A device fault diagnosis method based on time series data analysis

Abstract

The application discloses a kind of equipment fault diagnosis methods based on time series data analysis in industrial production in industrial scene, steps include: the data in time series database generated when industrial production system runs is preprocessed;Effective sample is extracted from the data after pre-processing, and multiple measuring point sample time is synchronized;Effective sample data is analyzed according to the characteristics of occurrence frequency, occurrence time, duration, and the correlation of the measuring points is obtained, so as to extract relevant measuring points;According to the data of relevant measuring points, the fault attribution factor graph is inferred.The application uses the data mining method related to big data to mine the correlation of equipment failure from massive sensor data, uses the related method in the field of artificial intelligence, carries out single-source, multi-source time series analysis on historical sensor data, locates one or more fault positions corresponding to abnormal data in combination with the correlation relationship mined.

Description

Technical Field

[0001] This invention relates to the characteristics, environment, and fault detection methods of industrial production in industrial settings, and particularly to a method for diagnosing equipment faults based on time-series data analysis. Background Technology

[0002] With the continuous development of computer technology and cross-industry integration, the monitoring technology of production equipment in industrial scenarios has gradually developed towards intelligence. Real-time monitoring information at the work site can be easily transmitted to the host computer for display through sensing devices. Abnormal changes in real-time measurement point information reflect the abnormal working status of local equipment. Accurately identifying abnormal points can help on-site personnel to discover equipment problems in a timely manner, thereby enabling alarm and fault analysis.

[0003] Equipment measurement point information is generally a series of time-series data. Existing time-series anomaly detection algorithms include those based on curve fitting, which assumes that for a time series, the data at time t is strongly dependent on the data at time t-1, so a curve can be used to fit the trend of data change. If new data breaks the trend, i.e., the curve is not smooth, it can be identified as an anomaly. Curve fitting methods mainly include linear regression, moving average, etc.; methods based on contemporaneous data utilize the periodicity of equipment operation in some projects, selecting the data of the most recent day of the detection item as a reference; methods based on contemporaneous amplitude utilize the similarity matching characteristics of two trend graphs, using the absolute value of the amplitude as a standard for anomaly judgment; detection algorithms based on month-on-month data target data within the most recent time window (T), using the fact that they follow a specific trend to detect anomalies. Detection algorithms based on neural networks include Convolutional Neural Network (CNN) algorithms, Recurrent Neural Network (RNN) algorithms, and Deep Neural Network (DNN) algorithms, which use neural networks to mine unknown potential information to identify anomalies. Tang Yongjun et al. proposed a method to extract image features of the unit shaft center trajectory using convolutional neural networks, thereby constructing a hybrid feature vector containing image features and waveform features, and using probabilistic neural networks for identification and classification to assist in fault diagnosis (Tang Yongjun, Liu Dong, Xiao Zhihuai, Hu Xiao, Lai Xu--Research on Fault Diagnosis Method of Hydropower Unit Based on Convolutional Neural Network and Singular Value Decomposition [J]. China Rural Water Resources and Hydropower, 2021(02):175-181.). Hou Jikun et al. used the state feature identification method to identify the massive state storage data of hydropower units according to certain rules, thereby assisting in fault diagnosis.

[0004] The aforementioned existing technologies have certain shortcomings, including: they do not utilize the logical correlations in the operation of production equipment in industrial scenarios; they do not utilize the correlation effects of multiple variables to identify anomalies; and they do not fully utilize historical information to learn and classify anomalous and normal samples from historical data. Therefore, they cannot identify multiple anomaly patterns and have the problem of missing anomalies that have different trends from those of the same period or amplitude. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to propose a device fault diagnosis method based on time series data analysis, which is used to accurately locate one or more fault locations corresponding to abnormal data.

