A process parameter anomaly detection method and system

Abstract

The present application relates to a kind of process parameter anomaly detection method and system, method includes the following steps: step one, process parameter collection;Step two, feature parameter extraction;Step three, multiple weight value anomaly determination;Step four, envelope dynamic line drawing;The present application can release data analysis post personnel from the heavy curve analysis, with the example of tightening curve, manual analysis of the past, only can carry out spot check to data every day, there is the problem of missing detection, real-time is poor, and needs to consume a lot of time for visual analysis.The present application can realize the online real-time detection of all production line tightening process parameter data, timely alarm push once data anomaly is found;Whether from the real-time of data monitoring, or accuracy, there is significant improvement.

Description

Technical Field

[0001] This invention relates to a method and system for detecting abnormal process parameters, specifically a method and system for detecting abnormal process parameters based on unsupervised technology. Background Technology

[0002] For determining process parameter anomalies, taking tightening data anomaly detection as an example, the industry currently has two main technical approaches:

[0003] One type of approach is triggered from a business perspective, mining features from standard data. Data exceeding these features is identified as anomalous. For example, an envelope curve is plotted using standard data, and the difference between the actual business curve and the envelope curve is compared to determine if the data is abnormal. Anomalies can also be identified by the relationship between tightening angles and tightening torques at different stages. This type of technical solution is described in patents such as "A Method for Generating and Applying the Envelope of Press-fitting and Tightening Curves" (Publication No.: CN201811532029.9) and "A Method for Anomaly Analysis of Screw Tightening Curves" (Publication No.: CN202211338334.0). While this technology fully utilizes process-related knowledge, it is also limited by the application of specialized knowledge. It requires pre-analysis and feature mining of a large number of curves and can only be applied to specific scenarios. Once the scenario changes, feature mining and analysis need to be performed again, resulting in significant limitations in fault tolerance and real-time performance.

[0004] Another approach is based on mathematical and statistical methods. It involves first building a model using a batch of standard curves, and then using the standard model to identify anomalies in the data. For example, in the patent "Construction of Automotive Bolt Tightening Curve Reconstruction Model and Tightening Quality Identification Method" (CN202311027177.6), this method first acquires tightening curve data for different types of bolts under different conditions (normal and abnormal), labels the data, and then inputs the tightening curves into an artificial neural network model for training. The trained model is then applied to anomaly curve detection. A similar approach is mentioned in the patent "Tightening Anomaly Identification Method, Device, and Medium Based on Dynamic Time Planning" (CN202310511106.7). All of these technical solutions are supervised modeling methods, requiring the collection of large amounts of data for each scenario and the labeling of normal / abnormal data, resulting in a massive workload. Taking tightening data as an example, a production line often has dozens of workstations corresponding to hundreds of tightening shafts. It requires comparing the data for each shaft one by one and manually filtering for abnormal / normal curves, resulting in a huge workload and high labor costs. If this is extended to scenarios such as press-fitting and leak testing, the cost will increase dramatically. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the above-mentioned problems existing in the prior art and to provide a method and system for detecting process parameter anomalies based on unsupervised technology.

[0006] This invention achieves unsupervised anomaly identification of process parameters through dynamic envelope generation and joint anomaly detection. It eliminates the need for manual analysis and observation of curves, as well as the need to pre-collect large amounts of labeled data for modeling and training, thus realizing intelligent identification of the entire process parameter curve.

[0007] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0008] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0009] A method for detecting abnormal process parameters includes the following steps:

[0010] Step 1: Process parameter collection;

[0011] Step 2: Feature parameter extraction;

[0012] Step 3: Determine the anomalies of multiple weights;

[0013] Step 4: Dynamically draw the envelope;

[0014] Step 5: Determine abnormal results.

[0015] Furthermore, the process parameter collection in step one includes the following:

[0016] The raw data of the production line process parameters is processed to generate an n×(m*k) matrix, and then stored in the data dictionary dict(.) using the axis number as the index, in the following form:

[0017]

[0018] Assuming there are m tightening data lines, each with a length of n, and k axes, then the total number of data lines for each workstation is m*k, a m(n*k) This represents the data in the m-th row and n*k-th column.

