plane. The target two-dimensional plane is spanned by the two monitoring dimensions most relevant to process anomaly determination; in this embodiment, the furnace space temperature difference and temperature rise rate are selected. The selection criteria for these two dimensions are as follows: During the calibration phase, the industrial control system statistically analyzes the correlation between each monitoring dimension and historical product quality defects, and calculates the distinguishing power of different dimension combinations for normal and abnormal samples. The two-dimensional combination with the highest distinguishing power is selected as the target two-dimensional plane. In this embodiment, the furnace space temperature difference reflects the uniformity of temperature distribution within the furnace, and the temperature rise rate reflects whether the heating process deviates from the target heating curve. The combination of these two dimensions has a high distinguishing power for typical quality risks such as underheating, overheating, and excessive local temperature differences. Therefore, the target two-dimensional plane is constructed using these two monitoring dimensions as coordinate axes.

[0105] Specifically, a Cartesian coordinate system is established with the temperature difference in the furnace space as the horizontal axis and the temperature rise rate as the vertical axis. The plane containing this coordinate system is the target two-dimensional plane. During projection, the components belonging to the above two dimensions are retained in the coordinates of each vertex of the multidimensional envelope region, while the components of other dimensions are directly discarded, resulting in the projected coordinates of each vertex on the target two-dimensional plane. Vertices with coincident coordinates or Euclidean distances less than a preset distance threshold after projection are merged. The merged projected vertices are sorted according to their polar angle relative to the geometric center. The sorted vertices are then connected sequentially to form a closed polygonal boundary, thus obtaining the two-dimensional boundary contour corresponding to the aforementioned process stage.

[0106] The preset distance threshold is set as follows: the measurement resolution of the two dimensions of the target two-dimensional plane is read respectively. In this embodiment, the measurement resolution of the furnace space temperature difference dimension is 0.5℃, and the measurement resolution of the temperature rise rate dimension is 0.1℃ / min. After converting the measurement resolution of the two dimensions according to the normalized scale of the corresponding dimension, the larger value of the two is taken as the preset distance threshold. When the coordinate difference between two vertices after projection is less than the preset distance threshold, the two vertices have no physical distinction and are regarded as the same vertex and merged. This process ensures that the target monitoring dimension used for the two-dimensional boundary contour is consistent with that used for the abnormal boundary.

[0107] It should be noted that all contour vertices of the two-dimensional boundary contour are extracted and connected sequentially in a clockwise direction to obtain a closed abnormal boundary. This abnormal boundary is bound to the process stage for storage, allowing the abnormal temperature rise rate stage, furnace temperature difference imbalance stage, holding temperature drift stage, and furnace exit temperature deviation stage to use their respective abnormal boundaries. When connecting contour vertices, the polar angle of each contour vertex relative to the geometric center is calculated using the geometric center of the two-dimensional boundary contour as a reference point, and the vertices are arranged in ascending order of polar angle. If two contour vertices have the same polar angle, the vertex farther from the geometric center is retained as the outer boundary point. After connection, a self-intersection check is performed on the closed boundary. If self-intersecting edges exist, they are reordered according to the adjacent edge exchange rule until a closed boundary without self-intersection is formed. The boundary edge set consists of edges formed by every two adjacent contour vertices after sorting and connection, and each edge records the starting vertex number and ending vertex number. The stage label is the category number of the process stage bound to the abnormal boundary, consistent with the stage label obtained in the stage mapping. The abnormal boundary includes the contour vertex coordinates, the boundary edge set, and the stage label. When storing, it also records the dimension name of the target two-dimensional plane to avoid subsequent steps misusing other monitoring dimensions for the abnormal boundary.

[0108] The technical function of the aforementioned abnormal boundaries is to convert historical normal operating trajectories into geometric boundaries that match the current process stage, rather than sharing a single numerical threshold across all process stages. Taking an abnormal temperature rise rate stage as an example, a higher temperature rise rate is normal process behavior, but the same temperature rise rate may indicate a control imbalance when the furnace exit temperature deviates. By using abnormal boundaries bound to process stages, the heating stage and the holding stage each have their own boundary range, thereby reducing false alarms during stage switching.

[0109] It should be noted that the extreme value range is not a uniform value taken from all the original sensor data, but is determined for the target monitoring dimension. In this embodiment, the coordinate values ​​of each contour vertex are extracted for the furnace space temperature difference dimension, and the upper and lower extreme values ​​are selected to form the spatial temperature difference extreme value range corresponding to the process stage. Therefore, the real-time values ​​in the subsequent over-limit judgment have the same target monitoring dimension as the extreme value range, avoiding the confusion of arbitrary values ​​in the original data stream with the boundary range.

