An abnormality alarm method and system for feed production control

Abstract

The present application belongs to the technical field of abnormal alarm, and particularly relates to a feed production control abnormal alarm method and system, comprising the following steps: obtaining feed production historical parameter data, filtering parameter points located in the forbidden hyperspace based on a production process constraint model, and the remaining parameter points constituting a candidate set; for each candidate point of the candidate set, extracting parameter time series within a preset time window, calculating the Frobenius norm of the time gradient matrix as a state fluctuation measure; calculating the number of points within a specified neighborhood radius of each candidate point as a local density, and calculating the Euclidean distance from the candidate point to the nearest point with a local density higher than itself as a dispersion degree. The present application can trigger an alarm for malignant boundary crossing phenomena that violate the rules, and can also identify hidden deviations according to the boundaries set by each normal sub-working condition, reducing false positives and false negatives, and providing a guarantee for the safe and reliable operation of the feed production process.

Description

Technical Field

[0001] This invention belongs to the field of abnormal alarm technology, specifically relating to an abnormal alarm method and system for feed production control. Background Technology

[0002] The feed mill's underlying control system can collect multi-dimensional production parameter data such as temperature, pressure, motor current, and material flow rate in real time. In actual production scenarios, due to frequent changes in processing formulas, differences in raw material properties, and changes in environmental factors, the feed production process exhibits multi-variable and multi-condition operating characteristics. By combining unsupervised clustering algorithms with anomaly detection technology, historical production parameter data is clustered to divide data clusters representing different steady-state operating conditions, and corresponding monitoring benchmarks are constructed based on normal sub-conditions.

[0003] Patent application CN116719309A discloses an anomaly alarm method and system for a feed production control system. This method not only performs cross-domain interaction between the fault description vector sequence and the first low-order knowledge description vector to expand the domain representation capability of feed anomaly production control data and generate significant low-order knowledge description vectors, but also performs intra-domain interaction and fusion between the second low-order knowledge description vector and the first high-order knowledge description vector to expand the first high-order knowledge description vector and generate significant high-order knowledge description vectors. Based on cross-domain and intra-domain interaction, fault point prediction can be performed by combining multi-domain description vectors, improving the accuracy of fault prediction for feed anomaly production control data.

[0004] However, most existing methods fail to incorporate production process constraints during data preprocessing, leading to the inclusion of sensor anomalies, dead values, or erroneous points that violate objective laws into the candidate dataset, thus polluting the operational condition model space. In the selection of initial cluster centers for clustering algorithms, evaluation is typically based solely on the local density and spatial distance of sample points in the static space, neglecting the temporal fluctuation characteristics implicit in industrial time-series data. When processing feed processing parameters with temporal correlation characteristics, parameter points in fluctuating transition states are easily mistakenly selected as cluster centers, causing the resulting sub-operating condition model clusters to deviate from the true process steady state. Using a uniform global distance metric for alarm judgment fails to establish control boundary mechanisms for the differences in the distribution of data from various normal sub-operating conditions and cannot organically separate exceeding limits from deviations, resulting in poor generalization ability of the alarm system for unknown operating conditions, a high false alarm rate, and an inability to meet the early warning requirements for high-quality feed production. Summary of the Invention

[0005] This invention provides an anomaly alarm method and system for feed production control to solve the technical problems in the prior art, such as data preprocessing not effectively filtering out anomalies, cluster initialization ignoring temporal steady-state characteristics, and lack of cluster adaptive boundaries and two-level alarm mechanisms.

[0006] In a first aspect, the present invention provides an abnormal alarm method for feed production control, comprising the following steps: S1. Obtain historical parameter data of feed production, filter out parameter points located in the prohibited hyperspace based on the production process constraint model, and form a candidate set of the remaining parameter points; for each candidate point in the candidate set, extract the parameter time series within a preset time window, calculate the Frobenius norm of the time gradient matrix as a measure of state fluctuation; calculate the number of points within a specified neighborhood radius of each candidate point as the local density, and calculate the Euclidean distance from the candidate point to the nearest point with a local density higher than itself as the dispersion. S2, multiply the local density, dispersion, and state stability metric obtained by inverse processing based on state fluctuation metric to obtain the initial center weight of the candidate point; select the candidate point with the highest initial center weight as the first initial cluster center, and iteratively select subsequent centers based on maximizing the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers; based on the determined initial cluster centers, cluster the candidate sets to construct multiple normal sub-condition model clusters, and set control boundaries based on the data distribution within the normal sub-condition model clusters; S3: Obtain real-time production parameters. If the real-time data point is located within the prohibited overspace, trigger an over-limit alarm; otherwise, determine the nearest model cluster to which the real-time data point belongs. If the real-time data point is located outside the control boundary of the nearest model cluster, trigger a statistical deviation alarm.

[0007] Its effects are as follows: First, by prohibiting hyperspace filtering, sensor dead values ​​and erroneous points that violate physical laws are eliminated. Then, the state stability metric obtained by fusing local density, dispersion, and the Frobenius norm of the time gradient matrix is ​​used to evaluate the weights of candidate points, avoiding the misselection of parameter points in fluctuating transition states as cluster centers. At the same time, the limit-crossing alarm triggered by prohibiting hyperspace and the statistical deviation alarm based on the cluster distribution are independent yet complementary. This enables the alarm mechanism to respond immediately to malicious limit crossings that violate process laws, and to accurately identify hidden deviations that are within the reasonable range of physical quantities but deviate from normal operating conditions. This significantly reduces the false alarm rate and the missed alarm rate, effectively ensuring the continuous and stable operation of the feed production process.

