A method for detecting minor process faults based on local probability sliding window

Abstract

This invention provides a method for detecting minor process faults based on a local probability sliding window, belonging to the field of industrial process fault detection technology. It solves the technical problems of insufficient local distribution modeling capability and inadequate utilization of sliding window information in traditional fault detection methods when processing industrial process data. The technical solution comprises two parts: offline training and online detection. The beneficial effects of this invention are: by aggregating industrial process time series data through a sliding window, the observations at multiple time points within a window of length are constructed into a histogram distribution, thereby accumulating fault information and facilitating the detection of minor faults; local divergence is used as the detection statistic, fully utilizing the local distribution information of anchor points. Simultaneously, the distributional differences in the detection process are considered.

Description

Technical Field

[0001] This invention relates to the field of industrial process fault detection technology, and in particular to a method for detecting minor process faults based on a local probability sliding window. Background Technology

[0002] Fault detection is a crucial link in ensuring the safety of industrial processes. Minor faults have low amplitude and evolve slowly; if they are not identified in time, they may develop into serious malfunctions, causing huge losses to industrial production processes. Therefore, improving the accuracy and sensitivity of minor fault detection is of paramount importance.

[0003] To improve fault detection performance, researchers have proposed various improvement methods. For example, Chinese patent application CN201810899623, entitled "Based on Survival Information Potential-Principal Component Analysis Algorithm,"... A process fault detection method is proposed, based on an improved... A new method for detecting faults in chemical processes is proposed, which expands the data dimension by introducing a sliding window technique to capture the temporal correlation of process data. However, this method is still based on... The current global linear modeling framework uses a single linear subspace to describe normal operation, making it difficult to capture the multimodal and nonlinear distribution characteristics commonly found in industrial processes. Furthermore, while this method utilizes a sliding window to expand the data dimensionality, it still employs a point-level approach. and While statistical methods are used for detection, they fail to characterize the statistical changes caused by faults at the distribution level, and the detection rate and sensitivity for minor faults still need improvement. Chinese patent application CN202010813193, invention title: A method based on complex industrial chemical processes... This paper proposes a deep learning-based fault detection method for chemical processes, utilizing the nonlinear feature extraction capability of deep autoencoders for fault detection. However, deep learning methods suffer from high model complexity, long training time, the need for large amounts of labeled data, and poor interpretability. Furthermore, reconstruction error, as a detection metric, mainly reflects the overall deviation of the data and is difficult to accurately characterize subtle changes in local distribution, making it prone to false alarms when handling fault detection under multiple operating conditions.

[0004] In recent years, local Mahalanobis distance ( The method determines the boundary of the healthy region by generating anchor points and calculates the Mahalanobis distance from the sample to the anchor points as the detection index. However, Using single-sample point-to-point distance metrics still lacks sufficient sensitivity for detecting minute faults. Because the amplitude of minute faults is close to noise levels, the statistical characteristics of single samples are unstable, instantaneous distance calculations are easily affected by noise, and there is a lack of effective utilization of time-series information.

[0005] Therefore, there is an urgent need for a method for detecting minor faults in industrial processes that can effectively aggregate timing information, improve the signal-to-noise ratio, and have a high detection rate and sensitivity for minor faults. Summary of the Invention

[0006] To address the technical problems of insufficient local distribution modeling capability and inadequate utilization of sliding window information in traditional fault detection methods when processing industrial process data, this invention provides a process micro-fault detection method based on local probability sliding window, which integrates the advantages of both anchor point local distribution modeling and sliding window distribution aggregation.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a method for detecting minor process faults based on a local probability sliding window, comprising the following steps:

[0008] Step 1) Obtain multivariate time series data of industrial processes, including training set data and test set data; the training set data is used for anchor point generation and distribution modeling, and the test set data is used for minor fault detection.

[0009] Step 2) Data standardization: Calculate the mean and standard deviation of the training set data, and standardize the training and test set data.

