An intermittent process measurement data anomaly detection method based on double support vector data description

Abstract

The application discloses a kind of intermittent process measurement data anomaly detection methods based on double support vector data description, first modal division is carried out to intermittent process using fuzzy clustering method, obtain each modal dataset, and fusion each stable mode and adjacent transition mode data construct fusion modal dataset;Then according to each modal dataset and fusion modal dataset, construct double SVDD process measurement data anomaly detection model respectively;Finally, the inner and outer control limits of double SVDD model and data anomaly discrimination strategy are used to realize the abnormal detection of intermittent process measurement data.The application fully considers the data change relation between different modes of intermittent process, detects the measurement data anomaly of intermittent process according to the inner and outer control limits of the double SVDD model and data anomaly discrimination strategy, which can effectively reduce the false detection rate of process measurement data anomaly detection and improve the accuracy of intermittent process measurement data anomaly detection.

Description

Technical Field

[0001] This invention belongs to the field of intermittent process monitoring technology, and in particular relates to a method for detecting anomalies in intermittent process measurement data based on dual support vector data description. Background Technology

[0002] Batch processes, capable of producing high-value-added products according to market demand changes and customer customization requirements, have been widely used in chemical, pharmaceutical, and food processing industries. Field measurement instruments in batch production processes provide a wealth of process measurement data, supporting data-driven process modeling. However, factors such as performance degradation of field measurement instruments or external environmental interference can cause anomalies in batch process measurement data, directly affecting the accuracy of data-driven process modeling. Therefore, anomaly detection in batch process measurement data can provide reliable data for data-driven process modeling and promote the application of batch process monitoring and optimization control methods and technologies.

[0003] Support Vector Data Description (SVDD) constructs a minimal closed hypersphere of all training data in a high-dimensional space, distinguishing normal and abnormal data based on the hypersphere's boundary to detect anomalies in intermittent process measurement data. However, anomaly detection methods based on a single SVDD model neglect the multimodal characteristics of intermittent processes, reducing the accuracy of anomaly detection. Anomaly detection methods based on a multimodal SVDD model exhibit higher detection performance compared to those based on a single SVDD model, but these methods only consider data features within each modality, ignoring the data variation relationships between adjacent modalities, leading to a higher false positive rate. Therefore, this paper proposes an anomaly detection method for intermittent process measurement data based on dual support vector data description. This method fully considers the data variation relationships between different modalities, constructs a dual SVDD anomaly detection model using datasets of stable and transitional modalities, and combines it with an anomaly discrimination strategy to detect anomalies in intermittent process measurement data. This effectively reduces the false positive rate of abnormal data and improves the accuracy of anomaly detection in intermittent process measurement data. Summary of the Invention

[0004] This invention aims to improve the accuracy of anomaly detection in intermittent process measurement data. It proposes a method for anomaly detection in intermittent process measurement data based on dual support vector data description, comprising the following steps:

[0005] Step 1: Collect multiple batches of intermittent process data, use the Fuzzy C-means (FCM) algorithm to perform mode segmentation, obtain the datasets of each stable mode and transition mode based on the mode segmentation information, and fuse the stable mode data and adjacent transition mode data to obtain the fused mode dataset;

[0006] Step 2: Train the inner SVDD model on all stable mode and transition mode datasets respectively, and train the outer SVDD model on the fused mode dataset to construct an anomaly detection model for intermittent process measurement data based on dual SVDD;

[0007] Step 3: Utilize the inner and outer hypersphere radii R of the dual SVDD model in and R out A data anomaly detection strategy was constructed by combining the concept of moving windows to perform online anomaly detection on intermittent process measurement data.

[0008] Step one specifically includes:

[0009] Collect X batches of intermittent process measurement data. i (K×J), where i (1≤i≤I) is the batch index, J is the number of variables, and K is the number of sampling points. The modal data X(K×J) is obtained by averaging the data from I batches along the batch direction. Then, the mean of each variable in X is subtracted from the mean, and the result is standardized by dividing by the standard deviation.

