Circuit breaker model generation system, circuit breaker model generation device, and circuit breaker model generation method

Abstract

To efficiently generate the operation model of a breaker.SOLUTION: A breaker model generation system comprises: a common model training unit 114 that outputs a common model 134 by common model training S1 that is training using as input data a feature quantity tensor generated from acoustic data acquired from first breakers which are a plurality of installed breakers 201a and 201b, and using as teaching data the activity ratio associated to a feature quantity tensor 411; and an individual model training unit 115 that outputs an individual model 136 by individual model training S2 that is training using as input data the feature quantity tensor generated on the basis of the acoustic data acquired from a breaker 201c different from the breakers 201a and 201b, and using as teaching data the activity ratio associated to the feature quantity tensor. In the individual model training S2, the common model 134 outputted as the result of the common model training S1 is used as an initial value.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a circuit breaker model generation system, a circuit breaker model generation apparatus, and a circuit breaker model generation method. [Background technology]

[0002] There is a growing need to perform maintenance on circuit breakers without dispatching workers by remotely monitoring their status. To meet this need, methods have been proposed that involve attaching various sensors to the circuit breakers to understand their operating status and determining their condition based on the difference from the normal state. In such diagnostics, attempts are being made to monitor the condition by installing current sensors inside the circuit breaker or attaching acceleration sensors to the circuit breaker housing.

[0003] One technique for attaching sensors and making decisions based on acquired data is described in Patent Document 1. Patent Document 1 discloses a power conversion system and circuit breaker diagnostic device comprising: a power conversion circuit that converts input DC power to AC power; output wiring that transmits the power converted by the power conversion circuit; and an AC circuit breaker provided in the middle of the output wiring and configured to switch on and off based on the voltage value of the input DC power; a sound detector configured to detect the sound emitted by the AC circuit breaker when the AC circuit breaker switches on and off; and a diagnostic device that diagnoses whether the AC circuit breaker that emitted the sound is abnormal based on the sound detected by the sound detector and predetermined judgment criterion information. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2019-184341 [Overview of the project] [Problems that the invention aims to solve]

[0005] However, the technology described in Patent Document 1 requires the circuit breaker to be operated and a large amount of data collected under normal conditions when applied to newly installed circuit breakers. However, as the number of operations increases, the circuit breaker deteriorates, and the amount of work required to collect the data becomes a challenge.

[0006] In light of this background, the present invention was made, and its objective is to efficiently generate an operating model of a circuit breaker. [Means for solving the problem]

[0007] To solve the aforementioned problems, the present invention includes: a sound collection device that collects the operating sound of a circuit breaker that interrupts power and outputs it as acoustic data; a first learning unit that outputs a first learning model by first learning, which is learning that takes first feature quantities, which are feature quantities generated from first acoustic data, which are acoustic data acquired from a first circuit breaker, which is one of a plurality of installed circuit breakers, as input data and uses a first acoustic model, which is an acoustic model associated with the first feature quantities, as training data; and a second learning unit that outputs a second learning model by second learning, which is learning that takes second feature quantities, which are feature quantities generated based on second acoustic data, which are acoustic data acquired from a second circuit breaker, which is a circuit breaker different from the first circuit breaker, as input data and uses a second acoustic model, which is an acoustic model associated with the second feature quantities, as training data, wherein the acoustic model is The activity level is expressed as a triangular wave that rises perpendicular to the time axis starting from the start time and end time of the circuit breaker's closing operation, the operating sound emitted when the circuit breaker's closing operation begins, the operating sound emitted when the circuit breaker's opening operation begins, and the operating sound emitted when the circuit breaker's opening operation ends, and is generated for each of these, and is expressed as a triangular wave that rises perpendicular to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively, and then decays with a predetermined slope, or a rectangular wave that rises perpendicular to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively. The second learning process is characterized by using the first learning model output as a result of the first learning process as the initial value of the second learning model in the second learning process. Other solutions will be described as appropriate in the embodiments. [Effects of the Invention]

[0008] According to the present invention, an operating model of a circuit breaker can be efficiently generated. [Brief explanation of the drawing]

[0009] [Figure 1] It is a diagram schematically showing circuit breaker diagnosis performed by a circuit breaker diagnosis system. [Figure 2] It is a diagram showing details of collection of acoustic data and stroke position data used in the present embodiment. [Figure 3] It is a diagram showing a configuration example of a diagnosis device according to the present embodiment. [Figure 4] It is a diagram showing each timing during a closing operation of a circuit breaker. [Figure 5] It is a diagram showing each timing during a closing operation of a circuit breaker. [Figure 6] It is a diagram showing an example of a generation procedure of a feature quantity tensor in common model learning and individual model learning. [Figure 7] It is a diagram showing a correspondence relationship between a feature quantity tensor and activity. [Figure 8] It is a diagram showing processing in a common model learning unit performed in the present embodiment. [Figure 9] is a diagram showing processing in an individual model learning unit performed in the present embodiment. [Figure 10] It is a diagram showing a processing procedure of an acoustic diagnosis unit performed in the present embodiment. [Figure 11] It is a diagram showing a configuration of common learning data and individual learning data. [Figure 12] It is a diagram showing a data configuration of teacher data in a common model. [Figure 13] It is a diagram showing a specific example of a data configuration of teacher data. [Figure 14] It is a diagram showing an example of a diagnosis result display screen output to an output device in the present embodiment. [Figure 15] It is a diagram showing another example of activity. DETAILED DESCRIPTION OF THE INVENTION

[0010] Next, modes for carrying out the present invention (referred to as "embodiments") will be described in detail with appropriate reference to the drawings.

[0011] <Overview> Figure 1 shows a schematic diagram of circuit breaker diagnosis performed by the circuit breaker diagnostic system (circuit breaker operation model generation system) Z. The circuit breaker diagnostic system Z includes a microphone (sound collection device) 202, a stroke position measuring device (contact position measuring device) 203, a common model learning unit 114, an individual model learning unit 115, and an acoustic diagnostic unit 118. Furthermore, the diagnostic process performed by the circuit breaker diagnostic system Z is carried out in three stages: common model learning (S1), individual model learning (S2), and diagnosis (S3).

[0012] (Common model learning (S1: First learning, first learning step)) Circuit breaker 201 is used to interrupt the power supply.

[0013] Furthermore, as shown in Figure 1, in common model learning (S1), microphone 202 is installed near circuit breaker 201a. Microphone 202 then converts the sound pressure during opening and closing of circuit breaker 201a (201: first circuit breaker) into an electrical signal. The converted signal is output as acoustic data 301 (see Figures 4 and 5). In this way, microphone 202 collects the operating sound of circuit breaker 201 and outputs it as acoustic data 301. In other words, acoustic data 301 is data of the operating sound of circuit breaker 201.

[0014] The stroke position measuring device 203 then measures the position of the contacts during the opening and closing operation using a laser or the like. Specifically, the stroke position measuring device 203 measures the distance of the opening and closing rod (not shown) of the circuit breaker 201a. In this way, the stroke position measuring device 203 measures the position of the contacts of the circuit breaker and outputs it as stroke position data 302 (contact position data, see Figures 4 and 5).

[0015] Furthermore, the sound collection by the microphone 202 when the circuit breaker 201a is opened and closed, and the measurement of the stroke position by the stroke position measuring device 203 are performed synchronously. The sound data (first sound data) 301, which is the sound data collected by the microphone 202, and the stroke position data 302, which is the stroke position data, are sent to the common model learning unit 114 as common learning data 133.

[0016] Any method of acquisition is acceptable as long as the acoustic data 301 and stroke position data 302 are acquired synchronously. For example, sound collection by the microphone 202 and measurement by the stroke position measuring device 203 may be performed continuously. Alternatively, sound collection by the microphone 202 and measurement by the stroke position measuring device 203 may start when an open or closed signal of the circuit breaker 201 is detected. Alternatively, when the opening or closing operation of the circuit breaker 201 is signaled by a person, the switches of the microphone 202 and the stroke position measuring device 203 may be manually turned ON.

[0017] Similarly, from a circuit breaker 201b (201: first circuit breaker), which is a different model from circuit breaker 201a, the opening and closing operation sound of circuit breaker 201b is measured synchronously by the microphone 202, and the stroke position is measured synchronously by the stroke position measuring device 203. The measured acoustic data (first acoustic data) 301 and stroke position data 302 of circuit breaker 201b are sent to the common model learning unit 114 as common learning data 133. In addition, acoustic data 301 and stroke position data 302 may also be measured for circuit breakers 201a of the same model but with different operating conditions. Different operating conditions may include, for example, different types of oil being used.