[0010] Let the training set data be ,in For the sample size, Number of variables; calculate the first... Mean of each variable and standard deviation :

[0011] , ;

[0012] The standardized data are:

[0013] ;

[0014] Step 3), Anchor point generation: using... Frame generates anchor point set ,in The number of anchor points; by traversing the local radius Choose the one that minimizes the total approximation error. value as optimal radius ;

[0015] Step 4) Local distribution modeling of anchor points: For each anchor point Establish Local probability distribution within the radius: ,in This is the local radius coefficient;

[0016] For anchor points Calculate the Mahalanobis distance between samples in its local neighborhood:

[0017] ;

[0018] in, Let be the covariance matrix of the training set data; take the condition that... For local samples, construct a histogram distribution for each variable. , Quantity is ;

[0019] Step 5) Sliding window data extraction: Apply a length of [length missing] to the test set data. Extract the sliding window data matrix. ;No. Each window contains samples ;

[0020] Step 6) Window Distribution Aggregation: Aggregate the distribution of windows Each variable within ,Will The observations are aggregated into a histogram distribution. ; used for anchor points with the same distribution Boundaries, frequency distribution of data within the statistical window and normalization;

[0021] Step 7) Calculate the local Divergence: for a window Calculate its relationship with the distribution of all anchor points. Divergence, take the minimum value as Statistic;

[0022] Step 8) Control Limit Determination and Fault Detection: Calculate based on training set or normal operation data. Control limits for the statistic; use empirical quantiles, based on the significance level. Determine detection control limits , making ;like Then determine the first A fault was detected in one window; if If so, the process is considered to be running normally.

[0023] Step 7) includes the following steps:

[0024] Step 71) For anchor points and window distribution ,calculate Divergence:

[0025] ;

[0026] Step 72) Divergence is then summed using variance-weighted methods:

[0027] ;

[0028] in, anchor point The The standard deviation of each variable in its local sample set;

[0029] Step 73) The statistic is defined as:

[0030] .

[0031] The fault detection method also includes online detection, comprising the following steps:

[0032] S1, Obtain the first After sampling the observations at each time point, a sliding window data matrix of length l is constructed by combining it with historical data. , which includes samples ;

[0033] S2, for the window Each variable within ,Will The observations are aggregated into a histogram distribution. Use the same distribution as the anchor points Boundaries, frequency distribution of data within the statistical window and normalization;

[0034] S3, Calculate the window distribution and the distribution of all anchor points. Divergence, take the minimum value as Statistic:

[0035] ;

[0036] S4. If This indicates that a fault has been detected; if This indicates that the process is running normally.

[0037] Step 3) Anchor point generation includes the following steps:

[0038] Step 31) Set the minimum number of local samples and local radius search range ;

[0039] Step 32) Calculate the covariance matrix of the training set data. For any two sample points and Its Mahalanobis distance is defined as:

[0040] ;

[0041] Step 33) Traverse the local radius For each To determine the value, perform the following operations: use the data mean as the first anchor point; iterate through the samples in ascending order of Mahalanobis distance, and for each unprocessed sample, find its... All unprocessed samples within the radius; if the number of samples within the radius is not less than Then, the mean of these samples is used as the new anchor point; the sum of the Mahalanobis distances from all samples to the nearest anchor point is calculated as the total approximation error.

[0042] ;

[0043] Step 34) Select the option that minimizes the total approximation error. The value is used as the optimal local radius. ;

[0044] Step 35) Use Generate the final anchor set ,in This represents the total number of anchor points.

[0045] Step 4) Local distribution modeling of anchor points includes the following steps:

[0046] Step 41) Calculate the local modeling radius ,in This is the local radius coefficient;

[0047] Step 42) For each anchor point Extract all samples from the training set that satisfy the following conditions: The samples constitute the local sample set of the anchor point. ;

[0048] Step 43) Determine the histogram Boundaries: for variables ,That The boundary is:

[0049] ;

[0050] in, and Variables The minimum and maximum values ​​in the training set, For boundary expansion, for Quantity; all anchor points share the same boundary;

[0051] Step 44) For anchor points Local sample set Each variable in Assign all observations of this variable to the corresponding In the middle, statistics for each The frequency of the numbers is normalized to obtain the probability distribution:

[0052] ,

[0053] in, Indicates anchor point Variables in local samples Falling into indivual The number of samples;

[0054] Step 45) Smooth the probability distribution;

[0055] ,

[0056] in, For smoothing parameters.

