Time sequence sampling data fault-tolerant mining system and method oriented to variation process

Abstract

The invention belongs to the technical field of data mining of artificial intelligence, and particularly provides a time sequence sampling data fault-tolerant mining system and method oriented to a variation process, and the system comprises a data processing module which is used for obtaining a decomposed sampling data sequence; the excavator construction module is used for constructing an excavator by taking a sliding window double median filter as a main part and combining a signal extraction accessory, a model fitting accessory and a residual error generation accessory; the fault-tolerant mining module is used for performing signal form fault-tolerant mining, model structure information fault-tolerant mining and noise statistical characteristic fault-tolerant mining on the decomposed sampling data sequence by using an excavator to obtain a fault-tolerant mining result; and the working condition identification module is used for monitoring the working condition of the industrial production process and performing fault diagnosis based on the fault-tolerant mining result. According to the method, the problems of state monitoring and safety early warning in the operation process of the complex system are effectively solved, and the method has important value for process system safety analysis, abnormal change detection, fault diagnosis and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data mining of artificial intelligence, and particularly relates to a time series sampling data fault-tolerant mining system and method for abnormal change process. BACKGROUND

[0002] In the operation process of a large oil and chemical unit, most of the first-hand data obtained are data sequences of various instruments or sensors sampled over time, which are referred to as process sampling data. Finding useful information from the sampling data, establishing the time-varying law between the system operation process states or parameters and the correlation with fault phenomena, and having wide application in state monitoring, process control, safety warning and fault diagnosis of the large unit operation.

[0003] From the method level, time series data mining, as an important branch of data mining and knowledge discovery research, is significantly different from the data mining of repeatable sampling objects. The core is to extract the internal information and knowledge such as the shape, evolution process, change law and abnormal change of data from the process sampling data sequence, usually including data change process shape mining, information mining and knowledge mining, etc., aiming to study the changes implied in the sequence, including trends, patterns, periodicity, similarity and correlation between sequences. Time series data mining is an object-oriented and purpose-driven activity. Due to the difference of research objects or mining targets, there are many different mining methods. However, the existing time series data mining methods mainly focus on anomaly point mining, and there are few mining methods for the trend characteristics of time series; fault-tolerant mining for data containing wild values is still blank.

[0004] The application takes the sampling data of abnormal change in the production process as the object, and provides a time series sampling data fault-tolerant mining system and method for abnormal change process. SUMMARY

[0005] In order to solve the problems in the prior art, the application provides a time series sampling data fault-tolerant mining system and method for abnormal change process, which establishes signal shape mining, signal change model structure mining and noise statistical property mining technologies with strong fault-tolerant capability around the key technologies of state monitoring and safety warning of complex system operation process.

[0006] In order to achieve the above object, the application provides the following scheme: A time series sampling data fault-tolerant mining system for abnormal change process, comprising: A data processing module is used for collecting time series sampling data sequences in the abnormal change period of industrial production process, and performing unified representation and model decomposition on the time series sampling data sequences to obtain decomposed sampling data sequences. The mining mechanism modeling module is mainly a sliding window double median filter, combined with a signal extraction accessory, a model fitting accessory, and a residual generation accessory, to construct a mining mechanism; The fault-tolerant mining module is configured to perform signal pattern fault-tolerant mining, model structure information fault-tolerant mining, and noise statistical characteristic fault-tolerant mining on the decomposed sampled data sequence by using the mining mechanism, to obtain a fault-tolerant mining result. The working condition recognition module is configured to monitor the working condition of the industrial production process and perform fault diagnosis based on the fault-tolerant mining result.

[0007] Preferably, in the mining mechanism, The signal extraction accessory is configured to perform signal pattern fault-tolerant mining on the decomposed sampled data sequence, and the signal pattern fault-tolerant mining includes constant value pattern mining, monotonic and convex-concave pattern mining, and periodic pattern mining. The model fitting accessory is configured to perform model structure information fault-tolerant mining on the decomposed sampled data sequence. The residual generation accessory is configured to perform noise statistical characteristic fault-tolerant mining on the decomposed sampled data sequence, and the noise statistical characteristic fault-tolerant mining includes first-order stationary information mining and second-order stationary information mining.