[0018] Furthermore, the circuit breakers 201a and 201b from which the common training data 133 is collected during common model training (S1) are circuit breakers 201 that are known not to have experienced any abnormalities.

[0019] The microphone 202 and stroke position measuring device 203 installed near the circuit breaker 201b may be different from or the same as the microphone 202 and stroke position measuring device 203 installed near the circuit breaker 201a.

[0020] In the common model learning (S1) shown in Figure 1, two sets of circuit breaker 201, microphone 202, and stroke position measuring device 203 are provided. However, the system is not limited to this; acoustic data 301 and stroke position data 302 may be measured for three or more sets of circuit breaker 201, microphone 202, and stroke position measuring device 203. In this way, in common model learning (S1), acoustic data 301 and stroke position data 302 are acquired from multiple installed circuit breakers 201.

[0021] The common model learning unit (first learning unit) 114 then learns the common model (first learning model) 134 using the common learning data 133 acquired from circuit breakers 201a and 201b, respectively. Specifically, the common model 134 is the connection strength between neurons in the deep neural network (neural network, first neural network) N1 (see Figure 8). The learning of the common model 134 will be described later. The common model 134 is the learning result by the common model learning unit 114.

[0022] (Individual model learning (S2: Second learning)) Next, when a new circuit breaker 201c (201: a second circuit breaker which is a different circuit breaker 201 from the first circuit breaker) is installed, individual learning (S2) is performed by the individual model learning unit (second learning unit) 115. First, an individual model (second learning model) 136 is created to diagnose the opening and closing operation of the newly installed circuit breaker 201c. Circuit breaker 201c is the same circuit breaker 201 that will be diagnosed later. Therefore, a microphone 202 installed on circuit breaker 201c collects sound when circuit breaker 201c opens and closes. In addition, a stroke position measuring device 203 installed on circuit breaker 201c measures the stroke position. The sound collection and stroke position measurement of circuit breaker 201c are performed synchronously. The acoustic data (second acoustic data) 301 of the sound collected by microphone 202 and the stroke position data 302 measured by stroke position measuring device 203 are passed to the individual model learning unit 115 as individual learning data 135. Note that the acoustic data 301 and stroke position data 302 collected as individual learning data 135 can be fewer in number compared to the common learning data 133.

[0023] Furthermore, the circuit breaker 201c from which the individual training data 135 is collected during individual model training (S2) is the circuit breaker 201, which is known not to have experienced any abnormalities.

[0024] The individual model learning unit 115 learns the individual model 136 by inputting the individual learning data 135 into the deep neural network (neural network, second neural network) N2 (see Figure 9), which uses the model (common model 134) already learned by the common model learning unit 114 as its initial value. The learning of the individual model 136 will be described later. The individual model 136 is the learning result by the individual model learning unit 115, and specifically, it is the connection strength of neurons in the deep neural network N2.

[0025] (Diagnosis (S3)) After training the individual model 136, the acoustic diagnostic unit (diagnostic unit) 118 performs a diagnosis of the circuit breaker 201c (S3). The acoustic diagnostic unit 118 corresponds to the activity estimation processing unit 116 and the state diagnosis processing unit 117 in Figure 3. When diagnosing the circuit breaker 201c, the acoustic diagnostic unit 118 uses acoustic data (third acoustic data, which is different from the second acoustic data) 301, which is sound data collected by the installed microphone 202, and the individual model 136 to diagnose the circuit breaker 201c. First, the acoustic diagnostic unit 118 inputs the acquired acoustic data 301 into the individual model 136. Based on the output results output from the individual model 136, the acoustic diagnostic unit 118 diagnoses the operation of the circuit breaker 201c.

[0026] <Details of the collection of acoustic data 301 and stroke position data 302> Figure 2 shows the details of the collection of acoustic data 301 and stroke position data 302 used in this embodiment. Refer to Figure 1 as appropriate. The collection of acoustic data 301 and stroke position data 302 in Figure 2 is a common process for collecting common training data 133 in common model training (S1) and for collecting individual training data 135 in individual model training (S2) in Figure 1.

[0027] As described above, a microphone 202 for collecting sound from the circuit breaker 201 and a stroke position measuring device 203 are installed near the circuit breaker 201 (distance L1). The logger 221 synchronously acquires acoustic data 301 (see Figures 4 and 5) output from the microphone 202 and stroke position data 302 (see Figures 4 and 5) output from the stroke position measuring device 203. Furthermore, the logger 221 converts the acquired acoustic data 301 and stroke position data 302 into digital signals and outputs them to the diagnostic device 1. The diagnostic device 1 includes a common model learning unit 114, an individual model learning unit 115, and an acoustic diagnostic unit 118, as shown in Figure 1. The diagnostic device 1 can also store common learning data 133, a common model 134, individual learning data 135, and an individual model 136 in an auxiliary storage device 130 (see Figure 3).

[0028] The diagnostic device 1 stores the digital signals of the acoustic data 301 and stroke position data 302 received from the logger 221 in the auxiliary storage device 130 as common learning data 133 and individual learning data 135, as shown in Figure 1.

[0029] Thus, the system shown in Figure 2 can be used to collect common learning data 133 when generating a common model 134. Furthermore, the system shown in Figure 2 can also be used to collect a small amount of individual learning data 135 when generating an individual model 136 for a new circuit breaker 201c. In addition, for diagnosis (S3 in Figure 1), the system shown in Figure 2 can be used, except that the stroke position measuring device 203 is not used.

[0030] <Diagnostic device 1> Figure 3 shows an example of the configuration of the diagnostic device (circuit breaker operation model generation device) 1 according to this embodiment. The diagnostic device 1 includes a main memory 100 composed of RAM (Random Access Memory), and an auxiliary memory 130 composed of HDD (Hard Disk Drive) or SDD (Solid State Drive). Furthermore, the diagnostic device 1 includes a central processing unit 101 composed of a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The diagnostic device 1 also includes an input device 102 composed of a keyboard or mouse, an output device (display device) 103 composed of a display, and a communication device 104 for communicating with the logger 221. Note that in Figure 3, the logger 221 is shown with a dashed line to indicate that in this embodiment, the logger 221 is not an element that constitutes the diagnostic device 1.

[0031] (Auxiliary storage device 130) The auxiliary storage device 130 stores the diagnostic software program 131, microphone position information 132, common learning data 133, common model 134, individual learning data 135, individual model 136, diagnostic parameters 137, and diagnostic history information 138. Diagnostic software program 131 is a program for executing diagnostic software 110. The microphone position information 132 is information regarding the positional relationship between the microphone 202 and the circuit breaker 201 (distance L1 in Figure 2). The common learning data 133 is the data used for common model learning (S1 in Figure 1) by the common model learning unit 114. The common model 134 is the model output by the common model learning unit 114 (S1 in Figure 1). As mentioned above, the common model 134 specifically represents the connection strength of each neuron in the neural network that constitutes the common model 134.

[0032] Individual training data 135 is data used in individual model training by the individual model training unit 115 (S2 in Figure 1). Individual model 136 is a model output by individual model learning (S2 in Figure 1) performed by the individual model learning unit 115. As described above, individual model 136 specifically represents the connection strength of each neuron in the neural network that constitutes individual model 136.

[0033] Diagnostic parameter 137 is a parameter used for diagnosis. For example, diagnostic parameter 137 could be a threshold value indicating what percentage deviation from the rated value the average of the estimated opening start timing and opening end timing will be considered abnormal. The diagnostic history information 138 is the diagnostic results 211 performed by the status diagnostic processing unit 117 to date.

[0034] The main memory 100 is equipped with diagnostic software 110 and a work area 120. (Diagnostic software 110) The diagnostic software 110 is realized when the diagnostic software program 131 stored in the auxiliary storage device 130 is loaded into the main storage device 100 and executed by the main storage device 100. The diagnostic software 110 includes an overall control processing unit 111, a spectrogram generation unit (feature generation unit) 112, a feature extraction processing unit (feature setting unit, feature generation unit) 113, a common model learning unit 114, an individual model learning unit 115, an activity estimation processing unit (diagnosis unit) 116, and a state diagnosis processing unit (diagnosis unit) 117.

[0035] The overall control processing unit 111 manages the control of the spectrogram generation unit 112, the feature extraction processing unit 113, the common model learning unit 114, the individual model learning unit 115, the activity level estimation processing unit 116, and the state diagnosis processing unit 117.

[0036] The spectrogram generation unit 112 generates a spectrogram 401 (see Figure 6) from the acoustic data 301 acquired from the microphone 202. The spectrogram 401 has time on the horizontal axis and frequency on the vertical axis. Details of the spectrogram 401 will be described later. The feature extraction processing unit 113 extracts a feature tensor (features) 411 (Figures 6 and 7) from the spectrogram 401 generated by the spectrogram generation unit 112. The feature tensor 411 is obtained by extracting the spectrogram 401 at a predetermined time interval. Details of the feature tensor 411 will be described later. Incidentally, a tensor is represented as a multidimensional array and includes vectors and matrices.