[0110] In this embodiment, if the target monitoring dimension is the furnace space temperature difference, the extreme value set is composed of the coordinates of each contour vertex of the abnormal boundary in the furnace space temperature difference dimension; if the target monitoring dimension is the temperature rise rate, the extreme value set is composed of the coordinates of each contour vertex in the temperature rise rate dimension. When generating the extreme value interval, the industrial control system writes the target monitoring dimension name, unit, upper limit extreme value, lower limit extreme value, and corresponding process stage into the boundary parameter table. Before calling the extreme value interval for the limit judgment, it first checks whether the dimension name and unit of the real-time value are consistent with the boundary parameter table. If they are inconsistent, no abnormal signal is generated, but an abnormal record of the dimension configuration is output. This dimension check makes the source of the abnormal signal clear and avoids mixing and comparing the absolute temperature value, space temperature difference, and heating rate.

[0111] In step S15, the real-time value is the target monitoring dimension value extracted from the data stream, corresponding to the current process stage. If the target monitoring dimension is the furnace space temperature difference, the average temperature of the upper, middle, and lower regions is calculated within the current window, and the difference between the maximum and minimum average temperatures is taken as the real-time value. If the target monitoring dimension is the temperature rise rate, the difference between the average temperatures of two adjacent windows is read and divided by the time interval between the center times of the two windows. If the target monitoring dimension is the deviation from the target insulation zone, the distance between the current average temperature and the nearest boundary of the target insulation zone is calculated. The above calculation rules are consistent with the target monitoring dimensions bound to the extreme value interval.

[0112] When the real-time value exceeds the extreme value range, the current process parameters have deviated from the normal fluctuation range, indicating a potential process anomaly. Further calculation of the deviation magnitude is required for analysis. To avoid invalid alarms caused by short-term noise, this embodiment sets a threshold for the number of consecutive out-of-limit windows. Specifically, when a real-time value exceeds the extreme value range, the industrial control system records the out-of-limit window. If the number of consecutive out-of-limit windows reaches three, a process anomaly is determined and anomaly scoring is initiated. If less than three out-of-limit windows are recorded, the out-of-limit record is stored as a candidate anomaly record and not entered into the anomaly scoring process. When the real-time value does not exceed the extreme value range, it indicates that the current monitored dimension value is within the normal fluctuation range of the corresponding process stage. The industrial control system does not calculate the deviation magnitude or enter anomaly scoring; it only continues to read the data stream from subsequent sampling windows.

[0113] In one implementation, the deviation magnitude is calculated based on the difference between the real-time value and the boundary of the extreme value interval, and an anomaly score is obtained by anomaly scoring, including:

[0114] Calculate the difference between the real-time value and the boundary of the extreme value interval to obtain the initial deviation;

[0115] The initial deviation is quantized to obtain the deviation magnitude;

[0116] Perform multi-dimensional feature mapping on the deviation magnitude to obtain the deviation feature vector;

[0117] The deviation direction is extracted based on the deviation feature vector, and the deviation direction and the deviation magnitude are weighted and fused to obtain the anomaly score.

[0118] It should be noted that the initial deviation is calculated as follows: when the real-time value is higher than the upper limit of the extreme value range, the difference between the real-time value and the upper limit is calculated; when the real-time value is lower than the lower limit of the extreme value range, the difference between the lower limit and the real-time value is calculated. The deviation amplitude is determined by the following formula: .

[0119] In the formula, A represents the deviation range. Let U be the initial deviation, U be the upper limit of the extreme value range, and L be the lower limit of the extreme value range. This is the amplitude quantification coefficient. The amplitude quantification coefficient is calibrated as follows: During the calibration phase, the ratio of the initial deviation to the extreme value range of the corresponding process stage is calculated for each historical abnormal sample. Based on the quality inspection records, the severity of quality defects in each sample is divided into three levels: minor, moderate, and severe. The distribution of the ratios of the three levels of samples is statistically analyzed. The coefficient value that causes the deviation amplitudes of the minor, moderate, and severe samples to fall into the three target ranges of 0 to 0.5, 0.5 to 1.0, and above 1.0, respectively, is selected as the amplitude quantification coefficient. In this embodiment, it is set to 1.2. This calibration makes the numerical levels of deviation amplitude correspond to the severity of quality defects, allowing for direct use in anomaly scoring.

