WO2016117453A1

WO2016117453A1 - Anomaly diagnosis/analysis apparatus

Info

Publication number: WO2016117453A1
Application number: PCT/JP2016/051065
Authority: WO
Inventors: 内田　貴之; 崎村　茂寿; 智昭蛭田; 藤城　孝宏
Original assignee: 株式会社日立製作所
Priority date: 2015-01-19
Filing date: 2016-01-15
Publication date: 2016-07-28
Also published as: JP2016133944A; JP6430263B2

Abstract

A diagnosis process focusing on a plurality of machines is required in order to extend the application of effective anomaly diagnosis to preventive maintenance. However, due to cost reduction efforts, etc., the specifications for the sensors and components of machines change. The result is that the number of false or missing diagnostic reports can vary among different machines of the same type. Therefore, it is necessary to detect and analyze the underlying causes for many false or missing reports in a particular machine. However, if there are many individual machines to be checked for causes or many sensors used in diagnosis, the number of operations performed by an analyst to specify the causes increases, and analysis becomes difficult. In order to address this problem, this anomaly diagnosis/analysis apparatus is characterized by being provided with a diagnosis unit for performing anomaly diagnosis on the basis of sensor data of a machine, a false/missing report detection unit for comparing the diagnosis results of the diagnosis unit with the anomaly history of the machine to detect false/missing reports, and a sensor difference detection unit for informing the analyst about sensors of differing specifications in individual machines having many or few false/missing reports.

Description

Abnormality diagnosis analyzer

The present invention relates to a technology that supports improvement in diagnosis accuracy of machine abnormality diagnosis.

In order for machines such as gas engines, elevators, mining and construction machinery to always operate, machine maintenance work is essential. One of the effective technologies for maintenance work is to collect sensor data from sensors attached to each part of the machine, and perform sensor abnormality diagnosis from the collected sensor data as sensor data. If there is an abnormality, analyze the cause. There is technology to do.

In order to implement this technique, there is a method of investigating machine abnormality from outliers of the appearance frequency distribution by expressing it with a scatter diagram or histogram showing the sensor data of the machine and the data appearance frequency.

For example, FIG. 14 is a drawing that expresses the balance between the engine temperature and the cooling water pressure of the machine in a scatter diagram. A circle 14100 in the figure represents a scatter diagram of temperature and pressure at the time of normal operation as a set of circles called clusters. A technique for creating such a cluster from a scatter diagram is called clustering, and is a known technique in the fields of machine learning and data mining. Creating a cluster is called “learning” in the field of machine learning. The distance 14120 from the cluster is calculated as the degree of abnormality, that is, the degree of abnormality, and if it is larger than the abnormality level threshold, the machine is diagnosed as abnormal.

The problem in applying the above-described diagnostic processing to a large number of machines is the variation in the accuracy of the diagnostic results for each machine. In general, the specifications of sensors and parts change from lot to lot due to cost reduction activities. For example, the sensor model number may change between the initial lot and the latest lot, and the frequency characteristics and dynamic range may be different. Therefore, even if the accuracy of diagnosis is high for individuals in the initial lot, the number of misreports / missing reports increases in the individual of the latest lot, and the accuracy may decrease. In order to improve the accuracy of abnormality diagnosis, it is necessary to correct the diagnosis process by identifying a sensor that causes a false alarm or a false alarm and removing it from the diagnosis target or changing the number of clusters.

(Patent Document 1) is an example of an invention that prevents a decrease in accuracy of abnormality diagnosis caused by a change in parts. This document is an invention for determining the cause of an abnormality as part of an abnormality diagnosis. If there is a change in the part specifications between the lot of the device to be diagnosed and the latest lot, it is assumed that the part has changed due to some problem with the part, and the cause is increased by increasing the probability that the changed part is the cause of the abnormality. Improve judgment accuracy.

Japanese Patent Laid-Open No. 5-101246

In the known technology, the cause of the abnormality is estimated using the change information of the parts. However, it was difficult to find the cause of false or misreporting, not the cause of the abnormality.

The present invention is to provide a technique for finding the cause of misreporting or misreporting and assisting in improving the diagnostic accuracy of machine abnormality diagnosis.

In order to solve the above problems, the abnormality diagnosis analyzer of the present invention detects a misreport / missing report by comparing a diagnosis unit that performs an abnormality diagnosis from sensor data of the machine and a diagnosis result of the diagnosis unit against a history of abnormality of the machine. And a sensor difference detection unit that presents an analyst with sensors having different specifications of individuals with a large number of individuals and a small number of individuals with the number of misinformation / missing reports.