[0010] for FCM mode partitioning can be considered as the following optimization problem

[0011]

[0012] In the formula, v p Represents the p-th mode center; 0≤u kp ≤1 indicates that the process data sample x k Membership degree of the p-th mode; P represents the number of modes in the intermittent process; d kp For process data sample x k and modal center v p The Euclidean distance between them; m is the fuzziness factor; U(K×P) is the membership matrix; This represents any sampling point k∈[1,K].

[0013] From equation (1), the iterative update formulas for membership degree u and mode center v can be obtained as follows:

[0014]

[0015]

[0016] In the formula, h is the index of the P modes; d kh For process data sample x k and corresponding mode center v h The Euclidean distance between them.

[0017] After the iterative update stops, based on the membership matrix U, there are l independent modes, including stable modes and transitional modes, with the right boundary of each mode being b = {b1, b2, ..., b}. l}, b l This represents the sampling time of the right boundary of the l-th mode, where the odd-numbered l is the sampling time of the right boundary of the l-th mode. odd Representing stable modes, even-numbered l even This represents the transition mode. Therefore, the data for the l-th mode can be represented as: x ik Let X represent the intermittent process data samples with sampling time k in the i-th batch. After expanding according to the variable direction, the corresponding modal dataset is obtained as X. l (N l ×J), where N l =I×(b) l -b l-1 ) indicates the number of samples in the corresponding dataset.

[0018] The stable mode data is fused with its adjacent transition mode data. For the first stable mode, only the data of the next subsequent transition mode needs to be fused; for the last stable mode, only the data of the preceding transition mode needs to be fused; and for all other stable modes, the data of the two adjacent transition modes before and after it needs to be fused. (Using X) st This represents the dataset obtained after fusing stable mode and transition mode data. Therefore, for the first stable mode, i.e., l odd When = 1, the corresponding merged data and the dataset expanded according to the variable direction can be represented as:

[0019]

[0020] For the last stable mode, i.e. l odd When =l, the corresponding merged data and dataset are as follows:

[0021]

[0022] When l odd ≠1 and l odd When ≠l, the corresponding fused data and dataset are:

[0023]

[0024] In the formula, N st This indicates the number of samples in the fused modality dataset.

[0025] Standardize the obtained independent modality datasets and fused datasets by subtracting the mean and standard deviation of the variable directions respectively.

[0026] Step two specifically includes:

[0027] The SVDD hypersphere model is determined by the center a and the hypersphere radius R, and it satisfies the following optimization problem.

[0028]

[0029] In the formula, C is the penalty parameter; ξ j φ is a slack variable; n is the number of samples; φ(.) is a nonlinear function.

[0030] Introducing the Lagrange multiplier α transforms it into the dual form as follows:

[0031]

[0032] In the formula, α j and α h Sample x j and x h The corresponding Lagrange multipliers.

[0033] Gaussian kernel function K(x) j x j ) is defined as

[0034]

[0035] In the formula, σ is the kernel width parameter; exp(.) is an exponential function with the natural constant e as the base. Solving equation (8) yields the solution set of the Lagrange multiplier α.

[0036] A dual SVDD model is constructed on the standardized dataset. The model construction diagram is shown below. Figure 1 As shown, odd-numbered data points represent stable modality datasets, and even-numbered data points represent transitional modality datasets. For the standardized l-th independent modality dataset... Training the inner SVDD model is

[0037]

[0038] In the formula, and x represents the center and radius of the hypersphere of the inner layer SVDD model in the l-th independent mode, respectively; sv This represents the support vectors of the corresponding model.

[0039] Similarly, for the standardized l-th... odd A fusion dataset Training the outer SVDD model is

[0040]

[0041] In the formula, and They respectively represent the fusion of the lth odd The hypersphere center and radius of the outer SVDD model are trained using stable mode data and adjacent transition mode data.

[0042] For each stable mode, the corresponding inner SVDD model control limits are: The control limits of the outer SVDD model are For each transition mode, the control limits of its inner SVDD model are: The control limits of the outer SVDD model are

[0043]

[0044] In the formula, and These are the outer model control limits for the stable modes before and after the corresponding transition modes.

[0045] The control limits of the inner and outer layers of all modes can be calculated from equations (10) to (12), thus completing the construction of the dual SVDD model.

