Operating state classification system and operating state classification method

Abstract

The present application realizes an operation state classification system whose classification accuracy is continuously improved in additional learning and which also appropriately prevents illegal copying of the classification function. An operation state classification system connects an edge device and a server through a communication network, inputs sensor data to a neural network, and outputs a state label, the edge device including: a first holding unit that holds a front stage side of the neural network; and a dimension compression unit that inputs the sensor data to the front stage side of the neural network and outputs intermediate data, the server including: a second holding unit that holds a rear stage side of the neural network; an inference execution unit that inputs the intermediate data to the rear stage side of the neural network and outputs the state label; and a learning unit that updates the rear stage side of the neural network through additional learning.

Description

TECHNICAL FIELD

[0001] The present application relates to an operation state classification system and an operation state classification method that classify an operation state of an edge device or the like. BACKGROUND

[0002] One of the problems in detecting abnormalities of factory equipment or the like is the setting of appropriate determination criteria. For example, when diagnosing a machine tool, the load on the driving motor differs depending on the type of process (for example, the load is large during machining, and the load is small during idling), and if different determination criteria are not set for each process, many false positives and false negatives will occur. Or, in order to improve the accuracy of abnormality detection, it can be considered to use only data of a specific process, or to use only data of a period in which the load varies little. Therefore, there is a need for an operation state classification system that can classify the operation state of a target device before abnormality detection.

[0003] On the other hand, recently, it has become possible to equip edge devices such as power conversion devices and industrial controllers with a reasoning execution function (neural network) of deep learning. Therefore, with this function, if an operation state classification function is mounted on an edge device, the operation state can be determined by the edge device, and the data used for abnormality detection can be transmitted to a server, which not only reduces the amount of communication data, but also enables more accurate abnormality detection.

[0004] A conceptual diagram of such a prior art is shown in Figure 15 First, in the edge device, data from various sensors is saved in association with state labels L (for example, L1, L2, L3) that represent each operation state. Then it is recovered in some way into a learning server, and a neural network is learned using the learning server. Then the resulting completed learning neural network is installed on the edge device. In this way, the edge device alone can achieve automation of state classification. After that, processing is performed according to the classification result, such as diagnosis using only data of a specific operation state.

[0005] However, current edge devices, particularly various control devices such as power conversion devices or industrial controllers, rather than devices such as industrial PCs that assume edge computing, are often difficult to perform learning of a neural network due to low performance of the microcomputer mounted and small storage space.

[0006] Therefore, in the case of an operation state classification system based on the prior art as described above, there is a problem that the function is fixed at the time of actual use start, and it is difficult to improve the classification accuracy through actual use. Also, the edge device itself can shut down the function, and therefore, the neural network function can be illegally copied by monitoring the input and output.

[0007] Therefore, a technique has been proposed in which a neural network is split and loaded on an edge device and a server, so as to make full use of the computational power of both the edge device and the server. For example, in Patent Literature 1, a technique is disclosed in which a process of a first half (from an input layer to a first half of an intermediate layer) of a neural network is executed on a smartphone, and a result thereof is output to a server, and a process of a second half (from a second half of the intermediate layer to an output layer) is executed by the server. Not only inference execution, but also learning is cooperatively executed by the smartphone and the server.

[0008] Prior Art Documents

[0009] Patent Literature

[0010] Patent Literature 1: International Publication No. WO2018 / 011842 SUMMARY

[0011] PROBLEMS TO BE SOLVED BY THE INVENTION

[0012] However, in various control devices such as power conversion devices and industrial controllers, the performance and storage capacity of a microcomputer are not as high as those of a smartphone, and thus there is a problem in that it is difficult to execute learning of a neural network using the technique disclosed in Patent Literature 1.

[0013] The present application has been achieved in order to solve the problems of the related art described above, and has an object to provide a running state classification system that can reduce the burden on an edge device side while continuously improving the accuracy of running state classification through additional learning when the edge device and a server cooperate.

