A method for mining time-dependent patterns of multiple anomaly modes of periodic operation components

Abstract

The application relates to a periodic operation component multivariate perception abnormal mode time sequence dependence mining method, which comprises the following steps: S1, a component multivariate perception variable abnormal event capture platform is constructed; S2, abnormal mode time sequence data of the perception variable is captured, and an abnormal time sequence database Event_seqDB is constructed; S3, a multivariate perception abnormal mode time sequence database KDD_SeqDB is constructed; and S4, multivariate perception abnormal mode time sequence dependence mining is carried out based on an improved PrefixSpan. Through the construction of a key component multitransmission variable perception mode time sequence dependence model and algorithm, the time sequence cause-effect relationship formed in the component operation and maintenance process can be effectively mined, thereby providing support for preventive maintenance and design iteration optimization.

Description

TECHNICAL FIELD

[0001] The application relates to a time sequence dependent relationship mining method, in particular to a periodic operation component multi-element perception abnormal mode time sequence dependent mining method. BACKGROUND

[0002] There are a large number of research results on component time sequence fault mining methods driven by perception big data, and most of these results are concentrated in fault diagnosis, life prediction and fault correlation mining based on time sequence data, and in the research process, the causal relationship between multi-element perception variables and faults is mostly only for the causal relationship between multi-element perception variables and faults, and the time sequence relationship of abnormal propagation of perception variables caused by the cooperation of multiple parts in the operation process of the component is ignored, resulting in that the analysis result is only local but not global for preventive maintenance and design iteration optimization.

[0003] The mining of the time sequence dependent relationship between the perception variable modes of each assembled part in the operation of the critical component is helpful for preventive maintenance and new round of design optimization of the culprit part and its weak link from the global perspective of the component, and has important reference significance for the closed-loop iteration of complex equipment products. SUMMARY

[0004] In order to achieve the above application purpose, the application provides the following technical scheme:

[0005] A periodic operation component multi-element perception abnormal mode time sequence dependent mining method comprises the following steps S1-S4:

[0006] Step S1: component multi-element perception variable abnormal event capture platform construction;

[0007] Step S2: capture the time sequence data of the perception variable abnormal mode, and construct an abnormal time sequence database Event_seqDB;

[0008] Step S3: construct a multi-element perception abnormal mode time sequence database KDD_SeqDB;

[0009] Step S4: multi-element perception abnormal mode time sequence dependent mining based on improved PrefixSpan; the PrefixSpan algorithm is used, the improved PrefixSpan algorithm is used when solving the frequent 2-item sequence, the projection database of the calculated frequent 1-item sequence is deleted, the abnormal mode of the same perception variable in the previous sequence is deleted from the projection database, and then the sequence count is calculated based on the deleted projection database to obtain a local 2-item frequent set, and then the PrefixSpan algorithm is used for recursive projection to obtain a multi-element frequent time sequence dependent item set;

[0010] Wherein:

[0011] In step S3, the time window of the component operation cycle is taken as the associated valid time window, and the valid time window is used to cut, sort and combine each sensing variable abnormal mode time series data to form a component cycle operation abnormal time series record. Step S3 specifically includes S31-S34:

[0012] S31: initialize the operation cycle k = 1;

[0013] S32: read the component work log to obtain the start time STIMEk and the end time ETIMEk of the cycle k;

[0014] S33: obtain the abnormal mode time series data of each sensing variable in the time period, and the relationship operation is as follows:

[0015] DATATIMEk = σ 时间≥STIMEk∧时间≤ETIMEk (Event_seqDB) (8)

[0016] In formula (8), DATATIMEk is the abnormal time series data pair set obtained by cutting through the operation cycle k time window, σ represents the conditional selection, ∧ represents the left and right side conditions, and the formula is the selection operation of the data set, that is, the abnormal time series data in the operation cycle k is selected as the condition of the start and end time;

[0017] S34: DATATIMEk data cleaning, removing null, single value and same sensing variable mode multiple continuous value, specifically including S341-S343:

[0018] S341: determine whether the abnormal data number DATATIMEk.count contained in DATATIMEk is greater than or equal to 2, if not: go to S35; if yes: continue to execute S342;

[0019] S342: integrate the abnormal mode time series data into operation project time series abnormal data, S342 specifically includes S3421-S3429:

[0020] S3421: initialize the record variable i = 1, and initialize the project time series data set variable KDD_Datak = null;