[0037] The common model learning unit 114 generates a common model 134 by inputting a feature tensor 411 obtained from acoustic data 301 acquired from multiple types of circuit breakers 201 with different operating conditions into a neural network and learning from it. The individual model learning unit 115 inputs a feature tensor 411 obtained from acoustic data 301 when the circuit breaker 201 to be diagnosed is normal into the common model 134 and generates an individual model 136 by learning from it.

[0038] The activity level estimation processing unit 116 obtains an estimated activity level 681 (see Figure 10), which is the result of inputting the feature tensor 411 obtained from the acoustic data 301 to be diagnosed into the individual model 136. The estimated activity level 681 is output as a result of inputting the acoustic data 301 into the individual model 136. The state diagnosis processing unit 117 performs a state diagnosis of the circuit breaker 201, which is the target of the diagnosis, based on the estimated activity level 681 obtained by the activity level estimation processing unit 116.

[0039] Incidentally, the activity level estimation processing unit 116 and the state diagnosis processing unit 117 constitute the acoustic diagnosis unit 118 in Figure 1.

[0040] (Work area 120) Furthermore, the work area 120 is an area used as a temporary storage unit when the diagnostic software 110 is running. The work area 120 includes an acoustic data storage area 121, a spectrogram storage area 122, a feature data storage area 123, a model storage area 124, an estimated activity level storage area 125, a diagnostic parameter storage area 126, and a diagnostic result storage area 127. The acoustic data storage area 121 is an area where acoustic data 301 acquired from the microphone 202 is temporarily stored. Spectrogram 401 is temporarily stored in spectrogram storage area 122. The feature tensor 411 is temporarily stored in the feature storage area 123. The model storage area 124 temporarily stores common models 134 and individual models 136.

[0041] The estimated activity level storage area 125 temporarily stores the estimated activity level 681 estimated by the individual model 136. The diagnostic parameter storage area 126 temporarily stores the diagnostic parameters 137 used when diagnosing the circuit breaker 201. The diagnostic result storage area 127 temporarily stores the diagnostic result 211 from the state diagnostic processing unit 117.

[0042] <Closing operation> Figure 4 shows the timings for each step during the closing operation of the circuit breaker 201 (see Figures 1 and 2). Figure 4 shows, from top to bottom, acoustic data 301a (301) and stroke position data 302b (302). Furthermore, Figure 4 shows the activity level 310 at the start of closing operation (activity level 311 at the start of closing operation), the activity level 310 at the end of closing operation (activity level 312 at the end of closing operation), the activity level 310 at the start of opening operation (activity level 313 at the start of opening operation), and the activity level 310 at the end of opening operation (activity level 314 at the end of opening operation). In addition, in Figure 4, each graph is aligned on the time axis. Note that activity level 310 is an acoustic model that models the sound generated when the circuit breaker 201 operates.

[0043] Furthermore, the activity level 311 at the start of closing operation is the activity level 310 generated for the operating sound emitted when the closing operation of the circuit breaker 201 begins. The activity level 312 at the end of closing operation is the activity level 310 generated for the operating sound emitted when the closing operation of the circuit breaker ends. Also, the activity level 313 at the start of opening operation is the activity level 310 generated for the operating sound emitted when the opening operation of the circuit breaker 201 begins. And the activity level 314 at the end of opening operation is the activity level 310 generated for the operating sound emitted when the opening operation of the circuit breaker ends.

[0044] As mentioned above, the activity level 310 is an acoustic model that models the sound generated when the circuit breaker 201 is closed, opened, or otherwise operated by the circuit breaker. Two types of activity levels 310 are set for the closing operation of the circuit breaker 201: activity level 311 at the start of the closing operation and activity level 312 at the end of the closing operation. This is because, in this embodiment, two sounds are generated when the circuit breaker 201 is closed.

[0045] Furthermore, two types of activity levels 310 are set for the opening operation of the circuit breaker 201: an activity level 313 at the start of the opening operation and an activity level 314 at the end of the opening operation. This is because two sounds are generated when the circuit breaker 201 is opened.

[0046] Therefore, the activity level 310 should be set according to the sound generated by the circuit breaker 201 during closing and opening operations. For example, if one sound is generated during closing and one sound during opening, then one type of activity level 310 should be provided for each. If three or more sounds are generated, then three types of activity levels 310 should be set for both closing and opening operations. Furthermore, if the number of sounds generated differs between closing and opening operations, a different number of activity levels 310 may be set. For example, if three sounds are generated during closing and two sounds are generated during opening, then three types of activity levels 310 should be set for closing and two types of activity levels 310 should be set for opening.

[0047] Furthermore, since Figure 4 shows the closing operation, the activity levels 313 at the start of the opening operation and 314 at the end of the opening operation are inactive. Note that the dashed lines in the graphs for activity levels 311 at the start of the closing operation, 312 at the end of the closing operation, 313 at the start of the opening operation, and 314 at the end of the opening operation indicate the peak values ​​of activity level 310 for each.

[0048] As described above, the acoustic data 301a is stored in synchronization with the stroke position data 302a. From the start time 321 and end time 322 of the change in the stroke position data 302a, the true values ​​of the start and end times of the closing operation can be determined. As a result, the activity level 311 at the start of the closing operation is set to rise at the start time of the closing operation (start time 321 of the change in the stroke position data 302a) and to decrease at a constant rate. The activity level 311 at the start of the closing operation, set in this way, serves as a model of the acoustic intensity at the start of the closing operation. Similarly, the activity level 312 at the end of the closing operation is also set to rise at the end time (end time 322 of the change in the stroke position data 302a) and to decrease at a constant rate. The activity level 312 at the end of the closing operation, set in this way, serves as a model of the acoustic intensity at the end of the closing operation.

[0049] As shown in Figure 4, each of the activity levels 310 is represented as a triangular wave that rises perpendicularly to the time axis, starting from the start time and end time of the closing or opening operation (start time of operation), and then decays at a predetermined slope.

[0050] <Opening operation> Figure 5 shows the timings for each step during the closing operation of the circuit breaker 201 (see Figures 1 and 2). In Figure 5, as in Figure 4, acoustic data 301b and stroke position data 302b are shown from top to bottom. Furthermore, Figure 5 shows the activity level at the start of the closing operation 310 (activity level at the start of the closing operation 311), the activity level at the end of the closing operation 310 (activity level at the end of the closing operation 312), the activity level at the start of the opening operation 310 (activity level at the start of the opening operation 313), and the activity level at the end of the opening operation 310 (activity level at the end of the opening operation 314). Note that in Figure 5, each graph is aligned on the time axis.

[0051] The acoustic data 301b and stroke position data 302b are the same as the acoustic data 301a and stroke position data 302a in Figure 4, except that they are data from the opening operation, so the explanation in Figure 5 is omitted. Note that since Figure 5 shows the data during the opening operation of the circuit breaker 201, the activity level 311 at the start of the closing operation and the activity level 312 at the end of the closing operation are inactive, while the activity level 313 at the start of the opening operation and the activity level 314 at the end of the opening operation are active.

[0052] Thus, in this embodiment, the operation of the circuit breaker 201 is described as opening and closing the circuit breaker 201.

[0053] Furthermore, in this embodiment, the activity level 310 is composed of an activity level 311 at the start of the closing operation, an activity level 312 at the end of the closing operation, an activity level 313 at the start of the opening operation, and an activity level 314 at the end of the opening operation. This configuration of the activity level 310 makes it possible to apply it to a circuit breaker 201 that sounds twice during both the opening and closing operations.

[0054] (Generation of a feature tensor 411) Figure 6 shows an example of the procedure for generating the feature tensor 411 in common model learning (S1 in Figure 1) and individual model learning (S2 in Figure 1). Refer to Figure 3 as appropriate. Figure 6 further explains in detail the generation process of the common learning data 133 in common model learning (S1) and the individual learning data 135 in individual model learning (S2) in Figure 1. The acoustic data 301 is frequency-analyzed (S101) by the spectrogram generation unit 112 and converted into a spectrogram 401. Hereafter, the acoustic data 301 collected for use in training will be appropriately referred to as training acoustic data 301 (first acoustic data, second acoustic data) A. As shown in Figure 6, the spectrogram 401 has frequency on the vertical axis and time on the horizontal axis. In other words, the spectrogram 401 is obtained by concatenating the frequency analysis results of each waveform extracted from the training acoustic data 301A at a fixed time window W1. That is, the spectrogram 401 is obtained by converting the acoustic waveform into a spectral sequence at fixed time intervals.

