Neural network learning method, computer program, and remaining life prediction system

Abstract

To properly learn a neural network that predicts a remaining life of a device.SOLUTION: A method of learning a neural network is the method of learning the neural network that predicts a remaining life of a target device that is a maintenance target. The neural network includes: (i) a first model for predicting a remaining life at an arbitrary time of maintenance cycle data that is time series operation data in a period from immediately after maintenance of the target device to immediately before next maintenance, as a value with reference to an arbitrary reference value; and (ii) a second model for predicting a remaining life at a final time of the maintenance cycle data, as a value with reference to the reference value. A weight parameter of the first model is updated so as to predict a remaining life with reference to an end of the maintenance cycle data, by using an output of the first model and an output of the second model that are obtained from learning data including a plurality of pieces of maintenance cycle data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] This disclosure relates to the technical fields of neural network learning methods, computer programs, and life expectancy prediction systems. [Background technology]

[0002] Methods for estimating the remaining lifespan of various devices (i.e., the period until failure occurs) are known for performing maintenance on such devices. For example, Patent Document 1 discloses predicting the remaining lifespan of NAND flash memory installed in a numerical control device of a machine tool using a machine learning model. Patent Document 2 discloses predicting the remaining lifespan of a monitored system using a functional deep network. Patent Document 3 discloses predicting the remaining lifespan by switching between two machine learning models (artificial intelligence unit). [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Patent No. 6386523 [Patent Document 2] Japanese Patent Publication No. 2020-198081 [Patent Document 3] Japanese Patent Publication No. 2021-056153 [Overview of the project] [Problems that the invention aims to solve]

[0004] This disclosure aims to improve upon the technology disclosed in prior art documents. [Means for solving the problem]

[0005] One aspect of the neural network learning method of this disclosure is a neural network learning method for predicting the remaining lifespan of a target device subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining lifespan at any time in maintenance cycle data, which is time-series operational data of the target device from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value, wherein the weight parameters of the first model are updated to predict the remaining lifespan based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from learning data including a plurality of the maintenance cycle data.

[0006] One aspect of the computer program of this disclosure is a neural network learning method for predicting the remaining lifespan of a target device subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining lifespan at any time in maintenance cycle data, which is time-series operational data of the target device from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value, wherein the computer is instructed to perform a neural network learning method that updates the weight parameters of the first model to predict the remaining lifespan based on the end of the maintenance cycle data, using the output of the first model and the output of the second model obtained from training data including a plurality of the maintenance cycle data.

[0007] One aspect of the remaining useful life prediction system of this disclosure is a remaining useful life prediction system including a neural network for predicting the remaining useful life of a target device to be maintained. The neural network includes: (i) a first model for predicting the remaining useful life at any time of maintenance cycle data, which is time-series operation data during the period from immediately after maintenance of the target device to immediately before the next maintenance, as a value based on an arbitrary reference value; and (ii) a second model for predicting the remaining useful life at the final time of the maintenance cycle data as a value based on the reference value. The first model is learned by updating weight parameters to predict the remaining useful life based on the end of the maintenance cycle data, using the output of the first model and the output of the second model obtained from learning data including a plurality of the maintenance cycle data.

Brief Description of Drawings

[0008] [Figure 1] It is a block diagram showing the hardware configuration of the remaining useful life prediction system according to the first embodiment. [Figure 2] It is a block diagram showing the functional configuration of the remaining useful life prediction system according to the first embodiment. [Figure 3] It is a flowchart showing the flow of the learning operation of the remaining useful life prediction system according to the first embodiment. [Figure 4] It is a flowchart showing the flow of the prediction operation of the remaining useful life prediction system according to the first embodiment. [Figure 5] It is a conceptual diagram showing the time change of the soundness of the target device in each maintenance cycle. [Figure 6] It is a conceptual diagram showing the learning method executed by the remaining useful life prediction system according to the first embodiment. [Figure 7] It is a flowchart showing the flow of the learning method executed by the remaining useful life prediction system according to the first embodiment. [Figure 8] It is a conceptual diagram showing the learning method executed by the remaining useful life prediction system according to the second embodiment. [Figure 9]This is a flowchart showing the flow of the learning method executed by the remaining life prediction system according to the second embodiment. [Figure 10] This flowchart shows the flow of the method for calculating the loss function using the remaining life prediction system according to the third embodiment. [Figure 11] This is a flowchart showing the learning operation flow of the remaining life prediction system according to the fourth embodiment. [Figure 12] This flowchart shows the flow of the learning operation of the remaining life prediction system according to the fifth embodiment. [Figure 13] This is a flowchart showing the flow of the learning operation of the remaining life prediction system according to the sixth embodiment. [Modes for carrying out the invention]

[0009] The following describes the neural network training method, computer program, and embodiment of the remaining life prediction system, with reference to the diagrams. The following explanation will use an example of how the neural network training method is implemented in the remaining life prediction system.

[0010] <First Embodiment> The remaining life prediction system according to the first embodiment will be described with reference to Figures 1 to 7.

[0011] (Hardware configuration) First, the hardware configuration of the remaining life prediction system according to the first embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing the hardware configuration of the remaining life prediction system according to the first embodiment.

[0012] As shown in Figure 1, the remaining life prediction system 10 according to the first embodiment includes a processor 11, RAM (Random Access Memory) 12, ROM (Read Only Memory) 13, and a storage device 14. The remaining life prediction system 10 may further include an input device 15 and an output device 16. The processor 11, RAM 12, ROM 13, storage device 14, input device 15, and output device 16 are connected via a data bus 17.

