JP7012086B2 - Artificial intelligence algorithm - Google Patents

Description

The invention disclosed herein relates to an artificial intelligence algorithm.

In recent years, a machine health monitoring system has been proposed that detects an abnormality in a monitored device using a sensor node.

As an example of the prior art related to the artificial intelligence algorithm, for example, Non-Patent Document 1 and Non-Patent Document 2 can be mentioned.

Diego FernaNdez-Francos, Automatic bearing fault diagnosis based on one-class ν-SVM, Journal Computers and Industrial Engineering archive Volume 64 Issue 1, January, 2013 Pages 357-365 Yingchao Xiao, Two methods of selecting Gaussian kernel parameters for one-class SVM and their application to fault detection, Knowledge-Based Systems, Volume 59, March 2014, Pages 75-84

However, in the conventional machine health monitoring system, abnormality detection of the monitored device is performed by collecting and analyzing the measurement data of the sensor node on the server. Therefore, the amount of communication between the sensor node and the server is very large, which is a factor that hinders the introduction of the system.

In view of the above-mentioned problems found by the inventors of the present application, the invention disclosed in the present specification includes an easy-to-introduce machine health monitoring system, and a sensor node and an artificial intelligence chip used therein. And, an object of the present invention is to provide an artificial intelligence algorithm.

The artificial intelligence algorithm disclosed in the present specification includes a step of generating a feature vector by extracting a feature amount for each frequency band from input data using a plurality of bandpass filters connected in parallel. The configuration (first configuration) includes a step of obtaining the value of the kernel function using the feature vector and the support vector.

The artificial intelligence algorithm having the first configuration may be configured to further include a step of detecting an abnormality in the input data from the value of the kernel function (second configuration).

Further, in the artificial intelligence algorithm having the first or second configuration, the input data may be configured to be vibration data (third configuration).

Further, in the artificial intelligence algorithm having any of the first to third configurations, if the kernel function is a linear kernel, a Gaussian kernel, or an RBF [radial base function] kernel (fourth configuration). good.

Further, in the artificial intelligence chip disclosed in the present specification, before generating a feature vector by extracting feature quantities for each frequency band from input data using a plurality of bandpass filters connected in parallel. It has a configuration (fifth configuration) including a processing unit and a classifier for obtaining the value of the kernel function using the feature vector and the support vector.

The artificial intelligence chip having the fifth configuration may be configured to further have a post-processing unit (sixth configuration) for detecting an abnormality in the input data from the value of the kernel function.

Further, in the artificial intelligence chip having the fifth or sixth configuration, the classifier may have an OCSVM [one class support vector machine] configured by hardware (seventh configuration).

Further, the sensor nodes disclosed in the present specification include a sensor, an artificial intelligence chip having the fifth to seventh configurations for receiving the input data from the sensor, and the artificial intelligence chip and the server. It is configured to have a communication unit for wireless communication between the two (eighth configuration).

In the sensor node having the eighth configuration, the sensor may be a vibration sensor (nineth configuration).

Further, the sensor node having the eighth or ninth configuration includes an environmental power generation unit, a power storage unit that stores the generated power of the environmental power generation unit, and each sensor node unit using the generated power or the stored power of the power storage unit. It is preferable to have a configuration (tenth configuration) further including a power management unit that supplies electric power to the power supply.

Further, the machine health monitoring system disclosed in the present specification has a configuration (No. 1) including a sensor node having the tenth configuration attached to the monitored device and a server that receives an abnormality flag from the sensor node. 11 configurations).

In the machine health monitoring system having the eleventh configuration, the server may be configured to receive the abnormality flag from the sensor node and notify the abnormality state (the twelfth configuration).

According to the invention disclosed in the present specification, it is possible to provide an easy-to-introduce machine health monitoring system, and a sensor node, an artificial intelligence chip, and an artificial intelligence algorithm used for the machine health monitoring system.