[0006] To address the aforementioned technical problems, this invention proposes a device fault diagnosis method based on time-series data analysis, comprising the following steps:

[0007] S1: Preprocess the data in the time-series database generated during the operation of the industrial production system, including processing the timestamps of the data, removing outliers, handling missing values, testing for stationarity, and visual analysis;

[0008] S2: For the preprocessed data, sample and extract valid samples, and use time windows to align them to ensure time synchronization of samples from multiple measurement points;

[0009] S3: Based on the characteristics of occurrence frequency, occurrence time, and duration, the extracted valid sample data are analyzed by association rules to obtain the lift matrix of correlation between measurement points, thereby extracting relevant measurement points;

[0010] S4: Based on the data from the relevant measurement points, and combined with the correlation inference of the measurement points, the fault attribution factor map is obtained.

[0011] Furthermore, the data preprocessing in step S1 includes:

[0012] S11: The timestamp processing of the data includes: sorting unordered timestamps, and converting the data time column into a date data type for timestamps stored in the default string data type.

[0013] S12: Missing value handling: The missing value is estimated by using spline interpolation or linear interpolation with two surrounding known data points. Data sets with a missing window size greater than 5 are directly discarded in their entirety to ensure the validity and continuity of the remaining data.

[0014] S13: Calculate the average value of the time series with a fixed window size, and make a preliminary observation of the noise situation in each window; transform the time series to the frequency domain through Fourier transform, and detect outliers with sudden peaks or drops in the trend line, thereby filtering out noise;

[0015] S14: Stationarity is tested using a visual method by plotting a line graph of the time series. On the one hand, observe whether the curve fluctuates around a certain value. If it does, the average value is stable; otherwise, the average value is unstable, and the entire data segment is discarded. On the other hand, observe the fluctuation range of the curve. If the difference between the preceding and following data in each group is within the range of 0-1, the variance is stable; otherwise, second-order differencing is performed to make the data stationary.

[0016] Furthermore, the data preprocessing in step S2 includes the following two sub-steps:

[0017] S21: For the preprocessed data, a minute-level downsampling method is used to extract complete and valid time-series measurement point data from all measurement points in the database;

[0018] S22: According to the system to which they belong, extract the measurement points with similar topological contexts, extract the time period data of key measurement points, and according to the offset of the time window, use left alignment based on the offset to ensure that the extracted data of each auxiliary system in the same time period correspond.

[0019] Furthermore, based on the data obtained from cleaning and sampling in steps S1 and S2, association rule analysis is performed to find device measurement points with correlations higher than a certain threshold, which are used for fault location. This process includes the following sub-steps:

[0020] S31: For each system, extract multiple alarm-triggered records from all data. For the alarm record time point t, extend it by one hour before and after, that is, take the equipment operation records of [t-60, t+60] minutes.

[0021] S32: For topological structures with sequential relationships, align them according to a certain time difference;

[0022] S33: Extract data from three to four months of the dry season after alignment, count the number of abnormal measurement points and alarm messages, and treat a measurement point of a device as an itemset;

[0023] S34: List itemsets that contain only one element;

[0024] S35: And so on, list itemsets containing 2 to n elements;

[0025] S36: Calculate the frequency of each item in the itemset set as the support. The support of itemset {A,B} is P(A&B). Keep the itemsets that meet the minimum support threshold condition and eliminate the itemsets that do not meet the condition to obtain the candidate itemsets.

[0026] S37: Add an element to the candidate itemset and generate all possible combinations of itemsets;

[0027] S38: Repeat steps S36 and S37 to determine the support for increasingly larger itemsets until there are no new itemsets to be checked;

[0028] S39: Calculate the lift of an itemset. The formula is: Lift {A,B} = Support {A,B} / Support {A} * Support {B}. Lift reflects the correlation between two transactions in an association rule. A lift greater than 1 indicates a higher positive correlation between the transactions. A lift less than 1, and the smaller the value, the higher the negative correlation. If the lift equals 1, it indicates no correlation. Since there are almost no mutually exclusive relationships between equipment in industrial production auxiliary systems, only transactions with a lift greater than 1 are considered valid association rules, and the itemset is retained.

[0029] S310: For the retained itemsets, calculate the confidence of the itemsets using the formula: Confidence {A,B} = Support {A,B} / Support {A}. The association matrix is ​​then calculated based on multiple itemsets.