[0014] TECHNICAL SOLUTION

[0015] In order to solve the above problems, a running state classification system of the present application connects an edge device and a server through a communication network, inputs sensor data to a neural network, and outputs a state label, the edge device including: a first holding unit that holds a front stage side of the neural network; and a dimension compression unit that inputs the sensor data to the front stage side of the neural network and outputs intermediate data, the server including: a second holding unit that holds a rear stage side of the neural network; an inference execution unit that inputs the intermediate data to the rear stage side of the neural network and outputs the state label; and a learning unit that updates the rear stage side of the neural network through additional learning.

[0016] Further, the operation state classification method of the present application includes: a learning step of learning, by a server, a neural network that inputs sensor data and outputs a state label; a first saving step of saving a front stage side of the neural network in an edge device; a second saving step of storing a back stage side of the neural network in the server; and an additional learning step of additionally learning the back stage side of the neural network by the server.

[0017] Effects of the Invention

[0018] According to the operation state classification system and the operation state classification method of the present application, even if the neural network installed in the edge device does not evolve, the neural network on the server side evolves as the edge device is continuously used, so that the classification accuracy as a whole system can be improved. Thus, actual use can be started without waiting for learning data to be sufficiently accumulated. Further, since the data for communication is dimensionally compressed, not only can the communication cost be reduced, but also even if the communication is stolen, it is not easily misused. Also, since the server cannot be connected to function, it is possible to suppress illegal copying. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a conceptual diagram at the time of learning in the operation state classification system of Example 1.

[0020] Figure 2 is a flowchart at the time of learning in the operation state classification system of Example 1.

[0021] Figure 3 is a conceptual diagram at the time of inference execution in the operation state classification system of Example 1.

[0022] Figure 4 is a flowchart at the time of inference execution in the operation state classification system of Example 1.

[0023] Figure 5 is a conceptual diagram at the time of additional learning in the operation state classification system of Example 1.

[0024] Figure 6 is a flowchart at the time of additional learning in the operation state classification system of Example 1.

[0025] Figure 7 is a functional structure diagram of the edge device and the server in the operation state classification system of Example 1.

[0026] Figure 8 is a conceptual diagram at the time of inference execution in the operation state classification system of Example 2.

[0027] Figure 9 is a conceptual diagram at the time of learning in the operation state classification system of Example 3.

[0028] Figure 10 is a flowchart at the time of learning in the operation state classification system of Example 3.

[0029] Figure 11 is a conceptual diagram at the time of inference execution in the operation state classification system of Example 3.

[0030] Figure 12 is a flowchart at the time of inference execution in the operation state classification system of Example 3.

[0031] Figure 13 is a basic configuration diagram in the case where the function on the edge device side that constitutes the operation state classification system to which the present application relates is built into an industrial controller.

[0032] Figure 14 is a basic configuration diagram in the case where the function on the edge device side that constitutes the operation state classification system to which the present application relates is built into a power conversion device.

[0033] Figure 15 is a conceptual diagram of a related art relating to the operation state classification system. DETAILED DESCRIPTION

[0034] Next, an embodiment of the present application will be described using the drawings.

[0035] Example 1

[0036] Referring to Figures 1 to 7 , an operation state classification system 100 relating to Example 1 of the present application will be described. The operation state classification system 100 of the present embodiment is a system that classifies the operation state of an edge device 2 using a neural network, and the server 1 and the edge device 2 are connected through a communication network. In addition, in the present embodiment, an autoencoder is cited as one example of the neural network, but the idea of the present application can also be applied to a system that uses other types of neural networks.

[0037] The server 1 is a computer including a computing device such as a CPU, a main storage device such as a semiconductor memory, an auxiliary storage device such as a hard disk, and a communication device and the like. Further, the edge device 2 is a power conversion device or an industrial controller or the like in which a microcomputer is built in. Such a server 1 or edge device 2 is able to realize each function described later by executing a program loaded in the main storage device by the computing device. In addition, it is assumed that the microcomputer of the edge device 2 has a computing capacity to use an installed neural network, but does not have a computing capacity to make the installed neural network evolve on its own.