The figure which shows the 1st example (TBM) of the equipment maintenance method The figure which shows the 2nd example (CBM) of the equipment maintenance method Diagram showing problems when adopting Wi-SUN Diagram showing the outline of the artificial intelligence chip The figure which shows an example of the monitored device Diagram showing the inverse proportion of anomaly detection operation by old and new artificial intelligence algorithms A diagram showing an example of a configuration of an artificial intelligence chip Diagram showing a configuration example of the kernel arithmetic processing unit The figure which shows one configuration example of a vector arithmetic unit Diagram showing a configuration example of a machine health monitoring system

<Applicable target and system specifications>
First, the background of the artificial intelligence chip disclosed in the present specification (hereinafter, appropriately referred to as AI-SNP [artificial intelligence --sensor node processor]) will be described. The motivation for developing AI-SNP comes from the smart factory strategy as advocated in Industry 4.0. In these next-generation factories, equipment and facilities must be checked in real time all day long without stopping their operation. However, on the other hand, in order to maintain the high profit of the factory, it is necessary to consider cost reduction.

1 and 2 are diagrams showing an example of equipment maintenance method in a factory, respectively. Note that FIG. 1 shows an example of adoption of the TBM [time-based-maintenance] method (= corresponding to the old factory). In the TBM method, the repair or replacement of the device is performed when the operating time of the device meets a predetermined standard. On the other hand, FIG. 2 shows an example of adoption of the CBM [condition-based-maintenance] method (= corresponding to a smart factory). In the CBM method, the repair or replacement of the device is performed when the condition of the device meets a predetermined standard.

In the TBM method of FIG. 1, it is necessary to maintain the equipment of the factory 200 on a regular basis, and if the equipment still fails, a staff member is dispatched from the headquarters 100 to the factory 200 after the fact. Therefore, the operation will be stopped until the repair or replacement of the device is completed. On the other hand, in the case of the CBM method of FIG. 2, when a sign of failure appears in the device of the factory 200, the factory 200 reports the abnormal state to the headquarters 100. Therefore, the staff can repair the device before it shuts down.

As described above, the CBM method (FIG. 2) is one of the best solutions that are in line with the above smart factory strategy as compared with the TBM method (FIG. 1). If the CBM method is adopted, regular maintenance of the device becomes unnecessary, and it is sufficient to check the state of the device only when an abnormal state (= sign of failure) of the device is detected.

In order to introduce a machine health monitoring system that uses the CBM method to an existing factory, it is necessary to add a sensor node for each device or replace the old device with a new device at a high cost. Must be. Nowadays, there are many devices in operation all over the world. Therefore, it is best if sensor nodes can be added to a running device without the need for additional wiring or power. However, the existing wireless module does not necessarily have a sufficient communication bandwidth as compared with the size of the measurement data (raw data) obtained by the sensor node.

FIG. 3 is a diagram showing problems when a Wi-SUN [wireless smart utility network] is adopted as a wireless module of a sensor node. Wi-SUN is one of the best solutions for realizing IIoT [industorial internet of things], and it is possible to perform good power saving and long-distance communication.

However, for example, when the vibration of the monitored device is detected by the sensor node and the vibration data (up to 10 kHz, 16 bits) is transmitted as raw data from the factory 200 to the headquarters 100, the amount of the transmitted data is 3 axes (3 axes). It is 60 kB / s on the X-axis, Y-axis, and Z-axis). This is a considerably large value as compared with the communication bandwidth of Wi-SUN (100 kbps → 12.5 kB / s). In this way, if an attempt is made to transmit the vibration data as raw data, the communication speed will be too slow with respect to the amount of transmitted data, and the power consumption of the Wi-SUN will also increase (about 180 mW). Therefore, data compression (= reduction of the amount of transmitted data) is required.

Further, for example, in order to perform real-time sensing by a sensor node without using a battery as in EN-OCEAN, the power consumption of the wireless module must be dramatically reduced. The AI-SNP disclosed herein can be used to provide a very useful method of deploying a machine health monitoring system that does not require additional wiring or batteries, and thus can be used. It will be possible to dramatically accelerate the IIoT of old and new devices all over the world.