[0120] To avoid extreme anomalies amplifying the scoring input infinitely, the deviation magnitude is divided by the maximum reference deviation magnitude and limited to a range of 0 to 1 to obtain the magnitude ratio. The maximum reference deviation magnitude is determined by the industrial control system during the calibration phase, which statistically analyzes the deviation magnitudes of confirmed abnormal samples at each process stage, sorts them from smallest to largest, and takes the 95th percentile value. This is then statistically analyzed according to the target monitoring dimension and linked to the process stage in the scoring parameter table. This process ensures that slight, moderate, and severe deviations have stable numerical levels in the anomaly scoring.

[0121] It should be noted that the deviation feature vector includes the amplitude ratio, deviation direction code, over-limit duration, target monitoring dimension number, process stage code, and sensor spatial region code. Specifically, the deviation direction code is determined by the out-of-bounds sign of the real-time value relative to the upper or lower limit of the extreme value range; the over-limit duration is obtained by multiplying the number of consecutive over-limit windows by the window step size; the process stage code is obtained by the stage label conversion in step S13; and the sensor spatial region code is determined by the deployment area of ​​the temperature sensor within the furnace.

[0122] When constructing the deviation feature vector, continuous and discrete features employ different encoding methods. The amplitude ratio and duration of exceeding limits are treated as continuous features and standardized using the mean and standard deviation fixed during the training phase. The target monitoring dimension number, process stage code, deviation direction code, and sensor spatial region code are treated as discrete features and converted into numerical codes according to a preset encoding table. If the abnormal signal involves multiple sensor spatial regions, the codes for each region are used to generate multi-bit region identifiers in a preset order, rather than arbitrarily selecting a single region. The resulting deviation feature vector contains not only the intensity and duration of the abnormal deviation but also the process stage, target monitoring dimension, and spatial location of the anomaly.

[0123] It should be noted that the dimensions and order of the deviation feature vectors are fixed into a feature order table during the calibration phase. When constructing the deviation feature vectors in the usage phase, the feature name, unit, and encoding method are checked item by item. If a feature is missing, the missing value defined in the calibration phase is written in, and the reason for the missing value is recorded in the anomaly signal. This process ensures the stability of the input structure received for subsequent extraction of deviation direction and deviation magnitude, avoiding scoring distortion due to changes in feature order.

[0124] It should be noted that the deviation direction is jointly characterized by the deviation direction code and the sensor spatial region code in the deviation feature vector, used to characterize the direction of the anomaly crossing the target monitoring dimension and the spatial distribution direction of the anomaly within the furnace; the deviation magnitude is characterized by the magnitude ratio in the deviation feature vector, used to characterize the relative magnitude of the current deviation relative to the maximum reference deviation magnitude. The deviation direction is mapped to a directional risk weight of 0 to 1, the directional risk weight and the magnitude ratio are weighted and summed according to a preset fusion coefficient, and the summation result is mapped to the interval of 0 to 100 to obtain the anomaly score, calculated using the following formula: .

[0125] In the formula, S is the anomaly score, d is the directional risk weight corresponding to the deviation direction, ranging from 0 to 1, and R is the amplitude ratio, ranging from 0 to 1. The preset fusion coefficient is the deviation from the direction. The preset fusion coefficient is the deviation amplitude, and and The sum is 1. A higher anomaly score indicates a higher overall risk in terms of both the direction of the anomaly and the magnitude of the deviation. The directional risk weight d is combined with a preset fusion coefficient. , All values ​​are determined by historical samples during the calibration phase, and are not adjusted in reverse based on current abnormal signals during the usage phase.

[0126] In this process, both the directional risk weight and the preset fusion coefficient are determined from historical samples during the calibration phase. Specifically, the industrial control system statistically analyzes the proportion of quality defects occurring in different out-of-bounds directions and different spatial distribution directions in historical samples, assigning higher directional risk weights to deviation directions with higher defect occurrence ratios. The preset fusion coefficient is obtained by calibrating the contribution ratio of deviation direction and deviation magnitude to the severity of quality defects in historical abnormal samples, ensuring that deviation direction factors and deviation magnitude factors have reasonable relative weights in the abnormality scores. The above calibration is completed on qualified batches and batches with confirmed quality abnormalities, and the directional risk weight table and preset fusion coefficient are permanently stored, without being adjusted in reverse based on current abnormal signals during the usage phase.

[0127] In step S16, the warning level is determined based on the anomaly score, a warning label is generated, and rule matching is performed based on the warning label to output a graded warning instruction, including:

[0128] Extract the risk gradient corresponding to the abnormal score, and determine the warning level based on the risk gradient;

[0129] Extract the status identifier corresponding to the warning level, and perform label encoding on the status identifier to generate a warning label;

[0130] Extract the rule index carried by the warning label, retrieve the matching matrix based on the rule index, and extract the response rule based on the matching matrix;

[0131] The response rules are parsed to determine the action to be taken, and the action is encapsulated into a tiered early warning instruction.