[0013] The processor 11 reads a computer program. For example, the processor 11 is configured to read a computer program stored in at least one of the RAM 12, ROM 13, and storage device 14. Alternatively, the processor 11 may read a computer program stored in a computer-readable storage medium using a storage medium reading device (not shown). The processor 11 may also obtain (i.e., read) a computer program from a device (not shown) located outside the remaining life prediction system 10 via a network interface. The processor 11 controls the RAM 12, storage device 14, input device 15, and output device 16 by executing the read computer program. In this embodiment in particular, when the processor 11 executes the read computer program, functional blocks for predicting the remaining life of the target device and functional blocks for learning a neural network are realized within the processor 11. That is, the processor 11 may function as a controller that executes each control in the remaining life prediction system 10 according to this embodiment.

[0014] The processor 11 may be configured as, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an FPGA (field-programmable gate array), a DSP (Demand-Side Platform), or an ASIC (Application Specific Integrated Circuit). The processor 11 may consist of one of these, or it may be configured to use multiple of them in parallel.

[0015] RAM12 temporarily stores computer programs executed by processor 11. RAM12 also temporarily stores data that processor 11 uses temporarily while executing computer programs. RAM12 may be, for example, D-RAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory). Alternatively, other types of volatile memory may be used instead of RAM12.

[0016] ROM 13 stores computer programs executed by processor 11. ROM 13 may also store other static data. ROM 13 may be, for example, P-ROM (Programmable Read Only Memory) or EPROM (Erasable Read Only Memory). Alternatively, other types of non-volatile memory may be used instead of ROM 13.

[0017] The storage device 14 stores data that the remaining life prediction system 10 will save for the long term. The storage device 14 may also operate as a temporary storage device for the processor 11. The storage device 14 may include, for example, at least one of a hard disk drive, a magneto-optical disk drive, an SSD (Solid State Drive), and a disk array device.

[0018] The input device 15 is a device that receives input instructions from the user of the remaining life prediction system 10. The input device 15 may include, for example, at least one of a keyboard, a mouse, and a touch panel. The input device 15 may be configured as a portable terminal such as a smartphone or tablet. The input device 15 may also be a device capable of voice input, for example, including a microphone.

[0019] The output device 16 is a device that outputs information related to the remaining life prediction system 10 to the outside. For example, the output device 16 may be a display device (e.g., a display) capable of displaying information related to the remaining life prediction system 10. Alternatively, the output device 16 may be a speaker or the like capable of outputting information related to the remaining life prediction system 10 as audio. The output device 16 may be configured as a mobile terminal such as a smartphone or tablet. Furthermore, the output device 16 may be a device that outputs information in a format other than images. For example, the output device 16 may be a speaker that outputs information related to the remaining life prediction system 10 as audio.

[0020] Although Figure 1 shows an example of a remaining life prediction system 10 comprising multiple devices, all or some of these functions may be implemented in a single device. In that case, the remaining life prediction system may be configured to include only the processor 11, RAM 12, and ROM 13 described above, while other components (i.e., storage device 14, input device 15, and output device 16) may be provided by external devices connected to the remaining life prediction system 10. Furthermore, the remaining life prediction system may implement some of its computational functions using external devices (e.g., external servers or cloud services).

[0021] (Functional configuration) Next, the functional configuration of the remaining life prediction system 10 according to the first embodiment will be described with reference to Figure 2. Figure 2 is a block diagram showing the functional configuration of the remaining life prediction system according to the first embodiment.

[0022] As shown in Figure 2, the remaining life prediction system 10 according to the first embodiment is configured to include a data acquisition unit 110, a learning unit 120, a prediction unit 130, an output unit 140, and a storage unit 150 as components for realizing its function. Each of the data acquisition unit 110, the learning unit 120, the prediction unit 130, and the output unit 140 may be a processing block realized by, for example, the processor 11 (see Figure 1) described above. The storage unit 150 may be realized by, for example, the storage device 14 (see Figure 1) described above.

[0023] The data acquisition unit 110 is configured to collect maintenance cycle data of the target equipment that is subject to maintenance. The maintenance cycle data is time-series operational data for the period from immediately after maintenance of the target equipment to immediately before the next maintenance. The target equipment is not particularly limited as long as it is equipment that is subject to maintenance, but examples include hard disks, NAND flash memory, and rotating equipment (e.g., pumps and fans). In the case of hard disks, the maintenance cycle data may include Write Count, Average Write Response Time, Max Write Response Time, Write Transfer Rate, Read Count, Average Read Response Time, Max Read Time, Read Transfer Rate, Busy Ratio, Busy Time, etc. In the case of NAND flash memory, the maintenance cycle data may include the number of writes, write interval, read count, temperature in the operating environment, error rate, information on the manufacturer and information on the manufacturing lot, as well as information on the ECC performance of the memory controller that performs error correction coding (ECC) processing on the NAND flash memory, information on the manufacturer and information on the manufacturing lot, etc. In the case of rotating machinery, maintenance cycle data may include output values ​​from acceleration sensors, ultrasonic (AE) sensors, current, motor torque, strain gauges, etc.

[0024] The learning unit 120 is configured to learn a model for predicting the remaining lifespan of the target equipment (i.e., the period until failure occurs) by using the maintenance cycle data collected by the data acquisition unit 110 as training data. In this embodiment, the remaining lifespan prediction system 10 includes a first model and a second model as models for predicting the remaining lifespan, but the learning unit 120 only needs to be configured to learn at least the first model. However, the learning unit 120 may be configured to learn both the first model and the second model. The learning of the second model will be described in other embodiments described later.

[0025] The prediction unit 130 is configured to predict the remaining lifespan of the target equipment using a model learned by the learning unit 120. Specifically, the prediction unit 130 is configured to predict the remaining lifespan of the target equipment corresponding to the maintenance cycle data by inputting the maintenance cycle data collected by the data collection unit 110 into a trained model as data for remaining lifespan prediction. The prediction operation by the prediction unit 130 will be explained in detail later.

[0026] The output unit 140 is configured to output various types of information from the remaining life prediction system 10. For example, the output unit 140 may be configured to output information regarding the remaining life of the target equipment predicted by the prediction unit 130. In this case, the output information may indicate the remaining life value, or it may be an alarm corresponding to the remaining life (for example, information to prompt maintenance). The output unit 140 may be configured to output various types of information via the output device 16 described above. For example, the output unit 140 may be configured to output various types of information via a monitor, speaker, etc.