[0021] S3422: sort the DATATIMEk data set in ascending order of time;

[0022] S3423: add the sensing abnormal mode of the first data to the KDD_Datak data set;

[0023] S3424: i = i + 1, point to the next data of DATATIMEk;

[0024] S3425: take the sensing abnormal mode of the i-th data;

[0025] S3426: judging whether the i-th and the i-1-th sensed abnormal pattern is same, if yes, going to S3427; if no, adding the i-th abnormal pattern into the KDD_Datak dataset with comma separating the items;

[0026] S3427: judging whether i>DATATIMEk.count is true, if no, going to S3424; if yes, going to S3428;

[0027] S3428: querying the abnormal pattern set in the KDD_Datak, and querying the sensed variable coding set according to the abnormal pattern association and storing into the variable sen_vars;

[0028] S3429: judging whether the number of sensed variables in sen_vars is greater than 1, if yes, writing the KDD_Datak into the multi-sensed abnormal timing database KDD_SeqDB; if no, discarding the KDD_Datak without saving;

[0029] S343: judging whether the records in the Event_seqDB are processed, if not, k=k+1, going to S32, if yes, ending.

[0030] Preferably, in step S4, the following steps S41-S44 are included:

[0031] S41: setting the minimum support minsup;

[0032] S42: obtaining the frequent 1-item sequence SK;

[0033] S43: for each item in the frequent 1-item sequence SK, calculating the projected database and deleting the same sensed variable abnormal pattern;

[0034] S44: outputting the frequent sequences of frequent 2-item set and above.

[0035] Preferably, S42 specifically includes S421-S423:

[0036] S421: scanning the whole database to obtain all different sensed abnormal patterns

[0037] S422: counting the occurrence of each alarm pattern to obtain all frequent 1-item sequences SK;

[0038] S423: deleting the alarm pattern with the support count less than the threshold from the KDD_SeqDB.

[0039] Preferably, S43 specifically includes S431-S436:

[0040] S431: initialize the count j = 1;

[0041] S432: project the database PJDj with the frequent pattern Sk[j] as prefix;

[0042] S433: remove the alarm pattern of the same sensing variable from PJDj;

[0043] S434: determine whether PJDj is empty, if yes, count the support count of each item in PJDj to obtain the frequent item PI, if no, go to S436;

[0044] S435: determine whether PI is empty, if no, go to S436, if yes, execute:

[0045] S4351: merge the single item satisfying the support count with the predecessor Sk[j] to obtain a new predecessor;

[0046] S4352: project the database PJDj with Sk[j] as prefix;

[0047] S4353: remove the non-frequent item from the projected database;

[0048] S4354: go to S434;

[0049] S436: j = j + 1, until j > SK.longth, where SK.longth is the number of frequent items.

[0050] Preferably, step S1 comprises steps S11-S14:

[0051] S11: encode the parts, sensors and sensing variables: each part of the component BOM structure tree node is uniquely identified by part_ID, each sensor is uniquely identified by source_ID, and the sensing variable is uniquely identified by sen_var_ID. Since one part can deploy multiple sensors and one sensor can collect multiple variables, the following conditions should be met:

[0052]

[0053] In formula (1), represents total functional dependence, that is, a corresponding sensor can be uniquely found through variable encoding, and a corresponding component can be uniquely found through the sensor;

[0054] S12: set data cleaning and abnormal diagnosis algorithm for the sensing variable class, specifically including:

[0055] S121: classify the perception variables corresponding to the components, and the perception variable class is uniquely identified by sen_Type_ID, which satisfies the constraint between the perception variable and the perception variable class:

[0056]

[0057] In formula (2), represents a total functional dependence, that is, the perception variable can uniquely determine the classification to which it belongs;

[0058] S122: data cleaning and abnormal diagnosis algorithm packaging: select algorithms or deep learning models for each perception variable class from the general machine learning library Scikit-learn and the deep learning library Pytorch;

[0059] The packaged algorithm is uniquely identified by algori_ID, and one algorithm corresponds to one normal output and multiple abnormal outputs, and the abnormal type is uniquely identified by Abn_ID, and the constraint between the abnormal type and the algorithm satisfies:

[0060]

[0061] In formula (3), represents a total functional dependence, that is, the abnormal type can uniquely determine the corresponding algorithm;