[0019] Furthermore, the feature parameter extraction in step two specifically includes:

[0020] Locate and truncate data points, and calculate the K value for each process parameter curve using the sliding window method to perform precise feature extraction; including:

[0021] (1) Process parameter truncation;

[0022] (2) K-value feature extraction;

[0023] Furthermore, the truncation of the process parameters specifically includes:

[0024] 1) For tightening curve data dict(k) with a total number of axes k, calculate the standard deviation of all tightening data at each tightening angle, and plot the standard deviation curve. For dict(k), the standardization formula for the j-th data axis and the t-th data point is as follows: is the sum of the standard deviation and the mean of the data:

[0025]

[0026] st j (t) is the sum of the standard deviation and the mean. The mean of all curves at point t.

[0027] a ti Equivalent to the elements in the dict matrix; m is the number of curves. If there are m curves, then there are m data points.

[0028] For process parameters with k axes and n data lengths, the standard deviation is calculated for each axis and each tightening angle, resulting in the following standard deviation sequence:

[0029] ST(j)={st j (1),st j (2),st j (3)....,st j (n)}

[0030] ST(j) represents the sequence of all curves on the j-th axis calculated by standard deviation plus mean;

[0031] 2) Perform second-order differencing on all standard deviation sequence data:

[0032]

[0033] This yields the difference sequence WT(j).

[0034]

[0035] t represents the t-th data point of the curve;

[0036] 3) Calculate the 1 σ value P(j) of the difference sequence WT. This value reflects the overall fluctuation of the difference sequence. Points exceeding this value are the nodes where the tightening torque begins to change.

[0037]

[0038] For the sequence WT(j), find the position t of the first point greater than P(j), and use it as the cutoff point. The calculation formula is as follows:

[0039]

[0040] The process parameters are truncated at position t to obtain the data to be processed.

[0041] Furthermore, the K-value feature extraction specifically includes:

[0042] 1) Define a sliding window of length u, where u takes values ​​that are n / 50 of the total data length. Starting from the cutoff point, the window slides gradually to the right, and within each window j:[a j ,a j+1 ...a j+u Find the maximum point w within the range. (j) max ;

[0043]

[0044] 2) Further obtain the maximum value w within the window. (j) max The two points that are closest, at 30% and 70%, are defined as... The corresponding x-coordinate is The slope of the line connecting the two points is calculated as the K value as follows:

[0045]

[0046] k j This represents the slope of the line connecting the 30% and 70% points of the j-th window;

[0047] Finally, the sequence of K-value points under all sliding windows is obtained;

[0048] K = [k t ,k t+1 ...k n-u ].

[0049] Furthermore, the multi-weighted anomaly detection in step three is specifically as follows:

[0050] (1) Sampling;

[0051] For each sequence of K values ​​obtained along axis j, a random sampling method is used, with a random sampling length not exceeding [a certain value]. The data E is given, where m is the length of the K sequence data; one set of sampled data is generated each time, until all feature points are included:

[0052] K j =E2∪E2∪..∪E i

[0053] E represents the sampled data set, E i This represents the data set from the i-th sample.

[0054] (2) Unsupervised testing;

[0055] A basic anomaly detection model is selected and defined as model f; for the sampling feature set of the data to be detected, E1E2...E n The data is sequentially input into model f. For the j-th sampled dataset of the i-th process parameter data, the detection result is obtained.

[0056]

[0057] in

[0058]

[0059] (3) Result classification;

[0060] The formula for determining potential anomaly curves and standard curves is as follows:

[0061]

[0062] If Pre (i) If |E| < 0, then the parameter curve is a potential outlier curve and is classified into dataset A1; otherwise, it is a standard curve and is classified into dataset B1; where |E| < 0. j |For feature set E j The length represents The larger the corresponding weight, the stronger the weight. The more authoritative the judgment, the more accurate the judgment of the same model in multiple weight values ​​of process parameters.

[0063] Furthermore, the drawing of the envelope dynamic line in step four is as follows:

[0064] Based on the standard curve family B, the upper and lower envelopes are calculated using the following formula:

[0065]

[0066] u (j) Let represent the mean and standard deviation of all standard curves calculated for the j-th axis, respectively;

[0067] in, Let represent the upper and lower envelopes of the data on the j-th axis, respectively; the upper and lower envelopes fluctuate around the mean μ and standard deviation σ of the data with a coefficient of ±λ; λ is the smoothing coefficient, as shown in the following formula:

[0068] λ=ln(x) / (1+e -x )

[0069] Where x is the standardized value of the x-coordinate of a data set of length n, as shown in the following formula:

[0070]

[0071] Where kd is the self-adjusting coefficient, which defaults to 50. The larger kd is, the larger λ is. Based on the change of λ, the envelope can be dynamically drawn.