[0027] The memory unit 150 is configured to store various types of information handled by the remaining life prediction system 10. The memory unit 150 may be configured to store, for example, a model learned by the learning unit 120. The memory unit 150 may also be configured to store maintenance cycle data collected by the data acquisition unit 110.

[0028] (Learning process) Next, with reference to Figure 3, the learning operation of the remaining life prediction system 10 according to the first embodiment (i.e., the operation when learning a model to predict remaining life) will be described. Figure 3 is a flowchart showing the flow of the learning operation of the remaining life prediction system according to the first embodiment.

[0029] As shown in Figure 3, when the learning operation of the remaining life prediction system 10 according to the first embodiment is started, the data acquisition unit 110 first acquires maintenance cycle data (step S101). At this time, the data acquisition unit 110 may acquire new maintenance cycle data from the target equipment, or it may acquire previously acquired maintenance cycle data from the storage unit 150. The maintenance cycle data acquired by the data acquisition unit 110 is output to the learning unit 120.

[0030] Next, the learning unit 120 uses the maintenance cycle data as training data to train a model that predicts the remaining lifespan of the target equipment (step S102). The model training operation by the learning unit 120 will be explained in detail later. Once training is complete, the learning unit 120 saves the trained model to the storage unit 150 (step S103).

[0031] (Predictive action) Next, with reference to Figure 4, the prediction operation of the remaining life prediction system 10 according to the first embodiment (i.e., the operation when predicting the remaining life using a trained model) will be described. Figure 4 is a flowchart showing the flow of the prediction operation of the remaining life prediction system according to the first embodiment.

[0032] As shown in Figure 4, when the prediction operation of the remaining life prediction system 10 according to the first embodiment is started, the prediction unit 130 first reads a learned model from the storage unit 150 (step S201). Note that the model read here may be only the first model.

[0033] Next, the data acquisition unit 110 acquires maintenance cycle data for remaining life prediction (step S202). At this time, the data acquisition unit 110 may acquire new maintenance cycle data from the target equipment, or it may acquire previously acquired maintenance cycle data from the storage unit 150. The maintenance cycle data acquired by the data acquisition unit 110 is output to the prediction unit 130.

[0034] Next, the prediction unit 130 predicts the remaining lifespan of the target equipment using the trained model (step S203). Then, the prediction unit 130 determines whether the predicted remaining lifespan is below a predetermined threshold (step S204). The "predetermined threshold" here is a threshold used to determine whether or not maintenance should be performed on the target equipment, and any value may be set in advance.

[0035] If the predicted remaining lifespan is below a predetermined threshold (step S204: YES), the output unit 140 outputs an alarm to the user (step S205). The alarm may include information such as prompting maintenance work. On the other hand, if the predicted remaining lifespan is not below a predetermined threshold (step S204: NO), the process in step S205 described above may be omitted. However, information indicating when the next maintenance work should be performed based on the predicted remaining lifespan may be output.

[0036] (Training data and model) Next, with reference to Figures 5 and 6, the learning data used in the remaining life prediction system 10 according to the first embodiment and the model learned using the learning data will be described in detail. Figure 5 is a conceptual diagram showing the change in the health of the target equipment over time in each maintenance cycle. Figure 6 is a conceptual diagram showing the learning method executed by the remaining life prediction system according to the first embodiment.

[0037] In Figure 5, the remaining life prediction system 10 according to the first embodiment uses multiple maintenance cycle data (for example, see data A to F in the figure) as training data. These multiple maintenance cycle data may be acquired from different devices, or they may be acquired from the same device at different times. The maintenance cycle data consists of values ​​acquired from sensors installed on or around the target device for each maintenance cycle, or statistical values ​​thereof (at regular intervals, not the entire maintenance cycle), and is typically multidimensional time-series data. Note that the health level in the figure is a hypothetical index and may be a value that cannot actually be observed. Here, the health level corresponds to the remaining life in the degraded region (i.e., the region where the health level is decreasing).

[0038] Multiple maintenance cycle data sets exhibit variations in the remaining life at the end time immediately preceding maintenance (i.e., the circled portion in the figure). Therefore, if the remaining life at each end time is trained using a value that is assumed to occur when maintenance is needed (for example, assumed to be 0), the training may be inappropriate. However, as described later, the remaining life prediction system 10 according to this embodiment performs model training using the remaining life at any time and the final time in the maintenance cycle data (relative remaining life relative to an arbitrary reference value).

[0039] As shown in Figure 6, the remaining life prediction system 10 according to the first embodiment includes a first model and a second model as models for predicting the remaining life of the target equipment. Each of the first and second models is configured to include a neural network, and may include TSS2Vec (a neural network that can convert sequential data into vectors, such as a recurrent neural network (RNN, LSTM, GRU, etc.), CNN, Transformer, etc.), a nonlinear transformation (a neural network that converts one vector to another, such as a Multilayer Perceptron), and Vec2HI (a neural network that converts a vector to a scalar value, such as a Multilayer Perceptron).

[0040] The first model takes partial time-series data at an arbitrary time point extracted from maintenance cycle data (i.e., data extracted from any portion of the total period included in the maintenance cycle data) as input. While the arbitrary time point here is not particularly limited, it is preferable to select a point relatively close to the end of the maintenance cycle, as a time point far from the end may not contain degradation information. The first model uses the partial time-series data at the input arbitrary time point to determine the remaining lifespan r of the target equipment at that arbitrary time point. ij The (i) maintenance cycle index and the time index within each maintenance cycle are estimated based on an arbitrary reference value. When predicting remaining life, another reference value, as described later, is calculated and used for the prediction. Therefore, the "reference value" here does not need to be defined.