[0008] Furthermore, based on the production process constraint model, parameter points located in the forbidden hyperspace are filtered out, and the remaining parameter points constitute a candidate set, including: Extract the absolute lower bound threshold and absolute upper bound threshold corresponding to each parameter dimension in the constraint model; Iterate through the data records in the historical parameter data of feed production one by one, and compare the parameter values ​​of each dimension with the corresponding absolute lower limit threshold and absolute upper limit threshold. If any parameter value in any dimension of a data record is less than the absolute lower threshold or greater than the absolute upper threshold, the data record is determined to be within the prohibited hyperspace and is removed. All data records that were not removed were retained to form a candidate set.

[0009] Its effect is that by extracting the absolute upper and lower limit thresholds in the constraint model dimension by dimension and comparing each historical data record, it can accurately remove abnormal dead values ​​that violate physical laws, such as negative pressure values ​​or temperature changes to the sensor limit value, during the data preprocessing stage. Compared with conventional clustering preprocessing methods that do not combine process constraints, it avoids invalid erroneous samples from contaminating the candidate dataset from the source, and provides a clean and reliable data foundation for the subsequent construction of working condition model clusters.

[0010] Furthermore, for each candidate point in the candidate set, the parameter time series within a preset time window is extracted, and the Frobenius norm of the time gradient matrix is ​​calculated as a measure of state fluctuation, including: Centered on the time of the candidate point, a preset number of sampling points adjacent to the parameter time series are extracted to form a multidimensional parameter time series matrix, and each dimension of the multidimensional parameter time series matrix is ​​standardized. Calculate the time difference vectors of adjacent sampling points in the standardized multidimensional parameter time series matrix, and combine the time difference vectors of all dimensions to form the time gradient matrix; The square root of the sum of squares of the elements of the time gradient matrix is ​​calculated to obtain the Frobenius norm, which is then used as a measure of the state fluctuation of candidate points.

[0011] Its effect is as follows: by extracting adjacent sampling points centered on the time of the candidate point to form a time series matrix, and then obtaining the first difference after Z-score standardization, the Frobenius norm of the time gradient matrix is ​​compressed into a single scalar characterizing the intensity of temporal fluctuations in the neighborhood of the candidate point. Compared with the traditional clustering method that only evaluates candidate points based on static spatial attributes, it can explicitly quantify the temporal difference between norms as low as 1.25 under steady-state conditions and norms as high as 8.60 under drastic changing conditions, providing highly discriminative temporal features for subsequent accurate identification of representative points of real working conditions in a steady state.

[0012] Furthermore, the number of points within a specified neighborhood radius for each candidate point is calculated as the local density, and the Euclidean distance from the candidate point to the nearest point with a local density higher than its own is calculated as the discreteness, including: The number of other candidate points whose distance from the current candidate point is less than the specified neighborhood radius is used as the local density of the current candidate point. Calculate the Euclidean distance between the current candidate point and all comparison candidate points with a local density higher than the current candidate point, and take the smallest Euclidean distance among them as the dispersion of the current candidate point; For candidate points where there is no local density higher than the candidate point itself, the maximum Euclidean distance between the candidate point and all other candidate points is taken as the discreteness.

[0013] Its effect is as follows: by presetting the neighborhood radius and combining it with the Euclidean distance of the nearest point with a density higher than itself to construct a dual-attribute mapping mechanism, the maximum Euclidean distance with other points is used as the dispersion for the core point with the highest global density. This avoids the boundary problem of undefined dispersion at high-density core points in the traditional density peak method, so that each candidate point can obtain an accurate quantitative description that has both local representativeness and global spatial distinguishability, thus improving the rationality of the candidate point ranking.

[0014] Furthermore, the local density, dispersion, and state stability metric obtained by inverse processing based on state fluctuation metric are multiplied to obtain the initial center weight of the candidate point. This includes multiplying the local density of the current candidate point by the dispersion of the current candidate point and then by the state stability metric of the current candidate point, and using the product result as the initial center weight of the current candidate point.

[0015] Its effect is that by directly multiplying the normalized local density and dispersion with the state stability measure obtained by means of 1 / (1+state fluctuation measure) into a three-factor product, the problem of repeated parameter tuning required by introducing complex weight coefficients is avoided. This allows the initial center weight to simultaneously reflect the spatial clustering degree, spatial distribution discrimination and temporal stability of the candidate points, ensuring that the selected candidate points have comprehensive characteristics that truly represent the steady-state working condition.

[0016] Furthermore, subsequent centers are selected iteratively, based on maximizing the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers, including: Calculate the minimum Euclidean distance from each candidate point in the candidate set that was not selected as an initial cluster center to all selected initial cluster centers; The score is obtained by multiplying the initial center weight of each candidate point that was not selected as the initial cluster center with the square of the corresponding minimum Euclidean distance; Select the candidate point with the highest score as the next initial cluster center; Repeat the distance calculation and score calculation steps until the total number of selected initial cluster centers reaches the preset clustering threshold.