[0072] Furthermore, the abnormal result determination in step five is as follows:

[0073] (1) Overflow point calculation;

[0074] For a dataset with axis j, based on the upper and lower envelopes Determine the sum of the number of data points that overflow the upper and lower envelopes for each process parameter curve. e When sum e If the curve exceeds a given threshold, it is identified as an abnormal curve and classified into dataset A2; sum e The calculation formula is as follows, where y represents the corresponding data point:

[0075]

[0076] (2) Classification of abnormal data;

[0077] For the abnormal data A1 and A2 obtained respectively, they are divided into 3 priority categories and dispatched for processing according to the priority:

[0078] 1) Similar to anomalous data, this type of data represents the intersection of anomalous data and belongs to the highest priority data for processing:

[0079] Class 1 = {A1 ∩ A2}

[0080] 2) Type 2 abnormal data: This type of data is overflow envelope data, but it does not belong to A1 and belongs to secondary priority data.

[0081] Class 2 = {A2-A1}

[0082] 3) Three types of anomalous data: This type of data does not overflow the envelope, but belongs to A1, which is the lowest priority data to process.

[0083] Category 3 = {A1-A2}.

[0084] A process parameter anomaly detection system, comprising:

[0085] The parameter collection module is used for collecting process parameters.

[0086] The parameter extraction module is used for feature parameter extraction.

[0087] The judgment module is used for multi-weighted anomaly detection;

[0088] The drawing module is used for dynamic drawing of the envelope.

[0089] The result determination module is used for determining abnormal results.

[0090] Furthermore, the parameter extraction module includes a process parameter truncation unit and a feature extraction unit; the process parameter truncation unit is used to truncate the process parameters to obtain the data to be processed; the feature extraction unit is used for K-value feature extraction to obtain the K-value point sequence under all sliding windows;

[0091] Compared with the prior art, the beneficial effects of the present invention are:

[0092] This invention frees data analysts from tedious curve analysis. Taking tightening curves as an example, traditional manual analysis only allows for daily spot checks, leading to missed detections, poor real-time performance, and significant time-consuming visualization analysis. This invention enables online real-time monitoring of all tightening process parameters across the production line, promptly sending alarms upon detecting data anomalies; it significantly improves both the real-time performance and accuracy of data monitoring.

[0093] This invention employs unsupervised detection technology, eliminating the need for manual collection of large amounts of labeled data for modeling and comparison. Compared with existing technologies in the industry, this invention features significant technological innovations in feature extraction, redundant data filtering, envelope generation, and multi-weighted detection. It is easier to deploy and use in the field, and is not limited to a particular type of data or constrained by expert knowledge. It is applicable to various process parameter data such as tightening, pressing, and leak testing. Attached Figure Description

[0094] The invention will now be further described with reference to the accompanying drawings:

[0095] Figure 1 This is a flowchart of a process parameter anomaly detection method according to the present invention;

[0096] Figure 2This is a schematic diagram of the tightening curve family;

[0097] Figure 3 This is a schematic diagram illustrating the trend of λ change.

[0098] Figure 4 This is a schematic diagram of the envelope variation;

[0099] Figure 5 This is an overall trend graph of the process parameter curves;

[0100] Figure 6 For standardized sequence diagrams;

[0101] Figure 7 This is a schematic diagram illustrating the comparison and judgment of a quadratic difference sequence and a standard threshold.

[0102] Figure 8 To truncate the curve;

[0103] Figure 9 K-value features are extracted to obtain the K-sequence map;

[0104] Figure 10 List of abnormal curves;

[0105] Figure 11 Envelope diagram

[0106] Figure 12 This is a diagram illustrating type 1 abnormal data.

[0107] Figure 13 This is a schematic diagram of the complete shape of the curve;

[0108] Figure 14 The plot is a curve of the ST(j) sequence;

[0109] Figure 15 This is the graph corresponding to the second-order difference. Detailed Implementation

[0110] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this invention. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0111] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0112] The present invention will now be described in detail with reference to the accompanying drawings:

[0113] This invention integrates unsupervised methods with envelope detection methods, achieving automatic feature extraction of process parameters, unsupervised parameter partitioning, dynamic envelope plotting, and abnormal parameter curve localization. Based on this method, real-time, all-weather detection and localization of process parameter data such as tightening curve data on the production line can be achieved, greatly improving work efficiency and quality compared to manual observation and analysis.