[0030] Furthermore, step S4 specifically includes the following steps:

[0031] S41: Based on the correlation matrix, extract the relevant measurement point itemsets with a lift greater than 1 for observation, and convert them into a causal topological subgraph to form a correlation diagram;

[0032] S42: Verify the extracted relationships and compare them with the relationship graph obtained by manual extraction and analysis;

[0033] Furthermore, the manually extracted correlation diagram uses a single measurement point as the source point and a range of one hour before and after the alarm time to extract time-related measurement points; a topological correlation diagram is then created based on the time sequence.

[0034] Furthermore, the automatically extracted association graph is used to update the final fault attribution graph;

[0035] This invention utilizes big data-related data mining methods to extract correlations of equipment failures from massive amounts of sensor data. It also employs artificial intelligence methods to perform single-source and multi-source time-series analysis on historical sensor data. By combining the extracted correlations, it locates one or more fault locations corresponding to abnormal data.

[0036] The present invention has the following beneficial effects:

[0037] (1) In view of the omission of data cleaning steps in existing methods and the negative impact of noisy data on the association mining results in actual processing, data preprocessing such as outlier removal, missing value handling, stationarity test and visualization analysis is adopted to preprocess the massive industrial production auxiliary system measurement point data, retain effective data and reduce analysis noise.

[0038] (2) To address the problem that existing methods do not fully utilize correlation relationships to predict multi-source faults, a relation mining algorithm is used to mine the correlation relationships of measurement points reflected in a large amount of historical data. Valid itemsets are extracted through an automated model and confidence scores are calculated to obtain a fault attribution factor map that conforms to the operating rules of equipment in general industrial production. Attached Figure Description

[0039] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Figure 1 This is the correlation matrix diagram of the present invention.

[0041] Figure 2 This is a schematic diagram illustrating the correlation of the present invention. Detailed Implementation

[0042] This invention proposes a device fault diagnosis method based on time-series data analysis, comprising the following steps:

[0043] S1: Preprocess the data in the time-series database generated during the operation of the industrial production system, including data timestamp processing, outlier removal, missing value completion, stationarity test, and visualization analysis.

[0044] S2: After processing all the collected data, a minute-level downsampling method is adopted to extract valid samples, and time window alignment is used to ensure time synchronization of samples from multiple measurement points.

[0045] S3: Based on the characteristics of occurrence frequency, occurrence time, and duration, the effective data extracted in S2 are analyzed by association rules to obtain the lift matrix of correlation between measurement points, thereby extracting relevant measurement points.

[0046] S4: Based on the relevant measurement point data obtained in S3, and combined with the correlation reasoning of measurement points, obtain the fault attribution factor map.

[0047] The data preprocessing in step S1 includes:

[0048] S11: For the collected time-series database data, first process the unstructured timestamps, including: sorting unordered timestamps, and converting the data time column to a date data type for timestamps stored in the default string data type.

[0049] S12: Handle missing values ​​by using spline interpolation or linear interpolation with two surrounding known data points to estimate the value of the missing point. Data sets with large missing windows are discarded entirely to ensure the validity and continuity of the remaining data.

[0050] S13: Calculate the average value of the time series with a fixed window size, and make a preliminary observation of the noise situation in each window; transform the time series to the frequency domain through Fourier transform, and detect outliers with sudden peaks or drops in the trend line, thereby filtering out noise;

[0051] S14: Stationarity is tested using a visual method by plotting a line graph of the time series. On the one hand, observe whether the curve fluctuates around a certain value. If it does, the average value is stable; otherwise, the average value is unstable, and the entire data segment is discarded. On the other hand, observe the fluctuation range of the curve. If the difference between the preceding and following data in each group is within the range of 0-1, the variance is stable; otherwise, second-order differencing is performed to make the data stationary.

[0052] Step S2 specifically includes:

[0053] S21: For the preprocessed data, a minute-level downsampling method is used to extract complete and valid time-series measurement point data from all measurement points in the database;

[0054] S22: According to the system to which they belong, extract the measurement points with similar topological contexts, extract the time period data of key measurement points, and according to the offset of the time window, use left alignment based on the offset to ensure that the data extracted from the four auxiliary systems in the same time period correspond.