[0017] Furthermore, control boundaries are set based on the data distribution within the normal sub-condition model cluster, including: Calculate the mean vector and covariance matrix of all data points within each normal sub-condition model cluster; Based on the mean vector and covariance matrix, a Mahalanobis distance control threshold corresponding to the control boundary is set to define the statistical stability control range of the normal sub-condition model cluster.

[0018] Furthermore, the nearest model cluster to which the real-time data point belongs is determined. If the real-time data point is outside the control boundary of the nearest model cluster, a statistical deviation alarm is triggered, including: Standardize real-time data points; Calculate the Mahalanobis distance of the standardized real-time data points in each normal sub-condition model cluster; The normal sub-condition model cluster with the smallest Mahalanobis distance is determined as the nearest model cluster to which the real-time data point belongs. If the Mahalanobis distance of a real-time data point in its nearest model cluster is greater than the Mahalanobis distance control threshold corresponding to the control boundary set by the nearest model cluster, then the real-time data point is determined to be outside the control boundary, and a statistical deviation alarm signal is generated and output to the control terminal.

[0019] Furthermore, the historical parameters of feed production include: pellet mill spindle speed, pellet mill main unit current, conditioner steam temperature, conditioner steam pressure, feeding auger speed, and steam valve opening.

[0020] Secondly, the present invention provides an abnormal alarm system for feed production control, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned abnormal alarm method for feed production control is implemented.

[0021] The beneficial effects are as follows: This invention pre-eliminates invalid parameters within the prohibited hyperspace through a production process constraint model, integrates spatial distribution characteristics and temporal variation patterns, calculates center weights by combining local density, dispersion, and a state stability metric obtained through inverse processing of the Frobenius norm of the time gradient matrix, and iteratively finds the initial cluster centers using a distance penalty principle. This solves the defects of blind initialization in clustering and its tendency to get trapped in local optima, enabling the division and representation of normal sub-conditions in feed production. A dual monitoring mechanism based on mechanism and data statistics is constructed, which can trigger alarms for malicious boundary violations and identify hidden deviations based on the boundaries set for each normal sub-condition. This reduces false alarms and missed alarms, providing a guarantee for the safe and reliable operation of the feed production process. Attached Figure Description

[0022] Figure 1 A flowchart for an abnormal alarm method in feed production control.

[0023] Figure 2 This is a schematic diagram illustrating the filtration effect of feed production parameters based on production process constraints.

[0024] Figure 3 This is a schematic diagram illustrating the real-time monitoring and alarm triggering effect for abnormal feed production.

[0025] Figure 4 A diagram showing the performance indicators of different feed production anomaly alarm schemes. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] An embodiment of the abnormal alarm method for feed production control provided by the present invention: like Figure 1 As shown, the abnormal alarm method for feed production control includes the following steps: S1 filters historical parameters and calculates fluctuation, density, and dispersion.

[0028] Historical parameter data of feed production is obtained, and parameter points located in the prohibited hyperspace are filtered out based on the production process constraint model. The remaining parameter points constitute a candidate set. For each candidate point in the candidate set, the parameter time series within a preset time window is extracted, and the Frobenius norm of the time gradient matrix is ​​calculated as a measure of state fluctuation. The number of points within a specified neighborhood radius of each candidate point is calculated as the local density, and the Euclidean distance from the candidate point to the nearest point with a local density higher than itself is calculated as the dispersion.

[0029] The data acquisition and monitoring control system reads time-series data of feed production parameters from the historical database and imports historical data, including parameters such as pellet mill spindle speed, pellet mill main unit current, conditioner steam temperature, conditioner steam pressure, feeding auger speed, and steam valve opening. A production process constraint model is constructed, setting multi-dimensional constraint thresholds based on production process rules, equipment rated operating range, and safety operation requirements, including upper and lower temperature limits and pressure non-negativity. The feature vectors of parameter points are input into the constraint model to filter out abnormal data that does not meet the thresholds. The region outside the threshold where abnormal data is located is designated as the prohibited hyperspace. The filtered data points are stored as parameter points in a candidate set list. For each parameter point in the candidate set list, a fixed-duration historical time window centered on that point is extracted, and all historical parameter points within that window are used to form a multi-dimensional time series matrix. Each parameter dimension in the multi-dimensional time series matrix is ​​standardized, and the difference vectors between adjacent moments in the standardized time series matrix are calculated. These difference vectors are then stacked in chronological order to obtain the time gradient matrix.

[0030] Calculate the square root of the sum of squares of all elements in the time gradient matrix, and use the resulting scalar result as the state fluctuation metric for the candidate point. Calculate the pairwise Euclidean distances between all data points in the candidate set, based on the standardized candidate point feature vectors. Generate pairwise distance matrices. Preset the local neighborhood radius. For the current candidate point, traverse the corresponding row in the distance matrix and count the number of points whose distance value is less than the local neighborhood radius; this count is the local density of the candidate point. Among all other candidate points with a local density greater than the current candidate point's local density, find the point closest to the current candidate point. Find the minimum Euclidean distance value of the set of points meeting the conditions in the corresponding row of the distance matrix, and use this minimum distance value as the dispersion of the candidate point. If the current candidate point has no other candidate points with a local density strictly higher than its own, then the maximum Euclidean distance between this point and all other candidate points is set as the dispersion.

