Failure prediction device and failure prediction method

Abstract

To provide a failure prediction technique capable of suppressing a reduction in prediction accuracy even when temporary data variations due to an external cause such as an ambient temperature or a vibration occur in monitored operation data for failure prediction.SOLUTION: A failure prediction device includes: an input unit that inputs time-series operation data for each core of two or two or more core semiconductor devices provided in an electronic apparatus; a data analysis unit that uses a value at the same time period or a statistical amount for the time-series operation data for each core to calculate the presence or absence of a failure or a failure sign of the electronic apparatus; and an output unit that outputs the presence or absence of a failure or a failure sign of the electronic apparatus.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a failure prediction device and a failure prediction method, and more particularly to a failure prediction device and a failure prediction method that are applicable to electronic devices having multiple semiconductor devices. [Background technology]

[0002] To ensure the long-term stable operation of infrastructure and IoT (Internet of Things) sensors, it is necessary to develop highly accurate failure prediction technology for semiconductor devices (LSI: Large-Scale Integrated Circuits), which are the heart of electronic devices used for control processing and edge processing. Such failure prediction technology typically monitors and acquires operational data from electronic devices and uses machine learning techniques to detect failures or signs of failure. Furthermore, such electronic devices typically have multi-core semiconductor devices to ensure safety and operational efficiency.

[0003] Japanese Patent Application Laid-Open Publication No. 2015-184818 discloses a server including a distribution difference calculation unit and a determination unit. To avoid applying a low-accuracy model to a terminal device, the server determines whether a model that predicts the likelihood of a failure in a second model of terminal device (a new-generation hard disk drive) is applicable to a first model of terminal device (a new-generation hard disk drive) based on feature values ​​calculated from operational data of the second model of terminal device. The distribution difference calculation unit calculates difference information between a first distribution of feature values ​​calculated from operational data of the first model of terminal device and a second distribution of feature values ​​calculated from operational data of the second model of terminal device. The determination unit determines whether the model is applicable to the first model of terminal device based on the difference information. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2015-184818 Summary of the Invention [Problem to be solved by the invention]

[0005] According to the research of the present inventors, in electronic devices (electronic devices) used for control processing, edge processing, etc., it is expected that the accuracy of failure prediction will be improved by appropriately combining per-core operation data acquired from multiple semiconductor devices (multi-core semiconductor devices) or by data processing. In addition, temporary data fluctuations occur in operation data monitored for failure prediction due to external factors such as outside temperature and vibration, which causes a decrease in prediction accuracy, and it is necessary to develop a means to solve this problem.

[0006] In JP 2015-184818 A, the first distribution of feature values ​​calculated from the operation data of a first model of terminal device and the second distribution of feature values ​​calculated from the operation data of a second model of terminal device are not operation data at the same time. Therefore, in electronic devices (electronic devices) used for control processing, edge processing, etc., it is thought that temporary data fluctuations in operation data at the same time due to external factors such as outside temperature and vibration are not taken into consideration.

[0007] The objective of the present disclosure is to provide a failure prediction technology that can suppress a decrease in prediction accuracy even when temporary data fluctuations occur in operational data monitored for failure prediction due to external factors such as outside temperature or vibration.

[0008] Other objects and novel features will become apparent from the description of this specification and the accompanying drawings. [Means for solving the problem]

[0009] A brief summary of representative aspects of this disclosure is as follows.

[0010] According to one embodiment, the failure prediction device includes an input unit that inputs time-series operation data for each core of two or more core semiconductor devices provided in an electronic device, a data analysis unit that calculates the presence or absence of a failure or signs of failure in the electronic device using the difference between values ​​or statistical quantities at the same time in the time-series operation data for each core, and an output unit that outputs the presence or absence of a failure or signs of failure in the electronic device.

[0011] Data fluctuations due to external factors appear simultaneously in multiple cores, so by taking the difference between the values ​​or statistics of the operating data for each core at the same time, the influence of the external factors can be canceled out, improving the accuracy of failure prediction. [Effects of the Invention]

