Signal processing device and signal processing method

Abstract

To provide a signal processing device and a signal processing method, capable of discovering unlearned noise.SOLUTION: According to an embodiment, a signal processing device includes an arithmetic unit and a determination unit. The arithmetic unit outputs feature data indicating a feature of a target signal when a signal measured from a target apparatus is input to a machine learning model. The machine learning model is learned to output the feature data by using as input a superposition signal obtained by superposing noise of an environment provided with the target apparatus on the known target signal generated by the target apparatus. The determination unit determines whether or not the input signal includes unlearned noise on the basis of the feature data.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] An embodiment of the present invention relates to a signal processing device and a signal processing method. [Background technology]

[0002] There is a known technique for measuring a signal emitted from a device and diagnosing the state of the device based on the signal. However, since there is noise in the environment in which the device is installed, the measurement signal measured by a sensor or the like contains both the target signal emitted from the device and noise. Therefore, denoising technology is being researched to remove noise from the measurement signal to obtain the target signal. In order to perform denoising with high accuracy, it is necessary to comprehensively learn the noises that may occur in the environment in which the device is installed. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 07-181218 Summary of the Invention [Problem to be solved by the invention]

[0004] On the other hand, noise in the installation environment may change, such as when a new noise source appears in the installation environment of the equipment. In this case, unlearned noise may remain in the signal after denoising, making it difficult to distinguish it from partial discharge. An object of the present invention is to provide a signal processing device and a signal processing method capable of discovering unlearned noise. [Means for solving the problem]

[0005] A signal processing device according to an embodiment includes a calculation unit and a determination unit. The calculation unit inputs a signal measured from a target device to a machine learning model, and outputs feature data representing features of a target signal. The machine learning model is trained to receive as input a superimposed signal in which noise in an environment in which the target device is installed and a known target signal generated by the target device are superimposed, and output the feature data. The determination unit determines whether or not the input signal includes unlearned noise based on the feature data. [Brief description of the drawings]

[0006] [Figure 1] FIG. 1 is a diagram showing the appearance of a signal processing device according to a first embodiment. [Diagram 2] FIG. 2 is a block diagram showing the software configuration of the signal processing device according to the first embodiment. [Diagram 3] FIG. 2 is a diagram showing an example of a ΦQN pattern according to the first embodiment. [Figure 4A] FIG. 11 is a diagram showing an example of a Φ QN pattern after denoising when a measured signal includes a learned partial discharge signal in the first embodiment. [Figure 4B] FIG. 11 is a diagram showing an example of a Φ QN pattern after denoising when an unlearned partial discharge signal is included in a measurement signal in the first embodiment. [Figure 4C] FIG. 11 is a diagram showing an example of a Φ QN pattern after denoising when a measurement signal contains a first unlearned noise in the first embodiment. [Figure 4D] FIG. 11 is a diagram showing an example of a Φ QN pattern after denoising when a measurement signal contains a second unlearned noise in the first embodiment. [Diagram 5] 5 is a diagram showing the distribution of the number of discharge pulses and the normalized standard deviation for each type of signal in the first embodiment. [Figure 6] 4 is a flowchart showing a learning process of the signal processing device according to the first embodiment. [Figure 7] 5 is a flowchart showing a monitoring process of the signal processing device according to the first embodiment. [Figure 8] 5A to 5C are diagrams showing experimental results of the unlearned noise determination method according to the first embodiment. [Figure 9] FIG. 2 is a schematic block diagram showing the configuration of a computer of the signal processing device according to the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0007] (First embodiment) FIG. 1 is a diagram showing the appearance of a signal processing device 100 according to the first embodiment. The signal processing device 100 according to the first embodiment monitors the occurrence of partial discharge in a target electric power device 200, which is an electric device to be monitored, and diagnoses the deterioration of the target electric power device 200. The signal processing device 100 diagnoses the deterioration of the target electric power device 200 by using a machine learning model. Noise is present in the installation environment of the target electric power device 200. The noise changes depending on the load situation, etc. Therefore, the signal processing device 100 monitors the partial discharge of the target electric power device 200 after learning the noise in the installation environment.