[0041] The second model takes partial time-series data of the termination time extracted from the maintenance cycle data (i.e., data extracted from the entire period included in the maintenance cycle data, including the termination time) as input. Then, the second model calculates the remaining life e of the target equipment at the final time (i.e., the time immediately before maintenance) from the input partial time-series data of the termination time. i This is estimated using the same baseline value used in the first model as the reference value.

[0042] The first model is the output of the first model, which is the remaining lifetime r at any given time. ij And the output of the second model is the remaining lifetime e at the final time. i It is learned using and. Specifically, the first model is trained using the remaining lifetime r at any given time. ij and the remaining lifetime e at the final time i The first model is trained to predict the remaining life (i.e., the period from the state immediately before maintenance until a failure occurs) relative to the end of the maintenance cycle data. ij and the remaining lifetime e at the final time iThe weight parameters may be changed using the backpropagation method so that the loss calculated from e is calculated, for example, as in the following formula (1).

[0043]

Equation

[0044] (Flow of learning method) Next, referring to FIG. 7, the flow of the learning method (specifically, the process of step S102 described in FIG. 3) executed by the remaining life prediction system 10 according to the first embodiment will be described in detail. FIG. 7 is a flowchart showing the flow of the learning method executed by the remaining life prediction system according to the first embodiment.

[0045] As shown in FIG. 7, in the learning method executed by the remaining life prediction system 10 according to the first embodiment, first, the learning unit 130 initializes an evaluation value (specifically described later) (step S301). Also, the learning unit 130 initializes the weight parameters of the first model and the second parameters (step S302). Here, the weight parameters may be initialized so that the first model and the second model are the same.

[0046] Subsequently, the learning unit 130 extracts a specified component of partial time series data at an arbitrary time, the corresponding relative remaining life value (the remaining life value with the remaining life at the end of the maintenance cycle data set to 0), and the partial time series data corresponding to the end time of the maintenance cycle data from the maintenance cycle data included in the learning data (step S303). That is, data used for learning is appropriately extracted from the maintenance cycle data.

[0047] Next, the learning unit 130 calculates the loss L using the outputs of the first model and the second model based on the training data (step S304). Then, the learning unit 130 updates the weight parameters of the first model to minimize the loss L (step S305). Finally, the learning unit 130 calculates an evaluation value using the loss L (step S306). The "evaluation value" here is an index used to determine whether to save the weight parameters of the first model as the best values, and may be a function that includes the loss L, or it may be the loss L itself.

[0048] Next, if the evaluation value has improved, the learning unit 130 overwrites the weight parameters of the first model in the storage unit 150 (step S307). The process from steps S303 to S307 described above is repeated for a predetermined number of iterations.

[0049] Once learning is complete, the learning unit 130 reads the weight parameters of the first model stored in the memory unit 150 and calculates the reference remaining life. Here, "reference remaining life" is the remaining life e at the final time of each maintenance cycle. i The minimum value of, or an estimate of the remaining lifetime at the final time based on multiple samples per cycle (e.g., e i This is the minimum value of (hat). The estimated remaining lifetime at the final time can be calculated, for example, using equation (2) below.

[0050]

number

[0051] In another form, the estimated remaining lifetime at the final time is given in the second model r ij Input the input data to the first model used in the calculation of r, and the output is r ij You may also use the value obtained by calculating equation (2).

[0052] (Technical effects) Next, we will explain the technical effects obtained by the remaining life prediction system 10 according to the first embodiment.

[0053] As explained in Figures 1 to 7, in the remaining life prediction system 10 according to the first embodiment, the remaining life r at any given time is ij and the remaining lifetime e at the final time i The system is trained to predict the remaining lifespan based on the end of the maintenance cycle data. This allows for proper learning even when there is variation in the end of each acquired maintenance cycle data. As a result, it becomes possible to predict the remaining lifespan of the target equipment with high accuracy, enabling, for example, maintenance to be performed at the appropriate time.

[0054] <Second Embodiment> The remaining life prediction system 10 according to the second embodiment will be described with reference to Figures 8 and 9. Note that the second embodiment differs from the first embodiment described above only in some configurations and operations; other parts may be identical to the first embodiment. Therefore, the following will describe in detail the parts that differ from the first embodiment already described, and will omit explanations of other overlapping parts as appropriate.

[0055] (Training the second model) First, with reference to Figure 8, the learning method performed by the remaining life prediction system 10 according to the second embodiment will be explained in detail. Figure 8 is a conceptual diagram showing the learning method performed by the remaining life prediction system according to the second embodiment.

[0056] As shown in Figure 8, in the remaining life prediction system 10 according to the second embodiment, the weight parameters of the second model are updated using the weight parameters of the first model (i.e., the weight parameters updated by learning). In other words, in the remaining life prediction system 10 according to the second embodiment, the second model is also learned in addition to the first model. When updating the weight parameters of the second model, it is sufficient to update at least some of them. That is, it is not necessary to update all the weight parameters included in the second model; it is acceptable to update only some of the weight parameters.

[0057] (Learning process flow) Next, with reference to Figure 9, the flow of the learning method executed by the remaining life prediction system 10 according to the second embodiment will be described. Figure 9 is a flowchart showing the flow of the learning method executed by the remaining life prediction system according to the second embodiment. Note that in Figure 9, the same reference numerals are used for processes as in Figure 7.

[0058] As shown in Figure 9, in the learning process of the remaining life prediction system 10 according to the second embodiment, the processing (see Figure 7) is performed in a flow that is generally the same as that of the first embodiment already described. However, in the remaining life system 10 according to the second embodiment, after the process of updating the weight parameters of the first model (step S305), the process of updating the weight parameters of the first model using the updated weight parameters of the first model is performed (step S311). Then, the process from S303 to S307, including step S311, is repeated a predetermined number of times.

