A data-driven process execution anomaly detection method

Abstract

The application relates to the field of anomaly detection, in particular to a process execution anomaly detection method based on data driving, which comprises the following steps: obtaining a historical reference parameter set by statistically modeling and analyzing historical normal process cycle characteristic data; obtaining a drift compensation vector by robustly estimating and stably evaluating the current batch process characteristic data; obtaining a structure relaxation amplification coefficient by jointly analyzing the local sparsity and the neighborhood eccentricity of the compensated process characteristic data; obtaining a process anomaly clustering result by dynamically and adaptively constructing an equivalent distance of the compensated process characteristic data; and obtaining a device anomaly detection result by comprehensively judging and analyzing the process anomaly clustering result and the drift compensation vector, so as to solve the problem that the existing anomaly detection algorithm based on static distance measurement is prone to false positives and false negatives under the coexistence of global reference drift and internal structure relaxation of the device.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and in particular to a data-driven method for detecting process execution anomalies. Background Technology

[0002] In modern manufacturing, automated equipment, represented by multi-station fixtures, CNC machine tools, and industrial robots, is widely used to perform high-precision, highly repetitive production processes. Taking multi-station rotary fixtures as an example, these devices use servo drives to achieve precise rotation, positioning, and locking of workpieces between different stations, serving as key execution units to ensure stable production cycles and consistent product processing. During equipment operation, multi-source sensor signals such as torque, current, and position of the drive motor can constitute time-series data reflecting the equipment's health status. To achieve real-time monitoring and early fault warning of equipment health status on the production floor, and to avoid processing quality degradation or production line downtime due to equipment malfunctions, existing technologies generally employ data-driven anomaly detection methods. Among these, density-based clustering algorithms (such as the DBSCAN algorithm) are common techniques. The basic idea is to first transform each complete process execution cycle (such as the process of a fixture rotating from one station and positioning to the next) into feature vectors in a feature space; then, using a large number of feature samples from the normal operation phase of the equipment to form high-density core clusters, and calculating the distance between new samples and the core clusters during the online monitoring phase to determine whether the sample belongs to a normal state. A new sample is identified as an anomalous sample when it cannot be classified into any known normal cluster.

[0003] However, the above methods have significant limitations in the applicability of equipment operating continuously for extended periods, such as multi-station fixtures. Firstly, during long-term service, equipment experiences global reference drift, where the overall operating characteristics of the equipment continuously shift in the feature space due to wear of mechanical components, changes in lubrication, etc., causing normal samples to deviate from their initial center. When the detection algorithm still uses a static reference center as the criterion, this overall shift will be misjudged as a large number of abnormal samples, resulting in significant false alarms. Secondly, due to differences in wear rates or stress states at different workstations, the equipment may also experience internal structural relaxation, where individual workstation features gradually deviate from the group center. This early individual deviation is usually small and not easily reflected under standard Euclidean distance metrics, making it easy to overlook and leading to missed detections. Therefore, existing anomaly detection methods based on static Euclidean distance metrics cannot simultaneously address both global reference drift and internal structural relaxation, which are caused by different factors in the feature space. In long-term industrial equipment scenarios, this defect forces the system to weigh false alarms against false negatives when accurately identifying anomalies, which seriously affects the stability and long-term reliability of the detection model and has become a core technical problem that urgently needs to be solved in the field of industrial intelligent monitoring. Summary of the Invention

[0004] In view of this, the present invention aims to propose a data-driven process execution anomaly detection method to solve the problem that existing anomaly detection algorithms based on static distance metrics are prone to false alarms and false negatives when there is both global reference drift and internal structural relaxation of the equipment.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] A data-driven method for detecting process execution anomalies, the method comprising:

[0007] Step S1: Obtain the historical baseline parameter set by performing statistical modeling analysis on the historical normal process cycle characteristic data;

[0008] Step S2: Obtain the drift compensation vector by performing robust center estimation and stability assessment analysis on the process characteristic data of the current batch;

[0009] Step S3: Obtain the structural relaxation amplification factor by performing a joint analysis of local sparsity and neighborhood eccentricity on the compensated process feature data;

[0010] Step S4: Obtain process anomaly clustering results by performing dynamic adaptive equivalent distance construction analysis on the compensated process feature data;

[0011] Step S5: By comprehensively analyzing the clustering results of process anomalies and the drift compensation vector, the equipment anomaly detection results are obtained.

[0012] Furthermore, the step of obtaining a historical baseline parameter set by statistical modeling and analysis of historical normal process cycle characteristic data includes:

[0013] When the equipment is in the initial healthy operation phase, the torque time series data of the drive servo motor is continuously collected, and combined with the motor position encoder signal or external trigger signal, the torque time series data is divided according to the complete process execution cycle to form time series data containing multiple process cycles.

[0014] Feature extraction processing is performed on the time series data of each process cycle to obtain a process cycle feature vector composed of the root mean square torque of the cycle, the kurtosis of the locking stage, and the torque sequence complexity. The root mean square torque of the cycle is used to characterize the overall energy consumption and average load level of the process cycle; the kurtosis of the locking stage is used to characterize the impact characteristics of the positioning and locking action during the process execution; and the torque sequence complexity is used to characterize the dynamic regularity of the torque signal change.

[0015] The feature vectors of all historical normal process cycles are used to form a historical feature set, and historical benchmark parameters are calculated and stored based on the historical feature set. The historical benchmark parameters include the geometric median of historical data, the mean and standard deviation of the distance from each feature vector to the geometric median, the mean and standard deviation of the internal distance standard deviation of each batch divided according to time order, the average value of the relative sparsity index, and the standard deviation of the distance from each feature vector to the median of the set.