<AI-SNP>
In the following, the AI-SNP for the machine health monitoring system and the artificial intelligence algorithm implemented in the AI-SNP (hereinafter referred to as the AI algorithm) will be described. The applicant of the present application is continuing research on AI algorithms for machine health monitoring systems using open data and real data of his own factory. In addition, through this research, the applicant of the present application has now gained a great deal of knowledge not only about the AI algorithm but also about the know-how of data collection.

As for the AI algorithm, a one-class support vector machine (hereinafter referred to as OCSVM [one class support vector machine]), a spiking neural network, a convolutional neural network, etc. are known, and the options thereof are limited. It's not something. In the following, based on the results of long-term research by the applicant of the present application, an example in which OCSVM is adopted as a candidate for an AI algorithm suitable for implementation in AI-SNP will be described. The details will be described later, and only the outline will be described here.

FIG. 4 is a diagram showing an outline of an artificial intelligence chip (= corresponding to the AI-SNP described above) mounted on a sensor node for a machine health monitoring system. The artificial intelligence chip 10 (hereinafter referred to as AI chip 10) of this configuration example is a semiconductor chip capable of executing an AI algorithm based on input data (vibration data, temperature data, etc.) from a sensor. It also has a pre-processing unit 11, a classifier 12, and a post-processing unit 13. The AI chip 10 is required to have low power consumption and a small area so that it can be mounted on a sensor node. Further, the AI algorithm described above is designed to realize functions such as abnormality diagnosis, toolware evaluation, and life prediction of the monitored device.

The pre-processor 11 extracts features for each frequency band from input data (for example, raw vibration data obtained by a vibration sensor) using a plurality of bandpass filters connected in parallel. By doing so, a feature vector is generated. As the above-mentioned feature quantity, for example, from 1 Hz to 20 kHz (200 dim), after acquiring the FFT [fast Fourier transform] amplitude of the frequency spectrum every 50 Hz, the root mean square of each is obtained. ) Can be calculated.

The classifier 12 uses a kernel function (for example, a linear kernel, a Gaussian kernel, or an RBF [radial base function]] using a feature vector input from the preprocessing unit 11 and a support vector stored in advance. Find the value of the kernel). As the classifier 12, for example, it is desirable to preferably use an OCSVM configured by hardware. Further, as the support vector, for example, input data less than 33 hours after the monitored device starts operation may be used for learning to create a hyperplane for OK / NG discrimination.

The post-processor 13 calculates an abnormality detection process of input data (= an abnormality flag indicating whether the monitored device is normal or abnormal) from the value of the kernel function obtained by the classifier 12. Processing).

As described above, in the AI chip 10 of this configuration example, the AI algorithm described above is implemented by using the pre-processing unit 11, the classifier 12, and the post-processing unit 13.

The OCSVM used as the classifier 12 is a kind of support vector machine, which is a lightweight and practical AI algorithm by unsupervised learning. OCSVM detects one-class problems on one piece of software and is similar and non-existent to a machine health monitoring system. This is because OCSVM itself has a poor affinity with complex and high-speed time-series data.

In recent years, studies have been conducted that enable the application of OCSVM by applying appropriate preprocessing to time-series data (see, for example, Non-Patent Documents 1 and 2). However, all studies require complicated preprocessing to extract features from time-series data, and there is room for further improvement in mounting on sensor nodes that require low power consumption and area saving. ..

Therefore, in the present specification, we propose an AI chip 10 to which OCSVM can be applied by performing a simple pretreatment as compared with the conventional one. The preprocessing unit 11 simply processes the raw input data using only the FFT operation in order to obtain the feature amount (= RMS of the frequency spectrum obtained every 50 Hz from 1 Hz to 20 kHz) regarding the frequency of the input data. Of course, it is difficult to perform the FFT calculation with a power and chip size reasonable for the AI chip 10. Therefore, in the preprocessing unit 11, a simple analog bandpass filter is used (details will be described later).

Also, paying attention to the classifier 12, it is a simple OCSVM using the kernel method. As the kernel function, as described above, a linear kernel, a Gaussian kernel, an RBF kernel, or the like can be selected. For example, if the RBF kernel can be selected, the AI algorithm by OCSVM will be powerful. However, for that purpose, an additional function circuit (logarithmic operation circuit) is required, and it is necessary to pay attention to this point.