[0008] The target electric power device 200 includes a box and a device body. The device body is housed in the grounded box. The device body is a device that may generate partial discharge, such as a switch gear, a power transformer, a gas-insulated switchgear, a generator, an electric motor, a capacitor, or a reactor. The device body is composed of devices such as a circuit breaker, a disconnecting switch, a current transformer, or a voltage transformer. An electrode 111 is attached to the box of the target electric power device 200. The electrode 111 and the signal processing device 100 are connected by a cable. The signal processing device 100 measures the voltage of the electrode 111 to measure the electric potential formed on the surface of the box via the stray capacitance between the box and the device body. That is, the electrode 111 according to the first embodiment is a TEV (Transient Earth Voltage) sensor. On the other hand, the electrode 111 according to other embodiments may be another electrode 111 that measures an electromagnetic signal derived from the operation of an electric device, such as a CT (Current Transformer) sensor, an AE (Acoustic Emission) sensor, or an antenna. The electrode 111 may be installed inside the box of the target electric power device 200.

[0009] FIG. 2 is a block diagram showing the software configuration of the signal processing device 100 according to the first embodiment. The signal processing device 100 includes an acquisition unit 101, a target signal storage unit 102, a generation unit 103, a model storage unit 104, a feature estimation unit 105, a learning unit 106, a progress calculation unit 107, a state identification unit 108, an output control unit 109, and a judgment unit 110.

[0010] The acquiring unit 101 acquires a measurement signal measured by the electrode 111. The measurement signal may include a target signal, which is a signal generated from the target electric power device 200, and noise in the environment of the target electric power device 200.

[0011] The target signal storage unit 102 stores a known partial discharge signal, a ΦQN pattern of the partial discharge signal, and a cause of occurrence of the partial discharge signal in association with each other. The ΦQN pattern is data indicating the relationship between the generation phase Φ of a discharge pulse, the charge amount Q, and the occurrence frequency N. The generation phase Φ corresponds to the phase of the power supply voltage applied to the target electric device 200. FIG. 3 is a diagram showing an example of a ΦQN pattern according to the first embodiment. The ΦQN pattern is represented by a two-dimensional table, image data, or the like. For example, the ΦQN pattern may be image data in which the horizontal axis represents the generation phase Φ, the vertical axis represents the charge amount Q, and the pixel value represents the occurrence frequency N. The ΦQN pattern shown in FIG. 3 is quantized by dividing the generation phase into 20 parts and the charge amount into 16 parts. The ΦQN pattern is an example of feature data indicating the features of the target signal. Examples of causes of occurrence of partial discharge signals include voids and peeling.

[0012] The generation unit 103 generates a superimposed signal by superimposing a known partial discharge signal stored in the target signal storage unit 102 and a measurement signal acquired by the acquisition unit 101 when no partial discharge is occurring, i.e., noise.

[0013] The model storage unit 104 stores a feature extraction model, which is a machine learning model that generates a ΦQN pattern from a measurement signal, and a cause estimation model that estimates the cause of partial discharge from the ΦQN pattern. The machine learning model according to the first embodiment is a neural network model. Other examples of the machine learning model include a convolutional neural network model, a random forest model, and a support vector machine. The machine learning model has structural data and parameter data. The structural data specifies the connection relationship between nodes, which are the calculation elements of the model. The parameter data includes coefficients used in the calculation of each node, hyperparameters that set the behavior during learning, and the like. The machine learning model is learned by updating the parameter data.

[0014] The feature estimation unit 105 calculates ΦQN data by inputting the signal to the feature extraction model stored in the model storage unit 104. The feature estimation unit 105 is an example of a calculation unit that outputs feature data by inputting a signal measured from a target device to a machine learning model.