[0132] The tiered early warning command is assigned a channel according to its warning level, and the tiered early warning command is output through the channel.

[0133] It should be noted that the industrial control system reads the risk score range from the scoring parameter table and maps abnormal scores to low-risk, medium-risk, or high-risk gradients. The scoring parameter table is generated and stored during the calibration phase. The risk score range is determined by the interval between the upper bound of normal fluctuation scores and the lower bound of abnormal quality sample scores in historical samples, and is stored separately according to process stages. During use, the table is directly consulted. Subsequently, the warning level mapping table is queried based on the risk gradient, process stage, and target monitoring dimension to obtain the warning level. This mapping table is calibrated by process engineers based on historical quality defects and on-site handling strategies and is fixed in parameter table form. The response rules for low-risk gradients primarily focus on record keeping and subsequent review, without triggering on-site handling actions, thus avoiding frequent on-site handling for single minor fluctuations.

[0134] After obtaining the warning level, the corresponding status identifier is further extracted. This status identifier is a status mark registered in the warning level definition table, corresponding one-to-one with the warning level, and uniquely represents the warning state of that level. Subsequently, the status identifier is tagged and encoded according to a preset tag encoding rule. This involves concatenating the status identifier along with its associated risk gradient, process stage, and target monitoring dimension into a fixed-length code using a fixed field order and bit width, generating the warning tag. This tag serves as the retrieval key for searching the matching matrix in the response rule base.

[0135] The warning label includes warning level, risk gradient, process stage, target monitoring dimension, associated sensor area, anomaly score, and rule index. The rule index is generated by combining the above fields according to preset coding rules and is used to retrieve the matching matrix in the response rule base. The rows of the matching matrix represent different warning levels, the columns represent different combinations of process stages and target monitoring dimensions, and the matrix elements point to the corresponding response rules. Response rules include at least the action, action parameters, permissible execution conditions, manual review requirements, and output channel. For actions that affect production safety or batch flow, a manual confirmation flag is set in the response rule; for actions that only involve alarm prompts or record keeping, an automatic output flag is set in the response rule.

[0136] It should be noted that the aforementioned preset encoding rules, response rule base, and matching matrix are all pre-set and stored during the calibration phase. The preset encoding rules determine their bit width and concatenation order based on the value range of each field, ensuring that the generated rule index is unique across all field combinations. The row and column dimensions of the matching matrix are determined by the number of combinations of warning levels, process stages, and target monitoring dimensions. The correspondence between matrix elements and response rules is calibrated item by item by process engineers based on historical quality defect types and on-site handling experience. During the usage phase, the rule base is directly retrieved using the rule index, without retraining or adjusting the rule base based on the current warning signal.

[0137] When parsing the response rules, the industrial control system extracts the handling actions and generates tiered early warning instructions. Each tiered early warning instruction includes at least an instruction identifier, an early warning tag, a handling action, a target device, a target process segment, a generation time, a validity period, and a checksum. The instruction identifier is generated by combining the batch number, sampling window number, and early warning tag code, ensuring that multiple identical handling instructions are not generated for the same abnormal event. Channels are allocated according to the early warning level of the tiered early warning instructions: high-level early warnings are allocated to a dedicated control channel with the largest bandwidth, medium-level early warnings are allocated to a general control channel, and low-level early warnings are allocated to a standard data channel. After channel allocation, the tiered early warning instructions are output to the production management personnel terminal and the field execution equipment through the allocated channels.

[0138] This invention also provides a real-time process anomaly monitoring system for quality control, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described real-time process anomaly monitoring method for quality control.

[0139] It should be noted that the real-time process anomaly monitoring system for quality control provided in this embodiment of the invention is used to execute all the process steps of the real-time process anomaly monitoring method for quality control in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0140] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0141] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for real-time monitoring of process anomalies for quality control, characterized in that, include: Acquire the data stream synchronously collected by multiple isomorphic sensors, extract the spatial collaborative features of the data stream, and obtain the spatial feature vector; The working condition is determined based on the spatial feature vector, the time series features of the working condition are extracted to obtain the time series features, and the state deviation value of the time series features from the preset historical distribution model is calculated. The operating conditions are divided into stages based on the state deviation values ​​to determine the process stages; Obtain the abnormal boundary corresponding to the process stage, extract the upper and lower limit extreme values ​​of the abnormal boundary, and obtain the extreme value range; When the real-time value in the data stream exceeds the extreme value range, the deviation range is calculated based on the difference between the real-time value and the boundary of the extreme value range, and an anomaly score is obtained. The warning level is determined based on the anomaly score, a warning label is generated, and rule matching is performed based on the warning label to output a graded warning instruction.

2. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The step of extracting the spatial co-location features of the data stream to obtain a spatial feature vector includes: Extract the amplitude of the data stream at different spatial locations and construct a spatial distribution matrix; The spatial distribution matrix is ​​centered to obtain a centered matrix; The spatial covariance matrix is ​​obtained by multiplying the centered matrix with its transpose. The spatial covariance matrix is ​​decomposed into eigenvalues ​​to obtain an initial set of eigenvectors; Based on the initial feature vector set, the main directional features are selected, and the main directional features are vectorized to obtain spatial feature vectors.

3. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The step of determining the operating condition state based on the spatial feature vector and extracting the time series features of the operating condition state to obtain time series features includes: Calculate the vector distance between the spatial feature vector and the preset reference feature vector, and determine the working condition based on the vector distance; Extract the timestamp information of the working condition status, construct a state time axis, and slice the state time axis to obtain multiple working condition status slices. Based on the statistical analysis of the state change relationships between adjacent slices of the multiple operating condition slices, a state transition matrix is ​​obtained; Based on the state transition matrix, the time series features of the working condition within the current production cycle are extracted to obtain the time series features.

4. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The calculation of the state deviation value of the time series features from the preset historical distribution model includes: The time-series features are input into the preset historical distribution model for feature mapping to obtain the mapped feature vector; Calculate the spatial distance between the mapped feature vector and the preset historical reference vector to obtain the deviation value; Calculate the difference between the deviation value and the preset anomaly determination threshold; The difference is multiplied by a preset weighting coefficient to obtain a weighted difference, and the weighted difference is mapped to a preset standard interval to obtain a state deviation value.

5. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The step of dividing the operating condition into stages based on the state deviation value and determining the process stage includes: When the state deviation value is greater than zero, multidimensional feature data of the working condition is obtained; The multidimensional feature data is normalized and mapped to obtain mapped operating condition data; Stage mapping is performed based on the mapped operating condition data to obtain stage labels; The process stage is determined based on the stage label.

6. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The step of obtaining the abnormal boundary corresponding to the process stage, extracting the upper and lower limits of the abnormal boundary, and obtaining the extreme value interval includes: Obtain the historical running trajectory corresponding to the process stage, and perform convex hull calculation on the historical running trajectory to obtain a multi-dimensional envelope region; Orthogonally project the multidimensional envelope region to obtain a two-dimensional boundary profile; Connect the vertices of the two-dimensional boundary contour to obtain the abnormal boundary corresponding to the process stage; The coordinate values ​​of each contour vertex of the abnormal boundary are extracted to generate an extreme value set. The upper limit extreme value and the lower limit extreme value are extracted from the extreme value set to obtain the extreme value interval.

7. The method for real-time monitoring of process anomalies for quality control according to claim 1, characterized in that, The step of calculating the deviation magnitude based on the difference between the real-time value and the boundary of the extreme value interval, and performing anomaly scoring to obtain an anomaly score, includes: Calculate the difference between the real-time value and the boundary of the extreme value interval to obtain the initial deviation; The initial deviation is quantized to obtain the deviation magnitude; Perform multi-dimensional feature mapping on the deviation magnitude to obtain the deviation feature vector; The deviation direction is extracted based on the deviation feature vector, and the deviation direction and the deviation magnitude are weighted and fused to obtain the anomaly score.

8. The method for real-time monitoring of process anomalies for quality control according to any one of claims 1-7, characterized in that, The process of determining the warning level based on the anomaly score, generating a warning label, and performing rule matching based on the warning label to output a tiered warning instruction includes: Extract the risk gradient corresponding to the abnormal score, and determine the warning level based on the risk gradient; Extract the status identifier corresponding to the warning level, and perform label encoding on the status identifier to generate a warning label; Extract the rule index carried by the warning label, retrieve the matching matrix based on the rule index, and extract the response rule based on the matching matrix; The response rules are parsed to determine the action to be taken, and the action is encapsulated into a tiered early warning instruction. The tiered early warning command is assigned a channel according to its warning level, and the tiered early warning command is output through the channel.

9. A real-time process anomaly monitoring system for quality control, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the real-time monitoring method for process anomalies for quality control as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the real-time monitoring method for process anomalies for quality control as described in any one of claims 1 to 8.

Images