[0008] Preferably, the process of performing constant value pattern mining on the decomposed sampled data sequence includes: A median algorithm is used to sort the decomposed sampled data sequence of a preset length by value and take the middle value to construct a double median filter. The double median filter is used to perform signal extraction on the decomposed sampled data sequence to construct a deviation sequence. A first allowable deviation amplitude constant is selected, and the first allowable deviation amplitude constant is combined with the deviation sequence to construct a first symbol sequence. The decomposed sampled data sequence of a preset length is divided into three types of data segments using the first symbol sequence, and the longest run length index of each type of data segment is counted. The longest run length index is compared with a preset threshold to determine whether the divided data segments belong to a constant value pattern, and the constant value pattern mining on the decomposed sampled data sequence is completed.

[0009] Preferably, the process of performing monotonic and convex-concave pattern mining on the decomposed sampled data sequence includes: The deviation sequence is used to perform first-order difference and second-order difference to obtain a first-order difference sequence and a second-order difference sequence. A second allowable deviation amplitude constant is selected, and the second allowable deviation amplitude constant is combined with the first-order difference sequence and the second-order difference sequence to construct a second symbol sequence and a third symbol sequence. The second symbol sequence and the third symbol sequence are used to divide the decomposed sampling data sequence of the preset length into six types of data segments, and the number of data points of each type of data segment is counted. The number of data points is compared with the preset threshold to determine whether the divided data segment belongs to a monotonic or convex-concave form, thereby completing the monotonic and convex-concave form mining of the decomposed sampling data sequence.

[0010] Preferably, the process of mining the periodic form of the decomposed sampling data sequence comprises: A sliding window periodogram of the deviation sequence is calculated, and a G-test statistic is constructed based on the sliding window periodogram. A threshold constant is obtained using a Fisher distribution table of G-test and a preset confidence level, the threshold constant is compared with the G-test statistic to determine whether the decomposed sampling data sequence has a periodic component, thereby completing the periodic form mining of the decomposed sampling data sequence.

[0011] Preferably, the process of model structure information fault-tolerant mining of the decomposed sampling data sequence comprises: An algebraic polynomial fitting model, a trigonometric polynomial fitting model, a periodic progressive model, an autoregressive AR model, and an autoregressive AR I model with a periodic component are used to construct a linear structure parameter model family. The AIC information criterion and the least square fitting algorithm are used to extract the signal sequence obtained by the double median filter to perform model order determination and residual sum of squares minimization on the linear structure parameter model family, thereby obtaining an optimal model family. Based on the optimal model family, the model structure information fault-tolerant mining of the decomposed sampling data sequence is completed.

[0012] The application also provides a time series sampling data fault-tolerant mining method for an abnormal process, which applies the system and comprises: Time series sampling data sequences of an abnormal change period of an industrial production process are collected, and the time series sampling data sequences are uniformly represented and model decomposed to obtain decomposed sampling data sequences. A sliding window double median filter is used as the main part, combined with a signal extraction accessory, a model fitting accessory, and a residual generation accessory to construct a mining machine. The mining machine is used to perform signal form fault-tolerant mining, model structure information fault-tolerant mining, and noise statistical characteristic fault-tolerant mining on the decomposed sampling data sequence to obtain fault-tolerant mining results. Based on the fault-tolerant mining results, the working conditions of the industrial production process are monitored and fault diagnosis is performed.