Furthermore, the abnormality diagnosis / analysis apparatus of the present invention includes a misreport / miss report difference determination unit that counts misreports / missing reports in groups such as machine manufacturing lots by comparing the diagnosis results with the machine malfunction history. It is what.

Furthermore, in the abnormality diagnosis / analysis apparatus of the present invention, the sensor difference detection unit searches the sensor information of lots with a large number and a small number of false alarms and / or misreports, and confirms the sensor specifications between lots. It is a feature.

Furthermore, the abnormality diagnosis analyzer of the present invention includes a sensor data database storing data measured by a sensor attached to the machine together with a measurement time, and the diagnosis unit performs diagnosis from the sensor data database. It is what.

Furthermore, the abnormality diagnosis analyzer of the present invention is characterized by including a sensor used for diagnosis and a diagnosis model database storing the processing contents of diagnosis.

Furthermore, the abnormality diagnosis / analysis apparatus of the present invention is characterized by comprising a parts information database storing information of parts including the sensors of the equipment.

Furthermore, the abnormality diagnosis analyzer of the present invention is characterized by comprising an abnormality period history database storing an abnormality period based on the maintenance history of the machine and the complaint information.

Furthermore, the abnormality diagnosis analyzer of the present invention is characterized by comprising a diagnosis result database for storing a period during which an abnormality is diagnosed by the diagnosis unit.

Analysts who create an abnormality diagnosis process can see differences in sensor specifications that cause false alarms and false alarms according to the present invention. The analyst can remove the sensor having the difference in the specification from the diagnosis target or change the number of clusters to improve the diagnosis accuracy.

1 is a system configuration diagram of an embodiment. It is a flowchart of an Example. It is a flowchart of an Example. It is a flowchart of an Example. It is a flowchart of an Example. It is a flowchart of an Example. It is a flowchart of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the data structure used with the system of an Example. It is drawing explaining the principle of abnormality diagnosis It is an example of a screen displayed in the embodiment It is an example of a screen displayed in the embodiment

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 shows the overall configuration of the present invention.

The sensor data database 100 is a database that stores data such as engine pressure, cooling water temperature, and rotation speed measured by sensors attached to various parts of equipment such as railways and construction machines, along with measurement times. FIG. 8 shows the internal table structure. Sensor values such as the engine pressure 805 and the engine speed 810 are stored in association with the measurement time 800, and each sensor value can be referred to from the measurement time 800. Hereinafter, it is assumed that data has already been measured and stored in the sensor data database 100.

The diagnostic model database 105 is a database that stores diagnostic processing contents such as sensors used for diagnosis and preprocessing. FIG. 9 shows the table structure. A diagnosis model ID 900 for specifying a diagnosis model, a sensor 910 and its sensor component ID 915 used for diagnosis processing in association with an abnormality name 905 detected by each diagnosis model, a pre-processing 920 applied to each sensor value, a cluster The number 925 stores the ID 930 of the device to which each diagnostic model is applied. The processing contents of the diagnostic model applied to each device can be acquired from the table. These processing contents are determined by an analyst at the time of designing the diagnostic process, and the processing contents are already stored in the diagnostic model database 105 at the start of the process of the present invention.

The parts information database 110 is a database that stores information on parts including sensors constituting the device. FIG. 10 shows the internal table structure. From the lot number 1000, it has a part name 1005, a part model number 1010, a dynamic range 1015, an attachment position 1020, and drawing data 1025 of the parts of the device of the lot. A dynamic range 1015 indicates a range of data values that can be measured when the component is a sensor. The attachment position 1020 is a number indicating the position where the component is attached, and data indicating the number exists in the drawing data 1025. The drawing data 1025 stores CAD and image data indicating a manufacturing drawing of the device. By reading and presenting the drawing data 1025, the analyst can confirm the attachment position 1020 in the form of an image or CAD data.

The device information database 120 is a database that manages the association between a device to be diagnosed and a lot. FIG. 11 shows the internal table structure. The table structure manages the device ID 1100 to be diagnosed and the lot number 1105 in association with each other.

The abnormal period history database 115 is a database that stores a truly abnormal period created based on the maintenance history of the machine to be diagnosed, customer complaints, and the like. If the abnormal period in the abnormal period history database 115 cannot be detected, it is regarded as a misreport, and conversely, an abnormality detected without being stored in the abnormal period history database 115 is regarded as a false report.

The diagnosis result database 125 is a database that stores a period during which the diagnosis unit 145 diagnoses an abnormality.