[0015] The learning unit 106 uses the first data set and the second data set to train the feature extraction model stored in the model storage unit 104. The first data set is a learning data set in which the superimposed signal generated by the generation unit 103 is used as an input sample, and a ΦQN pattern related to a partial discharge signal included in the superimposed signal is used as an output sample. The second data set is a learning data set in which the noise acquired by the acquisition unit 101 is used as an input sample, and a ΦQN pattern consisting of a predetermined pattern is used as an output sample. The predetermined pattern is a signal indicating that the input signal is noise, and is a pattern significantly different from the ΦQN pattern of the partial discharge signal. The predetermined pattern according to the first embodiment is a pattern in which all values ​​indicate zero (blank pattern). Examples of other patterns include a pattern in which all values ​​indicate a predetermined numerical value (e.g., 1), and a pattern in which predetermined numerical values ​​(e.g., 0 and 10) appear alternately. In another embodiment, the output data may be configured to include a ΦQN pattern and a noise flag indicating whether or not the signal is noise. In this case, the noise flag of the first data set is set to 0, and the noise flag of the second data set is set to 1. A noise flag indicating 1 is an example of a predetermined pattern.

[0016] The learning unit 106 trains the cause estimation model stored in the model storage unit 104 using the output data of the feature extraction model based on the first data set and the cause of partial discharge stored in the target signal storage unit 102. Specifically, the learning unit 106 trains the cause estimation model using a third data set, which is a learning data set in which the output data of the feature extraction model based on the first data set is used as an input sample and the cause of partial discharge associated with the partial discharge signal used to generate the first data set is used as an output sample.

[0017] The learning unit 106 inputs an input sample related to the learning dataset to the machine learning model, and updates the parameters by performing backpropagation calculations so as to reduce the difference between the calculation results of the machine learning model and the output sample.

[0018] The progress calculation unit 107 calculates the progress of learning of the feature extraction model based on the ΦQN pattern obtained by inputting the noise acquired by the acquisition unit 101 into the feature extraction model. Specifically, the progress calculation unit 107 calculates the rate at which the ΦQN pattern obtained by inputting the noise into the feature extraction model becomes a predetermined pattern as the progress of learning. The fact that the ΦQN pattern related to the noise shows a predetermined pattern indicates that the characteristics of the noise have been learned. On the other hand, the fact that the ΦQN pattern related to the noise does not show a predetermined pattern indicates that the noise having the characteristics has not yet been learned.

[0019] The state identifying unit 108 identifies the state of the target electric power device 200 based on the ΦQN pattern calculated by the feature estimation unit 105. Specifically, the state identifying unit 108 estimates the ΦQN pattern and the presence or absence of the occurrence of partial discharge. The state identifying unit 108 also infers the cause of the occurrence of partial discharge by inputting the ΦQN pattern to a cause estimation model. The inference result of the state identifying unit 108 is a vector representing the probability distribution of multiple occurrence causes.

[0020] The output control unit 109 controls the output of various information obtained by various processes performed by the signal processing device 100. For example, the output control unit 109 outputs the inference result by the state identification unit 108 and the feature quantity output from the feature quantity estimation unit 105 during inference to an external device that uses the inference result and the feature quantity, a display device, the model storage unit 104, and the like. The display device may be provided within the signal processing device 100.

[0021] The determination unit 110 determines whether or not unlearned noise is present in the measurement signal based on the ΦQN pattern calculated by the feature estimation unit 105. Specifically, the determination unit 110 obtains the number of discharge pulses and a value obtained by dividing the standard deviation of the number of discharge pulses for each phase by the total number of discharge pulses (hereinafter referred to as normalized standard deviation) from the ΦQN pattern, and determines whether or not unlearned noise is present in the measurement signal based on the number of discharge pulses and the normalized standard deviation. Note that in other embodiments, the determination unit 110 may make the determination using a degree of dispersion (variance, range, skewness, kurtosis, etc.) other than the standard deviation.