[0055] Based on the data obtained from cleaning and sampling in steps S1 and S2, association rule analysis is performed to find device measurement points with correlations higher than a certain threshold, which are used for fault location. This process includes the following sub-steps:

[0056] S31: For each system, extract multiple alarm-triggered records from all data. For the alarm record time point t, extend it by one hour before and after, that is, take the equipment operation records of [t-60, t+60] minutes.

[0057] S32: For topological structures with sequential relationships, align them according to a certain time difference;

[0058] S33: Extract data from three to four months of the dry season after alignment, count the number of abnormal measurement points and alarm messages, and treat a measurement point of a device as an itemset;

[0059] S34: List itemsets containing only one element, such as {201931231242}, where “201931231242” is the measurement point ID number;

[0060] Table 1 Alarm Item Set

[0061]

[0062] S35: Similarly, list itemsets containing 2 to n elements. The itemsets extracted based on the alarms are shown in Table 1.

[0063] S36: Calculate the frequency of each item in the itemset set as the support. The support of itemset {A,B} is P(A&B). Keep those itemsets that meet the minimum support threshold condition and discard those that do not.

[0064] S37: Add an element to the candidate itemset (the result after eliminating itemsets that do not meet the requirements in step S36), and use the itemsets retained in step S36 to generate all possible combinations;

[0065] S38: Repeat steps S36 and S37 to determine the support for increasingly larger itemsets until there are no new itemsets to be checked;

[0066] S39: Calculate the lift of an itemset. The formula is: Lift {A,B} = Support {A,B} / Support {A} * Support {B}. Lift reflects the correlation between two transactions in an association rule. A lift greater than 1 indicates a higher positive correlation between the transactions, while a lift less than 1, especially if it is smaller, indicates a higher negative correlation. If the lift equals 1, it indicates no correlation. Since there are almost no mutually exclusive relationships between equipment in industrial production auxiliary systems, only transactions with a lift greater than 1 are considered valid association rules, and the itemset is retained.

[0067] S310: For the retained itemsets, calculate the confidence score of the itemsets using the formula: Confidence score {A,B} = Support score {A,B} / Support score {A}. The association matrix results calculated from multiple itemsets are as follows: Figure 1 As shown;

[0068] Step S4 specifically includes the following steps:

[0069] S41: Based on the correlation matrix obtained in S310, extract the relevant measure point itemsets with a lift greater than 1 for observation, and convert them into a causal topological subgraph. The resulting correlation diagram is shown below. Figure 2 As shown;

[0070] S42: Verify the extracted relationships and compare them with the relationship graph obtained by manual extraction and analysis;

[0071] The manually extracted correlation diagram uses a single measurement point as the source point and a range of one hour before and after the alarm time to extract time-related measurement points; a topological correlation diagram is then created based on the time sequence.

[0072] The comparison reveals that the association maps automatically extracted using the association rule algorithm cover a wider range of measurement points and have richer association relationships, which can reflect the different mutual influences between measurement points in different time periods and scenarios.

[0073] The automatically extracted association graph is used to update the final fault attribution graph.

[0074] When the fault attribution factor map is obtained, the upstream and downstream related fault measurement points can be inferred from the fault attribution factor map when the next fault alarm information occurs, so as to carry out maintenance.

[0075] In summary, the key features of this invention are: (1) To address the problem of massive data redundancy, data cleaning technology is first used to preprocess the large amount of data generated by the operation of units in industrial production; (2) To address the problems of high noise, unbalanced samples, and redundancy in massive data, a minute-level downsampling method is adopted, and effective samples are used as input for relation mining; (3) To address the time difference in the logical sequence of equipment, a minute-level time window alignment technology is adopted to achieve relative alignment; (4) For the analysis of single-source and multi-source faults, data mining technology is used to extract reasonable multi-source time-series samples and mine the correlation relationships in the operation of industrial production equipment.