[0114] This invention provides an embodiment of a method for detecting abnormal process parameters, the method flow is as follows: Figure 1 As shown.

[0115] The main implementation technical steps include:

[0116] Step 1: Process parameter collection;

[0117] The raw data of the production line process parameters is processed to generate an n×(m*k) matrix. Using the number of axes as an index, this matrix is ​​further stored in the data dictionary `dict(.)`. For example, `dict(1)`, `dict(2)`, and `dict(k)` represent process parameter data with a total number of axes of 1, 2, and k, respectively. The format is as follows:

[0118]

[0119] The process parameter data is stored column-wise for each axis, with a length of n. There are m tightening data entries in total, resulting in a total of m*k columns. This data is then compiled into a dictionary.

[0120] Dict={'1':dict(1),'2':dict(2)....'k':dict(k)}

[0121] Step 2: Feature parameter extraction;

[0122] The data points were located and truncated. The K-values ​​of the process parameter curves were calculated using the sliding window method to extract precise features from the data.

[0123] In practical applications, taking tightening curve data as an example, a large amount of redundant data appears in the first half of the data, such as... Figure 2 The tightening curves shown have relatively small changes in the first half of the tightening process, resulting in a more uniform torque. This part of the data can affect the anomaly detection, so the curves need to be truncated.

[0124] (3) Process parameter truncation;

[0125] 4) For tightening curve data dict(k) with a total number of axes k, calculate the standard deviation of all tightening data at each tightening angle, and plot the standard deviation curve. For dict(k), the standardization formula for the j-th data axis and the t-th data point is as follows: is the sum of the standard deviation and the mean of the data:

[0126]

[0127] For process parameters with k axes and n data lengths, the standard deviation is calculated for each axis and each tightening angle, resulting in the following standard deviation sequence.

[0128] ST(j)={st j (1),st j (2),st j (3)....,st j (n)}

[0129] ST(j) represents the sequence of all curves along the j-th axis calculated using the standard deviation plus the mean; essentially, it is a curve corresponding to... Figure 14 .

[0130] 5) Perform second-order differencing on all standard deviation sequence data.

[0131]

[0132] This yields the difference sequence WT(j), which can effectively amplify the difference in tightening torque between adjacent points, making it easier to determine the cutoff point.

[0133]

[0134] t represents the t-th data point of the curve; for example, st j (t) and st j (t-1) are two adjacent points; The first difference of the t-th data point is the difference between the current point and the previous point; For second-order differences, the superscripts (1) and (2) represent "first-order" and "second-order" respectively, meaning that the difference is performed on top of the first-order difference. j represents the j-th axis;

[0135] The graph corresponding to the second-order difference is as follows: Figure 15 .

[0136] 6) After obtaining the difference sequence WT, calculate the 1 σ value P(j) of the difference sequence WT. This value reflects the overall fluctuation of the difference sequence. Points exceeding this value can be considered as nodes where the tightening torque begins to change.

[0137]

[0138] For the sequence WT(j), find the position t of the first point greater than P(j), and use it as the cutoff point. The calculation formula is as follows:

[0139]

[0140] The process parameters are truncated at position t to obtain the data to be processed.

[0141] (4) K-value feature extraction;

[0142] 3) Define a sliding window of length u, where the value of u can be n / 50 of the total data length. Starting from the cutoff point, slide the window gradually to the right, and within each window j:[a j ,a j+1 ...a j+u Find the maximum point w within the range. (j) max ;

[0143]

[0144] 4) Further obtain the maximum value w within the window. (j) max The two points that are closest, at 30% and 70%, are defined as... Its corresponding x-coordinate is (This can be an angle or a time sequence number). The slope of the line connecting the two points is then calculated as the K value as follows:

[0145]

[0146] k j This represents the slope of the line connecting the 30% and 70% points of the j-th window;

[0147] Finally, a sequence of K-value points under all sliding windows is obtained. The calculation of K-values ​​amplifies local features and can more accurately locate abnormal parameter curves.

[0148] K = [k t ,k t+1 ...k n-u ]

[0149] K is a sequence of the slopes of all sliding windows, where uppercase K represents the name and lowercase k represents the numerical variable.