[0022] 4A to 4D are diagrams showing examples of a ΦQN pattern after denoising according to the first embodiment. Fig. 4A is a diagram showing an example of a ΦQN pattern after denoising when a learned partial discharge signal is included in a measurement signal in the first embodiment. As shown in Fig. 4A, when a learned partial discharge signal is included, many discharge pulses occur near 90 degrees and 270 degrees of the power supply voltage.

[0023] FIG. 4B is a diagram showing an example of a ΦQN pattern after denoising when the measurement signal includes an unlearned partial discharge signal in the first embodiment. FIG. 4C is a diagram showing an example of a ΦQN pattern after denoising when the measurement signal includes a first unlearned noise in the first embodiment. FIG. 4D is a diagram showing an example of a ΦQN pattern after denoising when the measurement signal includes a second unlearned noise in the first embodiment. As shown in FIG. 4B to FIG. 4D, when the measurement signal includes an unlearned signal, it can be seen that a discharge pulse exists over the entire phase. Therefore, at first glance, it seems unclear whether the measurement signal includes an unlearned partial discharge signal or an unlearned noise.

[0024] Here, the inventors have observed the ΦQN pattern after denoising and have obtained the following findings. It can be seen that the total number of discharge pulses is not large for the ΦQN pattern after denoising when the first unlearned noise is included as shown in FIG. 4C, but many discharge pulses exist near 0 degrees. On the other hand, it can be seen that the number of discharge pulses is large and the variation of the discharge pulses across the entire phase is not large for the ΦQN pattern after denoising when the second unlearned noise is included as shown in FIG. 4D. From this, the inventors have considered that it is possible to classify the distribution of discharge pulses appearing in the ΦQN pattern as being due to partial discharge signals or due to unlearned noise, based on the relationship between the number of discharge pulses and the degree of dispersion of the discharge pulses for each phase.

[0025] Fig. 5 is a diagram showing the distribution of the number of discharge pulses and the normalized standard deviation for each type of signal in the first embodiment. From the distribution shown in Fig. 5, it can be seen that unlearned noise is present mainly in a first region where the number of discharge pulses is smaller than the first threshold value of 0.0475 and the normalized standard deviation is larger than the second threshold value of 130, and in a second region where the number of discharge pulses is larger than the third threshold value of 0.125 and the normalized standard deviation is smaller than the fourth threshold value of 10. Therefore, the determination unit 110 determines that the measurement signal contains unlearned noise when the number of discharge pulses and the normalized standard deviation are located in the first region or the second region. The determination unit 110 also determines that the measurement signal contains a partial discharge signal when the number of discharge pulses and the normalized standard deviation are not located in either the first region or the second region. The measurement signal containing a partial discharge signal includes both the presence of an unlearned partial discharge signal and the presence of both unlearned noise and a known partial discharge signal.

[0026] 6 is a flowchart showing the learning process of the signal processing device 100 according to the first embodiment. A site administrator executes the learning process by the signal processing device 100 in advance before the operation of the target electric power device 200 to be monitored. That is, at the time when the machine learning model is learned by the signal processing device 100, the target signal of the target electric power device 200 is not generated. Note that the signal processing device 100 does not permit the monitoring process of the target electric power device 200 until the learning process is completed.

[0027] When the signal processing device 100 starts the learning process, the acquisition unit 101 acquires a plurality of measurement signals from the electrode 111 (step S1). The plurality of measurement signals are signals of a certain length extracted in synchronization with the AC power input to the target electric power device 200. For example, the measurement signal may be a signal having a length of an integer multiple of the period of the AC power, starting from a zero cross point of the rising edge of the AC power. The generation unit 103 generates a plurality of superimposed signals by superimposing a plurality of target signals stored in the target signal storage unit 102 on each of the acquired plurality of measurement signals (step S2). For example, when N measurement signals and M target signals are acquired, the generation unit 103 generates N×M superimposed signals. The learning unit 106 generates a first data set by associating each superimposed signal with a ΦQN pattern related to the target signal included in the superimposed signal and a cause of occurrence (step S3).