[0031] In an optional embodiment, parameter points located in the forbidden hyperspace are filtered out based on the production process constraint model, and the remaining parameter points constitute a candidate set, including: Extract the absolute lower bound threshold and absolute upper bound threshold corresponding to each parameter dimension in the constraint model; Iterate through the data records in the historical parameter data of feed production one by one, and compare the parameter values ​​of each dimension with the corresponding absolute lower limit threshold and absolute upper limit threshold. If any parameter value in any dimension of a data record is less than the absolute lower threshold or greater than the absolute upper threshold, the data record is determined to be within the prohibited hyperspace and is removed. All data records that were not removed were retained to form a candidate set.

[0032] In a real-world IoT monitoring environment for feed production, the number of feature dimensions D is set, typically within the range of 10 to 20. A predefined upper and lower limit threshold matrix is ​​extracted from the configuration table; this matrix sets the constraint model. For example, for the pellet mill main unit current dimension, the preferred absolute lower limit threshold is 0A, and the preferred absolute upper limit threshold is 300A; for the conditioner steam temperature dimension, the absolute limit is 0℃ to 150℃, and for the conditioner steam pressure dimension, the preferred absolute lower limit threshold is 0MPa. N original multidimensional time series records are loaded into memory, and a parallel traversal and comparison operation is initiated.

[0033] During the row-by-row comparison process, conditional judgments are performed on each historical record. If any record contains an abnormal value for any characteristic dimension, such as the conditioner steam temperature reading, granulator spindle speed, or granulator main unit current exceeding the equipment's preset upper or lower limits, the record is determined to fall into the prohibited overspace, and that row of data is removed from the memory database. After the cleaning process is completed, the remaining M valid data records that meet all dimensional constraints are extracted and retained, forming an M-row, D-column candidate set matrix. The filtering effect based on the production process constraint model is as follows: Figure 2 As shown in the figure. The area inside the dashed box represents sample points that meet the full-dimensional constraints, while the area outside the dashed box represents invalid sample points that fall into the prohibited hyperspace.

[0034] In an optional embodiment, for each candidate point in the candidate set, the parameter time series within a preset time window is extracted, and the Frobenius norm of the time gradient matrix is ​​calculated as a measure of state fluctuation, including: Centered on the time of the candidate point, a preset number of sampling points adjacent to the parameter time series are extracted to form a multidimensional parameter time series matrix, and each dimension of the multidimensional parameter time series matrix is ​​standardized. Calculate the time difference vectors of adjacent sampling points in the standardized multidimensional parameter time series matrix, and combine the time difference vectors of all dimensions to form the time gradient matrix; The square root of the sum of squares of the elements of the time gradient matrix is ​​calculated to obtain the Frobenius norm, which is then used as a measure of the state fluctuation of candidate points.

[0035] Set the sampling window size W, preferably within the range of 5 to 15 sampling points. For example, at a sampling frequency of 1 Hz, extract the first 3 and last 3 sampling points centered on the current candidate point time t. Together with time t, these form a time window of W=7. Simultaneously, extract all D parameter dimensions, such as D=15, thereby constructing a 7×15 multidimensional parameter time series matrix. Apply the Z-score standardization formula to each column of the matrix, shifting the mean of each dimension to 0 and scaling the standard deviation to 1, generating a standardized sequence matrix.

[0036] The standardized matrix is ​​subjected to a row-wise successive differencing operation, that is, the standardized row vector of the previous time step is subtracted from the standardized row vector of the next time step to obtain the time difference vector of adjacent sampling points. Since there are 7 rows of data in the window, the differenced data is combined into a 6×15 time gradient matrix. All 90 elements of this gradient matrix are extracted, the sum of squares of each element is calculated, and the square root of the sum is taken to obtain the Frobenius norm. For example, under a certain locally stationary operating condition, the calculated norm is only 1.25, while under the drastic operating condition at startup, the value suddenly increases to 8.60. This calculation process compresses the degree of time series fluctuations in multiple dimensions and times into a scalar value with strong comparability, namely the state fluctuation metric, which reflects the intensity of changes in production parameters within the time window adjacent to the candidate point.

[0037] In an optional embodiment, the number of points within a specified neighborhood radius for each candidate point is calculated as the local density, and the Euclidean distance from the candidate point to the nearest point with a local density higher than its own is calculated as the discreteness, including: The number of other candidate points whose distance from the current candidate point is less than the specified neighborhood radius is used as the local density of the current candidate point. Calculate the Euclidean distance between the current candidate point and all comparison candidate points with a local density higher than the current candidate point, and take the smallest Euclidean distance among them as the dispersion of the current candidate point; For candidate points where there is no local density higher than the candidate point itself, the maximum Euclidean distance between the candidate point and all other candidate points is taken as the discreteness.

[0038] A specified neighborhood radius is determined by sorting all candidate points in ascending order of pairwise Euclidean distances and selecting the values ​​at the top 1% to 2% quantiles of the overall distance distribution as the specified neighborhood radius. Assuming there are 5000 candidate points, the distance value for the top 2% is set to 3.45. When calculating local density, a hard-threshold kernel function is used for traversal statistics. If there are 150 other points within a distance range of 3.45 around the current candidate point, the local density of the current candidate point is marked as 150. This parameter represents the degree of data aggregation between the candidate point and other normal operating condition points in the parameter feature space.