[0055] Next, the feature extraction processing unit 113 performs frame stacking (S102) on the generated spectrogram 401. That is, the feature extraction processing unit 113 extracts (stacks) the spectrogram 401 in a fixed time window W2 while shifting the time by a predetermined amount of time. Each extracted spectrogram 401 is called a feature tensor 411. In this way, the feature extraction processing unit 113 generates multiple feature tensors 411 by extracting the spectrogram 401 in a predetermined time window W2. In this way, the feature extraction processing unit 113 acquires the feature tensors 411 while shifting the start time of each feature tensor 411 by a predetermined time width (acquiring while shifting the predetermined time window by a predetermined time width). As a result, the feature extraction processing unit 113 generates multiple feature tensors 411. Thus, the feature tensors 411 are generated from the acoustic data 301 (in the example shown in Figure 6, the training acoustic data 301A). In this case, it is desirable to set the time shift to be smaller than the time width of the feature tensor 411. By doing so, the time resolution of the feature tensor 411 can be increased.

[0056] The feature tensor 411 shown in Figure 6 has a time width of 17 msec and a frequency width of 77 frames. The frequency width is divided into predetermined frequency units (e.g., 10 Hz), so a frequency width of 77 frames means 77 frequency units (770 Hz). The numbers assigned below the feature tensor 411 in Figure 6 are the feature tensor 411 numbers. In the example shown in Figure 6, nfrm feature tensors 411 are generated.

[0057] <Correspondence between feature tensor 411 and activity level 310> Figure 7 shows the correspondence between the feature tensor 411 and the activity level 310. For each feature tensor 411, a corresponding time interval at each activity level 310 (activity level 311 at the start of closing operation, activity level 312 at the end of closing operation, activity level 313 at the start of opening operation, and activity level 314 at the end of opening operation) is always associated. For example, the feature tensor 411z shown in Figure 7 is associated with the time interval T1 at each activity level 310. In other words, the feature tensor 411z corresponds to the period from before the rise to the decay period at the activity level 313 at the start of opening operation. That is, the feature tensor 411z represents the sound at the start of the opening operation of the circuit breaker 201. As shown in Figures 4 and 5, the acoustic data 301 and each activity level 310 are related to each other based on the acoustic data 301 and the stroke position data 302. Therefore, the feature tensor 411 generated based on the acoustic data 301 is associated with each activity level 310.

[0058] Other feature tensors 411 are similarly associated with time intervals in their respective activation levels 310. This association between feature tensors 411 and activation levels 310 is performed by the feature extraction processing unit 113 shown in Figure 3. In this way, the feature extraction processing unit 113 associates the activation levels 310, which model the sound generated when the circuit breaker 201 operates, with the feature tensor 411 generated based on the acoustic data 301 and stroke position data 302. Incidentally, the feature tensor (first feature) 411 used for training the common model (see Figure 1) 134 and the feature tensor (second feature) 411 used for training the individual model (see Figure 1) 136 are generated using the same procedure.

[0059] In the common model 134 and individual model 136, the feature tensor 411 is used as input data, and the corresponding activity levels 310 are used as training data for learning.

[0060] (Common Model Learning Unit 114) Figure 8 shows the processing performed in the common model learning unit 114 in this embodiment. Figure 8 provides a detailed explanation of the processing performed in the common model learning unit 114 in the common model learning (S1) of Figure 1. The common model learning unit 114 includes a convolutional neural network layer 511, a first pooling layer 512, and a second pooling layer 513. Furthermore, the common model learning unit 114 includes a first fully connected neural network layer 521 and a second fully connected neural network layer 522. Thus, the deep neural network (first neural network) N1 used in common model learning (S1: see Figure 1) is composed of multiple neural networks.

[0061] First, the common training data 133 contains a feature tensor (first feature) 411 (411a, 411b: first feature) generated based on training acoustic data (first acoustic data) 301A (see Figure 6) acquired from (multiple) circuit breakers 201 of various models and operating conditions. In Figure 8, the numbers written near the feature tensor 411, the first transformed data 502, the second transformed data 503, the third transformed data 504, the fourth transformed data 505, and the fifth transformed data 506 indicate the data size.

[0062] First, one of the feature tensor 411 is input to the convolutional neural network layer 511. As mentioned above, the feature tensor 411 consists of 17 × 77 × 1 data points.

[0063] The convolutional neural network layer 511 transforms the feature tensor 411 into a first transformed data 502 with a data size of 15 × 75 × 16. The number "16" in the data size of the first transformed data 502 represents the number of channels, indicating the number of features (filters) in the spectrogram 401. The number of channels is a value pre-set by the user.

[0064] Furthermore, this first transformed data 502 is compressed by the first pooling layer 512 into a second transformed data 503 with a data size of 6 × 36 × 16. In the number indicating the data size of the second transformed data 503, "16" represents the number of features (filters) of the spectrogram 401, similar to the first transformed data 502. Furthermore, the second transformed data 503 is further compressed by the second pooling layer 513 into a third transformed data 504 with a data size of 2 × 17 × 16. In the number indicating the data size of the third transformed data 504, "16" represents the number of features (filters) of the spectrogram 401, similar to the first transformed data 502 and the second transformed data 503.

[0065] Next, the common model learning unit 114 converts the third transformed data 504 into a fourth transformed data 505 having a data size of 1 × 544. The fourth transformed data 505 is simply the third transformed data 504 rearranged into 1 × 544 data based on predetermined arrangement rules.

[0066] Then, the fourth transformed data 505 is transformed into a fifth transformed data 506 having a data size of 1 × 128 by the first fully connected neural network layer 521. The fifth transformed data 506 is then transformed into common model output data 507 by the second fully connected neural network layer 522. The data size of the common model output data 507 is 4c, where c is the model of the circuit breaker 201 and the number of operating conditions. The "4" refers to the four activation levels described above: activation level 311 at the start of closing operation, activation level 312 at the end of closing operation, activation level 313 at the start of opening operation, and activation level 314 at the end of opening operation. In other words, the feature tensor 411 input to the convolutional neural network layer 511 is output as data corresponding to one of the activity levels 310, which are provided for each model and operating condition: activity level 311 at the start of closing operation, activity level 312 at the end of closing operation, activity level 313 at the start of opening operation, and activity level 314 at the end of opening operation.

[0067] On the other hand, the common learning data 133 retains the state of the activity level 310 (310a, 310b: first acoustic model) associated with each feature data. The common model learning unit 114 then calculates the difference between the common model output data 507 and the training data associated with the feature tensor 411 input to the convolutional neural network layer 511 as the error (S111). The common model learning unit 114 then performs a backpropagation process (S112) to update the connection strength of neurons in each layer based on the error. The initial values ​​of the connection strength of neurons in each layer are set using random numbers.

[0068] After backpropagation, the common model learning unit 114 inputs the next feature tensor 411 into the convolutional neural network layer 511. Then, once all feature tensors 411 have been input into the convolutional neural network layer 511, the feature tensor 411 that was initially input into the convolutional neural network layer 511 is input into the convolutional neural network layer 511 again.

[0069] The common model learning unit 114 continues the above process for a predetermined number of times, or until the error meets a predetermined condition (for example, less than or equal to a predetermined value). The common model learning unit 114 outputs the neuronal connection strengths in each layer calculated by the above process as the common model 134.

[0070] In this way, the common model learning unit 114 generates a common model 134 from a feature tensor 411 relating to numerous models and operating conditions, and an activity level 310 relating to the opening and closing sounds of the circuit breaker 201 under multiple models and operating conditions.

[0071] (Individual Model Learning Section 115) Figure 9 shows the processing performed in the individual model learning unit 115 in this embodiment. Figure 9 further explains in detail the processing of the individual model learning unit 115 in the individual model learning (S2) of Figure 1. The individual model learning unit 115, like the common model learning unit 114, includes a convolutional neural network layer 511, a first pooling layer 512, a second pooling layer 513, a first fully connected neural network layer 521, and a second fully connected neural network layer 522. Thus, the deep neural network (second neural network) N2 used in individual model learning (S2: see Figure 1) is composed of multiple stages of neural networks. Furthermore, the deep neural network N2 used in individual model learning (S2) has the same structure as the deep neural network N1 (see Figure 8) used in common model learning (S2).