[0028] The feature estimation unit 105 generates a ΦQN pattern corresponding to each of the multiple superimposed signals (step S4) by inputting the multiple superimposed signals related to the first data set generated in step S3 to the feature extraction model stored in the model storage unit 104. The learning unit 106 generates a third data set by associating each ΦQN pattern generated in step S4 with the cause of occurrence related to the superimposed signal related to the ΦQN pattern (step S5).

[0029] The state identifying unit 108 inputs the generated multiple ΦQN patterns into the cause estimation model stored in the model storage unit 104, thereby calculating the probability distribution of the cause of partial discharge occurrence corresponding to each of the multiple ΦQN patterns (step S6).

[0030] The learning unit 106 updates the parameters of the feature extraction model so that the difference between the ΦQN pattern generated in step S3 due to the input sample and the output sample is small for each combination of the input sample (superimposed signal) and the output sample (ΦQN pattern) related to the first data set (step S7). That is, the learning unit 106 trains the feature extraction model based on the first data set. The learning unit 106 also updates the parameters of the cause estimation model so that the difference between the probability distribution of the occurrence cause calculated in step S6 due to the input sample and the output sample is small for each combination of the input sample (ΦQN pattern) and the output sample (occurrence cause) related to the third data set (step S8). That is, the learning unit 106 trains the cause estimation model based on the third data set.

[0031] Next, the learning unit 106 generates a second data set by associating a ΦQN pattern related to a predetermined pattern with the multiple measurement signals acquired in step S1 (step S9). The feature estimation unit 105 generates a ΦQN pattern corresponding to each of the multiple measurement signals by inputting the multiple measurement signals related to the second data set generated in step S9 to a feature extraction model stored in the model storage unit 104 (step S10). The learning unit 106 updates the parameters of the feature extraction model so that the difference between the ΦQN pattern generated in step S9 due to the input sample and the output sample is reduced for each combination of an input sample (measurement signal) and an output sample (ΦQN pattern) related to the second data set (step S11). In other words, the learning unit 106 trains the feature extraction model based on the second data set. Steps S1 to S11 correspond to a learning period of the signal processing device 100.

[0032] Next, the acquisition unit 101 acquires a plurality of measurement signals from the electrode 111 (step S12). The feature estimation unit 105 generates a ΦQN pattern corresponding to each of the plurality of measurement signals by inputting the plurality of measurement signals acquired in step S12 to a feature extraction model stored in the model storage unit 104 (step S13). The progress calculation unit 107 calculates the ratio of the plurality of ΦQN patterns generated in step S13 that match a predetermined pattern as the progress of learning (step S14). Here, the pattern "matches" may mean that all values ​​match completely, or that the distance (e.g., L2 norm, etc.) is less than a predetermined threshold. The progress calculation unit 107 determines whether the progress of learning exceeds a predetermined threshold (step S15). If the progress of learning does not exceed the predetermined threshold (step S15: NO), the process returns to step S1, and learning of the machine learning model continues. Steps S12 to S15 correspond to a trial period of the signal processing device 100.

[0033] On the other hand, if the progress of the learning exceeds the predetermined threshold (step S15: YES), the output control unit 109 notifies the administrator that the learning process has been completed (step S16), and permits the execution of the monitoring process.

[0034] 7 is a flowchart showing the monitoring process of the signal processing device 100 according to the first embodiment. When the progress of learning exceeds a predetermined threshold, the signal processing device 100 permits execution of the monitoring process. When execution of the monitoring process is permitted, the site administrator operates the target electric power device 200 that is the monitoring target, and starts the monitoring process by the signal processing device 100.

[0035] When the signal processing device 100 starts a monitoring process, the acquiring unit 101 acquires a measurement signal from the electrode 111 (step S101). The measurement signal is a signal of a certain length extracted in synchronization with the AC power input to the target electric power device 200. For example, the measurement signal may be a signal having a length of an integer multiple of the period of the AC power, starting from the zero cross point of the rising edge of the AC power.