[0012] According to the failure prediction device of the above embodiment, the difference between values ​​or statistical quantities of the operational data at the same time is used, so even if temporary data fluctuations occur due to external factors such as outside temperature or vibration, it is possible to suppress a decrease in failure prediction accuracy and improve failure prediction accuracy. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram illustrating an example of a conceptual configuration of a failure prediction device according to the first embodiment. [Figure 2] FIG. 2 is a diagram illustrating the difference calculation function. [Figure 3] FIG. 3 is a diagram illustrating an example of an operation data processing method. [Figure 4] FIG. 4 is a diagram illustrating an example of a conceptual configuration of a failure prediction device according to the second embodiment. [Figure 5] FIG. 5 is a diagram illustrating an example of the configuration of the data analysis unit in the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments will be described with reference to the drawings. However, in the following description, the same components will be assigned the same reference numerals, and repeated description may be omitted. Note that the drawings may be more schematic than the actual embodiment to make the description clearer, but they are merely examples and do not limit the interpretation of the present invention. (Embodiment 1) First, a failure prediction device and a failure prediction method according to the first embodiment will be described in detail with reference to Fig. 1. Fig. 1 is a diagram illustrating an example of the conceptual configuration of the failure prediction device according to the first embodiment. Fig. 2 is a diagram illustrating a difference calculation function. Fig. 3 is a diagram illustrating an example of an operation data processing method. Fig. 5 is a diagram illustrating an example of the configuration of a data analysis unit according to the first embodiment.

[0015] FIG. 1 shows an electronic device 1 including a failure prediction device 10. The electronic device 1 is an electronic device (electronic device) used for control processing, edge processing, etc., provided for long-term stable operation of infrastructure and IoT (Internet of Things) sensors. The electronic device 1 is provided with multiple cores (core 1 (101), core 2 (102), ..., core n (103)). Here, each of core 1 (101), core 2 (102), ..., core n (103) is configured as a semiconductor device (LSI: Large Scale Integrated Circuit device). Therefore, the electronic device 1 is configured with multiple core semiconductor devices (multi-core semiconductor device).

[0016] Monitoring programs MP1, MP2, and MP3 are running on each of the multiple cores (101, 102, and 103). The monitoring programs MP1, MP2, and MP3 are configured to collect time-series operation data 1 to n (121, 122, and 123) for each core (101, 102, and 103) and send the collected time-series operation data 1 to n (121, 122, and 123) for each core to the failure prediction device 10. Hereinafter, the time-series operation data 1 to n (121, 122, and 123) for each core may be simply referred to as the operation data 1 to n for each core.

[0017] It is preferable that the monitoring programs MP1, MP2, and MP3 are similar monitoring programs. Here, similar monitoring programs mean that they have the function of acquiring and collecting data from the same sensors or the like as operational data at the same time or at the same point in time, and sending the collected operational data to the failure prediction device 10. When the semiconductor devices constituting each core have the same data processing operation unit, it is preferable that the monitoring programs executed by the same data processing operation unit be the same program, from the viewpoint of shortening program development time and reducing development costs. Even if the data processing operation units of the semiconductor devices constituting each core have different configurations, it is sufficient that the monitoring programs executed by each data processing operation unit are similar monitoring programs as described above.

[0018] The failure prediction device 10 includes an input unit 13, a data analysis unit 14, and an output unit 15. The data analysis unit 14 of the failure prediction device 10 performs data analysis and failure prediction. The failure prediction device 10 can be configured by a hardware circuit (e.g., an input circuit 13, a data analysis circuit 14, an output circuit 15, etc.) as a dedicated failure prediction circuit formed in a dedicated semiconductor device. The failure prediction device 10 can also be configured by a software program as a failure prediction program (e.g., an input program 13, a data analysis program 14, an output program 15, etc.). In this case, the failure prediction device 10 can be configured by a semiconductor device including a central processing unit (CPU) that executes the failure prediction program, a storage device, etc.

[0019] The input unit 13 receives operation data 1 to n (121, 122, 123) for each core.

[0020] The data analysis unit 14 calculates the difference between values ​​or statistical quantities at the same time points of the time-series operation data 1 to n for each core input to the input unit 13. The data analysis unit 14 also performs failure prediction based on the calculated difference, and calculates the presence or absence of an abnormality or a sign of an abnormality in the electronic device 1. The data analysis unit 14 also has a function of generating a failure prediction model based on the difference between values ​​or statistical quantities of the time-series operation data 1 to n (121, 122, 123) for each core at the same time points during normal operation, and calculating the probability of a failure occurring in the electronic device 1 within a certain period from the present time. The data analysis unit 14 also calculates the presence or absence of multiple failures or signs of failure based on the difference between values ​​or statistical quantities of the time-series operation data 1 to n (121, 122, 123) for each core, and can identify a core semiconductor device in which a failure or a sign of failure will occur by comparing the results of the difference between values ​​or statistical quantities of the time-series operation data 1 to n (121, 122, 123) for each core.