[0150] Step 3: Determine the anomalies of multiple weights;

[0151] (1) Sampling;

[0152] For each sequence of K values ​​obtained along axis j, a random sampling method is used, with a random sampling length not exceeding [a certain value]. The data is E, where m is the length of the K sequence data. One set of sampled data is generated each time, until all feature points are included:

[0153] K j =E2∪E2∪..∪E i

[0154] Simplified as:

[0155] K = E²∪E²∪..∪E i

[0156] Right now:

[0157] [k t ,k t+1 ...k n-u ]=E2∪E2∪..∪E i E represents the sampled data set, E i This represents the data set from the i-th sample.

[0158] (2) Unsupervised testing;

[0159] A basic anomaly detection model is selected, such as iForrest or DBSCAN, and is defined as model f. For the sampled feature set of the data to be detected, E1E2...E... i The data is sequentially input into model f. For the j-th sampled dataset of the i-th excessively long parameter data, the detection result is obtained.

[0160]

[0161] It is the output result of the j-th sampled data of the i-th curve input into the detection base model, with a value of -1 or 1;

[0162] in

[0163]

[0164] (3) Result classification;

[0165] The formula for determining potential anomaly curves and standard curves is as follows:

[0166]

[0167] It is a weighted calculation based on all sampling results, |E j | represents the corresponding data length; here we assume there are a total of n samples.

[0168] If Pre (i) If |E| < 0, then the parameter curve is a potential outlier curve and is classified into dataset A1; otherwise, it is a standard curve and is classified into dataset B1; where |E| < 0. j |For feature set E j The length represents The larger the corresponding weight, the stronger the weight. The more authoritative the judgment, the more accurate the judgment of the same model in multiple weight values ​​of process parameters.

[0169] Step 4: Draw the dynamic envelope line;

[0170] Based on the standard curve family B obtained in step three, calculate its upper and lower envelopes using the following formulas:

[0171]

[0172] u (j) Let represent the mean and standard deviation of all standard curves calculated for the j-th axis, respectively;

[0173] in, Let represent the upper and lower envelopes of the data on the j-th axis, respectively. The upper and lower envelopes fluctuate around the mean μ and standard deviation σ of the data with a coefficient ±λ. λ is a smoothing coefficient, ensuring that the range of variation is greater as the envelope approaches its tail. The formula is as follows:

[0174] λ=ln(x) / (1+e -x )

[0175] e is a natural constant in mathematics, and x is the input variable;

[0176] Where x is the standard value of the x-coordinate (tightening angle or time sequence number) of data with a length of n, and the formula is as follows:

[0177]

[0178] Where kd is the self-adjustment coefficient, which defaults to 50. The larger kd is, the larger λ is. The trend of λ is as follows: Figure 3 As shown, the envelope can be dynamically drawn based on the change of λ.

[0179] The corresponding envelope changes are as follows Figure 4 As shown.

[0180] Step 5: Determining Abnormal Results;

[0181] (3) Overflow point calculation;

[0182] For a dataset with axis j, based on the upper and lower envelopes Determine the sum of the data points that overflow the upper and lower envelopes of each process parameter curve. e When sum e If the curve exceeds a given threshold, it is identified as an abnormal curve and classified into dataset A2. e The calculation formula is as follows, where y represents the corresponding data point:

[0183]

[0184] (4) Classification of abnormal data;

[0185] For the two abnormal data points A1 and A2, they are divided into three priority categories, and the data is dispatched according to the priority level:

[0186] 1) Similar to anomalous data, this type of data represents the intersection of anomalous data and belongs to the highest priority data for processing:

[0187] Class 1 = {A1 ∩ A2}

[0188] 2) Type 2 abnormal data: This type of data is overflow envelope data, but it does not belong to A1 and belongs to secondary priority data.

[0189] Class 2 = {A2-A1}

[0190] 3) Three types of anomalous data: This type of data does not overflow the envelope, but belongs to A1, which is the lowest priority data to process.

[0191] Category 3 = {A1-A2}

[0192] Example

[0193] (1) Taking DDK tightening data as an example, 153 tightening process parameter data are collected from the workstation in real time. The number of axes in this data is 1. The data is converted into a data dict (1) stored by axis as follows:

[0194]

[0195] The overall trend of the process parameter curves is shown in the figure below. Figure 5 As shown, the method of this invention is used to identify abnormal curves:

[0196] (2) First, select the cutoff point for the tightening data according to the cutoff method described above:

[0197] The data from axis 1 are processed using the standardization method described in step two to obtain the standardized sequence ST, as follows: Figure 6 As shown.