If we focus only on the analysis, neglect, and learning of abnormal conditions, we should pay attention only to the adder, the multiplier, and the implementation of the simple kernel function f (X) shown in the following equation (1). Just do it.

Therefore, in the present specification, an example of configuring a classifier 12 (particularly its kernel arithmetic processing unit) by using an adder and a multiplier will be described. In addition, as a device to implement the above AI algorithm on the AI chip 10, many small-scale analog PEs are used to reduce AD / DA [analog-to-digital / digital-to-analog] and eliminate high-speed clocks. It is desirable to arrange a [processing engine] to make the arithmetic unit and memory an analog structure.

Next, an example using open data to detect an abnormality occurring in a bearing of a milling machine by the AI algorithm proposed by the applicant of the present application will be described.

FIG. 5 is a diagram showing an example of a monitored device (here, a milling machine). The milling machine 210 in this figure has a motor 211 and bearings 212 to 215. An accelerometer 216 and a thermocouple 217 are attached to the bearings 212 to 215, respectively. Further, the sensor node (not shown) equipped with the AI chip 10 described above constantly measures the vibration generated during the operation of the milling machine 210 to determine whether or not an abnormality has occurred in the bearing.

FIG. 6 is a diagram showing the inverse proportion of the abnormality detection operation by the old and new AI algorithms (indicated as “light” algorithm and “heavy” algorithm, respectively in the figure). On the left side of this figure, the waveforms of the vibration data obtained by the sensor node (after 1 hour and after 88 hours) are depicted. On the other hand, in the upper right part of this figure, a time chart showing the abnormality detection operation by the new AI algorithm is drawn. The vertical axis shows the distance from the hyperplane, and the horizontal axis shows the operating time. Further, in the lower right part of this figure, a time chart showing an abnormality detection operation by the old AI algorithm is drawn. The vertical axis shows the class output (0 = normal, 1 = abnormal), and the horizontal axis shows the operating time.

The various parameters of the new AI algorithm are as follows.
・ NASA IMS data ch1 (test2)
・ Using OCSVM ・ Kernel: Gauss ・ Gamma = 0.1
-Learning data: The first 200 data (33 hours of non-defective data is used for learning)
-Number of support vectors: 52
-Vector dimension: 99 dimensions-Feature quantity: RMS of FFT amplitude obtained every 50 Hz from 0 to 10 kHz

The various parameters of the old AI algorithm are as follows.
-NASA IMS data-AlexNet used (11-layer CNN)
・ Learning data Normal = 80 data (random selection up to 88h)
Abnormality = 80 data (random selection after 88h)
・ Mini batch size: 50
・ Number of learning: 1500

As shown in this figure, the waveforms of the vibration data obtained by the sensor node are almost the same in appearance after 1 hour and 88 hours. However, if the abnormality determination operation is performed by the new AI algorithm for such vibration data, the abnormality can be detected when the operating time of the milling machine 210 reaches 88 hours. Considering that the failure actually occurred when the operating time of the milling machine 210 reached 162 hours, it can be said that the above-mentioned abnormality determination operation is extremely useful for knowing the sign of the failure.

On the other hand, even in the abnormality determination operation by the old AI algorithm, the abnormality is detected for the first time 88 hours after the start of operation, and the abnormality is completely detected after 91 hours have passed.

In this way, as is clear from a comparison of the two, by adopting the new AI algorithm, the operational load is reduced compared to the old AI algorithm, and the abnormality of the monitored device is similar to or higher than this. Can be detected. Therefore, it can be said that the new AI algorithm is suitable for mounting on a sensor node that requires low power consumption and area saving.

<1 class (u) -SVM using bandpass filter algorithm>
When performing anomaly detection operations for machine health monitoring systems, (a) almost all are OK data and NG data is extremely rare, and (b) vibration data is useful for anomaly detection of monitored devices. Careful attention should be paid to the characteristic of presenting information. In view of the above features, we propose an unsupervised machine learning method and a simple preprocessing method.