[0038] Next, the details of the operation state classification system 100 of the present embodiment will be described in each of the "learning processing" implemented before actual use of the edge device 2, the "inference processing" implemented in the actual use of the edge device 2, and the "additional learning processing" for improving the neural network on the server 1 side, which is executed after the use of the edge device 2 for a predetermined period or more or a predetermined number of times.

[0039] <Learning Processing>

[0040] Figure 1 is a conceptual diagram of the learning processing of the operation state classification system 100 implemented before actual use of the edge device 2, Figure 2 is a flowchart showing the flow of the learning processing.

[0041] First, in step S10, the edge device 2 associates the sensor data detected by various sensors with the state label L (for example, L1, L2, L3) indicating the kind of the operation state input by the system designer or the like, and saves it in the storage device. In addition, although the operation state is classified into three kinds here, the kind of the operation state can be any integer of two or more.

[0042] Next, in step S11, the server 1 learns a neural network (hereinafter referred to as an autoencoder a) that reproduces itself using the sensor data of each operation state and the state label L saved in the edge device 2. For example, the autoencoder a1 learned by the sensor data of the state label L1 outputs the sensor data of the state label L1 when the sensor data of the state label L1 is input, but outputs different sensor data from the input when other sensor data is input. Therefore, the front stage portion (from the input layer to the front half of the intermediate layer) of the autoencoder a is a neural network (hereinafter referred to as an encoder A) that extracts the feature amount d of the sensor data for learning. In addition, the rear stage portion (from the rear half of the intermediate layer to the output layer) of the autoencoder a is a neural network (hereinafter referred to as a decoder A') that reproduces the sensor data based on the feature amount d.

[0043] Next, in step S12, the server 1 copies the encoder A obtained in step S11, and learns a neural network (hereinafter referred to as a state classifier B) that inputs the feature amount d extracted by the encoder A and outputs the state label L. In addition, since the encoder A extracts the feature amount d, the learning of the state classifier B is easy.

[0044] Finally, in step S13, the server 1 installs only the completed encoder A on the edge device 2. In addition, the encoder A installed on the edge device 2 is a part of the neural network of the autoencoder a, and thus has a small operation load at the time of execution and a small storage capacity. Therefore, even a microcomputer with relatively low performance built in the edge device 2 can perform the desired processing without delay.

[0045] <Inference processing>

[0046] Figure 3 is a conceptual diagram of inference processing of the operation state classification system 100 implemented in actual use of the edge device 2, Figure 4 is a flowchart showing the flow of the inference processing.

[0047] In step S20, the edge device 2 inputs sensor data collected by various sensors in actual use to the encoders A1 to A3 for each state, and extracts feature amounts d1 to d3 for each state.

[0048] In step S21, the edge device 2 transmits the obtained feature amounts d to the server 1. Since the feature amounts d are dimensionally compressed by the encoders, the data size is small while the features of the sensor data for each operation state are included. Therefore, it is possible to suppress the communication cost from the edge device 2 to the server 1.

[0049] In step S22, the server 1 inputs the received feature amounts d to the state classifier B, classifies the operation state, and generates a state label L.

[0050] In step S23, the server 1 transmits the generated state label L to the edge device 2.

[0051] Finally, in step S24, the edge device 2 changes the processing method after that according to the type of the input state label L. For example, it is possible to consider displaying the classification result, performing diagnosis only in a specific state, saving detailed data, and the like. In a case where it is necessary to immediately reflect the change in the operation state in the control of the edge device 2, it is desirable to use a high-speed, low-latency communication network as the communication network connecting the server 1 and the edge device 2.

[0052] In addition, the feature amounts d transmitted from the edge device 2 to the server 1 are intermediate data output by the encoders A as the front-stage part of the autoencoder a, and are meaningless data if the rear-stage part of the autoencoder a, that is, the decoder A' and the state classifier B, cannot be used, and even if the communication from the edge device 2 to the server 1 is intercepted, it is possible to maintain the confidentiality of the communication content.