[0072] Furthermore, for the convolutional neural network layer 511 to the first fully connected neural network layer 521 (other neural networks), the connection strength of neurons in each layer is set to the value of the common model 134 stored in the common model 134. In other words, the connection strength of neurons for the data shown in shaded areas in Figure 9 is initially set to the value of the common model 134 stored in the common model 134. However, in Figure 9, the initial value of the connection strength of neurons in the second fully connected neural network layer (the neural network one stage before the output) 522 is set to a random number. Thus, in individual model learning (S2), the common model 134 output as a result of common model learning (S1) is used as the initial value of the individual model 136 (deep neural network N2) in individual model learning (S2).

[0073] Furthermore, in the individual model learning unit 115, the final output, individual model output data 531, consists of four activity levels 310 for the circuit breaker 201c (see Figure 1) that is the subject of diagnosis. The four activity levels 310 are the activity level 311 at the start of closing operation, the activity level 312 at the end of closing operation, the activity level 313 at the start of opening operation, and the activity level 314 at the end of opening operation, as shown in Figures 4 and 5. Therefore, the individual model output data 531 consists of 4 × 1 data. Unlike the common model learning unit 114, the individual model learning unit 115 has only one target circuit breaker 201, so the individual model output data 531 consists of 1 × 4 data.

[0074] The feature tensor (second feature) 411c (411), extracted from the training acoustic data (second acoustic data) 301A (see Figure 6) acquired from the newly installed (to be diagnosed) circuit breaker 201c, is stored in the individual training data 135. In addition, the state of the activity level (second acoustic model) 310c (310) corresponding to the feature tensor 411c is stored in the individual training data 135 as training data.

[0075] The individual model learning unit 115 then inputs the initial feature tensor 411 into the convolutional neural network layer 511. Subsequently, processing is performed by the convolutional neural network layer 511, the first pooling layer 512, the second pooling layer 513, and the first fully connected neural network layer 521, resulting in the output of the fifth transformed data 506. The fifth transformed data 506 is then transformed into individual model output data 531 by the second fully connected neural network layer 522.

[0076] In other words, in the individual model learning unit 115, the input feature tensor 411 is output as one of the four activity levels 310. The four activity levels 310 are: activity level 311 at the start of closing operation, activity level 312 at the end of closing operation, activity level 313 at the start of opening operation, and activity level 314 at the end of opening operation.

[0077] On the other hand, as mentioned above, the individual training data 135 holds the state of activity level 310, which is associated with the feature tensor 411 of the circuit breaker 201c, as training data. Therefore, the individual model training unit 115 calculates the difference between the individual model output data 531 and the training data (activity level 310) associated with the feature tensor 411 of the input convolutional neural network layer 511 as the error (S211). Then, the common model training unit 114 performs a backpropagation process (S212) to update the connection strength of neurons in each layer based on the error.

[0078] When backpropagation is performed, the individual model learning unit 115 inputs the next feature tensor 411 into the convolutional neural network layer 511. Then, once all feature tensors 411 have been input into the convolutional neural network layer 511, the feature tensor 411 that was initially input into the convolutional neural network layer 511 is input into the convolutional neural network layer 511 again.

[0079] The individual model learning unit 115 continues the above process for a predetermined number of times, or until the error meets a predetermined condition (for example, less than or equal to a predetermined value). The individual model learning unit 115 stores the neuronal connection strengths in each layer calculated by the above process as an individual model 136 in the auxiliary storage device 130 (see Figure 3).

[0080] In this way, during the generation of the individual model 136, the common model 134 is first set as the initial value for the first fully connected neural network layer 521 from the convolutional neural network layer 511. Then, the individual model learning unit 115 calculates the individual model 136 using the newly installed feature tensor 411 of the circuit breaker 201c and the activation level 310 of the circuit breaker 201c. At this time, the individual model learning unit 115 inputs the activation level 310 of the circuit breaker 201c (training data) and the feature tensor 411 of the circuit breaker 201c into the convolutional neural network layer 511, and learns to minimize the difference between the output of the second fully connected neural network layer 522 (individual model output data 531) and the activation level 310 as training data. In this way, the individual model 136 is learned.

[0081] In this way, the initial values ​​of the connection strengths in the neurons from the convolutional neural network layer 511 to the first fully connected neural network layer 521 that constitute the individual model 136 are set with the values ​​of the common model 134 generated by the common model learning unit 114. As a result, the individual model learning unit 115 no longer needs to learn the connection strengths from the convolutional neural network layer 511 to the first fully connected neural network layer 521 from a random state. Therefore, even if the number of feature tensors 411 obtained for learning from the newly installed circuit breaker 201c is small, the learning of the individual model 136 can be stably converged.

[0082] (Processing by the acoustic diagnostic unit 118) Figure 10 shows the processing procedure of the acoustic diagnostic unit 118 as performed in this embodiment. Figure 10 further explains in detail the processing of the acoustic diagnostic unit 118 in the diagnosis (S3) of Figure 1. First, acoustic data (third acoustic data) 301 is acquired from the circuit breaker 201c to be diagnosed. As shown in Figure 10, the acoustic data 301 acquired during the diagnosis of the circuit breaker 201 will be appropriately referred to as diagnostic acoustic data (third acoustic data) 301B. The circuit breaker 201c is the circuit breaker 201c that has been trained by the individual model learning unit 115.

[0083] Next, the activity estimation processing unit 116 performs frequency analysis (S301) to convert the diagnostic acoustic data 301B into a spectrogram 401 composed of time-dependent spectra. Furthermore, the activity estimation processing unit 116 performs frame stacking (S302) to obtain multiple spectrograms 401 of a predetermined time width by extracting them while shifting the time. As a result, multiple feature tensors (third features) 411, which are spectrograms 401 with a predetermined time width, are generated. The data size of the feature tensor 411 is the same as the feature tensor 411 generated in Figure 6. The processing up to this point is the same as the processing shown in Figure 6.

[0084] Next, the activity estimation processing unit 116 sequentially inputs the feature tensor 411 to the individual model 136, which has been trained by the individual model learning unit 115. Each feature tensor 411 input to the individual model 136 is output as the estimated activity (estimated acoustic model) 681 of the four opening and closing operation sounds (S303). The estimated activity 681 is the activity 310 estimated by the individual model 136. As shown in Figure 10, the estimated activity 681, like the activity 310, has an estimated activity 681A at the start of the closing operation, an estimated activity 681B at the end of the closing operation, an estimated activity 681C at the start of the opening operation, and an estimated activity 681D at the end of the opening operation.

[0085] By converting all feature tensors 411 into estimated activity levels 681, and arranging the converted estimated activity levels 681 in chronological order (in the order of input to the feature tensors 411), four types of estimated activity level 681 state time series based on the diagnostic acoustic data 301B are obtained. Based on the peak values ​​of these estimated activity levels 681, the activity level estimation processing unit 116 can estimate the start and end times of the opening or closing operation of the circuit breaker 201c (estimate the operation timing). In this way, the activity level estimation processing unit 116 estimates the operation timing of the circuit breaker 201c based on the estimated activity levels 681 estimated from the diagnostic acoustic data 301B.

[0086] The state diagnosis processing unit 117 determines the health of the circuit breaker 201 by comparing the time taken for the switching operation with the rated timing (reference operating timing) based on the estimated start time and end time (estimated operating timing). The state diagnosis processing unit 117 determines whether or not there is an abnormality in the circuit breaker 201c based on the operating timing estimated from the results output when the feature quantity tensor 411 is input to the individual model 136.

[0087] Specifically, the status diagnosis processing unit 117 calculates the degree of deviation between the estimated start and end times and the rated timing, and if this degree of deviation is greater than or equal to a predetermined value, it determines that there is an abnormality in the circuit breaker 201c.

[0088] (Correspondence between feature tensor 411 and activity level 310) Figure 11 shows the configuration of the common learning data 133 and the individual learning data 135. Figure 11 shows an example of how to store the feature tensor 411 and the activity level 310 (see Figure 6) in the common training data 133 and the individual training data 135. However, the storage method for the feature tensor 411 and the activity level 310 in the common training data 133 and the individual training data 135 is not limited to the example shown in Figure 11. Refer to Figure 1 as appropriate in Figure 11. As shown in FIG. 11, common learning data 133 and individual learning data 135 store ID (item 701), feature tensor 411 (item 702), and activity 310 (item 703) in association with each other. The ID is an identifier assigned to a pair of the feature tensor 411 and the activity 310. As shown in FIG. 11, the feature tensor 411 is stored in the form of "float x i,j [nfrm i,j ,freq,nstack]". Here, "float" indicates that the value to be stored is a floating-point tensor. And x i、j i indicates the model of the circuit breaker 201 or an operating condition, and j indicates the number of the feature tensor 411. The number of the feature tensor 411 indicated by j is assigned in order from the oldest time. Hereinafter, for simplicity of explanation, i is assumed to indicate the model of the circuit breaker 201.