One class (u) -SVM (OCSVM) using a bandpass filter (BPF) can be adopted as an AI algorithm for performing an abnormality detection operation in view of the above characteristics. OCSVM is known as an unsupervised anomaly detection technique. Note that (u) is introduced to clarify the upper limit of the margin error and the lower limit of the ratio of the support vector to all the vectors. Further, the BPF is a means for simply obtaining frequency information without performing an FFT calculation, which is a heavy load for the AI chip 10.

FIG. 7 is a diagram showing an example of the configuration of the AI chip 10 (configuration without the post-processing unit 13). As shown in this figure, in the AI chip 10 of this configuration example, the preprocessing unit 11 includes a plurality of bandpass filters 11a connected in parallel. Further, the classifier 12 includes a support vector storage unit 12a and a kernel arithmetic processing unit 12b.

A plurality of preprocessing units 11 are provided for each predetermined bandwidth Δf (Δf = 50 Hz in this figure) from the lowest frequency fmin (= DC) of the input data (= raw data) to the highest frequency fmax (= 1 kHz). The bandpass filter 11a of the above is operated. Each of these bandpass filters 11a outputs feature values x ₀ to x _k . The preprocessing unit 11 generates a feature vector X (x 0, x 1, ..., X k) having feature values x ₀ to x _k as elements and outputs the feature vector X (x ₀ , x ₁ , ..., X _k ) to the classifier 12.

In the classifier 12 (OCSVM), the support vector storage unit 12a supports the support represented by the following equation (2) in order to form a hyperplane (= OK / NG determination boundary) for evaluating the feature vector X. It stores the vector X _n .

The above support vector Xn is a server having a higher arithmetic processing capacity than the AI chip 10 based on the input data (= raw data of non-defective products) obtained within a predetermined period after the monitored device is started for the first time. It is desirable to calculate.

In that case, the AI chip 10 sends the input data (= raw data of good products) necessary for the support vector calculation process to the server before starting the normal operation (= abnormality determination operation of the input data), and the server supports it. The vector X _n must be received.

Further, the kernel arithmetic processing unit 12b accepts the inputs of the feature vector X and the support vector X _n , and obtains the function value f (X) using the predetermined kernel function K (see the above equation (1)). ). Then, after the calculation, the AI chip 10 outputs the above-mentioned function value f (X) (= 0 to 1) as the evaluation result of the input data.

When the post-processing unit 13 (not shown in this figure) is provided, the above-mentioned function value f (X) is checked, and it is determined whether or not the abnormal state should be notified to the server.

FIG. 8 is a diagram showing a configuration example of the kernel arithmetic processing unit 12b. The kernel arithmetic processing unit 12b of this configuration example includes a plurality of vector arithmetic units b10 and an adder b20.

The vector arithmetic unit b10 generates a function value α _i K (X, X _i ) using the feature vector X, the support vector X _i , and the coefficient α _i , respectively (where i = 0, 1, ..., k).

The adder b20 generates a function value f (X) by adding a plurality of function values α _i K (X, X _i ).

FIG. 9 is a diagram showing a configuration example of the vector arithmetic unit b10. The vector arithmetic unit b10 of this configuration example includes a kernel arithmetic unit b11 and a multiplier b12.

The kernel arithmetic unit b11 accepts inputs of the feature vector X and the support vector X _i , and generates a function value K (X, X _i ) using a predetermined kernel function K.

For the kernel function K, there are several options such as a linear kernel, a Gaussian kernel, or an RBF kernel. For example, the linear kernel simply multiplies the feature vector X and the support vector Xi, and can be mounted on the AI chip 10 with an extremely simple hardware configuration. On the other hand, the RBF kernel can freely determine the boundary of OK / NG determination. The following equations (3a) to (3c) are arithmetic expressions of a linear kernel, a Gaussian kernel, and an RBF kernel, respectively.

The multiplier b12 generates the function value α _i K (X, X _i ) by multiplying the function value K (X, X _i ) by the coefficient α _i .