[0039] When determining the dispersion, a retrieval hierarchy is established based on local density. For points whose density is not the highest globally, all candidate points with local densities greater than that point are selected in memory to form a comparison set. The Euclidean distance from the current point to all points in this comparison set is calculated, and the minimum value is assigned to the current point as its dispersion. For example, if the current point has a density of 150, and there are 100 points with global densities higher than 150, the minimum distance between the current point and these 100 points is calculated to be 5.12, so the dispersion is 5.12. For core candidate points in the entire dataset where no local density is strictly higher than their own (e.g., the maximum density is 420), the distance between this point and all other 4999 candidate points is calculated, and the largest Euclidean distance among them is set as the dispersion of that point. This dual-attribute mapping mechanism separates parametric clustering points that possess both local representativeness and global spatial discriminativeness.

[0040] S2, calculate the weights to select cluster centers and construct the working condition model cluster.

[0041] The initial center weights of candidate points are obtained by multiplying the local density, dispersion, and state stability metric obtained by inverse processing based on state fluctuation metric. The candidate point with the highest initial center weight is selected as the first initial cluster center, and subsequent centers are selected iteratively. The selection criterion is to maximize the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers. Based on the determined initial cluster centers, the candidate sets are clustered to construct multiple normal sub-condition model clusters, and the control boundary is set based on the data distribution within the normal sub-condition model clusters.

[0042] For each candidate point, the local density value, dispersion value, and state stability metric are multiplied element-wise to generate a scalar result, which is the comprehensive initial center weight of that candidate point. The initial center weights of all candidate points are queried, and the candidate point with the highest weight is selected as the first initial cluster center and stored in the initial cluster center list. Iterative selection of the remaining cluster centers begins. For each remaining candidate point not yet selected as a cluster center, the Euclidean distance from that point to all existing center points in the initial cluster center list is calculated. The minimum Euclidean distance is obtained, and the square of this minimum distance is calculated based on the standardized candidate point feature vector.

[0043] The initial center weights of the remaining candidate points are multiplied by the square of the minimum distance using a multiplication function. The candidate point with the largest product is selected from all remaining candidate points and added to the initial cluster center list. This calculation and addition process is repeated until the length of the initial cluster center list reaches the preset total number of cluster categories. The KMeans clustering algorithm model from the Scikit-learn library is called, specifying the previously determined initial cluster center list array in the algorithm initialization parameter `init`. Iterative clustering fitting is performed using the standardized candidate set data points. Based on the cluster assignment labels, the candidate set is divided into multiple data subsets, each corresponding to a normal sub-condition model cluster. For each normal sub-condition model cluster, the Mahalanobis distance from all data points within the cluster to the cluster mean vector is calculated. The preset percentile of this distance is calculated as the control boundary threshold. The preset percentile includes the 95th percentile or the 99th percentile; each model cluster has an independent control boundary.

[0044] In an optional embodiment, the local density, dispersion, and state stability metric obtained by inverse processing based on the state fluctuation metric are multiplied to obtain the initial center weights of the candidate points, including: The product of the local density of the current candidate point, the dispersion of the current candidate point, and the state stability metric of the current candidate point is used as the initial center weight of the current candidate point.

[0045] Maintain a one-dimensional floating-point array of length equal to the total number of candidate points to store the center weights. For each candidate point, read in parallel the three scalar features calculated in the previous steps for its placement on the disk: local density, dispersion, and state fluctuation metric. Then, reverse-process the state fluctuation metric to obtain the state stability metric. Before the product operation, the local density, dispersion, and state stability metric can be normalized to ensure they are within a comparable range. Then, perform a product operation on the three values. For example, if a candidate point has a statistically calculated local density of 150, a spatial dispersion distance of 5.12, and a state fluctuation metric norm of 1.25 calculated within the time window, convert the state fluctuation metric to a state stability metric, for example, by using 1 / (1 + state fluctuation metric). Then, combine this with the normalized local density and dispersion for a product operation to obtain the initial center weight of the candidate point.

[0046] After the calculation is completed, the resulting floating-point number is written to the corresponding index position of the array as the initial center weight of the current candidate point. After all candidate points have been traversed and multiplied, all points are sorted in descending order according to the size of the initial center weight stored in the array. Candidate points with exceptionally prominent weights are often data nodes with dense operating conditions, far from other density peaks, and in a relatively stable temporal variation state.

[0047] In an optional embodiment, subsequent centers are iteratively selected based on maximizing the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers, including: Calculate the minimum Euclidean distance from each candidate point in the candidate set that was not selected as an initial cluster center to all selected initial cluster centers; The score is obtained by multiplying the initial center weight of each candidate point that was not selected as the initial cluster center with the square of the corresponding minimum Euclidean distance; Select the candidate point with the highest score as the next initial cluster center; Repeat the distance calculation and score calculation steps until the total number of selected initial cluster centers reaches the preset clustering threshold.

[0048] A pre-defined clustering threshold is set, corresponding to the number of clusters in a typical operating condition of the feed production line, with an optimal range of 3 to 8. Taking a clustering threshold of 5 as an example, an initial set of selected centers is initialized, currently containing only the first initial cluster centers selected based on the maximum initial center weight. For all remaining candidate points in the candidate pool, the Euclidean distances from each point to all known centers in the set are calculated iteratively, and a minimum search is performed, outputting the scalar of the shortest distance from the point to the existing set. Assuming there are two centers in the set at this point, and the distances from a remaining candidate point to them are 4.2 and 8.1 respectively, the minimum value of 4.2 is extracted as the minimum Euclidean distance.