[0062] S13: perception data access and abnormal event capture algorithm configuration, S13 includes S131-S134:

[0063] S131: build a perception big data access and processing platform basic environment using the open source message middleware Kafka and the open source stream data processing platform Flink;

[0064] S132: publish the real-time data corresponding to each perception variable as a message topic through Kafka;

[0065] S133: configure the tasks of each perception variable in the stream data processing platform Flink, including the time window size, the called algorithm and the consumed message topic data, and through the configuration, the following functional dependence is established:

[0066]

[0067] In formula (4), TWSize represents the time window size; algori_ID represents the called data processing algorithm, and TopicID represents the message topic, that is, the perception data processing method can be uniquely determined by the perception variable;

[0068] S134: perception abnormal mode coding: since each algorithm can output one normal and multiple abnormal types, for each perception variable sen_var_ID(i), use uniquely encoding the abnormal pattern; for any There are:

[0069]

[0070] In formula (5): It means that a set of perception variables and abnormal types can be uniquely determined by the perception abnormal pattern encoding, and vice versa, the perception abnormal pattern encoding can be uniquely determined by a set of perception variables and abnormal types;

[0071] S14: Write the corresponding component perception configuration data entities and contacts of (1), (2), (3), (4), (5) into the configuration database configDB.

[0072] Preferably, step S2 includes the following steps S21-S23:

[0073] S21: The stream data processing platform Flink calls the configured algorithm to consume the corresponding message topic data according to the time window, captures the abnormal events of each perception variable, and constructs the abnormal pattern sequence data. For any perception variable sen_var_ID(i), the corresponding abnormal perception time sequence can be expressed as:

[0074] Event_seq={<sen_var_ID(i),Timex,Abn_ID(j)>} (6)

[0075] In formula (6), Event_seq represents the abnormal event sequence data corresponding to sen_var_ID(i), {<sen_var_ID(i),Timex,Abn_ID(j)>} represents the perception variable time event triple record, Timex represents the abnormal occurrence time, and Abn_ID(j) is the corresponding abnormal type;

[0076] S22: Calculate by (5) using Abn_ID(j) and sen_var_ID(i) Transform the data into:

[0077]

[0078] S23: Write Event_seq into Event_seqDB.

[0079] Compared with the prior art, the beneficial effects of the present application are:

[0080] (1) The inventor found in practice that most of the key components of mechanical equipment are connected by interference fit or connecting pieces between multiple parts. In high-speed operation service conditions, the perceived variable anomalies between parts may have mutual influence and propagation. The application proposes a multi-element perceived variable anomaly event coding and capturing mechanism combined with a big data processing platform, establishes a correlation model of parts, sensors, perceived variables, algorithms, and perceived abnormal patterns, and obtains abnormal events of each perceived variable of each sensing device on each part in real time.

[0081] (2) The inventor found in practice that the data corresponding to the perceived variables of the parts have a streaming feature, but the abnormal patterns corresponding to each perceived variable are limited. The application extracts perceived variable anomaly events after processing the time series data of each perceived variable through a big data processing platform calling data cleaning and diagnostic algorithms, converts abnormal events into abnormal patterns, and constructs time series data of each perceived variable abnormal pattern.

[0082] (3) The inventor found in practice that the correlation between periodic operation component perceived abnormal patterns is only valid within one operation cycle, and will be detected and maintained after one cycle. The application proposes to use the component operation cycle time window as the correlation effective time window, and uses the effective time window to cut, sort, and integrate the time series data of each perceived variable abnormal pattern to form a component cycle operation abnormal time series record.

[0083] (4) The inventor found in practice that the component operation abnormal record intercepted according to the component operation time window contains a large number of null values, single values, and multiple consecutive values of the same perceived variable pattern, which is not suitable for association relationship mining. The application detects the component operation abnormal record formed by cutting and merging with the time window, retains only one consecutive perceived pattern value in the sequence, and deletes null value and single value records.

[0084] (5) The inventor found in practice that when directly applying the time series data generated by the above transformation for mining, there may also be frequent item interference between different patterns of the same variable. The application proposes a multi-element pattern sequence data screening algorithm to remove binary and below abnormal pattern sequence data within the time window, thereby constructing a periodic operation component multi-element perceived abnormal pattern time series database.