[0016] Furthermore, the step of obtaining the drift compensation vector by performing robust center estimation and stability assessment analysis on the current batch process characteristic data includes:

[0017] By performing centripetal force weight calculation on the current batch of process feature data, a robust weighted geometric center is obtained; and by performing internal stability evaluation on the current batch of process feature data, an adaptive learning rate parameter is obtained.

[0018] The drift compensation vector is obtained by dynamically and smoothly updating the weighted geometric center and historical benchmark parameters.

[0019] Furthermore, the process of obtaining a robust weighted geometric center by performing centripetal force weight calculation on the current batch of process feature data, and obtaining an adaptive learning rate parameter by performing internal stability evaluation on the current batch of process feature data, includes:

[0020] For any target feature vector in the current batch process cycle feature data, the numerator is calculated by subtracting the mean distance from the geometric median of the target feature vector to the geometric median of the historical data from the L2 norm of the target feature vector. The denominator is calculated by subtracting the mean distance from the geometric median of each feature vector in the historical data from the L2 norm of the target feature vector. The denominator is calculated by subtracting the mean distance from the geometric median of each feature vector in the historical data from the L2 norm of the target feature vector. The negative of the corresponding fraction is mapped by the Sigmoid function. The centripetal force weight of the target feature vector is calculated by subtracting the constant 1 from the corresponding mapping result.

[0021] All eigenvectors in the current batch are weighted and summed using the centripetal force weights of the eigenvectors. The weighted sum is used as the numerator, and the sum of the centripetal force weights of all eigenvectors is used as the denominator. The resulting fraction is used as the robust weighted geometric center of the current batch.

[0022] The numerator is calculated by subtracting the standard deviation of the Euclidean distance between all feature vectors in the current batch process cycle feature data from the mean of the internal distance standard deviations of each batch in the historical data. The denominator is the standard deviation of the internal distance standard deviations of each batch in the historical data. The corresponding fraction is mapped using the Sigmoid function, and the corresponding mapping result is used as the adaptive learning rate parameter for the current batch.

[0023] Furthermore, the process of obtaining the drift compensation vector by dynamically and smoothly updating the weighted geometric center and historical benchmark parameters includes:

[0024] Set the initial drift compensation vector; subtract the adaptive learning rate parameter of the current batch from the constant 1 as the first historical weight; multiply the first historical weight by the drift compensation vector of the previous time step as the historical drift compensation evaluation; subtract the geometric median of the historical data from the robust weighted geometric center of the current batch as the drift evaluation; multiply the adaptive learning rate parameter of the current batch by the current drift compensation evaluation as the current drift compensation evaluation; add the historical drift compensation evaluation and the current drift compensation evaluation as the drift compensation vector of the current batch.

[0025] Furthermore, the method of obtaining the structural relaxation amplification factor by jointly analyzing the local sparsity and neighborhood eccentricity of the compensated process feature data includes:

[0026] The relative sparsity index is obtained by calculating the local neighborhood distance of the compensated process feature data.

[0027] By performing neighborhood distribution directionality analysis on the compensated process feature data, the centrifugal projection gradient is obtained.

[0028] The structural relaxation amplification factor is obtained by nonlinearly coupling the relative sparsity index with the centrifugal projection gradient.

[0029] Furthermore, the step of obtaining the relative sparsity index by performing local neighborhood distance calculation on the compensated process feature data includes:

[0030] For any target feature vector in the current batch, the result of subtracting the drift compensation vector corresponding to the target time from the target feature vector is used as the compensated process feature data corresponding to the target feature vector;

[0031] Set the number of nearest neighbors; for any target compensated process feature data in the current batch, obtain the set of nearest neighbor data for the target compensated process feature data.

[0032] For any target nearest neighbor data in the nearest neighbor data set, the Euclidean distance between the target compensated process feature data and the target nearest neighbor data is used as the numerator, and the average Euclidean distance between the target compensated process feature data and all data in the nearest neighbor data set is used as the denominator. The resulting fraction is used as the relative sparsity index of the target feature vector.

[0033] Furthermore, the step of obtaining the centrifugal projection gradient by performing neighborhood distribution directionality analysis on the compensated process feature data includes:

[0034] The result of subtracting the corresponding drift compensation vector from the robust weighted geometric center of the current batch is used as the weighted geometric center of the process feature data after compensation.

[0035] For any target feature vector in the current batch, the result of subtracting the weighted geometric center of the compensated process feature data corresponding to the target feature vector from the target feature vector is used as the direction offset vector of the target feature vector.

[0036] For any first target feature vector in the current batch, obtain any second target feature vector in the current batch; subtract the first target feature vector from the second target feature vector as the neighboring gravitational vector of the first target feature vector; take the magnitude of the dot product of the direction offset vector of the first target feature vector and the neighboring gravitational vector as the numerator, take the square of the Euclidean distance between the first target feature vector and the second target feature vector as the denominator, and take the corresponding fraction as the first direction evaluation of the first target feature vector and the second target feature vector; take the sum of the first direction evaluations of the first target feature vector and all second target feature vectors as the local direction consistency evaluation of the first target feature vector.

[0037] The reciprocal of the Euclidean distance between the first target feature vector and the second target feature vector is used as the first center evaluation of the first target feature vector. The sum of the first center evaluations of the first target feature vector and all second target feature vectors is used as the geometric center evaluation of the first target feature vector. The local orientation consistency evaluation of the first target feature vector is used as the numerator, the geometric center evaluation of the first target feature vector is used as the denominator, and the corresponding fraction is used as the eccentric projection gradient of the first target feature vector.