[0198] Based on the ST sequence, calculate the quadratic difference sequence WT and the standard threshold P, and compare and judge as follows: Figure 7 As shown, the cutoff point position is obtained as t = 2610.

[0199] The data is truncated to obtain the truncated curve as shown below. Figure 8 As shown.

[0200] (3) Based on the truncated sequence, K-value features are extracted to obtain the K-sequence, such as... Figure 9 As shown.

[0201] (4) Based on the K-sequence, unsupervised anomaly detection was performed to obtain anomaly and standard curves, including 13 abnormal parameter curves and 140 normal curves. The abnormal curves are shown in the figure below. Figure 10 .

[0202] (5) Based on the standard envelope, where kd = 80, the smoothing coefficient λ of the envelope slowly approaches 3, resulting in the envelope as shown in the figure. Figure 11 The red part represents the upper and lower envelopes, and the blue part represents the process parameter data.

[0203] (6) Based on the above steps, we finally obtained 1 type of abnormal data, a total of 7 types, such as Figure 12 As shown.

[0204] The complete shape of the curve is as follows Figure 13 As shown, it has been confirmed that the curve is indeed an abnormal curve, and it has been verified that there are problems with the tightening process of the curves corresponding to codes 0316, 0317, 0329, and 0350.

[0205] This invention presents an innovative method and implementation for classifying process parameter curves using unsupervised techniques, drawing standard envelopes based on the classification results, and ultimately obtaining graded abnormal parameter curves.

[0206] The method and implementation of truncating process parameter curves based on second-order difference cutoff points to reduce data redundancy is innovative.

[0207] The method and implementation of calculating the K-value within the window based on a sliding window and extracting local features of process parameters are innovative.

[0208] The method and implementation of classifying process parameter curves based on weighted random sampling feature dimensions is innovative.

[0209] The present invention provides another embodiment of a process parameter anomaly detection system, comprising:

[0210] The parameter collection module is used for collecting process parameters.

[0211] The parameter extraction module is used for feature parameter extraction.

[0212] The judgment module is used for multi-weighted anomaly detection;

[0213] The drawing module is used for dynamic drawing of the envelope.

[0214] The result determination module is used for determining abnormal results.

[0215] The parameter extraction module includes a process parameter truncation unit and a feature extraction unit; the process parameter truncation unit is used to truncate the process parameters to obtain the data to be processed; the feature extraction unit is used for K-value feature extraction to obtain the K-value point sequence under all sliding windows;

[0216] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be included within the scope of protection of the present invention. Furthermore, all content not described in detail in this specification is prior art known to those skilled in the art.