[0085] (6) the inventor found in practice that when directly using a frequent item discovery sequential pattern mining algorithm, there are frequent sequences between different patterns of the same variable. The application applies the PrefixSpan algorithm to first obtain sequences of length 1, then constructs a projection database according to the 1-item sequences, and improves the calculation of frequent 2-item sequences, deletes abnormal patterns of the same sensing variable in the projection database, and then calculates the sequence count based on the deleted projection database to obtain a local 2-item frequent set, and then recursively projects according to the PrefixSpan basic algorithm to obtain a multivariate frequent time sequence dependent item set. BRIEF DESCRIPTION OF DRAWINGS

[0086] Figure 1 The overall technical route of the multivariate sensing abnormal pattern time sequence dependent mining for the periodic operation component is provided.

[0087] Figure 2 The configuration schematic diagram of the component multivariate sensing variable abnormal event capture is provided.

[0088] Figure 3 The abnormal time sequence data conversion algorithm schematic diagram of the component operation cycle time window cutting is provided.

[0089] Figure 4 The multivariate sensing abnormal pattern time sequence dependent mining algorithm schematic diagram based on the improved PrefixSpan is provided. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments.

[0091] Therefore, the following detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but only represents some embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0092] It should be noted that the embodiments in the present application and the features and technical solutions in the embodiments can be combined with each other without conflict.

[0093] A multivariate sensing abnormal pattern time sequence dependent mining method for a periodic operation component includes the following steps:

[0094] Step S1: component multivariate sensing variable abnormal event capture platform construction.

[0095] S11: Encode the parts, sensors and sensing variables: Each part of the component BOM structure tree node is uniquely identified by part_ID, each sensor is uniquely identified by source_ID, and the sensing variable is uniquely identified by sen_var_ID. Since one part may deploy multiple sensors and one sensor can collect multiple variables, the following conditions should be met:

[0096]

[0097] In the formula, It represents a complete functional dependence, that is, a corresponding sensor can be uniquely found through variable encoding, and a corresponding part can be uniquely found through the sensor.

[0098] S12: Set data cleaning and abnormal diagnosis algorithm for sensing variable class, including:

[0099] S121: Classify the sensing variables corresponding to the parts, and uniquely identify the sensing variable class by sen_Type_ID. The constraint between the sensing variable class and the sensing variable satisfies:

[0100]

[0101] In the formula, It represents a complete functional dependence, that is, the classification to which the sensing variable belongs can be uniquely determined through the sensing variable.

[0102] S122: Data cleaning and abnormal diagnosis algorithm packaging: Select algorithms or deep learning models for each sensing variable class from the general machine learning library Scikit-learn and the deep learning library Pytorch.

[0103] Preferably, if a supervised machine learning model is selected, it needs to be trained with a labeled scenario data set. The selected algorithm and the trained model are packaged using microservices for the sensing big data processing platform to call.

[0104] The packaged algorithm is uniquely identified by algori_ID. One algorithm corresponds to one normal output and multiple abnormal outputs. The abnormal type is uniquely identified by Abn_ID. The constraint between the abnormal type and the algorithm satisfies:

[0105]

[0106] In the formula, It represents a complete functional dependence, that is, the corresponding algorithm can be uniquely determined through the abnormal type.

[0107] S13: Sensing data access and abnormal event capture algorithm configuration, including:

[0108] S131: Use open source message middleware Kafka and open source stream data processing platform Flink to build a perception big data access and processing platform basic environment.

[0109] S132: Publish the real-time data corresponding to each perception variable as a message topic through Kafka.

[0110] S133: Configure the tasks of each perception variable in the stream data processing platform Flink, including the time window size, the called algorithm and the consumed message topic data. Through the configuration, the following function dependencies are established:

[0111]

[0112] In the formula, TWSize represents the time window size; algori_ID represents the called data processing algorithm, and TopicID represents the message topic, that is, the processing method of the perception data can be uniquely determined through the perception variable.

[0113] S134: Perception anomaly pattern coding: since each algorithm can output one normal and multiple abnormal types, for each perception variable sen_var_ID(i), use to uniquely encode the abnormal pattern. For example: for the monitoring i of the bogie bearing temperature variable, the abnormal pattern is divided into: to represent the temperature difference exceeding the threshold value alarm, to represent the temperature exceeding the threshold value alarm. For any there exists:

[0114]

[0115] In the formula: represents equivalence, that is, the perception abnormal pattern code can be uniquely determined through a group of perception variables and abnormal types, and vice versa, a group of perception variables and abnormal types can uniquely determine the perception abnormal pattern code.