[0046] Step three specifically includes:

[0047] Online sample x new Based on the sampling time information of the samples, the mode l to which the samples belong is determined, and the samples are standardized using the mean and standard deviation of the dataset of that mode. Will The inner SVDD model of input mode l obtains the distance between the sample and the center of the hypersphere. for

[0048]

[0049] Will and and When comparing, Less than or equal to At that time, x new It can be directly identified as a normal sample; while when Greater than or equal to At that time, x new They will also be directly identified as anomalous samples, that is

[0050]

[0051] when Greater than and smaller than At this time, x new Samples that cannot be directly classified as normal or abnormal are addressed in this invention by incorporating the concept of a moving window to determine their status. Let the left endpoint of the moving window be denoted as WIN. L The right endpoint is represented as WIN. R The selection of the moving window width W only needs to satisfy that it is less than or equal to the length of the shortest mode, and the movement of the window is restricted to the same mode. This ensures that the state of the sample at the current moment only refers to the state information of the data sample within the current mode. The window movement rule is: (1)x new When the sampling time is the right endpoint of the window and is not equal to the right boundary of the current mode, move backward by one width without exceeding the right boundary of the current mode; (2)x new When the sampling time is the right endpoint of the window and equal to the right boundary of the current mode, move directly backward by one width; (3) within the current mode, x new When the sampling time coincides with the left end of the window, the window moves forward until the right end coincides with the sampling time.

[0052] Let x new The sampling time is T. Combined with the established moving window, the sample state discrimination diagram is shown below. Figure 2 As shown. For Figure 2 (a), x new The distance from the center of the ball to the center of the ball is from the normal area ( (Internal) traveled to and Between these points, the calculation window is used to determine the time from the left endpoint of the window to the sample x at time T-1. T-1 The standard deviation of the distance between the centers of the spheres and the addition of x new After that, the standard deviations did not change significantly, so we assume x new These are also normal samples; Figure 2 In (b), add x new After that, the standard deviation changed by more than or equal to 2 times, at which point x is considered to be... new This is an abnormal sample; for Figure 2 (c), x new From the abnormal area ( (The outside) crossed to and Between these points, the mean distance between the left end of the calculation window and the center of the sample sphere at time T-1 is calculated. When the mean is less than At that time, it was believed that x new This is a normal sample; Figure 2 In (d), the mean distance from the left endpoint of the window to the center of the sample sphere at time T-1 is greater than or equal to... It is believed that x new It is still an anomalous sample. Furthermore, if x new Its state remains unchanged under the following two conditions, namely when x newThe left endpoint or x of the current mode new The sample was also in the previous moment and When x is between new The sample state remains unchanged from the previous time step. Therefore, the anomaly detection strategy for intermittent process data based on the dual SVDD model can be obtained as follows:

[0053]

[0054]

[0055]

[0056] In the formula, n1 = T-1 - (WIN) L -1); n2 = T - (WIN L -1); and These represent the distance from the left endpoint of the window to x. new The mean distance between the centers of the spheres of all samples from the previous moment to the current moment.

[0057] The current sample status can be judged by the above discrimination strategy, and the fault detection rate (FDR) and false alarm rate (FAR) can be calculated according to formulas (18) to (20). Then, the geometric mean g is calculated by the FDR and FAR values ​​to evaluate the detection results. The larger the FDR is, the smaller the FAR is. The larger the final geometric mean g is, the higher the accuracy of the data anomaly detection.

[0058]

[0059]

[0060]

[0061] Advantages of this invention: Addressing the problem that anomaly detection in intermittent process measurement data only considers data features within each modality and ignores the data variation relationships between adjacent modalities, leading to low accuracy in anomaly detection, this invention constructs a dual SVDD process measurement data anomaly detection model by training inner and outer layer SVDD models separately on each modal dataset and the fused modal dataset. Utilizing the model's inner and outer layer control limits and data anomaly discrimination strategy, anomaly detection in intermittent process measurement data is achieved, effectively reducing the false detection rate and improving the accuracy of anomaly detection in intermittent process measurement data. Attached Figure Description

[0062] Figure 1This is a schematic diagram illustrating the construction of an anomaly detection model for intermittent process measurement data based on dual SVDD;

[0063] Figure 2 This is a schematic diagram illustrating the state of the sample.