Claims

1. A method for detecting abnormal process parameters, characterized in that, Includes the following steps: Step 1: Process parameter collection; Step 2: Feature parameter extraction; Step 3: Determine the anomalies of multiple weights; Step 4: Dynamically draw the envelope; Step 5: Determining Abnormal Results; Step one involves collecting process parameters, specifically: The raw data of the production line process parameters is processed to generate an n×(m*k) matrix, and then stored in the data dictionary dict(.) using the axis number as the index, in the following form: Assuming there are m tightening data lines, each with a length of n, and k axes, then the total number of data lines for each workstation is m*k. This represents the data in the m-th row and n*k-th column; Step two involves feature parameter extraction, specifically as follows: Locate and truncate data points, and calculate the K value for each process parameter curve using the sliding window method to perform precise feature extraction; including: (1) Process parameter truncation; (2) K-value feature extraction; The process parameter truncation specifically includes: 1) For tightening curve data dict(k) with a total number of axes k, calculate the standard deviation of all tightening data at each tightening angle, and plot the standard deviation curve. For dict(k), the standardization formula for the j-th data axis and the t-th data point is as follows: is the sum of the standard deviation and the mean of the data: It is the sum of the standard deviation and the mean. The mean of all curves at point t. Equivalent to the elements in the dict matrix; m is the number of curves. If there are m curves, then there are m data points. For process parameters with k axes and n data lengths, the standard deviation is calculated for each axis and each tightening angle, resulting in the following standard deviation sequence: This represents the sequence of all curves on the j-th axis calculated using the standard deviation plus the mean. 2) Perform second-order differencing on all standard deviation sequence data: This yields the difference sequence. The curve represents the first One data point; 3) Calculate the difference sequence 1 times σ value This value reflects the overall fluctuation of the difference sequence. Points exceeding this value are the nodes where the tightening torque begins to change. For sequences Find greater than The location of the first point As the cutoff point, the calculation formula is as follows: By location The process parameters are truncated to obtain the data to be processed; The K-value feature extraction specifically includes: 1) Define a sliding window of length u, where u takes values ​​that are n / 50 of the total data length. Starting from the cutoff point, the window slides gradually to the right, and within each window j: Find the maximum point inside ; 2) Further obtain the maximum value within the window. The two points that are closest to each other at 30% and 70% are defined as... , The corresponding x-coordinate is , The slope of the line connecting the two points is calculated as the K value as follows: This represents the slope of the line connecting the 30% and 70% points of the j-th window. Finally, obtain all sliding windows. Value point sequence; ； Step three involves the multi-weight anomaly detection, specifically as follows: (1) Sampling; For each sequence of K values ​​obtained along axis j, a random sampling method is used, with a random sampling length not exceeding [a certain value]. The data E is given, where m is the length of the K sequence data; one set of sampled data is generated each time, until all feature points are included: Represents the sampled data set. This represents the data set from the i-th sampling. (2) Unsupervised testing; A basic anomaly detection model is selected and defined as model f; for the sampling feature set of the data to be detected, E1E2...E n The data is sequentially input into model f. For the j-th sampled dataset of the i-th process parameter data, the detection result is obtained. : in (3) Result classification; The formula for determining potential anomaly curves and standard curves is as follows: like If the parameter value is less than 0, the curve is considered a potential outlier and is classified into dataset A1; otherwise, it is a standard curve and is classified into dataset B1. For feature set E j The length represents The larger the corresponding weight, the stronger the weight. The more authoritative the judgment, the more accurate the judgment of the same model in multiple weight values ​​of process parameters.

2. The method for detecting abnormal process parameters according to claim 1, characterized in that: Step four involves drawing the dynamic envelope line, specifically as follows: Based on the standard curve family B, the upper and lower envelopes are calculated using the following formula: Let represent the mean and standard deviation of all standard curves calculated for the j-th axis, respectively; in, , These represent the upper and lower envelopes of the data on the j-th axis, respectively; the upper and lower envelopes fluctuate around the mean μ and standard deviation σ of the data with a coefficient of ±λ. The smoothing coefficient is given by the following formula: in For data length is The standardized value of the x-axis of the data is given by the following formula: in This is the self-adjustment coefficient, with a default value of 50. The larger λ is, the larger λ becomes. By adjusting the change in λ, the envelope can be dynamically drawn.

3. The method for detecting abnormal process parameters according to claim 2, characterized in that: The abnormal result determination in step five is as follows: (1) Overflow point calculation; For a dataset with axis j, based on the upper and lower envelopes Determine the number of data points that overflow the upper and lower envelopes for each process parameter curve. ,when If the curve exceeds a given threshold, it is identified as an abnormal curve and classified into dataset A2. The calculation formula is as follows, where y represents the corresponding data point: (2) Classification of abnormal data; For the abnormal data A1 and A2 obtained respectively, they are divided into 3 priority categories and dispatched for processing according to the priority: 1) Similar to anomalous data, this type of data represents the intersection of anomalous data and belongs to the highest priority data for processing: 2) Type 2 abnormal data: This type of data is overflow envelope data, but it does not belong to A1 and belongs to secondary priority data. 3) Three types of anomalous data: This type of data does not overflow the envelope, but belongs to A1, which is the lowest priority data to process. 。 4. The detection system of the process parameter anomaly detection method according to claim 1, characterized in that, include: The parameter collection module is used for collecting process parameters; The parameter extraction module is used for feature parameter extraction. The judgment module is used for multi-weighted anomaly detection; The drawing module is used for dynamic drawing of the envelope. The result determination module is used for determining abnormal results.

5. The detection system according to claim 4, characterized in that: The parameter extraction module includes a process parameter truncation unit and a feature extraction unit; the process parameter truncation unit is used to truncate the process parameters to obtain the data to be processed; the feature extraction unit is used to extract K-value features to obtain the K-value point sequence under all sliding windows.

Images