[0013] Compared with the prior art, the beneficial effects of the present application are: the present application takes abnormal change as the research object, focuses on the complex system operation process state monitoring and safety early warning problem, establishes signal shape mining over time, model mining of process signal structure and noise stationary or not statistical characteristic mining methods and algorithms with good fault tolerance ability to outliers and patchy outliers, gives a time series data mining system framework composed of three elements of process sampling time series data, model and mining machine, and verifies the correctness of the algorithm. The mining method and system can be used in different fields such as petrochemical industry, measurement and control engineering, navigation control, etc., and provide practical technical support for data-driven safety analysis, abnormal change detection and fault diagnosis technology research and engineering application. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments are briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0015] Figure 1 The structure diagram of the time series sampling data fault-tolerant mining system of the present application embodiment facing abnormal process is shown in the figure. Figure 2 The mining machine architecture diagram of the present application embodiment is shown in the figure. Figure 3 The sampling sequence variation scatter plot of the time-varying vibration component of the present application embodiment is shown in the figure. Figure 4 The difference sequence diagram of the present application embodiment is shown in the figure. Figure 5 The actual sampling diagram of the first group of experiments of the present application embodiment is shown in the figure. Figure 6 The first-order difference sequence diagram of the first group of experiments of the present application embodiment is shown in the figure. Figure 7 The actual sampling diagram of the second group of experiments of the present application embodiment is shown in the figure. Figure 8 The second-order difference sequence diagram of the second group of experiments of the present application embodiment is shown in the figure. Figure 9 The actual sampling diagram of the third group of experiments of the present application embodiment is shown in the figure. Figure 10 The second-order difference sequence diagram of the third group of experiments of the present application embodiment is shown in the figure. Figure 11 The actual sampling diagram of the fourth group of experiments of the present application embodiment is shown in the figure. Figure 12 The second-order difference sequence diagram of the fourth group of experiments of the present application embodiment is shown in the figure. Figure 13 This is a periodic diagram of an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1: like Figure 1 As shown, a fault-tolerant mining system for time-series sampling data oriented towards anomaly processes includes: a data processing module, an excavator construction module, a fault-tolerant mining module, and a working condition identification module.

[0019] The data processing module is used to collect time-series sampled data sequences during periods of abnormal changes in the industrial production process, and to perform unified representation and model-based decomposition on the time-series sampled data sequences to obtain decomposed sampled data sequences. In a further implementation, the decomposed sampled data sequences include the target real signal, random disturbance noise, and error components.

[0020] Time-series sampled data sequences are the most common form and format of instrument sampling data in the field of industrial process control, consisting of sampling time. and the measurement value at that moment It consists of two parts, forming a time-series sampling data sequence. The number of its sample points is The fragment can be summarized as Generally, time-series sampled data sequences any sample point All can be modeled and decomposed into the actual signal of the target being measured. Random disturbance noise and error components It consists of three parts: (1) If by time The horizontal axis represents the measured value. For the longitudinal axis, draw the sample data scatter plot, then with the measured object (such as temperature, pressure, flow rate, flow, vibration, displacement, etc.), noise changes and the probability of error occurs, the scatter plot will present a variety of complex shape. Production process will produce a large number of similar above-mentioned structure of the sample data sequence. In order to accurately perceive the production process and timely discovery of potential safety risks, the need to obtain regular knowledge from a large number of time series data. For this, information mining is the most direct technical approach. Specifically, for a set of unknown sources, background or physical meaning of the sampling data sequence, which information is necessary to prioritize mining, how to describe and mine these information, is an important problem that must be solved in time series sampling data sequence mining.

[0021] The information mining of time series sampling data sequence will focus on the signal form as important information, and mine the signal body information and noise information under the premise of effectively weakening the influence of wild value.

[0022] The mining mechanism modeling module is used to construct a mining machine mainly with a sliding window double median filter, combined with a signal extraction accessory, a model fitting accessory and a residual generation accessory, as shown in Figure 2 .

[0023] The fault-tolerant mining module is used to perform signal form fault-tolerant mining, model structure information fault-tolerant mining and noise statistical characteristic fault-tolerant mining on the decomposed sampling data sequence by using the mining machine, to obtain a fault-tolerant mining result.

[0024] Further embodiments are that in the mining machine, The signal extraction accessory is used to perform signal form fault-tolerant mining on the decomposed sampling data sequence; the signal form fault-tolerant mining includes constant value form mining, monotone and convex-concave form mining and periodic form mining; the measured target real signal The magnitude can be various (such as temperature, pressure, flow rate, flow, vibration, displacement, etc.), but the signal scatter plot presents a relatively stable form. From the information content point of view, the magnitude size will change with the unit of measurement, but the scatter plot does not, and the scatter plot form has good form invariance: time scale translation invariance and magnitude translation invariance. Form mining is an important part of information that discovers the inherent regularity of signal changes.

[0025] The typical performance form of the data sequence has two categories of constant value form and time-varying form. Among them, the time-varying form can be further divided into monotone convex, monotone concave, periodic change, periodic progressive change and other forms, and the present application briefly gives the first three form mining methods of constant value form and time-varying form.