[0049] In the scoring evaluation phase, the initial center weight of each candidate point is extracted. This initial center weight is then multiplied by the square of the minimum Euclidean distance to obtain the score. After one round of comprehensive calculation, the candidate point with the highest overall score is retrieved and added to the selected center set. The size of the selected set is updated, and the Euclidean distance minimization retrieval step is reset. This iterative process of distance calculation and score maximization screening continues until the total number of selected initial cluster centers reaches the preset five. This iterative logic, combining spatial dispersion constraints and positive weight incentives, prevents cluster centers from congregating in the parameter space of a single dominant condition.

[0050] In an optional embodiment, the control boundary is set based on the data distribution within the normal sub-condition model cluster, including: Calculate the mean vector and covariance matrix of all data points within each normal sub-condition model cluster; Based on the mean vector and covariance matrix, a Mahalanobis distance control threshold corresponding to the control boundary is set to define the statistical stability control range of the normal sub-condition model cluster.

[0051] After dividing the candidate set into multiple normal sub-condition model clusters at the selected cluster center, matrix statistical operations are initiated for the k-th cluster. Assuming this cluster contains 15,000 multi-dimensional historical normal data points with a parameter feature dimension of 15, the arithmetic mean of each of the 15,000 points is summed dimension-wise to obtain a 15-dimensional mean vector. Using this mean vector as the reference center, the deviation of each dimension of the samples within this cluster is calculated, and a 15×15 covariance matrix is ​​derived through matrix multiplication. This covariance matrix characterizes the variance and cross-correlation between different parameter dimensions, such as the pellet mill main unit current and the feeding auger speed, and the conditioner steam temperature and the conditioner steam pressure.

[0052] The inverse of the covariance matrix is ​​called, and the Mahalanobis distance formula is substituted to calculate the Mahalanobis distance from each of the 15,000 points within the cluster to the cluster mean vector. All calculated Mahalanobis distances are sorted in ascending order, and preferably truncated using the percentile method. The distance values ​​at the 95% or 99% confidence level are extracted as the Mahalanobis distance control threshold for the normal sub-condition. For example, in a high-load proton-controlled condition cluster, the calculated 99th percentile threshold is 12.4. This process, by incorporating the covariance distribution pattern, successfully transforms the traditional rigid hyperspherical alarm boundary into an ellipsoidal control boundary that conforms to the natural extension direction of the data. Preferably, when the covariance matrix is ​​non-invertible or nearly singular, the covariance matrix with added small regularization terms, or the pseudo-inverse of the covariance matrix, can be used for Mahalanobis distance calculation.

[0053] S3, real-time parameter dual threshold hierarchical alarm determination.

[0054] Obtain real-time production parameters. If the real-time data point is located within the prohibited overspace, trigger an over-limit alarm; otherwise, determine the nearest model cluster to which the real-time data point belongs. If the real-time data point is located outside the control boundary of the nearest model cluster, trigger a statistical deviation alarm.

[0055] The feed production parameters are collected in real time by an industrial programmable logic controller (PLC). The data is received via an interface communication protocol and converted into a real-time data point vector. This real-time data point vector is then substituted into the established production process constraint model for Boolean logic judgment. The Boolean logic judgment is based on the original quantities of the real-time production parameters, ensuring that they are within the same dimensional system as the absolute lower and absolute upper thresholds. If any one-dimensional parameter exceeds the set prior threshold and falls into the prohibited over-space, an alarm interface is invoked to trigger an over-limit alarm, and the subsequent statistical deviation judgment process is terminated. If the real-time data point does not fall into the prohibited over-space, the real-time data point vector is standardized using the standardized parameters saved during the historical modeling stage, and the Euclidean distance from the standardized real-time data point to the center of all constructed normal sub-condition model clusters is calculated.

[0056] Find the cluster center index with the smallest Euclidean distance and determine that the real-time data point belongs to the nearest model cluster with the smallest distance. Calculate the Mahalanobis distance from the real-time data point to the mean vector of the nearest model cluster using the covariance matrix and mean vector obtained from historical data within the nearest model cluster. Compare the calculated Mahalanobis distance with the control boundary threshold set for the nearest model cluster. If the Mahalanobis distance of the real-time data point is strictly greater than the boundary threshold, it is determined that the current operating state has deviated. The message publish / subscribe transmission protocol is invoked to push an abnormal notification to the terminal device of the production line management personnel, triggering a statistical deviation alarm. If the covariance matrix is ​​processed using a regularized inverse matrix or pseudo-inverse matrix during the modeling stage, the corresponding saved regularized inverse matrix or pseudo-inverse matrix is ​​called during the real-time determination stage to calculate the Mahalanobis distance.

[0057] In an optional embodiment, the nearest model cluster to which the real-time data point belongs is determined. If the real-time data point is outside the control boundary of the nearest model cluster, a statistical deviation alarm is triggered, including: Standardize real-time data points; Calculate the Mahalanobis distance of the standardized real-time data points in each normal sub-condition model cluster; The normal sub-condition model cluster with the smallest Mahalanobis distance is determined as the nearest model cluster to which the real-time data point belongs. If the Mahalanobis distance of a real-time data point in its nearest model cluster is greater than the Mahalanobis distance control threshold corresponding to the control boundary set by the nearest model cluster, then the real-time data point is determined to be outside the control boundary, and a statistical deviation alarm signal is generated and output to the control terminal.