[0021] The output unit 15 outputs the presence or absence of an abnormality or an abnormality sign of the electronic device 1 calculated by the data analysis unit 14. If the electronic device 1 has a display device, the electronic device 1 displays the presence or absence of an abnormality or an abnormality sign in the display area of ​​the display device. If the electronic device 1 has a communication circuit, the electronic device 1 notifies the presence or absence of an abnormality or an abnormality sign to a management device or a mobile terminal device of the administrator of the electronic device 1 via the communication circuit. In other words, when a failure or a failure sign for at least one core semiconductor device is calculated by at least one of the multiple calculations performed by the data analysis unit 14, the output unit 15 outputs the result.

[0022] In this embodiment, the operational data 1 to n (121, 122, 123) for each core is collected by monitoring programs MP1, MP2, and MP3 running on multiple cores (101, 102, 103). The monitoring method is free, and continuous monitoring is preferable, but is not limited to this. For example, the operational data 1 to n (121, 122, 123) for each core may be acquired at regular intervals. For example, the operational data 1 to n (121, 122, 123) for each core may be acquired for one minute every hour, which is the regular interval. In other words, the data processing calculation unit is operated intermittently to execute the monitoring program for one minute every hour. This allows for reduced power consumption in the electronic device 1 and the failure prediction device 10.

[0023] In this embodiment, the monitoring programs MP1, MP2, and MP3 running on the multiple cores (101, 102, and 103) are preferably similar monitoring programs, but are not limited to this. Different monitoring programs may be used as long as they can acquire the data necessary for data analysis. The monitoring programs MP1, MP2, and MP3 must at least have the function of acquiring the same data necessary for data analysis (for example, data that allows calculation of the difference between values ​​or statistical quantities at the same time point).

[0024] The input unit 13 has a function of inputting the operational data 1 to n (121, 122, 123) for each core to the data analysis unit 14, but is not limited to this. In some cases, the input unit 13 may have a function of transferring data to the outside or a fixed (or predetermined) data processing function.

[0025] 5 is a diagram illustrating the configuration of the data analysis unit 14 in the first embodiment. As shown in FIG. 5, the data analysis unit 14 has a difference calculation function 141, a failure prediction model generation function 142, and a failure probability calculation function 143. The difference calculation function 141, the failure prediction model generation function 142, and the failure probability calculation function 143 can be configured by hardware circuits (for example, the difference calculation circuit 141, the failure prediction model generation circuit 142, and the failure probability calculation circuit 143). Furthermore, the difference calculation function 141, the failure prediction model generation function 142, and the failure probability calculation function 143 can be configured by software programs (for example, the difference calculation program 141, the failure prediction model generation program 142, and the failure probability calculation program 143).

[0026] The difference calculation function 141 will be described using FIG. 2. FIG. 2 shows operational data 1 and 2 and the difference from time t1 to time tn. In FIG. 2, the horizontal axis represents time, and the vertical axis represents operational data 1 (121), operational data 2 (122), and the difference as feature quantities. For example, at time ta, there is a somewhat large data change (D1) in operational data 1 (121), and there is a somewhat small data change (D2) in operational data 2 (122). Here, it is assumed that the data change (D2) occurring in operational data 2 (122) is a temporary data fluctuation occurring due to external factors such as outside temperature or vibration. Therefore, since this temporary data fluctuation is also included in data change (D1), by subtracting data change (D2) from data change (D1) to obtain difference data (D3), it is possible to obtain data (here, difference data (D3)) in which the temporary data fluctuation occurring in data change (D1) and data change (D2) is canceled.

[0027] In this way, by calculating the difference between operation data 1 and operation data 2, temporary data fluctuations that appear in the operation data used to predict failures due to external factors such as outside temperature and vibrations can be canceled out, thereby improving the accuracy of failure predictions.

[0028] 2, when there are two cores (core 1 (101) and core 2 (102)), the difference calculation function 141 is characterized by having a function or difference calculation circuit that executes a difference calculation method that calculates the difference at the same point in time series values ​​of operational data 1 (121) acquired from core 1 and operational data (122) acquired from core 2. In other words, the difference calculation function 141 can also be considered as a function or circuit that executes data analysis of operational data.

[0029] Furthermore, if there are two or more cores, they are paired in pairs, and the difference between the time-series values ​​is calculated for each pair. That is, the difference between the values ​​or statistics of the time-series operational data for each of the multiple cores corresponding to two different core semiconductor devices is calculated. For example, if there are three cores, core 1 and core 2 are paired, and difference 1 is calculated using the time-series operational data of core 1 and the time-series operational data of core 2 using the difference calculation method. Then, core 2 and core 3 are paired, and difference 2 is calculated using the time-series operational data of core 2 and the time-series operational data of core 3 using the difference calculation method. Finally, core 1 and core 3 are paired, and difference 3 is calculated using the time-series operational data of core 1 and the time-series operational data of core 3 using the difference calculation method. Here, the difference between the time-series values ​​is shown as an example, but this is not limiting. As described in FIG. 3, the difference between the statistics of the time-series operational data for each core may also be calculated.