[0038] Furthermore, the method of obtaining the structural relaxation amplification factor by nonlinearly coupling the relative sparsity exponent with the centrifugal projection gradient includes:

[0039] For any target feature vector in the current batch, the calculation result of adding the eccentric projection gradient of the target feature vector to the constant 1 is logarithmically mapped, and the corresponding logarithmic mapping result is used as the first projection gradient evaluation of the target feature vector.

[0040] The average relative sparsity index of all historical data is used as the denominator, and the relative sparsity index of the target feature vector is used as the numerator. The resulting fraction is used as the first sparsity evaluation of the target feature vector.

[0041] The result of multiplying the first projection gradient evaluation of the target feature vector with the first sparsity factor is mapped by the hyperbolic tangent function, and the result of adding the constant 1 to the corresponding mapping result is used as the structural relaxation amplification coefficient of the target feature vector.

[0042] Furthermore, the step of obtaining process anomaly clustering results by performing dynamic adaptive equivalent distance construction analysis on the compensated process feature data includes:

[0043] For any first target feature vector and any second target feature vector in the current batch, the larger of the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector is used as the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector. The Euclidean distance between the first target feature vector and the second target feature vector after being compensated by the drift compensation vector is used as the basic distance evaluation between the first target vector and the second target vector. The result of multiplying the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector by the basic distance evaluation is used as the dynamic adaptive equivalent distance between the first target feature vector and the second target feature vector.

[0044] Set the neighborhood radius and minimum number of points threshold; use the dynamic adaptive equivalent distance between target vectors as the distance evaluation for DBSCAN clustering, and use the corresponding clustering results as the process anomaly clustering results.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] This invention discloses a data-driven process execution anomaly detection method. By introducing a dynamic adaptive benchmark compensation and structural amplification mechanism in the feature space, it achieves accurate modeling and stable monitoring of the long-term operating status of multi-station fixtures. Its core lies in dynamically aligning the true center position of the equipment's operating characteristics using a global drift compensation vector, thereby eliminating the systematic drift effects caused by factors such as mechanical wear and lubrication degradation, ensuring that the anomaly detection results only reflect the true structural changes within the equipment. Furthermore, by constructing an amplification factor based on local sparsity and neighborhood eccentricity, early weak deviation signals are nonlinearly enhanced. This allows the algorithm to capture early anomaly signs with high sensitivity even when facing load differences between workstations or local loosening, significantly improving the robustness and early warning capability of the detection model. In addition, this invention establishes a unified dynamic measurement system at the clustering decision level, incorporating global state and local deviations into the same equivalent distance framework for comprehensive evaluation, effectively solving the problem of difficulty in simultaneously addressing global false alarms and local false negatives in traditional methods. Through this mechanism, the system can not only maintain the stability of the detection boundary in the environment of feature drift, but also adaptively adjust the judgment threshold according to changes in the internal structure of the data, realizing continuous tracking and progressive degradation identification of equipment operating health. This method has high practical value in long-cycle operation scenarios such as multi-station fixtures, CNC machine tools, and industrial robots, and can achieve long-term reliable online anomaly detection and maintenance decision support without adding extra sensors or manual intervention. Attached Figure Description

[0047] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0048] Figure 1 This is a flowchart of a data-driven process execution anomaly detection method according to an embodiment of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] See Figure 1 This is a flowchart of a data-driven process execution anomaly detection method provided in Embodiment 1 of the present invention, as shown below. Figure 1 As shown, a data-driven method for detecting process execution anomalies may include:

[0051] Step S1: Obtain the historical baseline parameter set by performing statistical modeling analysis on the historical normal process cycle characteristic data.

[0052] This invention first performs offline training, extracts key features that can stably characterize the process execution status from a large amount of known historical data of normal equipment operation, and builds a benchmark parameter model for all the parameters required for subsequent online monitoring based on these key features.

[0053] First, from the continuously acquired raw sensor time-series data, independent process execution cycles are identified and segmented. Taking a multi-station fixture as an example, one process cycle corresponds to the complete process of rotating, positioning, and locking from one station to the next. Subsequently, feature engineering is performed on the data of each process cycle. Specifically, when the equipment is in the initial healthy operation phase, the torque time-series data of the drive servo motor is continuously acquired. Combined with the motor position encoder signal or external trigger signal, the torque time-series data is segmented according to the complete process execution cycle, forming time-series data containing multiple process cycles.

[0054] Feature extraction processing is performed on the time series data of each process cycle to obtain a process cycle feature vector composed of the root mean square torque of the cycle, the kurtosis of the locking stage, and the torque sequence complexity. The root mean square torque of the cycle is used to characterize the overall energy consumption and average load level of the process cycle; the kurtosis of the locking stage is used to characterize the impact characteristics of the positioning and locking action during the process execution; and the torque sequence complexity is used to characterize the dynamic regularity of the torque signal change.

[0055] The feature vectors of all historical normal process cycles are used to form a historical feature set, and historical benchmark parameters are calculated and stored based on the historical feature set. The historical benchmark parameters include the geometric median of historical data, the mean and standard deviation of the distance from each feature vector to the geometric median, the mean and standard deviation of the internal distance standard deviation of each batch divided according to time order, the average value of the relative sparsity index, and the standard deviation of the distance from each feature vector to the median of the set.

[0056] This completes the process of obtaining a historical baseline parameter set by statistical modeling and analysis of historical normal process cycle characteristic data.

[0057] Step S2: Obtain the drift compensation vector by performing robust center estimation and stability assessment analysis on the process characteristic data of the current batch.

[0058] The primary problem this invention addresses is that, during long-term operation, multi-station fixtures experience a global baseline drift in their operating characteristics due to physical factors such as overall mechanical wear and deterioration of lubrication. Specifically, in the feature space, this manifests as a slow and continuous shift in the overall position of the data points representing the operating states of all stations, forming a tightly packed group.