[0036] The feature estimation unit 105 generates a ΦQN pattern from the acquired measurement signal using the trained feature extraction model (step S102). The output control unit 109 determines whether the generated ΦQN pattern matches a predetermined pattern (step S103). If the generated ΦQN pattern matches the predetermined pattern (step S103: YES), the measurement signal does not include a target signal, so the output control unit 109 outputs the estimated ΦQN pattern and a message indicating that no partial discharge has occurred (step S104).

[0037] If the generated ΦQN pattern does not match the predetermined pattern (step S103: NO), the judgment unit 110 calculates the number of discharge pulses and the normalized standard deviation from the ΦQN pattern (step S105). The judgment unit 110 judges whether the number of discharge pulses is smaller than the first threshold and the normalized standard deviation is larger than the second threshold, that is, whether the number of discharge pulses and the normalized standard deviation are in the first region (step S106). If the number of discharge pulses is smaller than the first threshold and the normalized standard deviation is larger than the second threshold (step S106: YES), the judgment unit 110 judges that the measurement signal contains unlearned noise (step S108). If the number of discharge pulses is greater than the first threshold or the normalized standard deviation is smaller than the second threshold (step S106: NO), the determination unit 110 determines whether the number of discharge pulses is greater than a third threshold and the normalized standard deviation is smaller than a fourth threshold, that is, whether the number of discharge pulses and the normalized standard deviation are in the second region (step S107).If the number of discharge pulses is greater than the third threshold and the normalized standard deviation is smaller than the fourth threshold (step S107: YES), the determination unit 110 determines that the measurement signal contains unlearned noise (step S108).

[0038] When the determination unit 110 determines that unlearned noise exists, the output control unit 109 notifies the administrator of the restart of the learning process (step S109). Then, the signal processing device 100 starts re-learning the feature extraction model by the learning process shown in FIG. On the other hand, when the discharge pulse count and the normalized standard deviation are not present in either the first region or the second region (step S107: NO), the judgment unit 110 judges that the measurement signal includes a partial discharge signal (step S109). When the judgment unit 110 judges that a partial discharge signal is present, the state identification unit 108 calculates the probability distribution of the cause of partial discharge from the ΦQN pattern generated in step S102 using the learned cause estimation model (step S110). The output control unit 109 outputs the estimated ΦQN pattern and the probability distribution of the cause of occurrence (step S111). The output control unit 109 outputs the estimated ΦQN pattern and the probability distribution of the cause of occurrence to, for example, a display device. The administrator can judge the validity of the cause of occurrence inferred by the state identification unit 108 by referring to the displayed ΦQN pattern. Steps S101 to S111 correspond to a monitoring period of the signal processing device 100.

[0039] FIG. 8 is a diagram showing an experimental result of the unlearned noise determination method according to the first embodiment. Using the judgment flow shown in Fig. 7, we distinguished between partial discharge signals acquired by the TEV sensor, noise signals acquired from the on-site switchgear, and a measurement signal in which these were mixed. As a result, we were able to achieve a correct answer rate of over 90% for both partial discharge judgment and noise judgment, as shown in Fig. 8.

[0040] FIG. 9 is a schematic block diagram showing the configuration of a computer of the signal processing device 100 according to the first embodiment. The computer 500 includes a processor 510, a main memory 520, a storage 530, and an interface 540. The above-mentioned signal processing device 100 is implemented in a computer 500. The operations of the above-mentioned processing units are stored in the storage 530 in the form of a program. The processor 510 reads the program from the storage 530, loads it in the main memory 520, and executes the above-mentioned processing in accordance with the program. The processor 510 also secures storage areas in the main memory 520 corresponding to the above-mentioned storage units in accordance with the program. Examples of the processor 510 include a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), and a microprocessor.