[0053] <Additional learning processing>

[0054] Figure 5is a conceptual diagram of an additional learning process of the operation state classification system 100 performed after use of the edge device 2 for more than a prescribed period or more than a prescribed number of times, Figure 6 is a flowchart showing the flow of the additional learning process. Note that the additional learning process is a learning process for improving the state classifier B on the server 1 side, and the encoder A on the edge device 2 side continues to be used as is.

[0055] In step S30, the edge device 2 inputs sensor data collected by various sensors in actual use to the encoders Al to A3 for each state, and extracts feature amounts dl to d3 for each state.

[0056] In step S31, the correct state label L is associated with intermediate data identical to that used in inference execution, and stored in the server 1 as additional learning data. The addition of the correct state label L and the correction of the erroneous state label L are basically performed by a system designer or an operator of the edge device 2 or the like.

[0057] In addition, Figure 5 Both the correct state label L input to the server 1 by the system designer or the like and the correct state label L input to the edge device 2 by the operator are shown, and either one of them can be input.

[0058] Alternatively, in order to save the effort of people to label the correct state label, if the probability of being determined as the most reliable state in the classification of the state classifier B is higher than a preset value (for example, 80%), the state label can be automatically added as the correct state label. On the other hand, when the probability is lower than the preset value, that is, as long as the state classifier B cannot judge, the addition of the correct state label by people can be omitted, and the people can be relieved from the trouble of frequently performing the addition of the correct state label. In addition, if people completely stop adding the state label, in the case where the above probability is low, the state label can be added using the closeness to the labeled data as an index, or can not be used for additional learning.

[0059] In step S32, the server 1 performs additional learning at a stage where a certain amount or more of additional learning data is accumulated. This additional learning can be performed in parallel with Figure 3 or Figure 4 inference execution as shown, or can be performed while temporarily stopping inference execution.

[0060] Then, in step S33, the server 1 temporarily stops the inference execution, and replaces the state classifier B with the state classifier B obtained by the additional learning. Thereby, after the use of the edge device 2 for a prescribed period or more or a prescribed number of times or more, the state classifier B of the server 1 is updated to a state classifier B with higher classification accuracy based on human knowledge, and even if the encoder A installed in the edge device 2 does not evolve, the performance of the entire system is improved, and thus the inference processing performed thereafter can generate a more appropriate state label L.

[0061] <Configuration of server 1, edge device 2>

[0062] Figure 7 is a functional block diagram of the server 1 and the edge device 2 in the running state classification system 100 that implements the above-described each processing.

[0063] The running state classification pre- and post-processing section 20 provided in the edge device 2 includes a data input section 21 that inputs sensor data and a state label for additional learning, an encoder storage section 22 that stores the completed learning encoder A, a dimension compression section 23 that converts the sensor data into the feature quantity d using the encoder A, a data communication section 24 that transmits the feature quantity d to the server 1 and receives the state label L from the server 1, and a result response section 25 that displays the state label L, performs diagnosis only in a specific state, and stores the result of detailed data. In addition, as described above, these are implemented by executing a program by a CPU or the like.

[0064] On the other hand, the running state classification execution section 10 provided in the server 1 includes a data communication section 11 that receives the feature quantity d and the state label L for additional learning from the edge device 2 and transmits the classification result of the state classifier B, that is, the state label L, a state classifier storage section 12 that stores the completed learning state classifier B, an inference execution section 13 that performs running state classification using the state classifier B, a data storage section 14 that stores data for additional learning, and a learning section 15 that learns the autoencoder a or the state classifier B using the stored data at the time of learning processing or at the time of additional learning processing. In addition, as described above, these are implemented by executing a program by a CPU or the like.

[0065] With the system configuration as in this embodiment, as long as practical use continues, the data for additional learning is naturally stored in the server 1, and even if the classification accuracy of the running state at the time when the initial learning processing ends is not sufficient, the classification accuracy gradually improves over time. Furthermore, the data transmitted from the edge device 2 to the server 1 is dimensionally compressed, and thus the communication volume can be suppressed. Also, if the edge device 2 is not connected to the server 1 to construct an environment in which the state classifier B can be used, useful information (the state label L) cannot be obtained, and thus, in addition to the small risk of leaking information, even if the encoder A is improperly copied alone, there is little value in using it.