[0059] To address this issue, this invention departs from a static historical center and instead constructs a mechanism capable of dynamically tracking the current actual center position of the data group. The core of this mechanism lies in calculating a global drift compensation vector. This drift compensation vector transforms the coordinates of all data points in the current batch within the feature space before any distance calculation; that is, it subtracts the drift vector from the coordinates of each data point. This operation is equivalent to shifting the entire data group, now drifted, back to its historical center region. In this way, subsequent distance calculations are no longer affected by changes in the absolute position of the data group, but only by the relative structure within the group. However, when calculating the center of the current data group, it is necessary to consider that internal structural relaxation may generate early outliers, which can contaminate the center position calculation. Therefore, the center calculation method must be able to eliminate or reduce the influence of these potential outliers.

[0060] In summary, this invention first obtains a robust weighted geometric center by calculating the centripetal force weights of the current batch of process feature data, and then obtains an adaptive learning rate parameter by performing an internal stability evaluation on the current batch of process feature data. Specifically, for any target feature vector in the current batch of process cycle feature data, the numerator is calculated by subtracting the mean distance from the geometric median of the target feature vector to the geometric median of the historical data from the L2 norm of the target feature vector, the standard deviation of the distance from the geometric median of the historical data to the geometric median, and the denominator is calculated by applying a Sigmoid function to the negative of the corresponding fraction. The result of subtracting the constant 1 from the corresponding mapping result is used as the centripetal force weight of the target feature vector. All feature vectors in the current batch are weighted and summed using the centripetal force weights of the feature vectors, and the sum of the centripetal force weights of all feature vectors is used as the denominator. The corresponding fraction is used as the robust weighted geometric center of the current batch.

[0061] In one implementation, assume the current batch of data is... The first in the current batch The eigenvectors are The geometric median of historical data is The mean distance from each eigenvector to the geometric median in the historical data is ; The standard deviation of the distance from each eigenvector to the geometric median in the historical data is: The formula for calculating the robust weighted geometric center of the current batch is:

[0062]

[0063] in, Indicates the robust weighted geometric center of the current batch; Indicates the first in the current batch 1 eigenvector; This represents the data set in the current batch; This represents the geometric median of historical data. This represents the mean distance from each feature vector in historical data to the geometric median; It represents the standard deviation of the distance from each feature vector to the geometric median in historical data.

[0064] The numerator is calculated by subtracting the standard deviation of the Euclidean distance between all feature vectors in the current batch process cycle feature data from the mean of the internal distance standard deviations of each batch in the historical data. The denominator is the standard deviation of the internal distance standard deviations of each batch in the historical data. The corresponding fraction is mapped using the Sigmoid function, and the corresponding mapping result is used as the adaptive learning rate parameter for the current batch.

[0065] After obtaining the robust weighted geometric center and adaptive learning rate parameters of the current batch, the drift compensation vector is obtained by dynamically smoothing and updating the weighted geometric center and historical baseline parameters. Specifically, the initial drift compensation vector is set as a zero vector in this embodiment. The result of subtracting the adaptive learning rate parameter of the current batch from the constant 1 is used as the first historical weight. The result of multiplying the first historical weight with the drift compensation vector of the previous time step is used as the historical drift compensation evaluation. The result of subtracting the geometric median of the historical data from the robust weighted geometric center of the current batch is used as the drift evaluation. The result of multiplying the adaptive learning rate parameter of the current batch with the current drift compensation evaluation is used as the current drift compensation evaluation. The result of adding the historical drift compensation evaluation and the current drift compensation evaluation is used as the drift compensation vector of the current batch.

[0066] In one implementation, it is assumed that the adaptive learning rate for the current batch is... The drift compensation vector at the previous time step is The geometric median of historical data is The expression for calculating the drift compensation vector for the current batch is:

[0067]

[0068] in, This represents the drift compensation vector for the current batch; This represents the adaptive learning rate for the current batch; This represents the drift compensation vector from the previous time step. Indicates the robust weighted geometric center of the current batch; This represents the geometric median of historical data.

[0069] It should be noted that this invention addresses the robustness issue of drift estimation through a robust weighted geometric center approach. Its core lies in the design of the centripetal weights of the eigenvectors, which are based on the geometric center of each eigenvector relative to historical data. The deviation is nonlinearly attenuated. If a feature vector is too far from the historical normal region in the feature space, it is more likely to be an early outlier caused by a loose internal structure. By assigning such points a low weight close to 0, this invention can effectively suppress the adverse effects of these potential outliers when calculating the current robust weighted geometric center, thereby ensuring that the robust weighted geometric center accurately reflects the center position of the mainstream part of the data group and avoiding the pollution of drift estimation by outliers. Secondly, this invention solves the adaptive smoothness problem of drift tracking through adaptive learning rate and dynamic update process. The adaptive learning rate is dynamic, and its value is determined by the internal stability of the current data batch (measured by the internal distance standard deviation). When the internal structure of the current batch is compact and the state is stable, the adaptive learning rate will increase, enabling the system to quickly adopt the new center position and thus respond sensitively to the real global drift. Conversely, when the internal structure of the current batch is loose and the data fluctuates violently, the adaptive learning rate will decrease, and the system will rely more on the drift vector of the previous moment, thus showing stronger stability to the noise and disturbance of the current batch and ensuring the smooth and continuous drift vector update process.