[0041] The program may be for realizing a part of the functions to be performed by the computer 500. For example, the program may be for realizing the functions by combining with other programs already stored in the storage or by combining with other programs implemented in other devices. In another embodiment, the computer 500 may include a custom LSI (Large Scale Integrated Circuit) such as a PLD (Programmable Logic Device) in addition to the above configuration or instead of the above configuration. Examples of PLDs include PAL (Programmable Array Logic), GAL (Generic Array Logic), CPLD (Complex Programmable Logic Device), and FPGA (Field Programmable Gate Array). In this case, a part or all of the functions realized by the processor 510 may be realized by the integrated circuit. Such an integrated circuit is also included in an example of a processor. In another embodiment, the computer 500 may be virtualized on one or more computers.

[0042] Examples of storage 530 include a magnetic disk, a magneto-optical disk, an optical disk, and a semiconductor memory. Storage 530 may be an internal medium directly connected to the bus of computer 500, or an external medium connected to computer 500 via interface 540 or a communication line. In addition, when this program is distributed to computer 500 via a communication line, computer 500 that receives the program may load the program in main memory 520 and execute the above-mentioned process. In at least one embodiment, storage 530 is a non-transitory tangible storage medium.

[0043] The program may be for realizing part of the above-mentioned functions. Furthermore, the program may be a so-called difference file (difference program) that realizes the above-mentioned functions in combination with other programs already stored in storage 530.

[0044] As described above, the signal processing device 100 according to the first embodiment has a feature estimation unit 105 and a determination unit 110. The feature estimation unit 105 outputs feature data representing features of a target signal by inputting a signal measured from the target electric power device 200 to a feature extraction model. The feature extraction model is trained to output a ΦQN pattern by inputting a superimposed signal obtained by superimposing noise in the environment in which the target electric power device 200 is installed and a known target signal generated by the target electric power device 200. The determination unit 110 determines whether or not the input signal contains unlearned noise based on the ΦQN pattern. This allows the signal processing device 100 to find unlearned noise.

[0045] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents described in the claims, as well as in the scope and spirit of the invention.

[0046] The signal processing device 100 according to the first embodiment performs learning and inference of the machine learning model by a single device, but is not limited thereto. For example, in another embodiment, a learning device that performs learning of the machine learning model and a monitoring device that monitors the target electric power device 200 may be separately provided. The learning device is an example of a signal processing device. The monitoring device performs monitoring processing using the machine learning model learned by the learning device. The learning device and the monitoring device may be connected by a network, for example, and the trained machine learning model may be transmitted via the network. In another embodiment, the signal processing device 100 may be provided remotely from the site where the target electric power device 200 is provided. In this case, the signal processing device 100 may be provided near the target electric power device 200, receive an electromagnetic signal via a network or the like from an edge device that acquires the electromagnetic signal, and monitor partial discharge based on the electromagnetic signal.

[0047] The signal processing device 100 according to the first embodiment generates a ΦQN pattern as feature data of a target signal, but is not limited thereto. For example, a feature extraction model according to another embodiment may be trained to input a measurement signal and output a spectrogram of the target signal. Also, for example, a feature extraction model according to another embodiment may be trained to input a measurement signal and output a denoised target signal. The target signal and the spectrogram of the target signal are examples of feature data representing the features of the target signal.

[0048] The signal processing device 100 according to the first embodiment trains the feature extraction model so as to output a predetermined pattern when noise is input, but is not limited to this. For example, in another embodiment, the signal processing device 100 may be trained only by the first data set without using the second data set. The signal processing device 100 may calculate, as the progress, the rate at which the energy amount of the ΦQN pattern generated by inputting noise into the feature extraction model becomes equal to or less than a predetermined threshold.

[0049] The signal processing device 100 according to the first embodiment learns both the feature extraction model and the cause estimation model, but is not limited to this. For example, the cause estimation model according to other embodiments may be learned in advance based on a known ΦQN pattern and a cause of partial discharge occurrence.

[0050] The signal processing device 100 according to the first embodiment estimates the cause of partial discharge based on the ΦQN pattern, but is not limited to this. For example, the signal processing device 100 according to another embodiment may output a feature such as a ΦQN pattern and not estimate the cause of the occurrence. When the ΦQN pattern is output, the administrator determines the presence or absence of partial discharge in the target electric power device 200 from the output ΦQN pattern.