[0116] S14: Write the component perception configuration data entity and the connection corresponding to (1), (2), (3), (4) and (5) into the configuration database configDB. Since most of the key components of mechanical equipment are connected through interference fit or connecting pieces between multiple parts, under the high-speed running service working condition, the abnormality of the perception variables between the parts may have mutual influence and propagation. The present application proposes a multi-element perception variable abnormal event coding and capturing mechanism combined with a big data processing platform, establishes a correlation model of parts, sensors, perception variables, algorithms and perception abnormal patterns, and obtains the abnormal events of each perception variable of each sensing device on each part in real time.

[0117] Step S2: Capture the abnormal pattern time series data of the sensing variable, and construct the abnormal time series database Event_seqDB.

[0118] S21: The flow data processing platform Flume calls the configured algorithm according to the time window to consume the corresponding message topic data, captures the abnormal events of each sensing variable, and constructs the abnormal pattern sequence data. For any sensing variable sen_var_ID(i), the corresponding abnormal sensing time sequence can be expressed as:

[0119] Event_seq={<sen_var_ID(i), Timex, Abn_ID(j)>} (6)

[0120] In the formula, Event_seq represents the abnormal event sequence data corresponding to sen_var_ID(i), {<sen_var_ID(i), Timex, Abn_ID(j)>} represents a sensing variable time event triple record, Timex represents the abnormal occurrence time, and Abn_ID(j) is the corresponding abnormal type.

[0121] S22: Calculate Abn_ID(j) and sen_var_ID(i) through (5) Transform the data into:

[0122]

[0123] Since the data corresponding to the sensing variable of the component has the characteristics of streaming, but the abnormal pattern corresponding to each sensing variable is limited. After the time series data of each sensing variable is processed by the big data processing platform to call the data cleaning and diagnosis algorithm, the abnormal events of the sensing variable are extracted, the abnormal events are converted into abnormal patterns, and the abnormal pattern time series data of each sensing variable is constructed.

[0124] S23: Write Event_seq into Event_seqDB.

[0125] Step S3: Construct the multi-element sensing abnormal pattern time series database KDD_SeqDB.

[0126] The time window of the component running period is taken as the effective time window for association. The abnormal pattern time series data of each sensing variable is cut, sorted and combined by using the effective time window, to form the abnormal time series record of the component period running, which specifically includes:

[0127] S31: Initialize the running period k=1.

[0128] S32: Read the component work log to obtain the start time STIMEk and the end time ETIMEk of the period k.

[0129] S33: Obtain abnormal pattern time series data of each sensing variable in the time period, and the relationship operation is as follows:

[0130] DATATIMEk=σ 时间≥STIMEk∧时间≤ETIMEk (Event_seqDB) (8)

[0131] In the formula, DATATIMEk is a set of abnormal time series data obtained by intercepting a time window of a running period k, σ represents a conditional selection, ∧ represents a left and right condition combination, and the formula is a selection operation of a data set, that is, abnormal time series data in the running period k is selected as a condition.

[0132] In view of the fact that the correlation between the sensing abnormal patterns of the period running component is only valid within one running period, detection and maintenance will be performed after one period running. The component running period time window is proposed as a correlation valid time window, and the abnormal pattern time series data of each sensing variable is cut, sorted and integrated by using the valid time window, to form a component period running abnormal time series record.

[0133] S34: DATATIMEk data cleaning, removing null values, single values and multiple continuous values of the same sensing variable pattern, specifically including:

[0134] S341: Determine whether the number of abnormal data DATATIMEk.count contained in DATATIMEk is greater than or equal to 2, if not: go to S35; if yes: continue to execute S342.

[0135] S342: Integrate the abnormal pattern time series data into running project time series abnormal data, specifically including:

[0136] S3421: Initialize the record variable i = 1, and initialize the project time series data set variable KDD_Datak = null.

[0137] S3422: Sort the DATATIMEk data set in ascending order of time.

[0138] S3423: Take the sensing abnormal pattern of the first data and add it to the KDD_Datak data set.

[0139] S3424: i = i + 1, pointing to the next data of DATATIMEk.

[0140] S3425: Take the sensing abnormal pattern of the i-th data.