[0026] Further embodiments are that the process of performing constant value form mining on the decomposed sampling data sequence includes: The median algorithm is used to sort the decomposed sampling data sequence of the preset length by value and take the middle value to construct a double median filter. Specifically, for any form of decomposed sampling data sequence, a set of double median filters with good outliers-tolerant ability is constructed: the median algorithm is used to sort the decomposed sampling data sequence of the preset length by value and take the middle value to design the filter: , (2) j, i w , and k represent the point sequence number of the sampling point, h represent the radius of the sliding window,

[0027] The double median filter is used to extract the signal from the decomposed sampling data sequence to construct a deviation sequence. Specifically, the filter is used to extract the signal from the decomposed sampling data sequence with high fidelity , and a morphological test algorithm based on the deviation run length index is constructed. The deviation sequence is constructed as follows: , (3) n

[0028] A first allowable deviation constant is selected, and a first symbol sequence is constructed based on the deviation sequence: , (4) The first symbol sequence is used to divide the decomposed sampling data sequence of the preset length into three types of data segments, and the longest run length index of each type of data segment is counted. Specifically, the data segment is divided into three types of data segments , and according to the continuity of the symbols "+", "o" and "-", and the longest run length index (the number of data points of the same symbol) of each type of data segment , and is counted.

[0029] The longest run length index is compared with a preset threshold to determine whether the divided data segment belongs to a constant value form, and the constant value form mining of the decomposed sampling data sequence is completed. Specifically, if and , it can be considered that the sampling sequence approximately belongs to the constant value form.

[0030] ​​​​​​The monotonous interval and convex-concave interval of the time series change are important links for grasping the time series change rule and discovering abnormal change. Further implementation is that the process of monotonous and convex-concave form mining of the decomposed sampling data sequence comprises: using the deviation sequence , performing first-order difference and second-order difference to obtain first-order difference sequence and second-order difference sequence : , (5) , (6) selecting a second allowable amplitude constant , combining the first-order difference sequence and the second-order difference sequence to construct a second symbol sequence and a third symbol sequence : and . (7) using the second symbol sequence and the third symbol sequence , dividing the decomposed sampling data sequence of the preset length into six types of data segments , , , , and , and counting the number of data points , , , , and of each type of data segment.

[0031] comparing the number of data points with a preset threshold value to determine whether the divided data segment belongs to a monotonous or convex-concave form, and completing the monotonous and convex-concave form mining of the decomposed sampling data sequence. Specifically, if and , it is determined that the corresponding data segment is in a monotonous decreasing form; if and , it is determined that the corresponding data segment is in a monotonous increasing form. If and , it is determined that the corresponding data segment is in a concave form; if and , it is determined that the corresponding data segment is in a convex form.

[0032] Periodogram is a very effective mathematical tool to mine whether there is a periodic component in the time series data sequence of physical quantities with periodic variation characteristics (such as vibration, temperature variation).

[0033] Further embodiments are directed to a process for performing periodic pattern mining on the decomposed sampling data sequence, comprising: calculating a deviation sequence a sliding window periodogram , and constructing a G-test statistic based on the sliding window periodogram

[0034] The sliding window periodogram is defined as: , , (8) where s represents the sampling data, and represent the deviations at different times, k represents the radius of the sliding window, n and n is the number of sample points.

[0035] The G-test statistic is constructed as: , (9) The threshold constant is obtained using the Fisher distribution table and the preset significance level The threshold constant is compared with the G-test statistic to determine whether the decomposed sampling data sequence has a periodic component, thereby completing the periodic pattern mining on the decomposed sampling data sequence. Specifically, the discrimination rule includes: When , it means that the decomposed sampling data sequence does not have a significant periodic component; otherwise, a scatter plot with the abscissa and the ordinate is drawn, and each peak appearing in the scatter plot means that the time series has a periodic component with a frequency of The more significant the periodicity, the larger and steeper the corresponding peak in the sliding window periodogram.

[0036] The model fitting attachment is used to perform model structure information fault-tolerant mining on the decomposed sampling data sequence.