[0057] Step 5) includes the following steps:

[0058] Step 51) Set the length of the sliding window To ensure the first window contains the complete Each sample is concatenated with the last part of the training set before the test set. One sample as a prefix:

[0059]

[0060] Step 52), for the first Multiple windows are used to extract data from time series data. A window data matrix is ​​formed by consecutive samples:

[0061]

[0062] in, Indicates the first The observation vector at each time step;

[0063] Step 53) For the window Each variable in ,Will Each observation is assigned to the same distribution as the anchor point. Within the boundary, count the frequencies and normalize them:

[0064]

[0065] in, Display window Medium variables Falling into indivual The number of samples;

[0066] Step 54) Smooth the window distribution:

[0067]

[0068] in, For the smoothing parameter, use the same value as the anchor point distribution smoothing process.

[0069] Step 8) Determining the control limits includes the following steps:

[0070] Step 81) Apply a sliding window to the normal operation segment data in the test set and calculate the value of each window. Statistics constitute the normal state Sample set ,in This represents the number of normal segment windows;

[0071] Step 82) Set the significance level ;

[0072] Step 83) Determine the detection control limit using the empirical quantile method. for Sample Quantiles:

[0073] ;

[0074] Step 84), during online detection, if the first Each window The statistic satisfies If a fault is detected, then it is determined that a fault has been detected; if If so, the process is considered to be running normally;

[0075] Step 85), Under normal conditions, The probability that the statistic exceeds the control limit is approximately ,Right now .

[0076] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.

[0077] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.

[0078] Finally, the present invention proposes a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the method described in the present invention.

[0079] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0080] 1. This invention aggregates industrial process time series data through a sliding window, and aggregates data of length [missing information]. The observations at multiple time points within a window are constructed into a histogram distribution, thereby accumulating fault information and making it more conducive to the detection of minor faults. This approach utilizes local... Divergence is used as a detection statistic to make full use of the local distribution information of anchor points. At the same time, considering the distribution difference characteristics of the detection process, an adaptive control limit setting method based on kernel density estimation is proposed.

[0081] 2. This invention solves the problem of traditional sample point detection methods being insensitive to minor faults by using a sliding window distribution aggregation mechanism. and All methods are based on detection of a single sample point. When the fault amplitude is small, the deviation of a single point may be overwhelmed by noise, resulting in insufficient detection sensitivity. However, this invention uses a length of... The data within the window is aggregated into a distribution, and the statistical characteristics of the data within the window are captured by histogram estimation. Even if the deviation of a single sample point is small, the cumulative effect of the window distribution can amplify the fault signal, thereby achieving an upgrade from point-level detection to distribution-level detection, which significantly improves the detection capability of minor faults and slowly changing faults.

[0082] 3. This invention adopts The divergence detection index solves the problem that traditional distance measurement methods are insufficient in characterizing distribution differences. The method uses Mahalanobis distance to measure the deviation of sample points from anchor points, which can only reflect positional shifts and cannot characterize changes in distribution patterns. of and The statistics are also based on distance metrics and are insensitive to distribution drift; however, this invention uses... The divergence measure is the difference between the window distribution and the anchor point distribution. Divergence, as a classic metric in information theory, can accurately quantify the relative entropy between two probability distributions. It can not only detect mean shifts but also capture complex fault modes such as variance changes and distribution shape changes, providing richer fault information than traditional distance metrics.

[0083] 4. This invention uses a nonparametric histogram estimation method to solve the problem of dependence on distribution assumptions in traditional parameterization methods. The previous method assumed that the data followed a Gaussian distribution, and the detection performance deteriorated significantly when the actual data deviated from this assumption. In contrast, this invention uses a histogram method for nonparametric distribution estimation, which eliminates the need to make any assumptions about the data distribution. quantity It adaptively captures the true distribution of data and has good adaptability to complex distributions such as non-Gaussian, multimodal, and skewed distributions, thus improving the robustness and generalization ability of the method.