[0141] S3426: Determine whether the sensing abnormal patterns of the i-th and i-1-th are the same, if yes: go to S3427; if not: add the i-th abnormal pattern to the KDD_Datak data set, and separate the items with commas.

[0142] In view of the component operation abnormal record cut by the component operation time window contains a large number of null, single value and the same perception variable mode multiple continuous values, which is not suitable for association rule mining. The application detects the component operation abnormal record formed by the above-mentioned time window cutting and merging, and only retains one in the continuous perception mode value in the sequence, and deletes the null value and single value record.

[0143] S3427: judge whether i>DATATIMEk.count is true, if true, go to S3424; if false, continue to execute S3428.

[0144] S3428: query the abnormal mode set in KDD_Datak, and store the perception variable code set in the variable sen_vars according to the abnormal mode association query.

[0145] S3429: judge whether the number of perception variables in sen_vars is greater than 1, if true, write KDD_Datak into the multivariate perception abnormal time sequence database KDD_SeqDB; if false, discard KDD_Datak and do not save.

[0146] In view of the fact that when the above-mentioned transformed time sequence data is directly applied to mining, there may also be frequent item interference between different modes of the same variable. The application proposes a multivariate mode sequence data screening algorithm, which removes the abnormal mode sequence data below two within the time window, thereby constructing a multivariate perception abnormal mode time sequence database of the periodically running component.

[0147] S343: judge whether the record in Event_seqDB is processed, if not processed, k=k+1, go to S32, if processed, end.

[0148] Step S4: multivariate perception abnormal mode time sequence dependence mining based on improved PrefixSpan. The PrefixSpan algorithm is adopted, the frequent 2-item sequence is solved, the projection database of the calculated frequent 1-item sequence is improved, the abnormal mode of the same perception variable in the projection database is deleted, then the sequence count is calculated based on the deleted projection database, the local 2-item frequent set is obtained, then the multivariate frequent time sequence dependence item set is obtained by recursive projection according to the PrefixSpan algorithm. The advantage of deleting the abnormal mode of the same perception variable is: one is to avoid the interference of different abnormal modes of the same perception variable, and the other is to improve the efficiency of the algorithm.

[0149] S41: set the minimum support minsup.

[0150] S42: obtain the frequent 1-item sequence SK, specifically including:

[0151] S421: Scan the whole database to get all different perception abnormal patterns

[0152] S422: Count the number of occurrences of each alarm pattern to get all frequent 1-item sequences SK.

[0153] S423: Remove alarm patterns with support counts less than the threshold from KDD_SeqDB.

[0154] S43: For each item in the frequent 1-item sequence SK, calculate the projected database and remove the same perception variable abnormal patterns, which specifically includes:

[0155] S431: Initialize the count j = 1.

[0156] S432: Calculate the projected database PJDj with the frequent pattern Sk[j] as the prefix.

[0157] S433: Remove alarm patterns of the same perception variable Sk[j] from PJDj.

[0158] S434: Determine whether PJDj is empty, if yes: count the support counts of each item in PJDj to get the frequent items PI; if not, go to S436.

[0159] S435: Determine whether PI is empty, if not, go to S436, if yes, execute:

[0160] S4351: Merge the single item that meets the support count with the previous sequence Sk[j] to get a new sequence.

[0161] S4352: Calculate the projected database PJDj with Sk[j] as the prefix.

[0162] S4353: Remove non-frequent items from the projected database.

[0163] S4354: Go to S434.

[0164] S436: j = j + 1, until j > SK.longth, SK.longth is the number of frequent items.

[0165] S44: Output the frequent sequences of frequent 2-item sets and above.

[0166] In view of the frequent sequence pattern mining algorithm directly using frequent item discovery sequence, there are frequent sequences between different patterns of the same variable. The application applies the PrefixSpan algorithm, first obtains the sequence of length 1, then constructs the projection database according to the 1-item sequence, improves the solution of the frequent 2-item sequence, deletes the abnormal mode of the same kind of sensing variable in the previous sequence from the projection database, then calculates the sequence count based on the deleted projection database, obtains the local 2-item frequent set, and then recursively projects according to the PrefixSpan basic algorithm to obtain the multivariate frequent time sequence dependent item set.

[0167] The above examples are only used to illustrate the present application and not to limit the technical solutions described in the present application. Although the present application has been described in detail with reference to the above embodiments, the present application is not limited to the above specific embodiments. Therefore, any modification or equivalent replacement of the present application without departing from the spirit and scope of the present application, and the improvement thereof, are all included in the scope of the claims of the present application.