[0059] The weight parameters of the second model may be updated, for example, using the exponential moving average of the weight parameters of the first model. For example, the weight parameters W of the first model 1i , the weight parameter W of the second model 2i Therefore, in step S311 described above, W 2i =(1-a)W 1i +aW 2i The weight parameters of the second model may be updated to achieve the following: In this case, the past weight parameters W of the second model 2i The coefficient a (a value less than 1) is reduced and taken into consideration, and by repeating this process, an update using the exponential moving average is achieved.

[0060] (Technical effects) Next, we will explain the technical effects obtained by the remaining life prediction system 10 according to the second embodiment.

[0061] As explained in Figures 8 and 9, in the remaining life prediction system 10 according to the second embodiment, the weight parameters of the second model are updated using the weight parameters of the first model. In this way, both the first and second models are learned, allowing for more appropriate learning compared to updating only the weight parameters of the first model. Furthermore, by using the exponential moving average of the weight parameters of the first model, it is possible to update the weight parameters of the second model while appropriately considering the weight parameters of the first model.

[0062] <Third Embodiment> The remaining life prediction system 10 according to the third embodiment will be described with reference to Figure 10. Note that the third embodiment differs from the first and second embodiments described above only in some configurations and operations; other parts may be identical to those of the first and second embodiments. Therefore, the following will describe in detail the parts that differ from the embodiments already described, while other overlapping parts will be omitted as appropriate.

[0063] (Calculation of losses) First, the method for calculating losses by the remaining life prediction system 10 according to the third embodiment will be explained with reference to Figure 10. Figure 10 is a flowchart showing the flow of the method for calculating the loss function by the remaining life prediction system according to the third embodiment.

[0064] In the remaining life prediction system 10 according to the third embodiment, a model for predicting remaining life is learned by assuming that input data that are similar have similar remaining life values. Specifically, the loss L used when learning the model for predicting remaining life is calculated by adding a regularization term for feature matching. Specifically, the loss L may be calculated by the following equation (3).

[0065]

number

[0066]

number

[0067] As shown in Figure 10, when calculating Lf as described above, the multiple maintenance cycle data included in the training data are first divided into two sets (step S401). That is, the multiple maintenance cycle data are divided into data included in the first data group and data included in the second data group.

[0068] Next, a predetermined number of N partial time series data are extracted from the end of each cycle (step S402). Then, a predetermined number of M partial time series data similar to the extracted partial time series data are extracted from the other set (step S403). The N partial time series data extracted here and the M partial time series data are each other's potential pairs.

[0069] Next, when training the model, a predetermined number of K partial time series data are randomly selected from each set (step S404). Then, each of the K selected partial time series data is converted into a feature vector (step S405). The feature vectors converted here are the f in the above equation. k and f k It is.

[0070] Finally, the transformed feature vector f k and f k Calculate Lf using the expected value E of ' (step S406).

[0071] The regularization term for feature matching is f k and f kThe Maximum Mean Discrepancy of ' may be calculated using a kernel. In this case, for example, a Radial basis function kernel can be used.

[0072] (Technical effects) Next, the technical effects obtained by the remaining life prediction system 10 according to the third embodiment will be described.

[0073] As explained in Figure 10, in the remaining life prediction system 10 according to the third embodiment, the regularization term of feature matching is considered when calculating the loss L. This allows the model to be trained more effectively, making it possible to predict the remaining life of the target equipment with higher accuracy.

[0074] <Fourth Embodiment> The remaining life prediction system 10 according to the fourth embodiment will be described with reference to Figure 11. Note that the fourth embodiment differs from the first to third embodiments described above only in some configurations and operations; other parts may be identical to the first to third embodiments. Therefore, the following will describe in detail the parts that differ from each embodiment already described, and will omit explanations of other overlapping parts as appropriate.

[0075] (Setting initial values ​​through pre-training) First, with reference to Figure 11, the flow of the learning operation (specifically, the learning operation using pre-training) by the remaining life prediction system 10 according to the fourth embodiment will be explained. Figure 11 is a flowchart showing the flow of the learning operation of the remaining life prediction system according to the fourth embodiment. Note that in Figure 11, the same reference numerals are used for the same processes as shown in Figure 3.

[0076] As shown in Figure 11, in the learning operation of the remaining life prediction system 10 according to the fourth embodiment, the data acquisition unit 110 first acquires supervised learning data (step S501). The supervised learning data acquired here is run-to-failure data, which is time-series operational data of the target equipment from a healthy state to failure. The labels for learning are given, for example, the number of uses or operating time until failure.

[0077] Next, the learning unit 130 performs pre-training of the first model and the second model using supervised learning data (step S502). This is achieved by performing pre-training of the first model and initializing the second model with the weight parameters of the pre-trained first model. Then, the pre-trained weight parameters are set as the initial values ​​for the weight parameters of each model (step S503).

[0078] Subsequently, the same process as in the first embodiment described above (see Figure 3) is performed. That is, model training is performed using maintenance cycle data (i.e., unsupervised training data). However, in the fourth embodiment, as described above, training starts with the initial values ​​of the weight parameters of each model already set by pre-training. Therefore, the initialization of the weight parameters, as described in step S302 of Figure 7, does not need to be performed.

[0079] In pre-training, the first model is trained to determine the remaining lifetime r at any given time, which is the output of the first model. ij And, given label r ij’ It is learned using and. Specifically, the first model is r ij and, r ij’ The model is trained to match. The first model may have its weight parameters modified, for example, using backpropagation. The loss in this case is L es This can be calculated, for example, as shown in equation (5) below.

[0080]

number

[0081] (Technical effects) Next, the technical effects obtained by the remaining life prediction system 10 according to the fourth embodiment will be described.

[0082] As explained in Figure 11, in the remaining life prediction system 10 according to the fourth embodiment, the initial values ​​of the weight parameters of the first and second models are set by pre-training using supervised learning data. In this way, the initial values ​​of each model are set to appropriate values, so that subsequent training using unsupervised learning data can be performed efficiently. In this embodiment, a situation is assumed in which the number of supervised learning data is small compared to the number of unsupervised learning data (specifically, a situation in which there is not enough data to train using only supervised learning data, and it is necessary to use unsupervised learning data as well).