[0066] According to the above-described embodiment, even if the neural network installed in the edge device does not evolve, the neural network on the server side evolves through additional learning, and thus, the longer the running time, the better the performance of the neural network of the entire system. Furthermore, in practical use, the data amount of the feature amount (intermediate data) is small, which is transmitted from the edge device to the server, and thus the communication volume between the two can be suppressed. Also, since it is a system in which the feature amount (intermediate data) cannot be processed unless the edge device and the server are connected, the motivation to illegally copy the neural network on the edge device side and install it on other edge devices can be reduced.

[0067] Embodiment 2

[0068] Next, the inference processing of the running state classification system 100 of Embodiment 2 of the present application will be described with reference to the conceptual diagram of Figure 8 In addition, the repeated description of the common points with Embodiment 1 is omitted.

[0069] As shown in Figure 3 , Figure 4 , the inference processing of Embodiment 1 is a process in which the state classifier B generates the state label L, and the inference processing of this embodiment is characterized in that, in addition to this process, the latter stage portion (the decoder A1’) of the autoencoder obtained as a byproduct at the time of the learning processing of Figure 1 , Figure 2 is utilized to simulate the reproduction of the sensor data based on the feature amount d from the edge device 2.

[0070] For example, if the output of the state classifier B of the server 1 is the state label L1, the feature amount d1 is input to the decoder A1’, if it is the state label L2, the feature amount d2 is input to the decoder A2’, and if it is the state label L3, the feature amount d3 is input to the decoder A3’. In this way, data (hereinafter referred to as reproduced data) close to the sensor data before the dimension compression by the edge device 2 can be decoded. That is, by transmitting only the small data amount of the feature amount d from the edge device 2, the server 1 can simulate the reproduction of the large data amount of the sensor data.

[0071] The reproduction data obtained as described above can be used for a variety of purposes. For example, it can be used for anomaly detection based on a complex machine learning algorithm that is difficult for the edge device to mount, and it can be used for easily relearning an encoder mounted on the edge device.

[0072] Embodiment 3

[0073] Next, the learning process and the inference process of the operation state classification system 100 according to Embodiment 3 of the present application will be described with reference to Figures 9 to 12 The learning process and the inference process of the operation state classification system 100 according to Embodiment 3 of the present application will be described with reference to

[0074] <Learning Process>

[0075] In the learning process of Embodiment 1, as shown in Figure 1 , Figure 2 After learning a plurality of autoencoders a1 to a3, all of the encoders A1 to A3 as the preceding stage part thereof are installed in the edge device 2, whereas in the learning process of the present embodiment, as shown in the conceptual diagram of Figure 9 , the flowchart of Figure 10 , it is characterized by learning only one autoencoder a1 and installing the encoder A1 as the preceding stage part thereof on the edge device 2. Thereby, even in the case where the storage capacity of the edge device 2 is small and all of the encoders A cannot be stored, installation is possible.

[0076] Further, in the learning process of Embodiment 1, as shown in Figure 1 , Figure 2 The state classifier B that generates the state labels L1 to L3 based on the outputs, i.e., the feature amounts d1 to d3, of the encoders A1 to A3 has already been learned, but as shown in Figure 9 , Figure 10 The learning process of the present embodiment is characterized by learning the state classifier B that generates the state labels L1 to L3 based on only the feature amount d1 as the output of the encoder A1.

[0077] For example, in a case where the occurrence frequency of the operation state corresponding to the state label L1 is much higher than those of the other operation states (for example, in a case where the state label L1 corresponds to a normal state, and the state labels L2 and L3 correspond to abnormal states, respectively), at the start of actual use, it can be impossible to sufficiently collect abnormal data corresponding to the state label L2 and the state label L3. In this case, learning of the autoencoders a2 and a3 corresponding to the state labels L2 and L3 is not performed, and only learning of the autoencoder a1 corresponding to the state label L1 is performed. In this case, by taking measures such as allocating more time to learning of the state classifier B, increasing the number of layers of the state classifier B, and the like, based on the output of the autoencoder a1, that is, the feature quantity d1, it is possible to generate not only the state label L1 corresponding to the normal state but also the state labels L2 and L3 corresponding to the abnormal states.