FIG. 12 illustrates the table structure of the abnormal period history database 115. The table structure is managed by associating the ID 1200 of the device in which the abnormality actually occurred, the start time 1210 when the abnormality started, the time 1215 when the abnormality ended, and the type 1220 of the abnormality.

FIG. 13 illustrates the table structure of the diagnosis result database 125.

The table structure is managed by associating the ID 1230 of the device diagnosed as abnormal by the diagnosis unit 145, the abnormality start time 1235, the end time 1240 calculated by the diagnosis, and the abnormality name 1245 detected by the diagnosis.

The input unit 160 includes a keyboard, a mouse, a touch panel, and the like, and is used for pressing a button on the screen or inputting data into the diagnostic model database 105 in the present invention.

The display unit 155 is a device configured by a liquid crystal display or the like and displaying the screens of FIGS.

The temporary storage unit 150 is a volatile memory composed of a RAM or the like, and temporarily stores variables and a data table as shown in FIG.

The diagnosis unit 145 performs a diagnosis from the diagnosis model database 105 and the sensor database 100, and calculates a period when it is determined as abnormal. As a calculation method, for example, a period during which the degree of abnormality calculated by the clustering technique described in the above [Background Art] item exceeds a threshold may be calculated.

The false / missing report detection unit 140 checks whether the diagnosis unit 145 has correctly calculated an abnormal period, and if there is a false / missing report, calculates the number of occurrences. Specifically, if the true abnormal period in the abnormal period history database 115 and the abnormal period calculated by the diagnosis unit 145 are covered, it is considered that an abnormality has been detected correctly, and otherwise, it is regarded as a false or missing report. The flow of processing will be described later.

The misreport / miss report difference determination unit 135 counts the number of misreports and / or miss reports for each lot, and outputs a set of numbers of lots with many misreports / miss reports and lots with a small number. For this purpose, a search is made for a set of two lots that have a difference in the number of false / missing reports. If there is a difference in the number of occurrences between the two lots, it is highly possible that a change in the sensor specifications accompanying the change in lots is the cause of false or missing reports. The two lot numbers determined to have a difference are output to the sensor difference detection unit 130 as the numbers of lots with many false alarms and lots with few false alarms.
The sensor difference detection unit 130 detects whether there is a difference in the specifications of sensors used for diagnosis between lots with many false alarms / missing reports and lots with few. If there is a difference, it is considered that the cause of the false / missed report is the sensor specification. For this purpose, the sensor ID used for diagnosis is read from the diagnosis model database 105. A sensor specification is searched for each lot from the component information database using the sensor ID, and it is confirmed whether there is a difference in the specification between lots. If there is a difference, it is presented on the display unit 155.
Next, processing performed in the present invention will be described with reference to a flowchart. 2 will be described with reference to FIG. 3 to FIG.

FIG. 2 shows a main flow of processing in the present invention.

In step 202 (hereinafter referred to as “S202”), a list of types of abnormalities detected by diagnosis is displayed on the display unit 155 as shown in FIG.
As the abnormality name to be displayed, information obtained by reading the abnormality name (abnormality name 905 in FIG. 9) in the diagnostic model database 105 is displayed. The analyst selects an abnormality to be detected from FIG. 16 using the input unit.

In S205, all the devices to be diagnosed are diagnosed to detect the abnormality selected in S202, and the number of false / missing reports is counted for each device based on the result. SUB03, which is a subroutine for executing the processing, will be described with reference to FIG.

Hereinafter, in S300 in FIG. 3, the abnormality name selected by the analyst from the diagnosis model database 105 as a search key (905 in FIG. 9) is used as a search key (sensor 910 in FIG. 9) and the number of clusters (in FIG. 9). Information necessary for diagnosis such as the number of clusters (925) is read.

In order to acquire the ID of the device to be diagnosed in S305, the first line of the device ID (device ID 1100 in FIG. 11) is acquired from the device information database.

In S310, the sensor database 100 is searched using the device ID and sensor to be diagnosed as search keys, sensor data used for diagnosis is read, and abnormality diagnosis is performed. The abnormality diagnosis uses the above-mentioned (background art) and the clustering technique described in FIG. A period in which the degree of abnormality exceeds the threshold value is regarded as abnormal along FIG. The start time and end time of the abnormal period as output results are stored in the start time 1235 and end time 1240 of the diagnosis result database 125 shown in FIG.

In S315, the subroutine SUB01-01 for calculating the number Nf of false alarms of the diagnosis result is called. Hereinafter, SUB01-01 will be described with reference to FIG.