[0070] In summary, the drift compensation vector obtained through the above steps constructs a reference system capable of dynamically tracking equipment performance degradation. In subsequent distance calculations, by subtracting the drift compensation vector from each data point, the global, systematic data drift caused by normal equipment aging can be effectively eliminated. This allows the anomaly detection algorithm to focus on structural changes that truly indicate local faults, thus resolving the false alarm problem caused by global reference drift.

[0071] This completes the process of obtaining the drift compensation vector by performing robust center estimation and stability assessment on the process characteristic data of the current batch.

[0072] Step S3: By performing a joint analysis of local sparsity and neighborhood eccentricity on the compensated process feature data, the structural relaxation amplification factor is obtained.

[0073] After addressing the false alarm issue caused by global reference drift through the drift compensation vector in step S2, all data points are now macroscopically aligned to a unified reference frame. However, another degradation mode of the multi-station fixture's internal structure slacks—the gradual deviation of individual data points from the group due to differentiated wear or uneven load at specific stations—typically exhibits very small displacements in its early stages. The standard Euclidean distance responds linearly to such small displacements, meaning that the calculated distance increment for a data point that has just begun to deviate is insufficient to exceed the neighborhood radius threshold of the DBSCAN algorithm, leading to a false negative.

[0074] To address this missed detection problem, we can no longer rely solely on the original linear distance metric. Instead, we need to selectively amplify early deviation signals and calculate a structural relaxation amplification factor for each data point. This factor dynamically adjusts the equivalent distance between the data point and all other points. This amplification factor is designed to identify points that are leaving the group. Whether a point is leaving the group should be determined by two dimensions: first, whether the density of the local area where the point is located has become sparser compared to the normal state; second, whether the distribution of the point's neighbors shows a significant skew to one side, i.e., whether the point has moved from the inner part of the group to the edge. Only when a point simultaneously meets both the conditions of sparse neighborhood and marginalized location should it be given the maximum distance amplification effect to ensure that the system has the highest sensitivity to true structural deviations while remaining stable against normal random fluctuations within the group.

[0075] In summary, the present invention first obtains the relative sparsity index by performing local neighborhood distance calculation on the compensated process feature data. Specifically, for any target feature vector in the current batch, the result of subtracting the target feature vector from the drift compensation vector corresponding to the target time is used as the compensated process feature data corresponding to the target feature vector.

[0076] The number of nearest neighbors is set. In this embodiment of the invention, the number of nearest neighbors is set to 4. The number of nearest neighbors can be adjusted according to the actual situation and is not required. For any target compensated process feature data in the current batch, the set of nearest neighbor data of the target compensated process feature data is obtained. For any target nearest neighbor data in the set of nearest neighbor data, the Euclidean distance between the target compensated process feature data and the target nearest neighbor data is used as the numerator, and the average Euclidean distance between the target compensated process feature data and all data in the set of nearest neighbor data is used as the denominator. The corresponding fraction is used as the relative sparsity index of the target feature vector.

[0077] After obtaining the relative sparsity exponent of the target feature vector, the centrifugal projection gradient is obtained by performing neighborhood distribution directionality analysis on the compensated process feature data. Specifically, the weighted geometric center of the current batch is subtracted from the corresponding drift compensation vector as the weighted geometric center of the compensated process feature data. For any target feature vector in the current batch, the weighted geometric center of the compensated process feature data is subtracted from the weighted geometric center of the compensated process feature data corresponding to the target feature vector as the direction offset vector of the target feature vector. For any first target feature vector in the current batch, any second target feature vector in the current batch is obtained. The result of subtracting the first target feature vector from the second target feature vector is used as the neighbor attraction vector of the first target feature vector. The direction offset vector of the first target feature vector is then compared with the neighbor attraction vector. The magnitude of the dot product of the quantities is used as the numerator, and the square of the Euclidean distance between the first target feature vector and the second target feature vector is used as the denominator. The resulting fraction is used as the first direction evaluation of the first target feature vector and the second target feature vector. The sum of the first direction evaluations of the first target feature vector and all second target feature vectors is used as the local direction consistency evaluation of the first target feature vector. The reciprocal of the Euclidean distance between the first target feature vector and the second target feature vector is used as the first center evaluation of the first target feature vector. The sum of the first center evaluations of the first target feature vector and all second target feature vectors is used as the geometric center evaluation of the first target feature vector. The local direction consistency evaluation of the first target feature vector is used as the numerator, and the geometric center evaluation of the first target feature vector is used as the denominator. The resulting fraction is used as the eccentric projection gradient of the first target feature vector.

[0078] After obtaining the relative sparsity index and eccentric projection gradient of the target feature vector, the structural relaxation amplification coefficient is finally obtained by nonlinearly coupling the relative sparsity index and eccentric projection gradient. Specifically, for any target feature vector in the current batch, the calculation result of adding the eccentric projection gradient of the target feature vector to a constant 1 is logarithmically mapped, and the corresponding logarithmic mapping result is used as the first projection gradient evaluation of the target feature vector; the average relative sparsity index of all data in the historical data is used as the denominator, and the relative sparsity index of the target feature vector is used as the numerator, and the corresponding fraction is used as the first sparsity evaluation of the target feature vector; the calculation result of multiplying the first projection gradient evaluation of the target feature vector with the first sparsity index is mapped using a hyperbolic tangent function, and the calculation result of adding the constant 1 to the corresponding mapping result is used as the structural relaxation amplification coefficient of the target feature vector.

[0079] In one implementation, assume that the first in the current batch The relative sparsity index of the eigenvectors is The first in the current batch The gradient of the eccentric projection of the eigenvectors is The average relative sparsity index of all data in the historical data is: Then the first The expression for calculating the structural relaxation amplification factor of each eigenvector is:

[0080]

[0081] in, Indicates the first The structural relaxation amplification factor of each eigenvector; Indicates the first in the current batch The eccentric projection gradient of each eigenvector; Indicates the first in the current batch The relative sparsity index of each eigenvector; This represents the average relative sparsity index of all data in historical data. Represents the hyperbolic tangent function; This represents the logarithmic function with the natural constant as the base.