[0064] Figure 3 This is a flowchart of an anomaly detection method for intermittent process measurement data based on dual support vector data description, as described in this invention.

[0065] Figure 4 The results are based on the mode division of the penicillin fermentation process using FCM.

[0066] Figure 5 This is the result of abnormal detection of process measurement data of test batch 1 by the method described in this invention. Detailed Implementation

[0067] The present invention will be further described below with reference to examples and accompanying drawings. It should be noted that the embodiments do not limit the scope of protection claimed by the present invention.

[0068] Example

[0069] The penicillin fermentation process is a typical batch process. Using the Pensim simulation platform, 25 normal batch datasets were generated as the training set {X}. i (400×17)}, 1≤i≤25, each batch lasts for 400 hours, the sampling interval is 1 hour, the penicillin fermentation process variables are shown in Table 1, and 10 test batches containing abnormal measurement data are generated at the same time, the test batch information is shown in Table 2.

[0070] Table 1 Variables in the penicillin fermentation process

[0071]

[0072]

[0073] Table 2 Information on test batches containing abnormal measurement data

[0074]

[0075] The flowchart illustrating the application of this invention to the penicillin fermentation process is as follows: Figure 3 As shown, the specific steps are as follows:

[0076] Step 1: Calculate the average of 25 normal batches by batch direction and then standardize the result. The mode division results obtained from FCM are as follows Figure 4As shown, the first stable mode is 1–31 hours, the first transitional mode is 32–48 hours, the second stable mode is 49–101 hours, the second transitional mode is 102–117 hours, the third stable mode is 118–222 hours, the third transitional mode is 223–243 hours, and the fourth stable mode is 244–400 hours. Therefore, the right boundary b of each mode is {31, 48, 101, 117, 222, 243, 400}. Based on the mode partitioning results, the 25 training batch samples are divided into corresponding mode datasets. The obtained stable mode dataset and transitional mode dataset are then fused to obtain the fused mode dataset. For example, the first stable mode dataset is X. 1 (775×17), the fused dataset combining the first stable mode and the first transition mode is as follows:

[0077] Step 2: Train the inner SVDD model for each modality dataset separately, and train the outer SVDD model for the fused modality data to construct a dual SVDD intermittent process measurement data anomaly detection model. Combine this with the proposed data anomaly discrimination strategy and test it on 10 test batches, with a moving window width of 10 sampling times. The detection results for test batch 1 are as follows: Figure 5 As shown in Table 3, the results of anomaly detection for intermittent process measurement data for all test batches are presented.

[0078] Table 3. Anomaly detection results for intermittent process measurement data of all test batches.

[0079]

[0080] The results above show that the G-SVDD method, which does not consider the multimodal characteristics of the intermittent process, has the lowest accuracy in detecting anomalies in process measurement data. The FCM-SVDD method, which considers the multimodal characteristics, has improved detection accuracy, but still has a high false positive rate. Compared with the FCM-SVDD method, the method described in this invention reduces the average false positive rate of all test batches of intermittent process measurement data from 25.41% to 7.22%, and increases the geometric mean g from 85.23% to 94.87%, effectively improving the accuracy of anomaly detection in intermittent process measurement data.