[0084] 5. Local minima of the present invention The divergence strategy addresses the poor adaptability of a single reference model. Traditional methods typically establish a single normal operation model, which struggles to adapt to changes in operating conditions when multiple normal operating conditions exist. This invention, however, calculates the difference between the window distribution and the distribution of all anchor points. Divergence, take the minimum value as The statistics are equivalent to automatically selecting the best-matching local normal mode as a reference, realizing adaptive modeling of multi-condition processes, and maintaining a low false alarm rate even when the condition is switched.

[0085] 6. The method proposed in this invention is used in Tennessee Eastman School of Management (TES). Three different types of minute fault detection in chemical processes; simulation results show that, compared with , Compared with other fault detection methods, the proposed method for detecting minor process faults based on a local probability sliding window has a higher fault detection rate and a relatively lower false alarm rate. Attached Figure Description

[0086] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0087] Figure 1 This is a block diagram illustrating the principle of the minor fault detection method proposed in this invention.

[0088] Figure 2 The TE chemical process simulation flowchart used to verify the method of this invention is shown.

[0089] Figure 3 For the present invention Simulation comparison and detection diagram of process fault 3.

[0090] Figure 4 For the present invention Simulation comparison and detection diagram of process fault 9.

[0091] Figure 5 For the present invention Simulation comparison and detection diagram of process fault 15. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0093] Example 1

[0094] See Figures 1 to 5 This embodiment provides the following technical solution: a method for detecting minor process faults based on a local probability sliding window, utilizing the Tennessee Eastman (… The chemical process simulation platform generates various types of fault datasets to verify the effectiveness and performance of algorithms. The Tennessee Eastman chemical process simulation platform can simulate the nonlinear, non-Gaussian, time-varying, and multimodal characteristics of processes, providing a standard simulation model for validating various process modeling and control, fault detection and diagnosis methods. The process comprises five main operating units: reactor, condenser, compressor, separator, and stripper; it also includes four types of gas feedstocks, two main products generated from two gas-liquid exothermic reactions, and two byproducts generated from two derivative exothermic reactions. The process mechanism is complex with numerous variables, including various data fault types such as step fluctuations, random variations, slow drift, viscous fluctuations, and constant position. The process contains 41 measured variables and 12 control variables. Additive noise is superimposed on each measured variable to simulate noise in actual industrial processes. Each dataset contains 52 variables, with most variables sampled for 3 minutes; 14 variables are sampled for 6 minutes, and 5 variables are sampled for 15 minutes. Figure 2 Depicting Process flow diagram.

[0095] Figure 1 This is a block diagram illustrating the principle of the fault detection method of the present invention. The proposed method for detecting minor faults in industrial processes mainly includes three stages. Stage one involves employing... The framework generates anchor points and establishes a local distribution model for each anchor point; in the second stage, a sliding window technique is used to perform distribution-level aggregation on the time series data; in the third stage, local distribution models are calculated. Divergence statistics and adaptive detection control limits are set.

[0096] 1) Anchor point generation and local distribution modeling

[0097] Anchor point generation uses The framework determines the optimal set of anchor points by traversing local radius parameters. Anchor points are representative sample points in the normal operation data, and each anchor point represents a typical normal operation mode.

[0098] First, the normal operating data is standardized. Let the data matrix be... ,in For the sample size, For the number of variables. Process data, , Calculate the mean of each variable. and standard deviation :

[0099] ,

[0100] The standardized data are:

[0101]

[0102] in, Indicates the first The first sample Standardized values ​​of each variable.

[0103] Next, traverse the local radius. For each The value is used to calculate the local sample count and the total approximation error for all sample points. Let... Let be the covariance matrix of the data, for sample points Its local Mahalanobis distance is defined as:

[0104]

[0105] For each After generating the anchor point set, calculate the total approximation error:

[0106]

[0107] in, Let represent the Mahalanobis distance. Choose the value that minimizes the total approximation error. The value is used as the optimal local radius. In this embodiment, the following settings are provided. , The search scope is Step size is .