[0082] It's important to note that the structural relaxation amplification factor evaluates the isolation level of each eigenvector and transforms the evaluation result into a non-linear distance amplification factor, thus addressing the issue of insensitivity in detecting early internal structural relaxation. This design doesn't rely on a single dimension but rather uses the multiplicative coupling of the relative sparsity index and the centrifugal projection gradient to jointly determine the deviation of a point from both density and structural perspectives. Specifically, in the operation scenario of a multi-station fixture, when the positioning mechanism of a specific station (e.g., station 3) begins to show early wear, its corresponding eigenvector will gradually deviate from the dense data cloud composed of other healthy stations. At this time, the calculated relative sparsity index will respond to this change first: as the eigenvector corresponding to station 3 begins to move away from its nearest neighbor eigenvectors, its own nearest neighbor distance will increase. More importantly, the relative sparsity index of station 3 forms a relative metric by comparing it with the average nearest neighbor distance of its neighbors. If an eigenvector and its nearest neighbor eigenvectors are in a sparse region, the relative sparsity index will be close to 1; however, if a point becomes sparse while its neighbors remain dense (i.e., it is moving away from a dense cluster), the relative sparsity index will be significantly greater than 1. However, local density changes alone are insufficient for the most robust judgment. Therefore, this invention further introduces centrifugal projection gradient evaluation. The numerator of the centrifugal projection gradient quantifies the consistency between the eccentric vector of the eigenvector and the gravitational direction of its neighbors through vector dot product. In a real-world scenario, when the eigenvector at workstation 3 moves to the edge of the data cloud, the vast majority of its neighbors will be located on the side facing the center of the group. This eccentric distribution of neighbors leads to a large cumulative value in the numerator of the centrifugal projection gradient at workstation 3, significantly increasing the gradient and transforming the concept of whether a eigenvector is at the edge of the group into a calculable numerical index. Ultimately, this invention ensures that the strongest amplification effect is applied only to critical points that simultaneously satisfy relative sparsity and structural marginalization by multiplicatively coupling the relative sparsity exponent and the eccentric projection gradient and using function mapping for smoothing. This design can accurately transform a slight feature deviation due to a specific workstation into a significantly amplified equivalent distance, making that point easily identified as noise in subsequent DBSCAN clustering. This enables early warning during the fault development stage and solves the problem of missed detections caused by internal structural relaxation.

[0083] Thus, the structural relaxation amplification factor was obtained by jointly analyzing the local sparsity and neighborhood eccentricity of the compensated process feature data.

[0084] Step S4: Obtain process anomaly clustering results by performing dynamic adaptive equivalent distance construction analysis on the compensated process feature data.

[0085] In the DBSCAN clustering algorithm, it is necessary to determine whether any two data points are neighbors. In calculating the distance between data points, firstly, the feature vectors corresponding to the two data points are translated using a drift compensation vector to eliminate the common displacement effect caused by global drift. Then, the Euclidean distance between the compensated feature vectors is weighted and amplified using the structural relaxation amplification coefficients of each feature vector. Specifically, for any first target feature vector and any second target feature vector in the current batch, the larger of the structural relaxation amplification coefficients of the first and second target feature vectors is used as the structural relaxation amplification coefficient between the first and second target feature vectors. The Euclidean distance between the first and second target feature vectors after drift compensation is used as the basic distance evaluation between the first and second target feature vectors. The result of multiplying the structural relaxation amplification coefficients of the first and second target feature vectors by the basic distance evaluation is used as the dynamic adaptive equivalent distance between the first and second target feature vectors.

[0086] Set the neighborhood radius and minimum number of points threshold; use the dynamic adaptive equivalent distance between target vectors as the distance evaluation for DBSCAN clustering, and use the corresponding clustering results as the process anomaly clustering results.

[0087] It should be noted that in this embodiment of the invention, the neighborhood radius is set using a historical normal dataset. The average Euclidean distance between all feature vectors in the historical dataset and the nearest neighbor dataset is calculated, and the average Euclidean distance value of the 95th percentile is selected as the neighborhood radius. At the same time, the minimum number of points threshold is set to 6, which is twice the dimension of the feature vector. The parameter settings can be adjusted according to the actual scenario and are not required.

[0088] Thus, the analysis of dynamic adaptive equivalent distance construction based on the compensated process feature data was completed, and the results of process anomaly clustering were obtained.

[0089] Step S5: By comprehensively judging and analyzing the clustering results of process anomalies and the drift compensation vector, the equipment anomaly detection results are obtained.

[0090] In the final stage of this invention, a two-dimensional comprehensive evaluation of the equipment's operating status is performed to output the final anomaly detection result. This step is based on the structural anomaly detection result and the global degradation evaluation result, corresponding to two anomaly modes: local sudden failure and overall performance degradation, respectively. A comprehensive determination of the equipment's health status is achieved through parallel judgment and comprehensive decision-making.

[0091] In each monitoring cycle of equipment operation, the system first executes the structural anomaly judgment path. This judgment path, based on the process anomaly clustering results obtained through dynamic adaptive equivalent distance clustering analysis, performs a structural consistency check on the process cycle characteristic data of the current batch. When feature vectors marked as noise points appear in the clustering results, it indicates that, under the stable reference frame after drift compensation, there is a deviation in process execution caused by local structural changes or sudden failures. Such deviations are usually reflected in the fact that the operating status of individual workstations is significantly different from the characteristic distribution of their group, and the system thus judges it as a structural anomaly.