<Sensor node>
FIG. 10 is a diagram showing a configuration example of a sensor node on which the AI chip 10 is mounted. The sensor node 1 of this configuration example functions as one component of the machine health monitoring system 300 together with the server 2. In addition to the AI chip 10 described so far, the sensor 20 and the communication unit 30 are used. It has an energy harvesting unit 40, a power storage unit 50, and a power management unit 60.

The sensor node 1 may be understood as the sensor itself attached to the monitored device (not shown) or as a gateway connected to the sensor. That is, the sensor 20 may be externally attached to the sensor node 1.

The AI chip 10 is a semiconductor device that receives power from the power management unit 60 and operates at the edge of the sensor 20. It receives input data from the sensor 20 and performs abnormality detection processing, and the detection result is transmitted via the communication unit 30. Report to server 2. The communication between the AI chip 10 and the communication unit 30 may be performed, for example, via the UART [universal asynchronous receiver / transmitter] interface. Since the configuration and operation of the AI chip 10 have already been described, duplicate explanations here are omitted.

The sensor 20 is a means for measuring a predetermined measurement target (vibration, current, etc.) by receiving power supply from the power management unit 60, and for example, a vibration sensor can be preferably used. In order to eliminate the high-speed clock of the AI chip 10, it is desirable that the sensor 20 is an analog output type. Most of the most advanced vibration sensors are logic output type sensors equipped with a logic interface. This is because the analog output type is easily affected by noise and is fatal to a high-precision sensor. Therefore, when using the analog output type sensor 20, it is important to arrange the AI chip 10 in the vicinity of the sensor 20 so as not to be easily affected by noise.

The communication unit 30 is a module for receiving power supply from the power management unit 60 and performing wireless communication with the server 2. Normally, the AI chip 10 communicates with the server 2 only when error data is detected, but when sending learning data to the server 2, it is necessary to perform communication with a larger capacity than usual. In view of this, it can be said that it is desirable to adopt, for example, a Wi-SUN module capable of high-speed wireless communication as the communication unit 30.

The energy harvesting unit 40 is a means (so-called energy harvester) that receives energy (= vibration, light, heat, etc.) existing in the environment where the sensor node 1 is placed to generate power. When vibration is used as an energy source, a piezoelectric element such as a piezo element may be used as the power generation element. When sunlight or illumination light is used as an energy source, a silicon-based, compound-based, or organic-based photoelectric element may be used as the power generation element. When heat is used as an energy source, a thermoelectric element such as a Pelche element may be used as the power generation element.

In the sensor node 1, it is preferable that the measurement target of the sensor 20 and the energy source of the energy harvesting unit 40 are common. As one example, there is a case where the vibration is measured by the sensor 20 and the above vibration is used as an energy source by the energy harvesting unit 40. In this case, when the sensor 20 tries to measure the vibration, the energy harvesting unit 40 receives the vibration and generates electric power. Therefore, the electric power to the sensor 20 is more reliable than the case where the energy source is other than the vibration. It becomes possible to supply.

The power storage unit 50 is a means for storing the power generated by the energy harvesting unit 40 (about 100 μW), and for example, a supercapacitor (= a general term for electric double layer capacitors) can be preferably used. In particular, in order to consume a large amount of power and send a large amount of data to the server 2, a supercapacitor having a large capacity (about 1F) is required as the power storage unit 50.

The power management unit 60 uses the power generated by the energy harvesting unit 40 or the stored power of the power storage unit 50 to supply power to each unit of the sensor node 1 (AI chip 10, sensor 20, and communication unit 30). It is an internal power supply circuit (for example, a DC / DC converter with a DC 3.3V output). The energy harvesting unit 40 cannot stably supply the generated power. Therefore, in order to realize the stable operation of the sensor node 1, the function of the power management unit 60 is very important. That is, the power management unit 60 needs to perform not only storage control to the power storage unit 50 but also appropriate impedance matching control so that the maximum power can be obtained from the energy harvesting unit 40.

In the case of the sensor node 1 of this configuration example, since the power consumption is covered by the energy harvesting, it is not necessary to lay the power supply wiring or replace the battery. Further, since wireless communication is performed between the sensor node 1 and the server 2, signal wiring connecting the two is not required. Therefore, the sensor node 1 can be arranged at an arbitrary position.