Claims

1. A method for detecting anomalies in intermittent process measurement data based on dual support vector data description, characterized in that: The method includes the following steps: Step 1: Collect multiple batches of intermittent process data, perform mode segmentation using the fuzzy C-means FCM algorithm, obtain stable mode and transition mode datasets based on the mode segmentation information, and fuse stable mode and adjacent transition mode data to obtain a fused mode dataset; the intermittent process data are penicillin fermentation process variables, including aeration rate, stirring power, bottom stream acceleration rate, bottom stream temperature, substrate concentration, dissolved oxygen concentration, biomass concentration, penicillin concentration, reactor volume, carbon dioxide concentration, pH, reactor temperature, heat production, acid addition flow rate, alkali addition flow rate, cooling water addition flow rate, and heating water flow rate; Step 2: Train the inner SVDD model on all stable mode and transition mode datasets respectively, and then train the outer SVDD model on the fused mode dataset to construct an anomaly detection model for intermittent process measurement data based on dual SVDD; Step 3: Utilize the inner and outer hypersphere radii R of the dual SVDD model in and R out A data anomaly detection strategy was constructed by combining the concept of moving windows to perform online anomaly detection on intermittent process measurement data; Step three specifically includes: Online samples The mode to which the sample belongs is determined based on the sampling time information of the sample. The modality dataset was then standardized using its mean and standard deviation. ,Will Input mode The inner SVDD model yields the distance between the sample and the center of the hypersphere. ; Will and and When comparing, Less than or equal to hour, It was directly determined to be a normal sample; while when Greater than or equal to hour, These will be directly identified as abnormal samples; when Greater than and smaller than At this time Samples that cannot be directly classified as normal or abnormal are assessed using a moving window approach; let the left endpoint of the moving window be denoted as WIN. L The right endpoint is represented as WIN. R Move window width The selection of the window only needs to satisfy the requirement that it is less than or equal to the length of the shortest mode, and the movement of the window is restricted to the same mode, so as to ensure that the state of the sample at the current moment only refers to the state information of the data sample in the current mode. The window movement rule is: (1) When the sampling time is the right endpoint of the window and is not equal to the right boundary of the current mode, move backward by one width without exceeding the right boundary of the current mode; (2) When the sampling time is the right endpoint of the window and equal to the right boundary of the current mode, move directly backward by one width; (3) within the current mode, When the sampling time coincides with the left end of the window, the window moves forward until the right end coincides with the sampling time; set up The sampling time is The state of the sample is judged by combining the established moving window; The distance from the center of the ball is from The interior was traveled to and Between, at this time the left endpoint of the calculation window to Time sample The standard deviation of the distance between the centers of the balls and the addition After that, the standard deviations did not change significantly, which is considered... Also a normal sample; added After that, the standard deviation changed by more than or equal to 2 times, at which point it is considered that... This is an abnormal sample; from The outside has been crossed and Between, at this time the left end of the calculation window to The mean distance between the center of the sample spheres at time t, when the mean is less than At that time, it was believed For normal samples; from the left endpoint of the window to The mean distance between the center of the sample sphere at time t is greater than or equal to ,think It is still an anomalous sample; furthermore, if Its state remains unchanged under the following two conditions, namely when The left endpoint of the current mode or The sample was also in the previous moment and Between times, it is believed The sample state remains unchanged from the previous moment.

2. The method for detecting anomalies in intermittent process measurement data based on dual support vector data description according to claim 1, characterized in that: Step one specifically includes: Collect I batches of intermittent process measurement data. , Let J be the batch index, J be the number of variables, and K be the number of sampling points; average the data from I batches of data along the batch direction to obtain the modal segmentation data. and to Standardize by subtracting the mean from each variable and then dividing by the standard deviation. ;for FCM mode partitioning is considered as the following optimization problem ； In the formula, Indicates the first Modal centers; Represents process data samples Belongs to the Membership degree of each modality; Indicates the number of modes in the intermittent process; For process data samples and modal center The Euclidean distance between them; For fuzzy factors; This is the membership matrix; Represents arbitrary sampling points ; Membership degree is obtained from equation (1) and modal center The iterative update formula is: ； ； In the formula, To Index of each modality; For process data samples and corresponding modal centers The Euclidean distance between them; After the iterative update stops, the stable modes and transitional modes can be distinguished based on the membership matrix. There are 1 independent mode, and the right boundary of the mode is 1. , Indicates the first The right boundary sampling time of each modality, where the odd-numbered ones are... Represents stable modes, numbered evenly. Indicates the transition mode; therefore, the first Each modal data can be represented as , Indicates the first Batch sampling time is The intermittent process data samples, after being expanded along the variable direction, yield the corresponding modal dataset. ,in Indicates the number of samples in the corresponding dataset; The stable mode data and its adjacent transition mode data are fused together, and then... This represents the dataset obtained after fusing stable mode and transition mode data. Therefore, for the first stable mode, i.e. At that time, the corresponding fused data and the dataset expanded according to the variable direction can be represented as follows: ； For the last stable mode, i.e. At that time, the corresponding merged data and dataset are ； when and At that time, the corresponding merged data and dataset are ； In the formula, This indicates the number of samples in the fused modality dataset; Standardize the obtained independent modality datasets and fused datasets by subtracting the mean and standard deviation of the variable directions respectively.