The variable Nf for storing the number of false alarms calculated in S400 is secured in the temporary storage unit 150. Nf is initialized to 0.

In S405, one line is read from the diagnosis result database 125, which is an abnormal period of the diagnosis result, using the device ID acquired in S305 and the abnormal name specified in S202 for the start time 1235 and the end time 1240 shown in FIG. The read start time 1235 is ST, and the end time 1240 is EN.

Hereafter, it is determined in S410 to S420 whether the read abnormal period is a false alarm. The determination method includes a start time 1210 and an end time 1215 of the abnormality history database in FIG. 12, which are periods in which the machine is actually abnormal, and a start time 1235 indicating an abnormal period in the diagnosis result database shown in FIG. Judgment is made by whether or not there is a time zone covered by the end time 1240.

In S410, one line of the start time 1210 and the end time 1215 of the abnormality history database in FIG. 12 is read from the abnormality period history database 115. It is assumed that the read start time 1210 is ST_H and the end time 1215 is EN_H.
In S415, it is determined whether there is any time zone covered by the abnormal period ST to EN and the abnormal period history ST_H to EN_H of the diagnosis result. If ST is a time after EN_H or EN is a time before ST_H, it is determined that the time zone is not covered, and the process proceeds to S420. If it is covered, it is not a false alarm, so the process proceeds to S430 in order to determine the abnormal period of the diagnosis result of the next line.

In S420, it is determined whether all abnormal period histories have been read in S410. If there is still an abnormal period history, the process returns to S410. If everything has been read, the process proceeds to S425.

In S425, since it is determined that the abnormality period of the abnormality diagnosis read in S405 does not cover the abnormality period history, that is, it is determined to be misinformation, the number Nf of misinformation is incremented by 1 and the process proceeds to the next S430.

S430 reads all abnormal periods in S405 and determines whether or not it is misreported. If all are read, returns the number Nf of false reports and returns to S315 in FIG. Otherwise, return to S405. When SUB01-01 is completed as described above and S315 in FIG. 3 ends, the process proceeds to S320.

In S320, the number of unreported cases Nn is calculated. Of the abnormal period history, Nn is the number of cases that are not covered by the abnormal period that is the diagnosis result, that is, are actually abnormal but not detected.
Nn is calculated in subroutine SUB01-02. SUB01-02 is almost the same as SUB01-01 that calculates the number of false alarms already described. In SUB01-01, the abnormal period history loops S410 to S420 are turned inside the abnormal period loops S405 to S430. On the other hand, in SUB01-02, the loop of the abnormal period is turned inside the loop of the abnormal period history. The rest of the logic is exactly the same, so details are omitted. When the number of unreported cases Nn is returned from SUB01-02, the process proceeds from S320 to S330.

In S330, Nf and Nn counted in S315 and S320 are stored as a table in the temporary storage unit. The structure is shown in a table 1330 in FIG. The table 1330 includes a device ID 1300 and a diagnosis model ID 1310 that have issued a false / missed report, an abnormal name 1315, a false report count 1320, and a miss report count 1325. Record the diagnosis model ID, abnormality name, Nf and Nn in this table.

In S340, it is checked whether the diagnosis of all devices and the count of false / missed reports have been completed. If not completed, the process returns to S305, and if completed, SUB01 is terminated. After the completion, the process returns to S205 in FIG. 2 and then proceeds to S208.

S208 to S220 are processes for determining whether there is a difference in the number of false / missed reports for each device for each lot of devices. If two lots with differences are found, one of the lots has few false alarms / missing reports, and the other lot has many misreporting / missing alarms. If the difference in the number of misreports / missing reports appears in lot units, the cause of the misreporting / missing reports is likely to be a difference in lots, that is, a difference in specifications of sensor parts.

S208 calls a subroutine SUB02 that searches for two lots in which there is a difference in the number of false alarms.
SUB02 will be described with reference to S600 to S615 in FIG.

In S600, the lot number 1100 and the device ID 1105 are searched from the device information database 120, and the device ID list of the lot number (here A) is read. Using the device IDs as search keys, the table 1330 in the temporary storage unit 150 is searched to search for the number of misreports 1320 for all devices with the lot number A. The average value Ave (Nf_A) and variance Var (Nf_A) of the number of misreports retrieved are calculated.