[0078] Therefore, in step S10 of Figure 10 , the edge device 2 saves the sensor data from the various sensors in association with the state labels (for example, L1, L2, and L3) representing the respective operation states, as in Embodiments 1 and 2.

[0079] Next, in step Sll a, the server 1 selects the state with the highest occurrence frequency (in this case, the normal state = state label L1), and learns the autoencoder a1 reproducing the state using data of the state.

[0080] Then, in step S12a, the server 1 copies the encoder A1 obtained by learning in step Sll a, and learns the state classifier B, which inputs the feature quantity d1 extracted by the encoder A1 and outputs the state labels L1 to L3.

[0081] Further, finally in step S13a, the server 1 installs only the learned encoder A1 on the edge device 2.

[0082] <Inference Processing>

[0083] Figure 11 is a conceptual diagram of the inference processing in this embodiment, Figure 12 is a flowchart showing the flow of the inference processing.

[0084] In Embodiment 1, as shown in Figure 3 , Figure 4 , the feature quantities d1 to d3 are transmitted from the edge device 2 to the server 1, but in this embodiment, as shown in Figure 11 , Figure 12 , only the feature quantity d1 is transmitted from the edge device 2 to the server 1.

[0085] Further, in Embodiment 2, as shown in Figure 8As shown, according to the state label L output by the state classifier B, the decoder A' is used to generate reproduction data corresponding to the state label L, but in this embodiment, only when the state classifier B outputs the state label LI, the feature amount dl is input to the decoder Al', and reproduction data similar to the sensor data corresponding to the state label LI is output. As in Embodiment 2, the obtained data can be used for higher-precision control or the like.

[0086] Embodiment 4

[0087] Figure 13 is a basic configuration diagram when the operation state classification pre- and post-processing section 20 of Figure 7 is built in an industrial controller (programmable logic controller (PLC), sequencer, etc.) that is one of the edge devices 2. In addition to the above configuration, it includes the power conversion device 3, the motor 4, and the load machine 5. In addition, repeated explanations of common points with the above embodiments are omitted.

[0088] Since the industrial controller generally has a communication function with a higher-level system such as the server 1, if this communication function is utilized, the introduction cost of the operation state classification system 100 can be reduced. Further, since various data from a plurality of devices of the control target are collected in the industrial controller, there is an advantage that a network that takes the data of the plurality of devices as input is easily constructed and mounted. Also, it is possible to naturally perform reflection of the obtained classification result into control of the plurality of devices of the control target.

[0089] Embodiment 5

[0090] Figure 14 is a basic configuration diagram when the operation state classification pre- and post-processing section 20 of Figure 7 is built in the power conversion device 3 (general-purpose inverter, servo amplifier, DCBL controller, etc.). In addition, repeated explanations of common points with the above embodiments are omitted.

[0091] By building this function in the power conversion device 3, it is possible to take various parameters for motor control that are not normally extracted from the inside of the power conversion device as input of the operation state classification system 100. Further, if collection of various data is closed in the inside of the power conversion device, it is possible to improve the sampling speed regardless of the communication speed, so there is an advantage that data that captures a physical phenomenon at a higher speed can be used for operation state classification. Also, from the obtained classification result, in addition to being able to cut out detailed data for a very short time for diagnosis or saving, it is possible to issue an instruction to the motor control section and immediately reflect it into motor control.

[0092] The above describes the embodiments, but the present application is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments are detailed descriptions for facilitating understanding of the present application and are not limited to necessarily having all the structures described. Furthermore, a part of the structure of an embodiment can be replaced with the structure of another embodiment, and furthermore, the structure of another embodiment can be added to the structure of an embodiment. Furthermore, regarding a part of the structure of each embodiment, addition, deletion, or replacement of another structure can be performed.

[0093] Also, the present application is also applicable to industrial robots, machine tools, air compressors, industrial equipment such as conveyance tables in which power conversion devices are embedded, and wind power generation systems, hybrid cars, electric cars, fuel cell cars, railway vehicles, and the like.