[0089] And in FIG. 11, such a feature tensor 411 has a size of nfrm i,j ×freq×nstack. nfrm i,j is the number of frames (number of data) in the time direction for model i and number j. In the example shown in FIG. 6, nfrm i,j is "17". Also, freq is the number of frames in the frequency direction. In the example shown in FIG. 6, freq is "77". And nstack is the number of feature tensors 411 having a size of nfrmi,j×freq that are simultaneously input (bundled and input) to the common model 134 or the individual model 136. In a neural network, n feature tensors 411 each having a size of nfrm i,j ×freq may be collectively input. In the example shown in FIG. 6, since the feature tensors 411 are input one by one, nstack is "1".

[0090] Also, the activity 310 is "float y i,j [nfrm i,jIt is saved in the format ",4]". Here, "float" indicates that the value being saved is a floating-point tensor. And y i、j In this case, 'i' indicates the model or operating conditions, and 'j' indicates a number, similar to 'j' in the feature tensor 411. Activation level 310 is nfrm i,j This is a tensor (matrix) with a size of 4 × .

[0091] Also, nfrm i,j This indicates the number of time-direction frames for activity level 310, and contains the same number as the feature tensor 411 (in the example shown in Figure 6, it is "17"). The last "4" indicates that there are four types of activity level 310 (activity level 311 at the start of closing operation, activity level 312 at the end of closing operation, activity level 313 at the start of opening operation, and activity level 314 at the end of opening operation). The first column of the tensor representing activity level 310 stores the activity level 311 at the start of closing operation, the second column stores the activity level 312 at the end of closing operation, the third column stores the activity level 313 at the start of opening operation, and the fourth column stores the activity level 314 at the end of opening operation.

[0092] The feature tensor 411 and activity level 310 are both dependent on the length of the acoustic waveform of the original acoustic data 301, and represent the number of frames (time frame: nfrm) i,j The data exists for each of the following: Feature Tensor 411 stores data equal to the number of freq values ​​in the frequency direction, bundled together by adjacent nstacks. As a result, Feature Tensor 411 is a 3-dimensional array (tensor) where the number of elements (size) in each dimension is nfrm, freq, and nstack.

[0093] On the other hand, as mentioned above, the activity level 310 consists of four types of activity levels 310, which are combinations of the start and end of the two types of operations: closing and opening. The activity level 310 is stored for each frame, corresponding to the feature tensor 411. As a result, the activity level 310 becomes a two-dimensional array (tensor), where the number of elements in the first dimension (rows) is nfrm, and the second dimension (columns) stores the array of elements corresponding to each of the four types of activity level 310.

[0094] (Common model output data 507) Figure 12 shows the data structure of the training data in the common model 134. In Figure 12, the training data in the common model 134 is simply referred to as "training data". The training data ID (code 711) is a unique ID assigned to the training data. And, as shown in reference numeral 712, the training data is, Yi,j [nfrm i,j The data has the value ",4c". The training data may be represented as floating-point data, or as integers of type [1,0].

[0095] Y i,j In this context, 'i' indicates the model or operating conditions, and 'j' indicates the training data number. The numbers indicated by 'j' are assigned sequentially from oldest to newest. The training data is nfrm i,j It has been shown to be a tensor (matrix) with a size of ×4c. nfrm i,j This, as in Figure 11, represents the number of frames (data points) in the time direction for model i and number j. nfrm i,j Specifically, this is "17," as in Figure 11. c represents the number of models and operating conditions. In Figure 12, only the models are considered.

[0096] In other words, the training data is nfrm in the row direction. i,j It has elements (data) and has 4c data in the column direction. In the column direction, four types of activity levels 310 are stored side by side for each model.

[0097] The specific configuration of each training data is shown in Figure 12, labeled 720. i、j [nfrm i,j In ,4]", i, j and nfrm i,j is "Y i,j It is the same as ". And "4" is nfrm in the row direction i,jThis indicates that activity levels 310, which have the size of , are arranged in the column direction in the order of activity level 311 at the start of the closing operation, activity level 312 at the end of the closing operation, activity level 313 at the start of the opening operation, and activity level 314 at the end of the opening operation. i,j ,4]" indicates that all components are "0". And, "Y i,j [nfrm i,j In "4c", such activity levels 310 are arranged in a column in the order of "Activity level A for model A", "Activity level B for model B", ... "Activity level C for model C" (symbols 721-723).

[0098] In other words, as shown in Figure 13, one training data set (for a given time) has a configuration in which "Activity level at the start of closing operation of model A", "Activity level at the end of closing operation of model A", ... are arranged in the row direction. In Figure 13, the number of columns is 4c. Note that each of "Activity level at the start of closing operation of model A", "Activity level at the end of closing operation of model A", ... is 1 × nfrm i,j It has the number of elements.

[0099] The common model 134 is trained as a model to predict the activity level 310 value for multiple models. Therefore, as shown in Figure 12, the activity level 310 corresponding to each different model is set to zero (zero[nfrm i,j Activity levels 310 for training the common model, padded with [4]), are prepared as training data. The common model 134 is then trained to minimize the error between the common model output data 507 and the training data. When creating the common model 134, if type c is used, the training data is nfrm i,j This results in data with 4c dimensions.

[0100] (Diagnosis result display screen 800) Figure 14 shows an example of the diagnostic result display screen 800 output to the output device 103 in this embodiment. Refer to Figures 1, 4, and 6 as appropriate. The diagnostic result display screen 800 includes an acoustic data display area 801, a spectrogram display area 802, an estimated activity display area 810, an estimated operation display area 821, a rated deviation display area 822, and a diagnostic result display area 823. The acoustic data display area 801 displays the acoustic waveform shown by the diagnostic acoustic data 301B (see Figure 10) acquired via the microphone 202. The spectrogram display area 802 displays the spectrogram 401 (see Figure 10), which is the result of frequency analysis of the acoustic waveform (diagnostic acoustic data 301B) displayed in the acoustic data display area 801.

[0101] The estimated activity level display area 810 displays the estimated activity level (acoustic model estimated from the third acoustic data) 681, which is output as a result of the activity level estimation processing unit 116 inputting the diagnostic acoustic data 301B collected during diagnosis to the individual model 136. In the example shown in Figure 14, the estimated activity level display area 810 displays the estimated activity level at the start of closing operation 811, the estimated activity level at the end of closing operation 812, the estimated activity level at the start of opening operation 813, and the estimated activity level at the end of opening operation 814, respectively. The estimated activity level at the start of closing operation display area 811 displays the estimated activity level at the start of closing operation 681A shown in Figure 10. The estimated activity level at the end of closing operation display area 812 displays the estimated activity level at the end of closing operation 681B shown in Figure 10. The estimated activity level at the start of opening operation display area 813 displays the estimated activity level at the start of opening operation 681C shown in Figure 10. At the end of the opening operation, the estimated activity level 814 is replaced with the estimated activity level 681D shown in Figure 10.

[0102] The estimated operation display area 821 shows the result of the state diagnosis processing unit 117's determination of whether the acoustic waveform (diagnostic acoustic data 301B) displayed in the acoustic data display area 801 is from the closing operation or the opening operation of the circuit breaker 201. In the example shown in Figure 14, as shown in the estimated activity level display area 810, the estimated activity level 681 at the start of the opening operation and at the end of the closing operation is active. In other words, the estimated activity level 681C at the start of the opening operation displayed in the estimated activity level display area 813 and the estimated activity level 681D at the end of the opening operation displayed in the estimated activity level 814 are active. As a result, the state diagnosis processing unit 117 has determined that the acoustic waveform displayed in the acoustic data display area 801 is from the opening operation ("Open").

[0103] The rated deviation display area 822 displays a numerical value indicating how much the closing or opening timing estimated by the estimated activity level 681 deviates from the rated timing. In the example shown in Figure 14, the average of the opening start timing and opening end timing estimated from the estimated activity level 681 is deviating from the rated by +0.1%. In the rated deviation display area 822, "+" indicates that the timing is delayed, and "-" indicates that the timing is advanced.

[0104] The diagnostic result display area 823 displays the diagnostic results from the status diagnostic processing unit 117. Here, the status diagnostic processing unit 117 displays the result of its determination of whether the operation of the circuit breaker 201 is normal or abnormal, based on the deviation from the rating displayed in the rating deviation display area 822. In the example shown in Figure 14, it is indicated that the operation is normal.