[0051] The signal processing device 100 according to the first embodiment monitors partial discharge of electric power equipment, but is not limited thereto. For example, the signal processing device 100 according to another embodiment may monitor abnormal vibration of a rotating machine such as a motor or a generator. That is, the target signal may be a signal representing vibration of the rotating machine. Also, for example, the signal processing device 100 according to another embodiment may monitor abnormal vibration of a reactor. In this case, the target signal may be a signal representing vibration of the reactor or a sound wave signal emitted from the reactor.

[0052] The signal processing device 100 according to the first embodiment determines whether or not unlearned noise is included by determining whether or not the number of discharge pulses and the normalized standard deviation are in the first region or the second region shown in Fig. 5, but is not limited thereto. For example, the signal processing device 100 according to another embodiment may perform the determination using a trained model that determines whether or not unlearned noise is included from a combination of the number of discharge pulses and the normalized standard deviation. Furthermore, the signal processing device 100 according to another embodiment may perform the determination using a trained model that determines whether or not unlearned noise is included directly from the ΦQN pattern.

[0053] According to at least one of the embodiments described above, the signal processing device has a calculation unit and a determination unit. The calculation unit inputs a signal measured from a target device to a machine learning model, and outputs feature data representing features of a target signal. The machine learning model is trained to output feature data using a superimposed signal obtained by superimposing noise in the environment in which the target device is installed and a known target signal generated by the target device. The determination unit determines whether or not the input signal contains unlearned noise based on the feature data. This allows the signal processing device to discover unlearned noise. [Explanation of symbols]

[0054] 100: signal processing device 101: acquisition unit 102: target signal storage unit 103: generation unit 104: model storage unit 105: feature estimation unit 106: learning unit 107: progress calculation unit 108: state identification unit 109: output control unit 110: judgment unit 111: electrode 200: target electric power device 500: computer 510: processor 520: main memory 530: storage 540: interface

Claims

1. a calculation unit that receives as input a superimposed signal obtained by superimposing noise from an environment in which a target device is installed and a known target signal generated by the target device, and outputs feature data representing features of the target signal by inputting a signal measured from the target device into a machine learning model that has been trained to output the feature data; a determination unit that determines whether or not the input signal contains unlearned noise based on the feature data; A signal processing device comprising:

2. A learning unit that, when it is determined that the signal contains unlearned noise, trains the machine learning model using new noise. The signal processing device according to claim 1 .

3. the characteristic data represents a relationship between a phase of a power supply voltage applied to the target device and an occurrence frequency of a discharge pulse; The determination unit determines whether or not the input signal includes unlearned noise based on the number of discharge pulses in the feature data and a distribution degree of the number of discharge pulses for each phase. The signal processing device according to claim 1 .

4. The determination unit determines whether or not the input signal includes unlearned noise based on a relationship between the number of discharge pulses in the feature data and a value obtained by dividing a standard deviation of the number of discharge pulses for each phase by a total number of the discharge pulses. The signal processing device according to claim 3 .

5. The determination unit determines that the input signal contains unlearned noise when the number of discharge pulses in the feature data is smaller than a first threshold value and the value obtained by dividing the standard deviation by the total number is larger than a second threshold value, or when the number of discharge pulses in the feature data is larger than a third threshold value that is larger than the first threshold value and the value obtained by dividing the standard deviation by the total number is smaller than a fourth threshold value that is smaller than the second threshold value. The signal processing device according to claim 4.

6. The input signal is measured by a TEV sensor provided in the target device. The signal processing device according to claim 1 .

7. a step of inputting a superimposed signal obtained by superimposing noise in an environment in which a target device is installed and a known target signal generated by the target device into a machine learning model trained to output feature data representing features of the target signal, by inputting a signal measured from the target device into the machine learning model, and outputting the feature data; determining whether or not the input signal contains unlearned noise based on the feature data; A signal processing method comprising the steps of:

Images