[0058] During the real-time monitoring phase of the feed production line, when the PLC of the control system uploads a new real-time data point containing 15 dimensions according to a preset sampling period, such as the feeding auger speed of 30Hz, steam valve opening of 65%, and pellet mill main unit current of 210A, it reads the inverse matrix of the mean vector and covariance matrix of multiple pre-stored model clusters. First, it uses the original real-time data point to perform a no-over-limit judgment; after confirming that no over-limit alarm has been triggered, it generates a standardized real-time data point vector according to the same standardization rules as in the historical modeling phase. This standardized real-time data point vector is then substituted into the distance evaluation function of each cluster to calculate its Mahalanobis distance to, for example, five typical operating condition clusters. The minimum distance is extracted using a minimum-value logic, and the real-time data point is then assigned to the normal sub-operating condition model cluster corresponding to that distance.

[0059] After establishing the nearest model cluster, the Mahalanobis distance control threshold specific to that cluster is retrieved. Assuming the control threshold for this cluster is set previously, a fast comparison calculation is performed. If the real-time distance exceeds the set threshold, it is confirmed that while the current production status generally conforms to this type of operating condition, it has exceeded the allowable ellipsoidal boundary of the statistical stability control range in terms of multidimensional joint correlation. Once an out-of-bounds condition is determined, a statistical deviation alarm signal is immediately generated and packaged. The alarm severity level, the corresponding model cluster number, the real-time parameter vector, and the corresponding Mahalanobis distance value are assembled into a JSON format data packet, output in real-time via the MQTT industrial communication protocol, and pushed to the central control room's large screen or operator terminal for anomaly warning. The real-time monitoring and anomaly alarm effects are as follows: Figure 3 As shown in the figure. The solid line in the figure represents the Mahalanobis distance calculated based on real-time production parameters, and the dashed line represents the preset alarm control threshold for the corresponding operating condition cluster.

[0060] This experiment used 30 consecutive days of real IoT monitoring time-series data from the pelleting workshop of a feed production company for testing. The dataset contained 15 feature dimensions, with core parameters including pellet mill spindle speed, pellet mill main unit current, conditioner steam temperature, conditioner steam pressure, feeding auger speed, and steam valve opening, totaling 130,000 data records. 100,000 records were used as the historical model training set to construct a normal sub-condition model cluster, and the remaining 30,000 records were used as the test set. The test set, annotated by experts, contained 28,000 normal condition data records and 2,000 abnormal fluctuation data records. The hardware environment used a workstation equipped with a 64-core processor and 128GB of memory. The clustering threshold was set to 5, and the sampling window size was 7.

[0061] Four control groups were set up for validation in the experiment. The first group used a traditional basic clustering combined with a fixed threshold alarm model using Euclidean distance, achieving an anomaly detection accuracy of 78.5%, a false alarm rate of 15.2%, and a false negative rate of 18.4%. The second group, based on the first group, used density and dispersion for center initialization and Mahalanobis distance to control the boundary, but did not calculate the state fluctuation metric, improving the accuracy to 89.3%, reducing the false alarm rate to 6.1%, and the false negative rate to 12.5%. The third group included the state fluctuation metric but used traditional Euclidean distance boundary, achieving an accuracy of 87.6%, a false alarm rate of 9.4%, and a false negative rate of 11.2%. The fourth group used the complete scheme presented in this paper, achieving an accuracy of 97.8%, reducing the false alarm rate to 1.3%, and the false negative rate to only 0.9%. Performance metrics of each control group and the complete scheme of this invention are compared. Figure 4 As shown, a state stability metric obtained by inversely derived from a state fluctuation metric is used and explicitly multiplied and fused with local density and dispersion. This avoids the problem of complex weight tuning and allows cluster centers to select representative data nodes with dense operating conditions, high spatial discriminativeness, and relatively stable temporal changes, thus reducing the false alarm rate of abnormal states. Simultaneously, a control boundary constructed based on Mahalanobis distance and percentiles fully utilizes the covariance distribution and cross-correlation between different parameters, transforming the traditional hard boundary into an ellipsoidal control domain that conforms to the natural extension direction of the data. This solves the false alarm problem caused by production condition switching and multi-dimensional joint offset, improving the overall accuracy of production line monitoring.

[0062] An embodiment of the abnormal alarm system for feed production control provided by the present invention: The abnormal alarm system for feed production control includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned abnormal alarm method for feed production control is implemented.

[0063] The abnormal alarm system for feed production control also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0064] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An abnormal alarm method for feed production control, characterized in that, Includes the following steps: S1. Obtain historical parameter data of feed production, filter out parameter points located in the prohibited hyperspace based on the production process constraint model, and form a candidate set of the remaining parameter points; for each candidate point in the candidate set, extract the parameter time series within a preset time window, calculate the Frobenius norm of the time gradient matrix as a measure of state fluctuation; calculate the number of points within a specified neighborhood radius of each candidate point as the local density, and calculate the Euclidean distance from the candidate point to the nearest point with a local density higher than itself as the dispersion. S2, multiply the local density, dispersion, and state stability metric obtained by inverse processing based on state fluctuation metric to obtain the initial center weight of the candidate point; select the candidate point with the highest initial center weight as the first initial cluster center, and iteratively select subsequent centers based on maximizing the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers; based on the determined initial cluster centers, cluster the candidate sets to construct multiple normal sub-condition model clusters, and set control boundaries based on the data distribution within the normal sub-condition model clusters; S3: Obtain real-time production parameters. If the real-time data point is located within the prohibited overspace, trigger an over-limit alarm; otherwise, determine the nearest model cluster to which the real-time data point belongs. If the real-time data point is located outside the control boundary of the nearest model cluster, trigger a statistical deviation alarm.