[0108] Sure Then, all sample points that meet the minimum local sample number requirement are used as anchor points to form an anchor point set. ,in This represents the total number of anchor points.

[0109] For each anchor point Establish its local distribution model. First, calculate the local modeling radius:

[0110]

[0111] in, The local radius coefficient is set to [value] in this embodiment. For anchor points Extract all that satisfy The samples constitute the local sample set of the anchor point.

[0112] For each variable Construct a histogram distribution of all observations of this variable in a local sample set. Histogram Quantity set to , The boundaries are determined when processing the first anchor point, and all subsequent anchor points share the same bin boundaries to ensure comparability of distributions. Specifically, for variables... ,That The boundary is:

[0113]

[0114] in, and Variables The minimum and maximum values, This is the boundary extension amount, used to cover any new data that may be added.

[0115] variables in local samples The observations were assigned to the corresponding In the middle, statistics for each The frequency of the numbers is then normalized to obtain the probability distribution. :

[0116]

[0117] in, Indicates anchor point In local samples, variables Falling into indivual The number of samples. To avoid numerical problems caused by zero probability, the probability distribution is smoothed:

[0118]

[0119] in, This is a smoothing parameter. In this embodiment, quantity Set as .

[0120] 2) Sliding window distribution aggregation

[0121] A sliding window technique is used to aggregate multiple consecutive observations. Let the length of the sliding window be... For the first Multiple windows are used to extract data from time series data. A series of samples:

[0122]

[0123] in, Indicates the first The observation vector at each time step. In this embodiment, the window length... Set as .

[0124] For windows Each variable in ,Will The observations are aggregated into a histogram distribution. Use the same local distribution as the anchor points. Boundaries, variables within the window of Each observation is assigned to the corresponding In the middle, the frequency is counted and normalized:

[0125]

[0126] in, Display window Medium variables Falling into indivual The number of samples. The window distribution is also smoothed accordingly:

[0127] ,in, For smoothing parameters

[0128] 3) Local Divergence calculation and control limit determination

[0129] For the Each window is used to calculate its relationship with the local distribution of each anchor point. Divergence. For anchor points and variables , Divergence is defined as:

[0130]

[0131] when At that time, the contribution was ;when or At that time, skip that Because both anchor point distribution and window distribution have been smoothed, adding... In actual calculations, the zero probability case rarely occurs.

[0132] For all 1 variable The variance-weighted summation of the divergence yields the window. With anchor point The total between Divergence:

[0133]

[0134] in, anchor point The The standard deviation of a variable in its local sample set. The variance weighting mechanism gives higher weights to variables with greater volatility, while the influence of variables with less volatility (which may be noise or constant) is automatically reduced.

[0135] For all Repeat the above calculation for each anchor point to obtain... divergence vector Select all anchor points The minimum value of the divergence is taken as Statistic:

[0136]

[0137] The physical meaning of this strategy is: to select the local normal pattern that is closest to the current window distribution for comparison, thereby achieving adaptive multimodal modeling. If the current window differs significantly from the local distribution of all anchor points, then... A large value indicates a possible malfunction.

[0138] The empirical quantile method is used to determine the detection control limits. Assume the limits under normal operating conditions are... The sample is According to the significance level (Set to 0.01 in this embodiment), determine the detection control limit. for Sample Quantiles:

[0139]

[0140] That is, under normal conditions, The probability that the statistic exceeds the control limit is approximately When the first window The statistic satisfies If the condition is met, it is considered a fault; otherwise, it is considered normal.

[0141] Based on the above detection method, the specific detection steps of this invention can be divided into two aspects: offline training and online detection, specifically including:

[0142] A. Offline training

[0143] 1) Acquire multivariate time series data of industrial processes, including normal data from the training set. and test set data ,in The number of training samples, For the number of test samples, Number of variables;

[0144] 2) Standardize the normal data in the training set and calculate the mean vector. and standard deviation vector :

[0145] ,

[0146] Standardize the training and test sets using the mean and standard deviation of the training set:

[0147]

[0148] 3) Adopt Frame generates anchor point set Traverse the local radius For each Values ​​are used to generate the corresponding set of anchor points, and the total approximation error of all training samples is calculated. Let... Let the covariance matrix of the training set be denoted by the Mahalanobis distance, which is defined as:

[0149]

[0150] For each Value, calculate the total approximation error:

[0151]

[0152] Choose the option that minimizes the total approximation error. Value as Then, the final set of anchor points is generated: first, the data mean is used as the first anchor point; then, the samples are traversed in ascending order according to Mahalanobis distance. For each unprocessed sample, its... All unprocessed samples within the radius, if the sample size is not less than Then the mean of these samples is used as the new anchor point. In this embodiment, the following is set: , The search scope is Step size is ;

[0153] 4) For each anchor point Establish a local distribution model. Calculate the local modeling radius. ,in For anchor points Extract all samples from the training set that satisfy the following conditions: The samples constitute the local sample set for that anchor point. For each variable Construct a histogram distribution of all observations of this variable in a local sample set. , quantity , The boundary is determined based on the global minimum and maximum values ​​of the training set;

[0154] 5) Apply the sliding window technique to the normal segments of the training and test sets. Window length. To ensure the first window contains the complete... Each sample is concatenated with the last part of the training set before the test set. One sample as a prefix:

[0155]

[0156] For each window Calculate the histogram distribution for each variable. ;

[0157] 6) Calculate the value of each window. Statistics. For windows and anchor points ,calculate Divergence:

[0158]

[0159] Among them, the summation is only performed on conditions that satisfy and Proceed. For all 1 variable Divergence is then summed using variance-weighted methods:

[0160]

[0161] in, anchor point The The standard deviation of a variable in its local sample set.

[0162] Choose the minimum value as Statistic:

[0163]

[0164] 7) Determine the detection control limits using the empirical quantile method. Calculate the window size for the normal segment of the test set. Statistics constitute the normal operating conditions Sample set. Assume the sample set is in the normal state. The sample is According to the significance level Determine detection control limits for Quantiles:

[0165]

[0166] 8) Save the set of anchor points Local distribution of anchor points , Boundary, standardized parameters and Detection and control limits Used for online detection.

[0167] B. Online detection

[0168] 1) Obtain the first Observations at each sampling time Then, in conjunction with history A sliding window is constructed from the observations. ;

[0169] 2) Standardize the observations within the window, using the mean value saved during the offline training phase. and standard deviation ;

[0170] 3) For each variable , in the window The observations are aggregated into a histogram distribution. Use the data saved during the offline training phase boundary;

[0171] 4) Calculation window With each anchor point Between divergence Choose the minimum value as Statistic ;

[0172] 5) Judgment: If This indicates that the process is running normally; if This indicates that a fault has been detected.

[0173] Simulation Validation: Various fault datasets of different types were generated using the Eastman Chemical Process Simulation Platform in Tennessee to verify the effectiveness and performance of the algorithm. The Eastman Chemical Process Simulation Platform can simulate the nonlinear, non-Gaussian, time-varying, and multimodal characteristics of processes, providing a standard simulation model for validating various process modeling and control, fault detection, and diagnosis methods. The process comprises five main operating units: reactor, condenser, compressor, separator, and stripper; it also includes four types of gas feedstocks, two main products generated from two gas-liquid exothermic reactions, and two byproducts generated from two derivative exothermic reactions. The process mechanism is complex with numerous variables, including various data fault types such as step fluctuations, random variations, slow drift, viscous fluctuations, and constant position. The process contains 41 measured variables and 12 control variables. Additive noise is superimposed on each measured variable to simulate noise in actual industrial processes. Each dataset contains 52 variables, with most variables sampled for 3 minutes; 14 variables are sampled for 6 minutes, and 5 variables are sampled for 15 minutes. Figure 2 Depicting Process flow diagram.

[0174] The simulation example used test data for faults 3, 9, and 15 for verification. The training data contained 500 normal operation samples, and the test data contained 960 samples, of which the first 160 sampling times were the normal operation segment, and the last 800 sampling times were the minor fault segment (the fault was introduced at the 161st sample time).