[0030] For example, as shown in FIG. 3, the difference calculation function 141 may divide the time-series operational data (operational data 1 (121) and operational data 2 (122)) of X hours into groups at regular intervals of one hour, calculate the statistics of the average value (or variance, etc.) of the operational data in each group, and calculate the difference between the statistics at the same time. In other words, it calculates the difference (difference 1, difference 2, . . ., difference n) between "the 1-hour statistics (average value) of operational data 1 (121) at the same time: average value 1, average value 2, . . ., average value n" and "the 1-hour statistics (average value) of operational data 2 (122) at the same time: average value 1, average value 2, . . ., average value n". For example, difference 1 is the difference between average value 1 of operational data 1 (121) and average value 1 of operational data 2 (122). Since the difference between the hourly average value of the operation data 1 (121) and the hourly average value of the operation data 2 (122) is calculated, the amount of data used for calculation can be reduced, which can reduce the power consumption of the electronic device 1 and the failure prediction device 10. In addition, the scale of the failure prediction device 10 can be reduced.

[0031] In the example shown in FIG. 3, the time-series operational data for X hours is grouped every hour, and intervals are set, but this is not limiting and any interval (for example, grouping every 2 hours, grouping every 3 hours) may also be used.

[0032] The failure prediction model generation function 142 is characterized by using the difference or a portion of the difference calculated by the difference calculation function 141 to generate one or more failure prediction models using one or more machine learning algorithms. The failure prediction model generation function 142 generates a failure prediction model based on, for example, the difference between values ​​or statistical quantities at the same time point in time of the time-series operation data 1 to n (121, 122, 123) for each core during normal operation. In other words, the failure prediction model generation function 142 generates multiple failure prediction models by combining a failure prediction algorithm as a machine learning algorithm with the time-series operation data 1 to n (121, 122, 123) for each core, and calculates the presence or absence of multiple failures or signs of failure. The failure prediction model generation function 142 is not limited to generating a failure prediction model using only the difference. The failure prediction model generation function 142 may generate a failure prediction model using, for example, any combination of the difference and the operation data.

[0033] The failure probability calculation function 143 is characterized in that it uses the most recent operation data or the difference between the operation data acquired from the multiple cores (101, 102, 103) to calculate the probability that a failure will occur within a certain period from the present time using a failure prediction model. The failure prediction model generation function 142 and the failure probability calculation function 143 can also be considered as functions or circuits that execute failure prediction. The failure prediction model generation function 142 and the failure probability calculation function 143 calculate the presence or absence of multiple failures or signs of failure based on the difference between the values ​​or statistics of the time-series operation data 1 to n (121, 122, 123) for each core, and can identify a core semiconductor device in which a failure or a sign of failure will occur by comparing the difference results between the values ​​or statistics of the time-series operation data 1 to n (121, 122, 123) for each core.

[0034] Note that the machine learning algorithms, failure prediction algorithms, generation of failure prediction models, calculation of failure occurrence probability based on failure prediction models, and determination of the presence or absence of failures or signs of failure are themselves well known to those skilled in the art, and well-known methods can be used, so detailed explanations of them will be omitted here.

[0035] In this way, by calculating the difference between operation data 1 and operation data 2, temporary data fluctuations that appear in the operation data used to predict failures due to external factors such as outside temperature and vibrations can be canceled out, improving the accuracy of failure predictions. (Embodiment 2) FIG. 4 is a diagram illustrating an example of a conceptual configuration of a failure prediction device according to the second embodiment.

[0036] In the first embodiment, a case has been described in which data analysis and failure prediction are all performed by the failure prediction device 10 provided inside the electronic device 1. In this embodiment, referring to FIG. 4, the operational data (121, 122, 123) is transferred to the external device 24 as log data, and the data analysis and failure prediction are performed by the external device 24.