[0105] According to this embodiment, a diagnostic device 1 can be provided that diagnoses the opening and closing timing state of the circuit breaker 201 during operation from acoustic data 301. The diagnostic device 1 acquires the true value of the stroke position (stroke position data 302) measured in advance by a stroke position measuring device 203 at the position of the contacts in the opening and closing rod, and acoustic data 301 measured in synchronization with that true value. Then, the diagnostic device 1 estimates the start and end timings of the opening and closing operation of the circuit breaker 201 based on the stroke position data 302. Then, the diagnostic device 1 generates an activity level 310 represented by a triangular wave that rises vertically for each opening and closing and start / end. Furthermore, the diagnostic device 1 performs learning using a common model 134 as an estimation model that performs regression estimation from a feature tensor 411 generated based on the acoustic data 301 to the activity level 310. Furthermore, if a new circuit breaker 201 to be diagnosed is installed, the diagnostic device 1 learns the individual model 136 of the newly installed circuit breaker 201, using the common model 134 as the initial value.

[0106] During the diagnosis of circuit breaker 201, the diagnostic device 1 estimates the activity level 310 of the circuit breaker 201 being diagnosed (estimated activity level 681) based on acoustic data 301 and using a pre-trained individual model 136. The diagnostic device 1 then estimates the switching operation timing of the circuit breaker 201 being diagnosed from the peak value of the estimated activity level 310 (estimated activity level 681). Furthermore, the diagnostic device 1 calculates the deviation of the estimated switching operation timing from the time rating. The diagnostic device 1 then diagnoses the operating state of the circuit breaker 201 based on this deviation.

[0107] When training a new circuit breaker 201, it has been necessary to collect a large amount of new training acoustic data 301 and contact stroke position data 302. Therefore, with conventional technology, the amount of work required to collect training data for a new circuit breaker 201 is high. In addition, the circuit breaker 201 must be actually operated in order to collect training data. As a result, the opening and closing operation of the circuit breaker 201 during measurement causes excessive deterioration of the circuit breaker 201.

[0108] In this embodiment, a common model 134 is created in advance from a large number of circuit breakers 201. Then, when creating an individual model 136 for a new circuit breaker 201, the individual model 136 for the new circuit breaker 201 is trained using the trained common model 134 as the initial value. This makes it possible to significantly reduce the amount of training data required to train the individual model 136 for the new circuit breaker 201. As a result, the number of times the circuit breaker 201 is opened and closed to collect training data can be reduced, and deterioration of the circuit breaker 201 can be prevented.

[0109] For example, a pair is prepared consisting of a feature tensor 411 extracted from acoustic data 301 of multiple types of circuit breakers 201, and the true value of the activity level 310 related to the opening and closing operation sound. The diagnostic device 1 then transforms the feature tensor 411 into a latent space of a predetermined number of dimensions. This latent space of a predetermined number of dimensions consists of four types of activity levels 310.

[0110] The diagnostic device 1 then classifies the timing of the opening and closing operations of multiple types of circuit breakers 201 based on the converted representations (four types of activity levels 310). Furthermore, the diagnostic device 1 learns a model (common model 134) that regressively predicts the activity levels 310 of the opening and closing sounds of the circuit breakers 201. As a result, the diagnostic device 1 learns to determine four types of activity levels 310 that are suitable for predicting the timing of the opening and closing sounds.

[0111] To construct a model (individual model 136) for the new circuit breaker 201, the diagnostic device 1 generates a feature tensor 411 using the same procedure as when generating the common model 134. Then, by reusing the model and weights (neuron connection strength) of the deep neural network N1 of the common model 134, the diagnostic device 1 adds a layer (second fully connected neural network layer 522) for regression estimation of the activation level 310 of the opening and closing operation sound from the fifth transformation data 506 to train the individual model 136. For the individual model 136, it is sufficient to regression estimate only the activation level 310 at the start and end timings of the opening and closing operation for the new circuit breaker 201 to be diagnosed. Therefore, the output of the individual model 136 will be four nodes (outputs corresponding to the activation level 311 at the start of closing operation, the activation level 312 at the end of closing operation, the activation level 313 at the start of opening operation, and the activation level 314 at the end of opening operation). By performing individual model learning (S1) using these procedures, the convergence of learning for the individual models 136 can be improved even with a small number of training samples.

[0112] When the circuit breaker 201 is diagnosed, only sound is collected by the microphone 202, and the estimated activity level 681 is calculated using a pre-trained individual model 136 based on the feature tensor 411 extracted from the acquired acoustic data 301. The diagnostic device 1 can then determine the opening and closing timing of the circuit breaker 201 from the position of the peak of the estimated activity level 681. Subsequently, the diagnostic device 1 checks the health of the circuit breaker 201 by comparing the estimated opening and closing timing with the rated timing. In this way, it becomes possible to diagnose the circuit breaker 201 quickly and using an individual model 136 with improved learning convergence.

[0113] In other words, by using the individual model 136 in the diagnostic device 1 of this embodiment, it becomes possible to accurately measure the timing information of the opening and closing operation events of the circuit breaker 201 using only the microphone 202 that measures sound. Furthermore, the diagnostic device 1 of this embodiment can diagnose the state of the circuit breaker 201 from the changes in the timing of important opening and closing operation events (opening and closing operations in this embodiment) estimated using the individual model 136.

[0114] In this way, it becomes possible to diagnose the circuit breaker 201 using only the microphone 202, without using the stroke position measuring device 203. Furthermore, based on the estimated activity level 310 (estimated activity level 682) obtained by inputting the diagnostic feature tensor 411 into the individual model 136, the operating timing is estimated. Specifically, based on the estimated operating timing, the circuit breaker 201 is diagnosed based on the degree of deviation between the estimated operating timing and the rated operating timing. This allows for a quantitative diagnosis of the circuit breaker 201.

[0115] Thus, according to this embodiment, a common model 134, which has been trained on a large number of circuit breakers 201 in advance, is prepared. When diagnosing a new circuit breaker 201, the model is not created from scratch, but additional training is performed based on the common model 134. This makes it possible to effectively create individual models 136 from a small number of training samples. In particular, in individual model training (S2), only the second fully connected neural network 522 differs from the configuration of the common model 134. This makes it possible to construct a model for diagnosing a new circuit breaker 201 at low cost. In addition, in this embodiment, in individual model training (S2), only the second fully connected neural network 522 has random initial values ​​for the connection strength of neurons. This ensures that the convergence of the individual model 136 can be performed reliably and quickly.

[0116] In this embodiment, the start time and end time of operation are estimated based on the stroke position data 302 from the stroke position measuring device 203. Then, the activity level 310 is associated with the estimated start time and end time of operation. This allows the association between training data (activity level 310) and input data (feature tensor 411) to be made without manual work by the user.

[0117] Furthermore, in this embodiment, a feature tensor 411 is generated by extracting the spectrogram 401 through a time window W2. This allows learning to be performed based on the operating sound generated by the circuit breaker 201, and diagnosis is performed using the learning results. This enables diagnosis based on the frequency distribution specific to the circuit breaker 201, thereby improving the accuracy of the diagnosis.

[0118] In this embodiment, when generating the common model 134, the distinction between circuit breakers 201 is not considered. When generating the common model 134, the common model 134 may be learned based on acoustic data 301 and stroke position data 302 collected from circuit breakers 201 of the same type as the circuit breaker 201 to be diagnosed. The types of circuit breakers 201 include circuit breakers 201 in substations, circuit breakers 201 in power receiving equipment, circuit breakers 201 on Shinkansen lines, and circuit breakers found in ordinary homes. Furthermore, the types of circuit breakers 201 may be distinguished by the time between opening and closing operations, etc.

[0119] The deep neural network N1 shown in Figure 8 is often used in the image field. When deep neural network N1 is used in the image field, images are used as training data. In this embodiment, an activity level 310 is used, which is a triangular wave model of the sound emitted when the circuit breaker 201 opens and closes. Thus, the deep neural network N1 used in this embodiment is used in a completely different way from deep neural network N1 used in the image field.

[0120] Furthermore, in this embodiment, a triangular wave is used as the activity level 310, which rises perpendicularly to the time when sound is generated during the operation of the circuit breaker 201 and then decays at a predetermined slope. However, this is not the only option. For example, as shown in Figure 15, activity levels 310A and 310B, which are rectangular waves, may be used as the activity level 310. The activity levels 310A and 310B shown in Figure 15 are represented by rectangular waves that rise perpendicularly to the time axis at the start of operation of the circuit breaker 201. In Figure 15, activity level 310A is a model of the start sound of the opening or closing operation, and activity level 310B is a model of the start sound of the opening or closing operation.

[0121] Furthermore, in this embodiment, the common model 134, the individual model 136, and the circuit breaker 201 are diagnosed based on the sounds generated when the circuit breaker 201 is opened and closed, but this is not limited to this. For example, the common model 134, the individual model 136, and the circuit breaker 201 may be diagnosed based on the sound of a fuse blowing, etc.