[0037] Further embodiments are directed to a process for performing model structure information fault-tolerant mining on the decomposed sampling data sequence, comprising: ​​A family of linear structural parameter models is constructed using algebraic polynomial fitting models, triangular polynomial fitting models, periodic progressive models, autoregressive (AR) models, and autoregressive ARI models with periodic components. Specifically, based on Weierstrass's first and second approximation theorems, the most intuitive time-domain fitting models for one-dimensional sampled sequences are algebraic polynomial fitting models for trend components, triangular polynomial fitting models for periodic components, and periodic progressive models that combine the above two cases. Furthermore, in the field of time series analysis, this can reflect the decomposition of sampled data sequence components. Other mathematical models for correlation include autoregressive (AR) models and autoregressive ARI models with periodic components. Therefore, a family of linear structural parameter models is constructed. (10) (11) (12) (13) (14) In the formula For a time-step (one sampling interval) shift operator, satisfying the rule , Represents the set of algebraic polynomials. The set representing trigonometric polynomials, A set representing a periodic progression model. Represents the set of autoregressive models. This represents the set of autoregressive models with periodic components. , and Represents the model coefficients. Represents the set of real numbers. and Represents the order of the polynomial. Represents the set of positive integers. Represents frequency parameters. Sample data representing different times, d This represents the order of difference.

[0038] Signal sequence extracted using a dual median filter The AIC information criterion and least squares fitting algorithm are used to determine the order of the family of linear structure parameter models. and the sum of squared residuals Minimize to obtain the optimal family of models. Specifically, the sum of squared residuals. Minimum value The expression is as follows: (15) Based on the optimal model family, the fault-tolerant mining of model structure information of the decomposed sampled data sequence is completed. Specifically, the model family corresponding to the minimum value is determined based on equation (15), and then the model structure information is mined. For example, if Therefore, it can be inferred that, given a family of linear structure parameter models... In the middle, the sampling object The changes conform to the structural characteristics of a periodic progressive model.

[0039] The residual generation appendix is ​​used to perform noise statistical characteristic tolerance mining on the decomposed sampled data sequence; noise statistical characteristic tolerance mining includes first-order stationarity information mining and second-order stationarity information mining. Specifically, whether random noise is stationary is important information about whether the production process is normal. Considering random noise... Excessive noise can weaken or overwhelm the signal; non-stationary noise can alter or even distort the shape of a scatter plot (for example, first-order non-stationary noise distorts the signal's shape, while second-order non-stationary high-intensity noise overwhelms the signal's shape). The statistical characteristics of noise are important indicators for data mining. The stationarity of noise sequences can be categorized as strong or weak. Weak stationarity is further subdivided into first-order, second-order, and higher-order stationarity. Therefore, this invention establishes methods for mining first-order and second-order stationary information.

[0040] Specifically, the process of first-order stationary information mining is as follows: For the residual sequence of double median fault-tolerant filtering It mainly consists of two parts: disturbance noise and gross error. A suitable sliding window radius should be selected. (Default value 5), calculate the sliding window tolerance median: (16) And detection statistics: (17) when When the noise sequence is determined to have first-order stationary statistical characteristics, it is determined that the noise sequence has these characteristics.

[0041] The process of second-order stationary information mining is as follows: For the residual sequence of double median fault-tolerant filtering Select an appropriate window radius. (Default value 5), calculate the sliding window tolerance deviation: (18) And detection statistics: (19) when When the noise sequence is determined to have second-order stationary statistical characteristics, it is determined that the noise sequence has second-order stationary statistical characteristics.

[0042] The operating condition identification module is used to monitor the operating conditions of industrial production processes and perform fault diagnosis based on the results of fault-tolerant mining.

[0043] Example 2: This invention also provides a fault-tolerant mining method for time-series sampling data oriented towards anomaly processes, and an application system comprising: The time-series sampling data sequence of abnormal changes in the industrial production process is collected, and the time-series sampling data sequence is uniformly represented and decomposed into a model to obtain the decomposed sampling data sequence.

[0044] An excavator was constructed primarily using a sliding window dual median filter, combined with signal extraction attachments, model fitting attachments, and residual generation attachments.

[0045] An excavator was used to perform fault tolerance mining on the decomposed sampled data sequence, including signal morphology fault tolerance mining, model structure information fault tolerance mining, and noise statistical characteristic fault tolerance mining, to obtain fault tolerance mining results.

[0046] Based on the results of fault tolerance mining, the operating conditions of industrial production processes are monitored and fault diagnosis is performed.

[0047] Example 3: This embodiment provides a specific time-series sampling data fault-tolerant mining process.

[0048] (1) Description of time-series sampled data sequence.