2. The abnormal alarm method for feed production control according to claim 1, characterized in that, Based on the production process constraint model, parameter points located in the prohibited hyperspace are filtered out, and the remaining parameter points constitute a candidate set, including: Extract the absolute lower bound threshold and absolute upper bound threshold corresponding to each parameter dimension in the constraint model; Iterate through the data records in the historical parameter data of feed production one by one, and compare the parameter values ​​of each dimension with the corresponding absolute lower limit threshold and absolute upper limit threshold. If any parameter value in any dimension of a data record is less than the absolute lower threshold or greater than the absolute upper threshold, the data record is determined to be within the prohibited hyperspace and is removed. All data records that were not removed were retained to form a candidate set.

3. The abnormal alarm method for feed production control according to claim 1, characterized in that, For each candidate point in the candidate set, extract the parameter time series within a preset time window, and calculate the Frobenius norm of the time gradient matrix as a measure of state fluctuation, including: Centered on the time of the candidate point, a preset number of sampling points adjacent to the parameter time series are extracted to form a multidimensional parameter time series matrix, and each dimension of the multidimensional parameter time series matrix is ​​standardized. Calculate the time difference vectors of adjacent sampling points in the standardized multidimensional parameter time series matrix, and combine the time difference vectors of all dimensions to form the time gradient matrix; The square root of the sum of squares of the elements of the time gradient matrix is ​​calculated to obtain the Frobenius norm, which is then used as a measure of the state fluctuation of candidate points.

4. The abnormal alarm method for feed production control according to claim 2, characterized in that, Calculate the number of points within a specified neighborhood radius for each candidate point as the local density, and calculate the Euclidean distance from each candidate point to the nearest point with a higher local density as the discreteness, including: The number of other candidate points whose distance from the current candidate point is less than the specified neighborhood radius is used as the local density of the current candidate point. Calculate the Euclidean distance between the current candidate point and all comparison candidate points with a local density higher than the current candidate point, and take the smallest Euclidean distance among them as the dispersion of the current candidate point; For candidate points where there is no local density higher than the candidate point itself, the maximum Euclidean distance between the candidate point and all other candidate points is taken as the discreteness.

5. The abnormal alarm method for feed production control according to claim 1, characterized in that, The initial center weight of the candidate point is obtained by multiplying the local density, dispersion and the state stability metric obtained by inverse processing based on the state fluctuation metric. This includes multiplying the local density of the current candidate point by the dispersion of the current candidate point and then by the state stability metric of the current candidate point, and the product result is used as the initial center weight of the current candidate point.

6. The abnormal alarm method for feed production control according to claim 1, characterized in that, The subsequent centers are selected iteratively based on maximizing the product of the initial center weight of the candidate point and the square of the minimum distance from the candidate point to all selected initial cluster centers, including: Calculate the minimum Euclidean distance from each candidate point in the candidate set that was not selected as an initial cluster center to all selected initial cluster centers; The score is obtained by multiplying the initial center weight of each candidate point that was not selected as the initial cluster center with the square of the corresponding minimum Euclidean distance; Select the candidate point with the highest score as the next initial cluster center; Repeat the distance calculation and score calculation steps until the total number of selected initial cluster centers reaches the preset clustering threshold.

7. The abnormal alarm method for feed production control according to claim 1, characterized in that, Control boundaries are set based on the data distribution within the normal sub-condition model cluster, including: Calculate the mean vector and covariance matrix of all data points within each normal sub-condition model cluster; Based on the mean vector and covariance matrix, a Mahalanobis distance control threshold corresponding to the control boundary is set to define the statistical stability control range of the normal sub-condition model cluster.

8. The abnormal alarm method for feed production control according to claim 1, characterized in that, Determine the nearest model cluster to which the real-time data point belongs. If the real-time data point is outside the control boundary of the nearest model cluster, a statistical deviation alarm is triggered, including: Standardize real-time data points; Calculate the Mahalanobis distance of the standardized real-time data points in each normal sub-condition model cluster; The normal sub-condition model cluster with the smallest Mahalanobis distance is determined as the nearest model cluster to which the real-time data point belongs. If the Mahalanobis distance of a real-time data point in its nearest model cluster is greater than the Mahalanobis distance control threshold corresponding to the control boundary set by the nearest model cluster, then the real-time data point is determined to be outside the control boundary, and a statistical deviation alarm signal is generated and output to the control terminal.

9. The abnormal alarm method for feed production control according to claim 1, characterized in that, Historical parameters for feed production include: pellet mill spindle speed, pellet mill main unit current, conditioner steam temperature, conditioner steam pressure, feeding auger speed, and steam valve opening.

10. An abnormal alarm system for feed production control, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the abnormal alarm method for feed production control as described in any one of claims 1-9 is implemented.

Images