[0175] To compare under the same conditions method, Methods and The detection performance of the three methods was assessed using the same training and testing data. The number of principal components in the method is determined by the cumulative contribution rate; The method of anchor point generation and The method is the same; The method parameter is set to: sliding window length histogram quantity Local radius coefficient Minimum number of local samples significance level .

[0176] Figure 3 (a)(b) Figure 4(a)(b) and Figure 5 (a) and (b) respectively drew The method generated and Statistical test graph Figure 3 (c) Figure 4 (c) and Figure 5 (c) Drawn The statistical test chart, Figure 3 (d) Figure 4 (d) and Figure 5 (d) Drawn The statistical test plot shows that... The method exhibits higher fault detection sensitivity and a lower false alarm rate. During normal operation (samples 1-160). The statistics remained below the control limits, and the number of false alarms was low; in the fault segment (samples 161-960). A rapid increase in the statistic and its sustained exceedance of the control limit indicates that the fault has been successfully detected.

[0177] To more accurately compare detection performance, the example uses two metrics: fault detection rate and false alarm rate. The specific calculation formulas are as follows:

[0178]

[0179]

[0180] in, This indicates the number of correctly detected fault samples; This indicates the number of undetected fault samples; This represents the number of normal samples that were false positives. This represents the number of correctly identified normal samples.

[0181] Table 1

[0182]

[0183] Table 1 presents comparative data on the fault detection rate and false alarm rate for the three fault types. On the one hand, it can be seen from Table 1 that, compared with... , In comparison, the proposed The method has the highest fault detection rate. In the case of fault 3, The method's detection rate was 99.0%. The method's detection rate was 0.6%, while The method's detection rate was 4.5%. In case of fault 9, The method's detection rate was 70.9%. The method's detection rate was 0.6%, while The method's detection rate was 3.9%. In case of fault 15, The method's detection rate was 100.0%. The method's detection rate was 0.6%, while The method achieved a detection rate of 6.5%. In other words, compared to the other two methods, the proposed method... The method uses sliding window distribution aggregation and local Divergence metrics can more effectively capture distribution changes caused by faults. Furthermore, as can be seen from Table 1, A lower false alarm rate was achieved compared to the other two fault types. Simulation results show that the proposed method... The method not only has relatively high fault detection sensitivity, but also a low false alarm rate.

[0184] Example 2: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.

[0185] Example 3: This example proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this invention, which will not be repeated here.

[0186] Example 4: This example proposes a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the method described in this invention, which will not be repeated here.

[0187] It should be noted that the processing flow of embodiments 2-4 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.

[0188] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0189] 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 method for detecting minute process faults based on a local probability sliding window, characterized in that, Includes the following steps: Step 1) Obtain multivariate time series data of industrial processes, including training set data and test set data; the training set data is used for anchor point generation and distribution modeling, and the test set data is used for minor fault detection. Step 2) Data standardization: Calculate the mean and standard deviation of the training set data, and standardize the training and test set data; Let the training set data be ,in For the sample size, Number of variables; calculate the first... Mean of each variable and standard deviation : ， ； The standardized data are: ； Step 3), Anchor point generation: using... Frame generates anchor point set ,in The number of anchor points; by traversing the local radius Choose the one that minimizes the total approximation error. value as optimal radius ; Step 4) Local distribution modeling of anchor points: For each anchor point Establish Local probability distribution within the radius: ,in This is the local radius coefficient; For anchor points Calculate the Mahalanobis distance between samples in its local neighborhood: ； in, Let be the covariance matrix of the training set data; take the condition that... For local samples, construct a histogram distribution for each variable. , Quantity is ; Step 5) Sliding window data extraction: Apply a length of [length missing] to the test set data. Extract the sliding window data matrix. ;No. Each window contains samples ; Step 6) Window Distribution Aggregation: Aggregate the distribution of windows Each variable within ,Will The observations are aggregated into a histogram distribution. ; used for anchor points with the same distribution Boundaries, frequency distribution of data within the statistical window and normalization; Step 7) Calculate the local Divergence: for a window Calculate its relationship with the distribution of all anchor points. Divergence, take the minimum value as Statistic; Step 8) Control Limit Determination and Fault Detection: Calculate based on training set or normal operation data. Control limits for the statistic; use empirical quantiles, based on the significance level. Determine detection control limits , making ;like Then determine the first A fault was detected in one window; if If so, the process is considered to be running normally.

2. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, Step 7) includes the following steps: Step 71) For anchor points and window distribution ,calculate Divergence: ； Step 72) Divergence is then summed using variance-weighted methods: ； in, anchor point The The standard deviation of each variable in its local sample set; Step 73) The statistic is defined as: 。 3. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, The fault detection method also includes online detection, comprising the following steps: S1, Obtain the first After sampling the observations at each time point, a sliding window data matrix of length l is constructed by combining it with historical data. , which includes samples ; S2, for the window Each variable within ,Will The observations are aggregated into a histogram distribution. Use the same distribution as the anchor points Boundaries, frequency distribution of data within the statistical window and normalization; S3, Calculate the window distribution and the distribution of all anchor points. Divergence, take the minimum value as Statistic: ； S4. If This indicates that a fault has been detected; if This indicates that the process is running normally.

4. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, Step 3) Anchor point generation includes the following steps: Step 31) Set the minimum number of local samples and local radius search range ; Step 32) Calculate the covariance matrix of the training set data. For any two sample points and Its Mahalanobis distance is defined as: ； Step 33) Traverse the local radius For each To determine the value, perform the following operations: use the data mean as the first anchor point; iterate through the samples in ascending order of Mahalanobis distance, and for each unprocessed sample, find its... All unprocessed samples within the radius; if the number of samples within the radius is not less than Then the mean of these samples will be used as the new anchor point; The sum of the Mahalanobis distances from all samples to the nearest anchor point is calculated as the total approximation error: ； Step 34) Select the option that minimizes the total approximation error. value as the optimal local radius ; Step 35) Use Generate the final anchor set ,in This represents the total number of anchor points.

5. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, Step 4) Local distribution modeling of anchor points includes the following steps: Step 41) Calculate the local modeling radius ,in This is the local radius coefficient; Step 42) For each anchor point Extract all data in the training set that satisfy the following conditions: The samples constitute the local sample set of the anchor point. ; Step 43) Determine the histogram Boundaries: for variables ,That The boundary is: ； in, and Variables The minimum and maximum values ​​in the training set. For boundary expansion, for Quantity; all anchor points share the same boundary; Step 44) For anchor points Local sample set Each variable in Assign all observations of this variable to the corresponding In the middle, statistics for each The frequency of the numbers is normalized to obtain the probability distribution: ； in, Indicates anchor point Variables in local samples Falling into indivual The number of samples; Step 45) Smooth the probability distribution; ； in, For smoothing parameters.

6. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, Step 5) includes the following steps: Step 51) Set the length of the sliding window To ensure the first window contains the complete Each sample is concatenated with the last part of the training set before the test set. One sample as a prefix: ； Step 52), for the first Multiple windows are used to extract data from time series data. A window data matrix is ​​formed by consecutive samples: ； in, Indicates the first The observation vector at each time step; Step 53) For the window Each variable in ,Will Each observation is assigned to the same distribution as the anchor point. Within the boundary, count the frequencies and normalize them: ； in, Display window Medium variables Falling into indivual The number of samples; Step 54) Smooth the window distribution: ； in, For the smoothing parameter, use the same value as the anchor point distribution smoothing process.

7. The method for detecting minor process faults based on a local probability sliding window according to claim 1, characterized in that, Step 8) Determining the control limits includes the following steps: Step 81) Apply a sliding window to the normal operation segment data in the test set and calculate the value of each window. Statistics constitute the normal state Sample set ,in This represents the number of normal segment windows; Step 82) Set the significance level ; Step 83) Determine the detection control limit using the empirical quantile method. for Sample Quantiles: ； Step 84), during online detection, if the first Each window The statistic satisfies If a fault is detected, then it is determined that a fault has been detected; if If so, the process is considered to be running normally; Step 85), Under normal conditions, The probability that the statistic exceeds the control limit is approximately ,Right now .

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 7 when invoked by a processor.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images