[0083] <Fifth Embodiment> The remaining life prediction system 10 according to the fifth embodiment will be described with reference to Figure 12. Note that the fifth embodiment differs from the first to fourth embodiments described above only in some configurations and operations; other parts may be identical to the first to fourth embodiments. Therefore, the following will describe in detail the parts that differ from the embodiments already described, and will omit explanations of other overlapping parts as appropriate. In this embodiment, the training data includes two types: supervised data and unsupervised data (maintenance cycle data).

[0084] (Learning using training data) First, with reference to Figure 12, the learning operation (learning operation depending on the presence or absence of training data) of the remaining life prediction system 10 according to the fifth embodiment will be explained. Figure 12 is a flowchart showing the flow of the learning operation of the remaining life prediction system according to the fifth embodiment. Each process shown in Figure 12 corresponds to steps S303 to S307 in Figure 7, and is a process that is repeated a predetermined number of times as explained using Figure 7.

[0085] As shown in Figure 12, in the learning operation of the remaining life prediction system 10 according to the fifth embodiment, the learning unit 130 first performs learning using supervised learning data (i.e., time-series operating data of the target equipment from a healthy state to failure). Specifically, the learning unit 130 first samples data from the supervised learning data (step S601). Then, the learning unit 130 calculates the loss using the output of the first model based on the supervised learning data (step S602). In learning using supervised learning data, the loss can be calculated using the formula (5) described above, as explained in the fourth embodiment.

[0086] Next, the learning unit 130 updates the weight parameters of the first model to minimize the loss (step S603). Then, the learning unit 130 calculates an evaluation value using the loss (step S604). After that, if the evaluation value has improved, the learning unit 130 overwrites the weight parameters of the first model in the storage unit 150 (step S605).

[0087] The processes described in steps S601 to S605 are repeated a predetermined number of times, and when learning using supervised learning data is completed (step S606: YES), the learning unit 130 initializes the weight parameters of the second model with the weight parameters of the first model (step S607) and performs learning using unsupervised learning data (i.e., maintenance data). Specifically, the learning unit 130 samples data from the unsupervised learning data (step S608). Then, the learning unit 130 calculates the loss using the outputs of the first model and the second model based on the unsupervised learning data (step S609). In learning using unsupervised learning data, the loss can be calculated using the formula (1) described above, as explained in the first embodiment.

[0088] Next, the learning unit 130 updates the weight parameters of the first and second models to minimize the loss (step S610). Then, the learning unit 130 calculates an evaluation value using the loss (step S611). After that, if the evaluation value has improved, the learning unit 130 overwrites the weight parameters of the first and second models in the storage unit 150 (step S612). After repeating the processes from steps S608 to S612 described above a predetermined number of times, it is determined that the learning using unsupervised learning data has also finished (step S613: YES).

[0089] Here, we have given an example of training with unsupervised training data immediately after training with supervised training data, but these training processes can also be performed simultaneously. In other words, training can be performed simultaneously using both supervised and unsupervised training data. When performing training simultaneously, after sampling the data for each training dataset, calculate the corresponding loss for each and update the parameters using the sum of the calculated losses. Subsequently, the processes of calculating the evaluation value and saving the best parameters should be performed in that order.

[0090] (Technical effects) Next, the technical effects obtained by the remaining life prediction system 10 according to the fifth embodiment will be described.

[0091] As explained in Figure 12, in the remaining life prediction system 10 according to the fifth embodiment, when the weight parameters of the first and second models are learned, the final weight parameters are determined using both supervised and unsupervised data. In this way, a model that predicts the remaining life relative to the time of failure is learned, even when there is little supervised data, by utilizing the information from the unsupervised data. As a result, a more appropriate maintenance plan can be made.

[0092] Furthermore, the fourth embodiment (see Figure 11) and the fifth embodiment (see Figure 12) described above may be implemented in combination. Specifically, as explained in the fourth embodiment, initial values ​​may be set by pre-training, and then, as explained in the fifth embodiment, training may be performed using supervised training data and unsupervised training data. In this way, supervised training data can be utilized in both pre-training and subsequent training, making it possible to perform training more appropriately.

[0093] <Sixth Embodiment> The remaining life prediction system 10 according to the sixth embodiment will be described with reference to Figure 13. Note that the sixth embodiment differs from the first to fifth embodiments described above only in some operations; other parts may be identical to the first to fifth embodiments. Therefore, the following will describe in detail the parts that differ from each embodiment already described, and will omit explanations of other overlapping parts as appropriate.

[0094] (Training a different model) First, with reference to Figure 13, we will explain the learning operation of the remaining life prediction system 10 according to the sixth embodiment (specifically, the learning operation of another model using a previously trained model). Figure 13 is a flowchart showing the flow of the learning operation of the remaining life prediction system according to the sixth embodiment.

[0095] As shown in Figure 13, in the remaining life prediction system 10 according to the sixth embodiment, the prediction unit 130 predicts the remaining life at the end of each maintenance cycle data using a model learned using the method described in each embodiment above (step S701).

[0096] Next, the prediction unit 130 corrects the remaining lifespan at each time point for each maintenance cycle data based on the predicted remaining lifespan of the end (step S702). The corrected remaining lifespan is then treated as training data for each maintenance cycle data.

[0097] Next, the learning unit 130 uses the maintenance cycle data with corrected remaining lifespan as supervised learning data to train a new machine learning model (step S703). In this training, the weight parameters may be updated so that the model's predicted values ​​and the training data approach each other. The machine learning model trained here may not include a neural network.

[0098] The machine learning model trained as described above may be used in a different system (for example, a similar system installed in a different location). That is, the trained model used in step S701 and the machine learning model trained in step S703 may be used in different remaining life prediction systems.