[0094] Explanation of Reference Signs

[0095] 1 server, 10 operation state classification execution unit, 11 data communication unit, 12 state classifier storage unit, 13 inference execution unit, 14 data storage unit, 15 learning unit, 2 edge device, 20 operation state classification pre- and post-processing unit, 21 data input unit, 22 encoder storage unit, 3 power conversion device, 4 motor, 5 load machine, a, a1, a2, a3 autoencoder, A, A1, A2, A3 encoder, A', A1', A2', A3' decoder, B state classifier, d, d1, d2, d3 feature quantity (intermediate data), L, L1, L2, L3 state label.

Claims

1. An operational status classification system, wherein the operational status classification system connects edge devices and a server via a communication network, inputs sensor data into a neural network, and outputs status labels, characterized in that, The edge device includes: The system includes a first storage unit that stores the front-end of the neural network; and a dimensionality compression unit that inputs the sensor data into the front-end of the neural network and outputs intermediate data. The server includes: A second storage unit is used to store the subsequent stage of the neural network; An inference execution unit that inputs the intermediate data to the back-end of the neural network and outputs the state labels; and The learning unit on the later stage of the neural network is updated through append-learning. The learning unit supplements the learning of the later stages of the neural network based on the intermediate data stored in the server and the correct state labels associated with that intermediate data. When the probability of the inference execution unit determining the most reliable state is higher than a preset value, the correct state label is the state label output by the neural network.

2. An operational status classification system, wherein the operational status classification system connects edge devices and a server via a communication network, inputs sensor data into a neural network, and outputs status labels, characterized in that, The edge device includes: The system includes a first storage unit that stores the front-end of the neural network; and a dimensionality compression unit that inputs the sensor data into the front-end of the neural network and outputs intermediate data. The server includes: A second storage unit is used to store the subsequent stage of the neural network; An inference execution unit that inputs the intermediate data to the back-end of the neural network and outputs the state labels; and The learning unit on the later stage of the neural network is updated through append-learning. The learning unit supplements the learning of the later stages of the neural network based on the intermediate data stored in the server and the correct state labels associated with that intermediate data. When the probability of the inference execution unit determining the most reliable state is lower than a preset value, the correct state label is the state label input by the human.

3. The operational status classification system as described in claim 1 or 2, characterized in that, The front-end of the neural network stored in the edge device is the front-end part of the autoencoder learned in each classification state, i.e., the encoder. The back-end of the neural network stored in the server is a state classifier, which takes in the intermediate data output by the encoder and outputs the state label.

4. The operational status classification system as described in claim 3, characterized in that, The server also stores the post-encoder portion, namely the decoder.

5. The operational status classification system as described in claim 1 or 2, characterized in that, The edge device is an industrial controller or power conversion device.

6. A method for classifying operating states, characterized in that, include: The learning steps involve using a server to learn a neural network that takes input sensor data and outputs state labels. The first saving step involves storing the front-end side of the neural network in the edge device; The second saving step involves storing the back-end of the neural network in the server. as well as The server is used to perform additional learning steps on the back-end of the neural network. It also includes a step of saving additional learning data, which is obtained by assigning the state label output by the neural network to the intermediate data output by the previous stage of the neural network when the accuracy of the neural network in determining the most reliable state is higher than a preset value. In the additional learning step, the additional learning data is used to additionally learn the later stages of the neural network.

7. A method for classifying operating states, characterized in that, include: The learning steps involve using a server to learn a neural network that takes input sensor data and outputs state labels. The first saving step involves storing the front-end side of the neural network in the edge device; The second saving step involves storing the back-end of the neural network in the server. as well as The server is used to perform additional learning steps on the back-end of the neural network. It also includes a step of saving additional learning data, which is obtained by assigning the correct state label input by the human to the output of the front-end of the neural network when the accuracy of the neural network in determining the most reliable state is lower than a preset value. In the additional learning step, the additional learning data is used to additionally learn the later stages of the neural network.

Images