3. The method for detecting anomalies in intermittent process measurement data based on dual support vector data description according to claim 1, characterized in that: Step two specifically includes: The SVDD hypersphere model consists of a sphere center. and hypersphere radius It is determined that it satisfies the following optimization problem. ； In the formula, For penalty parameters; These are slack variables; The number of samples; It is a nonlinear function; Introducing Lagrange multipliers Transformed into dual form as ； In the formula, and Samples and The corresponding Lagrange multipliers; Gaussian kernel function Defined as ； In the formula, This refers to the kernel width parameter; For the natural constant An exponential function with base 0; For the standardized first A dataset of independent modalities The inner SVDD model is trained as follows: ； In the formula, and The first The center and radius of the hypersphere in an independent modal inner-layer SVDD model; This represents the support vectors of the corresponding model; Similarly, for the standardized first... A fusion dataset Training the outer SVDD model is ； In the formula, and They represent the fusion of the first The hypersphere center and radius of the outer SVDD model are trained using stable mode data and adjacent transition mode data; For each stable mode, the corresponding inner SVDD model control limits are: The control limits of the outer SVDD model are For each transition mode, the control limits of its inner SVDD model are: The control limits of the outer SVDD model are ； In the formula, and These are the outer model control limits for the stable modes before and after the corresponding transition modes; The control limits of the inner and outer layers of all modes can be calculated from equations (10) to (12), thus completing the construction of the dual SVDD model.

4. The method for detecting anomalies in intermittent process measurement data based on dual support vector data description according to claim 1, characterized in that: Step three specifically includes: Online samples The mode to which the sample belongs is determined based on the sampling time information of the sample. The modality dataset was then standardized using its mean and standard deviation. ,Will Input mode The inner SVDD model yields the distance between the sample and the center of the hypersphere. for ； Will and and When comparing, Less than or equal to hour, It was directly determined to be a normal sample; while when Greater than or equal to hour, These will be directly identified as abnormal samples, i.e. ； when Greater than and smaller than At this time Samples that cannot be directly classified as normal or abnormal are assessed using a moving window approach; let the left endpoint of the moving window be denoted as WIN. L The right endpoint is represented as WIN. R Move window width The selection of the window only needs to satisfy the requirement that it is less than or equal to the length of the shortest mode, and the movement of the window is restricted to the same mode. This ensures that the state of the sample at the current moment only refers to the state information of the data sample within the current mode. The window movement rule is: (1) When the sampling time is the right endpoint of the window and is not equal to the right boundary of the current mode, move backward by one width without exceeding the right boundary of the current mode; (2) When the sampling time is the right endpoint of the window and equal to the right boundary of the current mode, move directly backward by one width; (3) within the current mode, When the sampling time coincides with the left end of the window, the window moves forward until the right end coincides with the sampling time; set up The sampling time is The state of the sample is judged by combining the established moving window; The distance from the center of the ball is from The interior was traveled to and Between, at this time the left endpoint of the calculation window to Time sample The standard deviation of the distance between the centers of the balls and the addition After that, the standard deviations did not change significantly, which is considered... Also a normal sample; added After that, the standard deviation changed by more than or equal to 2 times, at which point it is considered that... This is an abnormal sample; from The outside has been crossed and Between, at this time the left end of the calculation window to The mean distance between the center of the sample spheres at time t, when the mean is less than At that time, it was believed For normal samples; from the left endpoint of the window to The mean distance between the center of the sample sphere at time t is greater than or equal to ,think It is still an anomalous sample; furthermore, if Its state remains unchanged under the following two conditions, namely when The left endpoint of the current mode or The sample was also in the previous moment and Between times, it is believed The sample state remained unchanged from the previous time step; therefore, the anomaly detection strategy for intermittent process data based on the dual SVDD model is as follows: ； ； ； In the formula, ; ; and These represent the distance from the left end of the window to... The state of the current sample can be determined by using the average distance between the center of the sphere of all samples from the previous moment to the current moment and the mean distance between the center of the sphere of all samples from the previous moment to the current moment, through the above discrimination strategy.

Images