[0099] (Technical effects) Next, the technical effects obtained by the remaining life prediction system 10 according to the sixth embodiment will be described.

[0100] As explained in Figure 13, in the remaining life prediction system 10 according to the sixth embodiment, a new machine learning model is learned to predict the remaining life after correcting the remaining life at the end of each maintenance cycle using a trained model that predicts the remaining life at the end of each maintenance cycle, which was learned using the methods of each embodiment described above. In this way, the learning to correct for the differences in the remaining life at the end of each maintenance cycle and the learning to predict the remaining life over the entire maintenance cycle data can be configured separately. As a result, even if there is a period in the remaining life when the operational data that forms the basis of the prediction does not contain information for predicting the remaining life, a model that predicts a wider range of remaining life can be learned by selecting the data to be used for learning based on the corrected remaining life, learning the model, and evaluating it.

[0101] The processing method of recording a program that operates the configuration of each embodiment in order to realize the functions of each embodiment described above on a recording medium, reading the program recorded on the recording medium as code, and executing it on a computer is also included in the scope of each embodiment. In other words, a computer-readable recording medium is also included in the scope of each embodiment. Furthermore, not only the recording medium on which the above-mentioned program is recorded, but also the program itself is included in each embodiment.

[0102] Examples of recording media that can be used include floppy disks, hard disks, optical disks, magneto-optical disks, CD-ROMs, magnetic tapes, non-volatile memory cards, and ROMs. Furthermore, the scope of each embodiment is not limited to programs that perform processing on the recording media alone, but also includes programs that operate on the OS and perform processing in cooperation with other software and the functions of expansion boards. In addition, the program itself may be stored on a server, and part or all of the program may be made available for download from the server to the user terminal.

[0103] <Note> The embodiments described above may also be described in the following appendix, but are not limited to these.

[0104] (Note 1) The neural network learning method described in Appendix 1 is a neural network learning method for predicting the remaining lifespan of a target device that is subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining lifespan at any time in maintenance cycle data, which is time-series operational data of the target device from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time of the maintenance cycle data as a value based on the reference value, and the weight parameters of the first model are updated to predict the remaining lifespan based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from learning data including a plurality of the maintenance cycle data.

[0105] (Note 2) The neural network learning method described in Appendix 2 is the neural network learning method described in Appendix 1, which updates at least a portion of the weight parameters of the second model using the updated weight parameters of the first model.

[0106] (Note 3) The neural network learning method described in Appendix 3 is the neural network learning method described in Appendix 2, wherein at least a portion of the weight parameters of the second model are updated using the exponentially smoothed moving average of the updated weight parameters of the first model.

[0107] (Note 4) The neural network learning method described in Appendix 4 is a neural network learning method described in any one of Appendix 1 to 3, which involves dividing the learning data into a first data group and a second data group, extracting a first partial data set included in the first data group and a second partial data set included in the second data group that is similar to the first partial data set, and updating the weight parameters of the first model so that the feature vectors of the first partial data set and the feature vectors of the second partial data set are similar to each other.

[0108] (Note 5) The neural network learning method described in Appendix 5 is the neural network learning method described in any one of Appendix 1 to 4, wherein the learning data includes a plurality of supervised learning data which are time-series operational data of the target device from a healthy state to failure, and a plurality of unsupervised learning data which are maintenance cycle data, the neural network is pre-trained using the supervised learning data, the weight parameters of the pre-trained neural network are set to the initial values ​​of the first model and the second model, and the weight parameters of the first model are updated using the output of the first model and the output of the second model obtained from the unsupervised learning data to predict the remaining lifespan based on the end of the maintenance cycle data.

[0109] (Note 6) The neural network learning method described in Appendix 6 is a neural network learning method described in any one of Appendix 1 to 5, wherein the learning data includes a plurality of supervised learning data which are time-series operational data of the target device from a healthy state to failure, and a plurality of unsupervised learning data which are maintenance cycle data, and when the supervised learning data is used, the weight parameters of the first model are updated so that the predicted value predicted by the first model and the measured value of the remaining life indicated by the supervised learning data approach each other.

[0110] (Note 7) The neural network learning method described in Appendix 7 is the neural network learning method described in any one of Appendix 1 to 6, wherein the learning data includes a plurality of supervised learning data which are time-series operational data of the target device from a healthy state to failure, and a plurality of unsupervised learning data which are maintenance cycle data, the neural network is pre-trained using the supervised learning data, the weight parameters of the pre-trained neural network are set to the initial values ​​of the first model and the second model, the weight parameters of the first model are updated to predict the remaining lifespan based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from the learning data, and if supervised learning data is used, the weight parameters of the first model are updated so that the predicted value predicted by the first model and the measured value of the remaining lifespan indicated by the supervised learning data approach each other.

[0111] (Note 8) The neural network learning method described in Appendix 8 is a neural network learning method according to any one of claims 1 to 7, wherein the first model or the second model, whose weight parameters have been updated, predicts the remaining lifespan at the end of each maintenance cycle data, modifies the remaining lifespan at each time point of each maintenance cycle data based on the predicted remaining lifespan at the end of each maintenance cycle data, and learns a new machine learning model that predicts the remaining lifespan of the target equipment using training data that includes the modified remaining lifespan as training data.

[0112] (Note 9) The computer program described in Appendix 9 is a neural network learning method for predicting the remaining lifespan of equipment subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining lifespan at any time in maintenance cycle data, which is time-series operational data of the equipment subject to maintenance from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value, and the computer program causes a computer to execute a neural network learning method that updates the weight parameters of the first model to predict the remaining lifespan based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from learning data including a plurality of the maintenance cycle data.

[0113] (Note 10) The recording medium described in Appendix 10 is a recording medium on which a computer program is recorded that causes a computer to execute a neural network learning method for predicting the remaining lifespan of a target device that is subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining lifespan at any time in maintenance cycle data, which is time-series operational data of the target device from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time of the maintenance cycle data as a value based on the reference value, and updates the weight parameters of the first model to predict the remaining lifespan based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from learning data including a plurality of the maintenance cycle data.