[0037] As shown in FIG. 4, an electronic device 1a having two or more cores (core 1 (101), core 2 (102), ..., core n (103)) has similar monitoring programs MP1, MP2, and MP3 running on the multiple cores (101, 102, 103) as in the first embodiment. Time-series operation data 1 to n (121, 122, 123) for each core collected by the monitoring programs MP1, MP2, and MP3 is transferred as a log to an external device 24 provided outside the electronic device 1a via a data processing unit 23 provided inside the electronic device 1a. The external device 24 is characterized in that it uses the time-series operation data 1 to n (121, 122, 123) for each core to perform data analysis and failure prediction performed by the data analysis unit 14 of the failure prediction device 10 described in the first embodiment. 1. The data processing unit 23 corresponds to the input unit 13 of the failure prediction device 10 described in the first embodiment, and the data processing unit 23 and the external device 24 can be considered to correspond to the failure prediction device 10 of FIG.

[0038] The data processing unit 23 has a function of transferring the operational data 1 to n (121, 122, 123) for each core to the external device 24, but is not limited to this. For example, the data processing unit 23 may be provided with a difference calculation function 141 that includes the difference calculation method described in the first embodiment, as needed. In this case, the difference (difference result) calculated by the difference calculation method is transferred as a log to the external device 24. Since the calculated difference has a small data volume, the transfer to the external device 24 takes a short time, and the power consumed by the data processing unit 23 for data transfer and the power consumed by the external device 24 can be reduced.

[0039] The invention made by the inventor has been specifically described above based on examples, but it goes without saying that the present invention is not limited to the above-described embodiments and examples, and various modifications are possible. [Explanation of symbols]

[0040] 1:Electronic equipment 10: Failure prediction device 13: Input section 13 14: Data analysis section 15: Output section 101, 102, 103: Core (core semiconductor device) 121, 122, 123: Operating data for each core 1 to n 141: Difference calculation function 142:Failure prediction model generation function 143:Failure probability calculation function

Claims

1. an input unit for inputting time-series operation data for each of two or more core semiconductor devices provided in the electronic device; a data analysis unit that calculates whether or not there is a failure or a sign of a failure in the electronic device by using a difference between values ​​or statistics at the same time of the time-series operation data for each core; an output unit that outputs whether or not the electronic device has a malfunction or a malfunction symptom; the data analysis unit generates a plurality of failure prediction models by combining a plurality of failure prediction algorithms and a plurality of time-series operation data for each of the cores, and calculates the presence or absence of a plurality of failures or signs of failure; The output unit outputs a result of calculation of a failure or a sign of failure of the core semiconductor device by at least one of the plurality of calculations performed by the data analysis unit.

2. 2. The failure prediction device according to claim 1, The data analysis unit is a failure prediction device having a function of calculating the difference between values ​​or statistics at the same time of the time-series operation data for each core.

3. 2. The failure prediction device according to claim 1, the input unit has a function of calculating a difference between values ​​or statistics at the same time of the time-series operation data for each core, The failure prediction device is characterized in that the calculation result of the difference is output to the data analysis unit or to an external device as a log.

4. 2. The failure prediction device according to claim 1, The data analysis unit generates a failure prediction model based on the difference in values ​​or statistical quantities at the same time in the time-series operational data for each core during normal operation, and calculates the probability that a failure will occur within a certain period of time from the present time.

5. 2. The failure prediction device according to claim 1, When the core semiconductor device has three or more cores, the data analysis unit calculates a difference between values ​​or statistics of time-series operational data for each of the plurality of cores corresponding to two different core semiconductor devices, The data analysis unit calculates the presence or absence of multiple failures or signs of failure based on the difference in values ​​or statistical quantities of time-series operation data for each of the multiple cores, and identifies the core semiconductor device in which a failure or signs of failure will occur by comparing the results.

6. A failure prediction method performed by a failure prediction device, comprising: inputting time-series operation data for each core of two or more core semiconductor devices provided in an electronic device into an input unit of the failure prediction device; a data analysis unit of the failure prediction device calculates whether or not there is a failure or a sign of a failure of the electronic device by using a difference between values ​​or statistics at the same time of the time-series operation data for each core; an output unit of the failure prediction device outputs whether or not there is a failure or a sign of a failure in the electronic device; The data analysis unit generates a plurality of failure prediction models by combining a plurality of failure prediction algorithms and a plurality of time-series operation data for each of the cores, and calculates the presence or absence of a plurality of failures or signs of failure; a failure prediction method, wherein when a failure or a sign of failure of the core semiconductor device is calculated by at least one of the plurality of calculations performed by the data analysis unit, the output unit outputs the result.

7. 7. The failure prediction method according to claim 6, A failure prediction method in which the data analysis unit generates a failure prediction model based on the difference in values ​​or statistical quantities at the same time in the time-series operation data for each core during normal operation, and calculates the probability that a failure will occur within a certain period of time from the present time.

Images