Claims

1. A method for periodic operation component multi-sensor anomaly pattern temporal dependency mining, characterized in that: Comprise the following steps S1-S4: Step S1: component multi-element perception variable abnormal event capture platform construction; Step S2: capture perception variable abnormal mode timing data, build abnormal timing database Event_seqDB; Step S3: build multi-element perception abnormal mode timing database KDD_SeqDB; Step S4: based on the improved PrefixSpan multi-element perception abnormal mode timing dependence mining; using PrefixSpan algorithm, when solving frequent 2-item sequence, improvement is made, the projection database obtained by calculation is deleted from the projection database, and then the sequence count is calculated based on the deleted projection database, the local 2-item frequent set is obtained, and then the recursive projection is performed according to the PrefixSpan algorithm, and the multi-element frequent timing dependence item set is obtained; Wherein: In step S3, the time window of the component running period is taken as the associated valid time window, the perception variable abnormal mode timing data is cut, sorted and combined by using the valid time window, the component period running abnormal timing record is formed, and step S3 specifically comprises S31-S34: S31: initialize the running period k=1; S32: read the component work log, obtain the start time STIMEk and the end time ETIMEk of the period k; S33: obtain the abnormal mode timing data of each perception variable in the time period, and the relationship operation is as follows: DATATIMEk= (8) In formula (8), DATATIMEk is a set of abnormal time series data pairs obtained by intercepting the time window of the running period k, represents condition selection, represents the left and right conditions, and the formula is a selection operation of the data set, that is, the abnormal time series data in the running period k is selected by taking the start and end time of the running period k as the condition; S34: DATATIMEk data cleaning, remove null, single value and multiple continuous values of the same perception variable mode, specifically comprising S341-S343: S341: judge whether the number of abnormal data DATATIMEk.count contained in DATATIMEk is greater than or equal to 2, if not: go to S35; if yes: continue to execute S342; S342: integrate the abnormal mode timing data into the running item timing abnormal data, S342 specifically comprises S3421-S3429: S3421: initialize the record variable i=1, and initialize the item timing data set variable KDD_Datak=null; S3422: sort DATATIMEk data set according to time from small to large; S3423: take the perception abnormal mode of the first data and add it to the KDD_Datak data set; S3424: i=i+1, point to the next data of DATATIMEk; S3425: take the perception abnormal mode of the i-th data; S3426: judge whether the i-th and i-1-th perception abnormal mode is same, if yes: go to S3427; if not: add the i-th abnormal mode to the KDD_Datak data set, and separate the items with comma; S3427: judge whether i> DATATIMEk.count is true, if not: go to S3424; if yes: continue to execute S3428; S3428: query the abnormal mode set in KDD_Datak, and store the perception variable coding set in variable sen_vars according to the abnormal mode association query; S3429: judging whether the number of perception variables in sen_vars is greater than 1, if yes: writing KDD_Datak into the multi-perception abnormal timing database KDD_SeqDB; if no: discarding KDD_Datak and not saving; S343: judging whether the records in Event_seqDB are processed, if not: k=k+1, turning to S32, if yes: ending.

2. The method of claim 1, wherein the method is a periodic operation component multi-element anomaly pattern temporal dependence mining method. In step S4, the following steps S41-S44 are included: S41: setting the minimum support minsup; S42: obtaining the frequent 1-item sequence SK; S43: for each item in the frequent 1-item sequence SK, calculating the projected database and deleting the same perception variable abnormal pattern; S44: outputting the frequent sequences of frequent 2-item sets and above.

3. The method of claim 2, wherein the method is a periodic operation component multi-element anomaly pattern temporal dependence mining method. S42 specifically includes S421-S423: S421: scan the entire database to obtain all different perception anomaly patterns } S422: counting the occurrence of each alarm pattern to obtain all frequent 1-item sequences SK; S423: deleting the alarm patterns with support counts less than the threshold from KDD_SeqDB.