[0049] Taking the compressor rotor vibration component as the object, the variation relationship of its periodic component can be modeled and represented as a sine function without loss of generality. Discrete-time sampling sequence, where, For amplitude, Let be the vibration frequency. To facilitate a more intuitive demonstration of the algorithm's effect, a dimensionless transformation is performed on the amplitude and frequency, and the transformed value is taken. and The Gaussian noise, compressed by 100 times, is then superimposed, as shown in equation (20). 10,000 sets of sampling data with Gaussian noise were collected. . Figure 3 This is a scatter plot showing the changes in the sampling sequence of vibration components over time.

[0050] , (20) Where, (N=10000).

[0051] (2) Signal morphology fault tolerance mining.

[0052] A) In the verification algorithm for fixed-value pattern mining, an appropriate allowable deviation threshold constant c=0.06 is selected. The difference sequence obtained by algorithm (3) is as follows: Figure 4 .

[0053] Depend on Figure 4 It can be seen that the number of data greater than c and the number of data less than -c are both greater than 6. Therefore, according to equations (3) and (4), this sequence is not a fixed sampling sequence.

[0054] B) Monotonicity and Concave / Concave Pattern Mining: Using different numbers of sampling points as units, five sets of data with a total of 3300 sampling points were tested. An appropriate allowable skewness constant c=5.8 was selected, and the aforementioned algorithm was applied to simulate and calculate the second-order difference matrix. The results obtained are as follows Figures 5-13 As shown.

[0055] (1) Take The first group of experiments used 450 sampling points from [1300 to 1750]. The actual sampling graph and the first-order difference sequence graph are shown below. Figure 5 and Figure 6 As shown.

[0056] The number of sampling points whose first-order difference is greater than c The number of points with a difference less than -c If so, then the sampling sequence is monotonically decreasing.

[0057] (2) Take The second group of experiments was conducted using 450 sampling points from [1751 to 2200]. The actual sampling graphs and first-order difference sequence graphs are shown below. Figure 7 and Figure 8 As shown.

[0058] The number of sampling points whose first-order difference is greater than c The number of points with a difference less than -c The sampling sequence in this segment is monotonically increasing.

[0059] (3) Take The third group of experiments was conducted using 900 sampling points from [1300 to 2200]. The actual sampling graphs and the graphs of the two-difference sequences are shown below. Figure 9 and Figure 10 As shown. The number of sampling points whose second-order difference is greater than c. The number of points with a difference less than -c The sampling sequence in this segment is convex upwards.

[0060] (4) Take The fourth group of experiments was conducted using 500 sampling points from [6500 to 7000]. The actual sampling graphs and the graphs of the two-difference sequence are shown below. Figure 11 andFigure 12 As shown.

[0061] The number of sampling points whose second-order difference is greater than c The number of points with a difference less than -c The sampling sequence in this segment is concave.

[0062] C) Periodic Pattern Mining: Plot the x-axis according to the algorithm mentioned in equation (8). The vertical axis is Scatter plot, such as Figure 13 As shown.

[0063] Depend on Figures 4-13 As can be seen, by establishing signal morphology mining through this invention, the morphological characteristics, variation model, and disturbance status of the original sampling sequence can be determined.

[0064] This invention focuses on abnormal changes and addresses the problem of monitoring the operational status and providing safety early warning for complex systems. It establishes methods and algorithms for morphological mining of signals over time that exhibit good fault tolerance for outliers and patchy outliers, model mining of process signal structures, and mining of statistical characteristics such as noise stationarity. A time-series data mining system framework consisting of process sampling time-series data, a model, and a mining machine is presented, and the correctness of the algorithms is verified. The mining methods and systems can be applied to various fields such as petrochemicals, measurement and control engineering, and navigation control, providing practical technical support for data-driven safety analysis, anomaly detection, and fault diagnosis technology research and engineering applications.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A time series sampled data fault-tolerant mining system for a hetero-process, characterized in that, The method comprises the following steps: a data processing module is used to collect time series sampling data sequences of an abnormal change period of an industrial production process, and to uniformly represent and model the time series sampling data sequences to obtain decomposed sampling data sequences; a mining mechanism modeling module is used to model a sliding window double median filter as the main part, and to combine a signal extraction accessory, a model fitting accessory and a residual error generation accessory to build a mining mechanism; a fault-tolerant mining module is used to use the mining mechanism to perform signal form fault-tolerant mining, model structure information fault-tolerant mining and noise statistical characteristic fault-tolerant mining on the decomposed sampling data sequences to obtain fault-tolerant mining results; a working condition recognition module is used to monitor the working condition of the industrial production process and perform fault diagnosis based on the fault-tolerant mining results.