[0092] Simultaneously, the system executes a global degradation judgment path. This judgment path uses the magnitude of the drift compensation vector as the core indicator to reflect the cumulative deviation of the overall operating characteristics of the equipment relative to historical health benchmarks. In each monitoring cycle, the system calculates the magnitude of the current drift compensation vector and compares this magnitude with a preset global degradation alarm threshold. When the magnitude of the drift compensation vector exceeds this threshold, it indicates that the equipment has experienced significant overall performance degradation during long-term operation, and is judged as a global degradation anomaly. The global degradation alarm threshold is set based on the statistical distribution characteristics of historical normal operation data. Specifically, it is calculated during the offline training phase using the standard deviation of the distance from all historical feature vectors to their geometric median as a basic measure, and then multiplied by a confidence factor commonly used in industrial statistical control (preferably six times) to form the alarm threshold, thereby ensuring high-confidence identification of severe degradation trends.

[0093] After independently determining the two types of anomalies mentioned above, the system performs comprehensive logical operations on the judgment results to obtain the final anomaly detection output. If neither structural anomaly nor global degradation anomaly occurs, it indicates that the equipment is operating within the normal range; if only structural anomaly exists, a local process fault alarm is output, and the workstation number corresponding to the anomaly feature vector can be identified simultaneously for maintenance location; if only global degradation anomaly exists, an equipment performance degradation warning is output, prompting maintenance personnel to perform planned maintenance; if both occur simultaneously, the highest-level composite anomaly alarm is output, indicating that there is a cumulative risk to the equipment status, and immediate intervention measures are required.

[0094] Through the aforementioned dual-path parallel judgment and integrated decision-making mechanism, this invention achieves comprehensive identification of equipment anomalies. This method can not only effectively detect structural anomalies caused by local process disturbances or mechanical failures, but also simultaneously monitor global degradation trends caused by long-term wear and performance decline. Thus, while eliminating false alarms caused by normal aging drift, it improves sensitivity to early latent faults, achieving high-precision judgment and long-term stable monitoring of complex equipment anomaly modes.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven procedure execution anomaly detection method, characterized by, The method includes: Step S1: Obtain the historical baseline parameter set by performing statistical modeling analysis on the historical normal process cycle characteristic data; Step S2: Obtain the drift compensation vector by performing robust center estimation and stability assessment analysis on the process characteristic data of the current batch; Step S3: Obtain the structural relaxation amplification factor by performing a joint analysis of local sparsity and neighborhood eccentricity on the compensated process feature data; Step S4: Obtain process anomaly clustering results by performing dynamic adaptive equivalent distance construction analysis on the compensated process feature data; Step S5: By comprehensively judging and analyzing the clustering results of process anomalies and the drift compensation vector, the equipment anomaly detection results are obtained; The step of obtaining a drift compensation vector by performing robust center estimation and stability assessment analysis on the current batch of process feature data includes: obtaining a robust weighted geometric center by performing centripetal force weight calculation on the current batch of process feature data, obtaining an adaptive learning rate parameter by performing internal stability assessment on the current batch of process feature data, and obtaining a drift compensation vector by performing dynamic smoothing update processing on the weighted geometric center and historical benchmark parameters. The method of obtaining the structural relaxation amplification coefficient by performing joint analysis of local sparsity and neighborhood eccentricity on the compensated process feature data includes: obtaining the relative sparsity index by performing local neighborhood distance calculation on the compensated process feature data; obtaining the centrifugal projection gradient by performing neighborhood distribution directionality analysis on the compensated process feature data; and obtaining the structural relaxation amplification coefficient by performing nonlinear coupling processing on the relative sparsity index and the centrifugal projection gradient.

2. The data-driven process execution anomaly detection method of claim 1, wherein, The process involves statistical modeling and analysis of historical normal process cycle characteristic data to obtain a historical baseline parameter set, including: When the equipment is in the initial healthy operation phase, the torque time series data of the drive servo motor is continuously collected, and combined with the motor position encoder signal or external trigger signal, the torque time series data is divided according to the complete process execution cycle to form time series data containing multiple process cycles. Feature extraction processing is performed on the time series data of each process cycle to obtain a process cycle feature vector composed of the root mean square torque of the cycle, the kurtosis of the locking stage, and the torque sequence complexity. The root mean square torque of the cycle is used to characterize the overall energy consumption and average load level of the process cycle; the kurtosis of the locking stage is used to characterize the impact characteristics of the positioning and locking action during the process execution; and the torque sequence complexity is used to characterize the dynamic regularity of the torque signal change. The feature vectors of all historical normal process cycles are used to form a historical feature set, and historical benchmark parameters are calculated and stored based on the historical feature set. The historical benchmark parameters include the geometric median of historical data, the mean and standard deviation of the distance from each feature vector to the geometric median, the mean and standard deviation of the internal distance standard deviation of each batch divided according to time order, the average value of the relative sparsity index, and the standard deviation of the distance from each feature vector to the median of the set.