In S605, as in S600, the device ID list of the lot number (here, B) in the device information database 120 is read, and the number of erroneous reports of lot number B is retrieved from the table 1330. The average value Ave (Nf_B) and variance Var (Nf_B) of the number of detected false alarms are calculated.
In S610, it is statistically determined whether there is a difference in the number of false reports from the average value Ave (Nf_A), variance Var (Nf_A), and the false report count Ave (Nf_B) and variance Var (Nf_B) of lot B.
There is a well-known t-test as a method for statistically judging whether there is a difference in values between data lists, but in this example, for the sake of simplicity, Ave (Nf_A) and Ave (Nf_B) If the square of the difference is greater than the sum of Var (Nf_A) and Var (Nf_B), it is considered that there is a difference.

If there is a difference, the lot number A, lot number B, Ave (Nf_A) and Ave (Nf_B) are returned, and this subroutine SUB02 is completed. Otherwise, go to S615.

In S615, the combination of lot numbers A and B is changed to determine whether there is another combination.
For example, if lot number A = 1 and B = 2, B is changed to A = 1 and B = 3. If there is a combination of lot numbers not yet applied to A and B, the numbers are assigned to A and B, and the process returns to S600. Otherwise, this subroutine SUB02 is terminated without returning anything. After SUB02 ends, the process proceeds to S210 in FIG.

In S210, check whether lot numbers A, B, Ave (Nf_A), and Ave (Nf_B) are returned from SUB02.If nothing is returned, there is no difference in the number of false reports for each lot. The process moves to determinations S215 and S220. Otherwise, the return value is created and stored in the temporary storage unit 150 in the table 1335 of FIG. Then move to S225.

S215 searches for lots with a large number of missed reports and lots with a small number of reports in the same procedure as the number of false reports in S210.
The subroutine SUB03 of S215 is shown in FIG. 7, but the description is omitted because the number of erroneous reports of SUB02 is only the number of unreported cases.
S220 stores lot A and lot B with a large number of missed reports in the table 1370 shown in FIG.

In S225, the sensor part ID of the sensor used for the diagnosis of lot numbers A and B is acquired from the diagnosis model database 105. The sensor component ID can be acquired from the sensor component ID 915 of FIG. 9 using the diagnosis model ID of the diagnosis model specified in S202 as a search key.

In S230, the sensor specifications of lot numbers A and B are searched from the sensor part ID.
Specifically, the sensor specifications of A and B are searched from the component information database 110 using the sensor component ID and lot numbers A and B acquired in S225 as search keys.

In S235, it is determined whether there is a difference in sensor specifications between lots A and B from the search result in S230. Specifically, it is determined whether there is a difference between the lot A and B in the part model number 1010, the dynamic range 1015, and the mounting position 1020 in FIG. If there is a difference, go to S240. Otherwise, this process ends.

In S240, the difference in the specifications of the sensor parts of lots A and B is presented to the analyst as the cause of the false alarm or misreport. FIG. 19 shows an example of a screen to be presented, in which a difference in sensor model number is presented as a cause of false alarm.
This completes the processing performed in the present invention.

100 ... Sensor data database
105 ... Diagnosis model database
110… Part information database
115… Abnormal period history database
120… Device information database
125 ... Diagnosis result database

Claims

A diagnostic unit for performing abnormality diagnosis from the sensor data of the machine;
An error / missing alarm detection unit that detects an error / missing error by comparing the diagnosis result of the diagnosis unit with an abnormality history of the machine;
An abnormality diagnosis analyzer characterized by having a sensor difference detection unit that presents to the analyst sensors with different specifications between individuals with a large number of false alarms and missed reports
In claim 1,
An abnormality diagnosis / analysis apparatus comprising a misreport / miss report difference determination unit that counts misreports / misses in a group such as a machine manufacturing lot by comparing diagnosis results with a machine malfunction history.
In claim 2,
The abnormality detection analysis apparatus characterized in that the sensor difference detection unit searches the sensor information of lots with a large number and a small number of false and / or misreports, and confirms sensor specifications between lots.
In claim 1,
It is equipped with a sensor data database that stores data measured by the sensors attached to the machine along with the measurement time,
The abnormality diagnosis analyzer according to claim 1, wherein the diagnosis unit makes a diagnosis from the sensor database.
In claim 1,
An abnormality diagnosis analysis apparatus comprising a sensor used for diagnosis and a diagnosis model database storing diagnosis processing contents.
In claim 1,
An abnormality diagnosis and analysis apparatus comprising a component information database that stores information on components including sensors included in the device.
In claim 1,
An abnormality diagnosis / analysis apparatus comprising an abnormality period history database storing an abnormality period based on maintenance history and claim information of the machine.
In claim 1,
An abnormality diagnosis analyzer comprising a diagnosis result database for storing a period during which an abnormality is diagnosed by the diagnosis unit.