[0076] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing equipment faults based on time-series data analysis, characterized in that, Includes the following steps: S1: Preprocess the data in the time-series database generated during the operation of the industrial production system, including timestamp processing, missing value handling, visualization analysis, and stationarity testing. S2: For the preprocessed data, sample and extract valid samples, and use time windows to align them to ensure time synchronization of samples from multiple measurement points; S3: Based on the characteristics of occurrence frequency, occurrence time, and duration, association rule analysis is performed on the extracted valid sample data to obtain the lift matrix of correlation between measurement points, thereby extracting relevant measurement points; step S3 specifically includes: S31: For each system, extract multiple alarm-triggered records from all data. For the alarm record time point t, extend it by one hour before and after, that is, take the equipment operation records of [t-60, t+60] minutes. S32: For topological structures with sequential relationships, align them according to a certain time difference; S33: For the aligned data, extract data from three to four months of the dry season, count the number of abnormal measurement points and alarm messages, and treat a device measurement point as an itemset; S34: List itemsets that contain only one element; S35: And so on, list itemsets containing 2 to n elements; S36: Calculate the frequency of each item in the itemset set as the support. The support of itemset {A,B} is P(A&B). Keep the itemsets that meet the minimum support threshold condition and eliminate the itemsets that do not meet the condition to obtain the candidate itemsets. S37: Add an element to the candidate itemset and generate all possible combinations of itemsets; S38: Repeat steps S36 and S37 to determine the support for increasingly larger itemsets until there are no new itemsets to be checked; S39: Calculate the lift of an itemset. The formula is: Lift {A,B} = Support {A,B} / Support {A} * Support {B}. Lift reflects the correlation between two transactions in an association rule. A lift greater than 1 indicates a higher positive correlation between the transactions. A lift less than 1, and the smaller the value, the higher the negative correlation. If the lift equals 1, it indicates no correlation. Since there are almost no mutually exclusive relationships between equipment in industrial production auxiliary systems, only transactions with a lift greater than 1 are considered valid association rules, and the itemset is retained. S310: For the retained itemsets, calculate the confidence of the itemsets using the formula: Confidence {A,B} = Support {A,B} / Support {A}. The association matrix is ​​then calculated based on multiple itemsets. S4: Based on the data from relevant measurement points and combined with the correlation inference of measurement points, a fault attribution factor map is obtained; step S4 specifically includes: S41: Based on the aforementioned correlation matrix, extract the relevant measurement point itemsets with a lift greater than 1 for observation, and convert them into a causal topological subgraph to form a correlation diagram; S42: Verify the extracted relationships and compare them with the relationship graph obtained by manual extraction and analysis; The manually extracted correlation graph is generated by taking a measurement point as the source point and extracting time-related measurement points within a one-hour range before and after the alarm time. A topological correlation graph is then created based on the time sequence. The automatically extracted correlation graph is used to update the graph, which serves as the final fault attribution factor graph.

2. The equipment fault diagnosis method based on time-series data analysis according to claim 1, characterized in that, The data preprocessing in S1 includes: S11: The timestamp processing of the data includes: sorting unordered timestamps, and converting the data time column into a date data type for timestamps stored in the default string data type. S12: Handling missing values: Use spline interpolation or linear interpolation with two surrounding known data points to estimate the value of the missing point. Data sets with a missing window size greater than 5 are discarded entirely to ensure the validity and continuity of the remaining data. S13: Calculate the average value of the time series with a fixed window size, and make a preliminary observation of the noise situation in each window; transform the time series to the frequency domain through Fourier transform, and detect outliers with sudden peaks or drops in the trend line, thereby filtering out noise; S14: Stationarity is tested using a visual method by plotting a line graph of the time series. On the one hand, observe whether the curve fluctuates around a certain value. If it does, the average value is stable; otherwise, the average value is unstable, and the entire data segment is discarded. On the other hand, observe the fluctuation range of the curve. If the difference between the preceding and following data in each group is within the range of 0-1, the variance is stable; otherwise, second-order differencing is performed to make the data stationary.

3. The equipment fault diagnosis method based on time-series data analysis according to claim 2, characterized in that, Step S2 specifically includes: S21: For the preprocessed data, a minute-level downsampling method is used to extract complete and valid time-series measurement point data from all measurement points in the database; S22: According to the system to which they belong, extract the measurement points with similar topological contexts, extract the time period data of key measurement points, and according to the offset of the time window, use left alignment based on the offset to ensure that the extracted data of each auxiliary system in the same time period correspond.

Images