3. The data-driven process execution anomaly detection method of claim 1, wherein, The process involves calculating centripetal force weights on the current batch of process feature data to obtain a robust weighted geometric center, and performing internal stability evaluation on the current batch of process feature data to obtain an adaptive learning rate parameter, including: For any target feature vector in the current batch process cycle feature data, the numerator is calculated by subtracting the mean distance from the geometric median of the target feature vector to the geometric median of the historical data from the L2 norm of the target feature vector. The denominator is calculated by subtracting the mean distance from the geometric median of each feature vector in the historical data from the L2 norm of the target feature vector. The denominator is calculated by subtracting the mean distance from the geometric median of each feature vector in the historical data from the L2 norm of the target feature vector. The negative of the corresponding fraction is mapped by the Sigmoid function. The centripetal force weight of the target feature vector is calculated by subtracting the constant 1 from the corresponding mapping result. All eigenvectors in the current batch are weighted and summed using the centripetal force weights of the eigenvectors. The weighted sum is used as the numerator, and the sum of the centripetal force weights of all eigenvectors is used as the denominator. The resulting fraction is used as the robust weighted geometric center of the current batch. The numerator is calculated by subtracting the standard deviation of the Euclidean distance between all feature vectors in the current batch process cycle feature data from the mean of the internal distance standard deviations of each batch in the historical data. The denominator is the standard deviation of the internal distance standard deviations of each batch in the historical data. The corresponding fraction is mapped using the Sigmoid function, and the corresponding mapping result is used as the adaptive learning rate parameter for the current batch.

4. The data-driven process execution anomaly detection method of claim 1, wherein, The process of obtaining the drift compensation vector by dynamically and smoothly updating the weighted geometric center and historical benchmark parameters includes: Set the initial drift compensation vector; subtract the adaptive learning rate parameter of the current batch from the constant 1 as the first historical weight; multiply the first historical weight by the drift compensation vector of the previous time step as the historical drift compensation evaluation; subtract the geometric median of the historical data from the robust weighted geometric center of the current batch as the drift evaluation; multiply the adaptive learning rate parameter of the current batch by the current drift compensation evaluation as the current drift compensation evaluation; add the historical drift compensation evaluation and the current drift compensation evaluation as the drift compensation vector of the current batch.

5. The data-driven process execution anomaly detection method according to claim 1, characterized in that, The process of obtaining a relative sparsity index by calculating local neighborhood distances on the compensated process feature data includes: For any target feature vector in the current batch, the result of subtracting the drift compensation vector corresponding to the target time is used as the compensated process feature data corresponding to the target feature vector; Set the number of nearest neighbors; for any target compensated process feature data in the current batch, obtain the set of nearest neighbor data for the target compensated process feature data. For any target nearest neighbor data in the nearest neighbor data set, the Euclidean distance between the target compensated process feature data and the target nearest neighbor data is used as the numerator, and the average Euclidean distance between the target compensated process feature data and all data in the nearest neighbor data set is used as the denominator. The resulting fraction is used as the relative sparsity index of the target feature vector.

6. The data-driven process execution anomaly detection method of claim 1, wherein, The step of obtaining the centrifugal projection gradient by performing neighborhood distribution directionality analysis on the compensated process feature data includes: The result of subtracting the corresponding drift compensation vector from the robust weighted geometric center of the current batch is used as the weighted geometric center of the process feature data after compensation. For any target feature vector in the current batch, the result of subtracting the weighted geometric center of the compensated process feature data corresponding to the target feature vector from the target feature vector is used as the direction offset vector of the target feature vector. For any first target feature vector in the current batch, obtain any second target feature vector in the current batch; subtract the first target feature vector from the second target feature vector as the neighboring gravitational vector of the first target feature vector; take the magnitude of the dot product of the direction offset vector of the first target feature vector and the neighboring gravitational vector as the numerator, take the square of the Euclidean distance between the first target feature vector and the second target feature vector as the denominator, and take the corresponding fraction as the first direction evaluation of the first target feature vector and the second target feature vector; take the sum of the first direction evaluations of the first target feature vector and all second target feature vectors as the local direction consistency evaluation of the first target feature vector. The reciprocal of the Euclidean distance between the first target feature vector and the second target feature vector is used as the first center evaluation of the first target feature vector. The sum of the first center evaluations of the first target feature vector and all second target feature vectors is used as the geometric center evaluation of the first target feature vector. The local orientation consistency evaluation of the first target feature vector is used as the numerator, the geometric center evaluation of the first target feature vector is used as the denominator, and the corresponding fraction is used as the eccentric projection gradient of the first target feature vector.

7. The data-driven process execution anomaly detection method of claim 1, wherein, The process of obtaining the structural relaxation amplification factor by nonlinearly coupling the relative sparsity index with the centrifugal projection gradient includes: For any target feature vector in the current batch, the calculation result of adding the eccentric projection gradient of the target feature vector to the constant 1 is logarithmically mapped, and the corresponding logarithmic mapping result is used as the first projection gradient evaluation of the target feature vector. The average relative sparsity index of all historical data is used as the denominator, and the relative sparsity index of the target feature vector is used as the numerator. The resulting fraction is used as the first sparsity evaluation of the target feature vector. The result of multiplying the first projection gradient evaluation of the target feature vector with the first sparsity factor is mapped by the hyperbolic tangent function, and the result of adding the constant 1 to the corresponding mapping result is used as the structural relaxation amplification coefficient of the target feature vector.

8. The data-driven process execution anomaly detection method of claim 1, wherein, The process involves dynamically adaptively constructing and analyzing equivalent distances on the compensated process feature data to obtain process anomaly clustering results, including: For any first target feature vector and any second target feature vector in the current batch, the larger of the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector is used as the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector. The Euclidean distance between the first target feature vector and the second target feature vector after being compensated by the drift compensation vector is used as the basic distance evaluation between the first target vector and the second target vector. The result of multiplying the structural relaxation amplification coefficients of the first target feature vector and the second target feature vector by the basic distance evaluation is used as the dynamic adaptive equivalent distance between the first target feature vector and the second target feature vector. Set the neighborhood radius and minimum number of points threshold; use the dynamic adaptive equivalent distance between target vectors as the distance evaluation for DBSCAN clustering, and use the corresponding clustering results as the process anomaly clustering results.

Images