4. The method of claim 3, wherein the method is characterized by: S43 specifically includes S431-S436: S431: initializing the count j=1; S432: taking the frequent pattern Sk[j] as a prefix to calculate its projected database PJDj; S433: deleting the alarm patterns of the same perception variable of Sk[j] from PJDj; S434: judging whether PJDj is not empty, if yes: counting the support counts of each item in PJDj to obtain the frequent item PI; if no, turning to S436; S435: judging whether PI is not empty, if no, turning to S436, if yes, performing: S4351: merging the single item satisfying the support count with the previous sequence Sk[j] to obtain a new sequence; S4352: taking Sk[j] as a prefix to calculate its projected database PJDj; S4353: deleting the non-frequent items in the projected database; S4354: turning to S434; S436: j=j+1, until j>SK.longth, where SK.longth is the number of frequent items.

5. The method of claim 4, wherein the method is a periodic operation component multi-sensor anomaly pattern temporal dependency mining method, characterized by: Step S1 includes the following steps S11-S14: S11: encoding the parts, sensors and perception variables: each part of the part BOM structure tree node is uniquely identified by part_ID, each sensor is uniquely identified by source_ID, and the perception variable is uniquely identified by sen_var_ID. Since one part can deploy multiple sensors and one sensor can collect multiple variables, the following conditions should be met: (1) In formula (1), represents a total functional dependency, i.e. one corresponding sensor can be uniquely found through variable coding, and one corresponding component can be uniquely found through the sensor; S12: setting data cleaning and abnormal diagnosis algorithms for the perception variable class, specifically including: S121: classifying the perception variables corresponding to the parts, and uniquely identifying the perception variable class by sen_Type_ID, which satisfies the following constraints with the perception variables: (2) In formula (2), represents a total functional dependency, i.e. the perception variable is able to uniquely determine the classification it belongs to; S122: data cleaning and abnormal diagnosis algorithm packaging: selecting algorithms or deep learning models for each perception variable class from the general machine learning library Scikit-learn and the deep learning library Pytorch; The encapsulated algorithm is uniquely identified by algori_ID, one algorithm corresponds to one normal output and multiple abnormal outputs, the abnormal type is uniquely identified by Abn_ID, and the constraint between the abnormal type and the algorithm satisfies: (3) In formula (3), represents a total functional dependency, i.e. the corresponding algorithm can be uniquely determined by the exception type; S13: Perception data access and abnormal event capture algorithm configuration, S13 includes S131-S134: S131: Build a perception big data access and processing platform basic environment by using open source message middleware Kafka and open source stream data processing platform Flink; S132: Publish the real-time data corresponding to each perception variable as a message topic through Kafka; S133: Configure the tasks of each perception variable in the stream data processing platform Flink, including the time window size, the called algorithm and the consumed message topic data, and establish the following function dependencies through the configuration: (4) In formula (4), TWSize represents the time window size; algori_ID represents the called data processing algorithm, and TopicID represents the message topic, that is, the processing method of the perception data can be uniquely determined through the perception variable; S134: Perception anomaly pattern encoding: Since each algorithm is able to output one normal and multiple abnormal types, for each perception variable sen_var_ID(i), use { , ,…, } to uniquely encode its anomaly pattern representation; for any , there exists: (sen_var_ID, Abn_ID) (5) In formula (5): represents equivalence, that is, a set of perception variables and an abnormal type can be uniquely determined by a perception abnormal pattern code, and vice versa, a perception abnormal pattern code can be uniquely determined by a set of perception variables and an abnormal type. S14: Write the corresponding component perception configuration data entity and contact of (1), (2), (3), (4) and (5) into the configuration database configDB.

6. The method of claim 5, wherein the method is a periodic operation component multi-sensor anomaly pattern temporal dependency mining method, characterized by: Step S2 includes the following steps S21-S23: S21: The stream data processing platform Flink consumes the corresponding message topic data according to the time window, calls the configured algorithm, captures the abnormal events of each perception variable, and constructs an abnormal mode sequence data, for any perception variable sen_var_ID(i), the corresponding abnormal perception time sequence can be expressed as: Event_seq={<sen_var_ID(i),Timex,Abn_ID(j)>} (6) In formula (6), Event_seq represents the abnormal event sequence data corresponding to sen_var_ID(i), {<sen_var_ID(i),Timex,Abn_ID(j)>} represents a perception variable time event triple record, Timex represents the abnormal occurrence time, and Abn_ID(j) is the corresponding abnormal type; S22: Calculate using Abn_ID(j) and sen_var_ID(i) by (5) Transform the data to: Event_seq={<Timex, >,..., <Timex, >,..., <Timex, >,...} (7) S23: Write Event_seq into Event_seqDB.

Images