[0114] (Note 11) The remaining life prediction system described in Appendix 11 is a remaining life prediction system comprising a neural network that predicts the remaining life of a target equipment subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining life at any time in maintenance cycle data, which is time-series operational data of the target equipment from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining life at the final time of the maintenance cycle data as a value based on the reference value, wherein the first model is learned by updating its weight parameters to predict the remaining life based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from training data including a plurality of the maintenance cycle data.

[0115] (Note 12) The remaining life prediction device described in Appendix 12 is a remaining life prediction system that includes a neural network for predicting the remaining life of a target device subject to maintenance, wherein the neural network includes (i) a first model that predicts the remaining life at any time in maintenance cycle data, which is time-series operational data of the target device from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining life at the final time of the maintenance cycle data as a value based on the reference value, wherein the first model is learned by updating its weight parameters to predict the remaining life based on the end of the maintenance cycle data using the output of the first model and the output of the second model obtained from training data including a plurality of the maintenance cycle data. That is the case.

[0116] This disclosure may be modified as appropriate, without contradicting the gist or idea of ​​the invention as can be inferred from the claims and the specification as a whole, and methods for learning neural networks, computer programs, and remaining life prediction systems, with such modifications, are also included in the technical idea of ​​this disclosure. [Explanation of symbols]

[0117] 10. Remaining Life Prediction System 11 processors 14 Storage device 14 110 Data Acquisition Unit 120 Learning Department 130 Prediction Section 140 Output section 150 Storage section r ij Remaining life at any given time e i Remaining lifespan at the final time

Claims

1. A neural network training method for predicting the remaining lifespan of equipment subject to maintenance, The neural network includes (i) a first model that predicts the remaining lifespan at any given time in the maintenance cycle data, which is time-series operational data of the target equipment from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value. The weight parameters of the first model are updated to predict the remaining lifespan relative to the end of the maintenance cycle data, using the output of the first model and the output of the second model obtained from training data including multiple maintenance cycle data. The learning method used by neural networks executed by computers.

2. The updated weight parameters of the first model are used to update at least a portion of the weight parameters of the second model. The method for training a neural network according to claim 1.

3. The weight parameters of the second model are updated using the exponentially smoothed moving average of the updated weight parameters of the first model. The method for training a neural network according to claim 2.

4. The aforementioned training data is divided into a first data set and a second data set. Extract the first partial data included in the first data group and the second partial data included in the second data group that is similar to the first partial data. The weight parameters of the first model are updated so that the feature vector of the first partial data and the feature vector of the second partial data are similar to each other. A method for training a neural network according to any one of claims 1 to 3.

5. The learning data includes multiple supervised learning data sets, which are time-series operational data of the target device from a healthy state to a failure, and multiple unsupervised learning data sets, which are maintenance cycle data. The neural network is pre-trained using the supervised learning data, The weight parameters of the pre-trained neural network are set to the initial values ​​of the first and second models, and the weight parameters of the first model are updated to predict the remaining lifespan relative to the end of the maintenance cycle data using the outputs of the first and second models obtained from the unsupervised training data. A method for training a neural network according to any one of claims 1 to 3.

6. The learning data includes multiple supervised learning data sets, which are time-series operational data of the target device from a healthy state to a failure, and multiple unsupervised learning data sets, which are maintenance cycle data. When using the supervised learning data, the weight parameters of the first model are updated so that the predicted value predicted by the first model and the measured value of remaining lifetime indicated by the supervised learning data approach each other. A method for training a neural network according to any one of claims 1 to 3.

7. The learning data includes multiple supervised learning data sets, which are time-series operational data of the target device from a healthy state to a failure, and multiple unsupervised learning data sets, which are maintenance cycle data. The neural network is pre-trained using the supervised learning data, After setting the weight parameters of the pre-trained neural network to the initial values ​​of the first and second models, the weight parameters of the first model are updated to predict the remaining lifespan relative to the end of the maintenance cycle data, using the outputs of the first and second models obtained from the training data. When using the supervised learning data, the weight parameters of the first model are updated so that the predicted value predicted by the first model and the measured value of remaining lifetime indicated by the supervised learning data approach each other. A method for training a neural network according to any one of claims 1 to 3.

8. The remaining lifespan at the end of each maintenance cycle data is predicted using the first or second model in which the weight parameters have been updated. Based on the predicted remaining lifespan at the end of each maintenance cycle data, the remaining lifespan at each time point in each maintenance cycle data is corrected. A new machine learning model is trained to predict the remaining lifespan of the target device using training data that includes the corrected remaining lifespan as training data. A method for training a neural network according to any one of claims 1 to 3.

9. A neural network training method for predicting the remaining lifespan of equipment subject to maintenance, The neural network includes (i) a first model that predicts the remaining lifespan at any given time in the maintenance cycle data, which is time-series operational data of the target equipment from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value. The weight parameters of the first model are updated to predict the remaining lifespan relative to the end of the maintenance cycle data, using the output of the first model and the output of the second model obtained from training data including multiple maintenance cycle data. A computer program that instructs a computer to perform the learning process of a neural network.

10. A remaining life prediction system comprising a neural network that predicts the remaining life of equipment subject to maintenance, The neural network includes (i) a first model that predicts the remaining lifespan at any given time in the maintenance cycle data, which is time-series operational data of the target equipment from immediately after maintenance to immediately before the next maintenance, as a value based on an arbitrary reference value, and (ii) a second model that predicts the remaining lifespan at the final time in the maintenance cycle data as a value based on the reference value. The first model is trained by updating its weight parameters to predict the remaining life relative to the end of the maintenance cycle data, using the output of the first model and the output of the second model obtained from training data including a plurality of the maintenance cycle data. Remaining life prediction system.

Images