When the server 2 receives the abnormality flag from the sensor node 1, the server 2 notifies the staff of the headquarters of the abnormality state. By constructing such a machine health monitoring system 300, equipment maintenance by the CBM method (FIG. 2) becomes possible.

<Other variants>
In the above embodiment, the machine health monitoring system is taken as an example, but the application target of the artificial intelligence algorithm (or the artificial intelligence chip on which the artificial intelligence algorithm is implemented) is not limited to this, and for example, the patient. It can also be applied to a biological health monitoring system for managing physical condition.

As described above, various technical features disclosed in the present specification can be modified in addition to the above-described embodiment without departing from the spirit of the technical creation thereof. That is, it should be considered that the above-described embodiment is exemplary in all respects and is not restrictive, and the technical scope of the present invention is not limited to the above-described embodiment and claims for a patent. It should be understood that the meaning of the scope and equality and all changes belonging to the scope are included.

The invention disclosed herein can be used, for example, in a machine health monitoring system for smart factories.

1 Sensor node 10 Artificial intelligence chip (AI chip)
11 Pre-processing unit 11a Bandpass filter 12 Classifier (OCSVM)
12a Support vector storage unit 12b Kernel arithmetic processing unit b10 Vector arithmetic processing unit b11 Kernel arithmetic unit b12 Multiplier b20 Adder 13 Post-processing unit 20 Sensor 30 Communication unit (Wi-SUN)
40 Energy harvesting unit 50 Power storage unit (supercapacitor)
60 Power Management Department 100 Headquarters 200 Factory 210 Milling Machine 211 Motor 212-215 Bearing 216 Accelerometer 217 Thermocouple 300 Machine Health Monitoring System

Claims (7)

Vibration sensor and
An artificial intelligence chip that accepts raw vibration data obtained by the vibration sensor,
A communication unit that performs wireless communication between the artificial intelligence chip and the server,
Is a sensor node with
The artificial intelligence chip is
Using k bandpass filters connected in parallel, the FFT amplitude of the frequency spectrum is acquired for each fixed frequency band from the lowest frequency to the highest frequency of the vibration data, and then the root mean square of each is calculated. , A preprocessing unit that generates a feature vector X (x ₀ , x ₁ , ..., X _k ) by extracting feature quantities for each frequency band from the vibration data .
The feature vector X (x ₀ , x ₁ , ..., X _k ) and the vibration data within a predetermined time from the start of operation of the monitored device to which the sensor node is attached are used for learning. The input of the support vector Xi (however, _i = 0,1, ..., K) corresponding to the hyperplane for OK / NG discrimination is accepted, and the function value f (X) is used by using the predetermined kernel function K. With a classifier that finds
Have,
The classifier is
A kernel arithmetic unit that accepts inputs of the feature vector X and the support vector Xi and generates k function values K (X, X _i ) using the kernel function K.
A multiplier that generates k function values αiK (X, Xi) by multiplying k function values K (X, Xi) by _k coefficients αi.
An adder that generates the function value f (X) by adding k of the function values αiK (X, Xi), and
Including sensor nodes . The sensor node according to claim 1 , further comprising a post-processing unit that detects an abnormality in the vibration data from the function value f (X). The sensor node according to claim 1 or 2 , wherein the classifier is an OCSVM [one class support vector machine] configured by hardware. The sensor node according to any one of claims 1 to 3, wherein the kernel function K is a linear kernel, a Gaussian kernel, or an RBF [radial base function] kernel. Energy harvesting department and
The power storage unit that stores the generated power of the energy harvesting unit and
A power management unit that supplies power to each sensor node using the generated power or the stored power of the power storage unit.
The sensor node according to any one of claims 1 to 4 , further comprising. The sensor node according to any one of claims 1 to 5, which is attached to the monitored device.
A server that accepts an error flag from the sensor node and
Has a machine health monitoring system. The machine health monitoring system according to claim 6 , wherein the server receives the abnormality flag from the sensor node and notifies the abnormality state.

Images