2. The system of claim 1, wherein, In the mining mechanism, the signal extraction accessory is used to perform signal form fault-tolerant mining on the decomposed sampling data sequences; the signal form fault-tolerant mining includes constant value form mining, monotonic and convex-concave form mining and periodic form mining; the model fitting accessory is used to perform model structure information fault-tolerant mining on the decomposed sampling data sequences; the residual error generation accessory is used to perform noise statistical characteristic fault-tolerant mining on the decomposed sampling data sequences; the noise statistical characteristic fault-tolerant mining includes first-order stationary information mining and second-order stationary information mining.

3. The system of claim 2, wherein, The process of performing constant value form mining on the decomposed sampling data sequences comprises the following steps: a median algorithm is used to sort the decomposed sampling data sequences in a preset length sliding window by value and take the median value to build a double median filter; the double median filter is used to perform signal extraction on the decomposed sampling data sequences to build a deviation sequence; a first allowable deviation constant is selected, and the first allowable deviation constant is combined with the deviation sequence to build a first symbol sequence; the first symbol sequence is used to divide the decomposed sampling data sequences in a preset length sliding window into three types of data segments, and the longest run length index of each type of data segment is counted; the longest run length index is compared with a preset threshold value to determine whether the divided data segments belong to a constant value form, and the constant value form mining of the decomposed sampling data sequences is completed.

4. The system of claim 3, wherein, The process of performing monotonic and convex-concave form mining on the decomposed sampling data sequences comprises the following steps: the deviation sequence is used to perform first-order difference and second-order difference to obtain a first-order difference sequence and a second-order difference sequence; a second allowable deviation constant is selected, and the second allowable deviation constant is combined with the first-order difference sequence and the second-order difference sequence to build a second symbol sequence and a third symbol sequence; the second symbol sequence and the third symbol sequence are used to divide the decomposed sampling data sequences of a preset length into six types of data segments, and the number of data points of each type of data segment is counted; the number of data points is compared with the preset threshold value to determine whether the divided data segments belong to a monotonic or convex-concave form, and the monotonic and convex-concave form mining of the decomposed sampling data sequences is completed.

5. The system of claim 3, wherein, The process of performing periodic form mining on the decomposed sampling data sequences comprises the following steps: a sliding window periodogram of the deviation sequence is calculated, and a G-test statistic is built based on the sliding window periodogram; The Fisher distribution table and preset reliability of G-test are used to obtain a threshold constant, the threshold constant is compared with the G-test statistic, whether there is a periodic component in the decomposed sampling data sequence is determined, and the periodic pattern mining of the decomposed sampling data sequence is completed.

6. The system of claim 3, wherein, The process of model structure information fault-tolerant mining of the decomposed sampling data sequence comprises: A linear structure parameter model family is constructed by using an algebraic polynomial fitting model, a trigonometric polynomial fitting model, a periodic progressive model, an autoregressive AR model and an autoregressive ARI model with a periodic component; The AIC information criterion and a least square fitting algorithm are used to perform model order determination and residual sum of squares minimization on the linear structure parameter model family to obtain an optimal model family by using the signal sequence extracted by the double median filter. Based on the optimal model family, the model structure information fault-tolerant mining of the decomposed sampling data sequence is completed.

7. A time series data fault-tolerant mining method for a heterogeneous process, applying the system of any one of claims 1-6, characterized in that, It comprises: Time series sampling data sequences of abnormal change periods of an industrial production process are collected, and the time series sampling data sequences are uniformly represented and model decomposed to obtain decomposed sampling data sequences; A mining machine is constructed mainly by using a sliding window double median filter, combined with a signal extraction accessory, a model fitting accessory and a residual generation accessory; The decomposed sampling data sequences are subjected to signal pattern fault-tolerant mining, model structure information fault-tolerant mining and noise statistical characteristic fault-tolerant mining by using the mining machine to obtain fault-tolerant mining results; Based on the fault-tolerant mining results, the working conditions of the industrial production process are monitored and fault diagnosis is performed.

Images