[0122] The present invention is not limited to the embodiments described above, and includes various modifications. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to add, delete, or replace parts of the configuration of each embodiment with other configurations.

[0123] Furthermore, each of the above-mentioned configurations, functions, parts 111-117, auxiliary storage device 130, etc., may be implemented in hardware, either partially or entirely, by designing them as integrated circuits, for example. Alternatively, as shown in Figure 4, each of the above-mentioned configurations, functions, etc., may be implemented in software by having a central processing unit 101, such as a CPU, interpret and execute programs that realize each function. Information such as programs, tables, and files that realize each function can be stored not only in the HD, but also in memory, a recording device such as an SSD, or a recording medium such as an IC (Integrated Circuit) card, an SD (Secure Digital) card, or a DVD (Digital Versatile Disc). Furthermore, in each embodiment, only those control lines and information lines deemed necessary for explanation are shown, and not all control lines and information lines are necessarily shown in the actual product. In practice, it can be assumed that almost all components are interconnected. [Explanation of Symbols]

[0124] 1. Diagnostic device (circuit breaker operation model generation device) 103 Output device (display device) 110 Diagnostic Software 112 Spectrogram generation unit (feature generation unit) 113 Feature Extraction Processing Unit (Feature Setting Unit) 114 Common Model Learning Unit (First Learning Unit) 115 Individual Model Learning Section (Second Learning Section) 116 Activity Estimation Processing Unit (Diagnostic Unit) 117 Condition Diagnosis Processing Unit (Diagnostic Unit) 118 Acoustic Diagnostics Department (Diagnostic Department) 133 Common training data 134 Common Model (First Learning Model) 135 Individual training data 136 Individual Models (Second Learning Model) 137 Diagnostic Parameters 138 Diagnostic History Information 201 Circuit breaker 201a Circuit breaker (first circuit breaker) 201b Circuit breaker (first circuit breaker) 201c Circuit breaker (second circuit breaker) 202 Microphone (sound collection device) 203 Stroke position measuring device (contact position measuring device) 211 Diagnosis Result 301 Audio data (First audio data, Second audio data, Third audio data) 301a Acoustic data 301b Acoustic data 301A Training audio data (First audio data, Second audio data) 301B Diagnostic acoustic data (third acoustic data) 302 Stroke position data (contact position data) 302a Stroke position data 302b Stroke position data 310 Activity level (acoustic model, triangular wave) 310a Activity level (first acoustic model, triangular wave) 310b Activity level (first acoustic model, triangular wave) 310c Activity (Second Acoustic Model, Triangular Wave) 310A Activity (Acoustic Model, Square Wave) 310B Activity (Acoustic Model, Square Wave) 311 Activity level at the start of closing operation (activity level generated for the operating sound emitted when the circuit breaker starts closing operation) 312 Activity level at the end of closing operation (Activity level generated for the operating sound emitted when the closing operation of the circuit breaker is completed) 313 Activity level at the start of opening operation (Activity level generated for the operating sound emitted when the circuit breaker starts opening operation) 314 Activity level at the end of the opening operation (Activity level generated for the operating sound emitted when the opening operation of the circuit breaker ends) 401 Spectrogram 411 Feature Tensor (Feature, First Feature, Second Feature) 411a Feature Tensor (Features, First Feature) 411b Feature Tensor (Features, First Feature) 411c Feature Tensor (Feature, Second Feature) 411z Feature Tensor (Features) 502 First conversion data 503 Second conversion data 504 Third conversion data 505 Fourth conversion data 506 Fifth conversion data 507 Common Model Output Data 511 Convolutional Neural Network Layers 512 First pooling layer 513 Second pooling layer 521 First Full Connection Neural Network Layer 522 Second Full Connection Neural Network Layer 531 Individual Model Output Data 682 Estimated activity level (estimated acoustic model) 682A Estimated activity at the start of closing operation 682B Estimated activity level at the end of closing operation 682C Estimated activity at the start of opening operation 682D Estimated activity level at the end of opening operation 801 Audio data display area 802 Spectrogram display area 810 Estimated Activity Display Area (Displays the acoustic model estimated from the third acoustic data) 811 Estimated activity display area at the start of closing operation 812 Display area for estimated activity level at the end of closing operation 813 Estimated activity display area at the start of opening operation 814 Estimated activity level at the end of opening operation 821 Estimated motion display area 822 Rated deviation display area 823 Diagnostic result display area (displays diagnostic results) N1 Deep Neural Network (Neural Network, First Neural Network) N2 Deep Neural Network (Neural Network, Second Neural Network) W1 hour window W2 Time Window Z Circuit Breaker Diagnostic System (Circuit Breaker Operation Model Generation System) S1 Common Model Learning (First Learning, First Learning Step) S2 Individual Model Learning (Second Learning, Second Learning Step)

Claims

1. A sound collection device that collects the operating sound of a circuit breaker that cuts off power and outputs it as acoustic data, A first learning unit outputs a first learning model by performing a first learning process, which takes first feature quantities, which are generated from first acoustic data obtained from a first circuit breaker, which is one of several installed circuit breakers, as input data, and a first acoustic model, which is an acoustic model associated with the first feature quantities, as training data. A second learning unit outputs a second learning model by performing a second learning process, which takes a second feature quantity, which is generated based on the second acoustic data obtained from a second circuit breaker that is a different circuit breaker from the first circuit breaker, as input data, and uses the second acoustic model, which is the acoustic model associated with the second feature quantity, as training data. It has, The acoustic model is an activity level that is generated for each of the operating sounds emitted when the circuit breaker starts closing, when the circuit breaker finishes closing, when the circuit breaker starts opening, and when the circuit breaker finishes opening, and is represented by a triangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively, and then decays with a predetermined slope, or a rectangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively. In the second learning process, the first learning model output as a result of the first learning process is used as the initial value of the second learning model in the second learning process. A circuit breaker operation model generation system characterized by the following:

2. The neural network used in the first learning performed in the first learning unit and the second learning performed in the second learning unit is composed of a multi-stage neural network. The configuration of the second neural network used for the second learning process has the same structure as the first neural network used for the first learning process. The first learning unit described above is, When performing the first learning, the initial values ​​of the connection strengths of the neurons constituting the first neural network are set to random numbers, and the first neural network is trained, thereby outputting the connection strengths of the neurons as the first learning model. The second learning unit described above is, In the second neural network, the initial value of the connection strength of the neurons in the neural network immediately preceding the output stage is set to a random number, and the initial value of the connection strength of the neurons in the other neural networks is set to the connection strength determined by the first learning model. The circuit breaker operation model generation system according to feature 1.

3. A diagnostic unit inputs a third feature quantity, which is generated based on third acoustic data that is different from the second acoustic data and is acquired from the second circuit breaker by the sound collection device, into the second learning model that is output as a result of the second learning, and performs a diagnosis regarding the operation of the second circuit breaker based on the output result output from the second learning model. The circuit breaker operation model generation system according to claim 1, characterized by having the following features.

4. The aforementioned diagnostic unit, Based on the operation timing estimated from the output result obtained by inputting the third feature quantity, which is a feature quantity generated based on the third acoustic data, into the second learning model, it is determined whether or not there is an abnormality in the second circuit breaker. The circuit breaker operation model generation system according to feature 3.

5. The aforementioned diagnostic unit, The degree of deviation between the estimated operating timing and the reference operating timing is calculated, and if the degree of deviation is greater than or equal to a predetermined value, it is determined that there is an abnormality in the second circuit breaker. The circuit breaker operation model generation system according to feature 4.

6. In the diagnostic unit, the output result is the acoustic model estimated by inputting the third acoustic data into the second learning model, The aforementioned diagnostic unit, Based on the acoustic model estimated from the third acoustic data, the operation timing is estimated. The circuit breaker operation model generation system according to feature 4.

7. The acoustic model estimated from the third acoustic data and the diagnostic results of the second circuit breaker by the diagnostic unit are displayed on at least the display device. The circuit breaker operation model generation system according to feature 6.

8. A contact position measuring device that measures the position of the contacts of the circuit breaker and outputs it as contact position data, A feature quantity setting unit associates the acoustic model with the feature quantities generated based on the acoustic data, based on the acoustic data and the contact position data output from the contact position measuring device. The circuit breaker operation model generation system according to claim 1, characterized by having the following features.

9. A feature generation unit generates a spectrogram from the aforementioned acoustic data and acquires the spectrogram while shifting a predetermined time window by a predetermined time width to generate the aforementioned feature quantities. The circuit breaker operation model generation system according to claim 1, characterized by having the following features.

10. Regarding a circuit breaker that interrupts power, a first learning unit outputs a first learning model by performing a first learning process that takes a first feature quantity, which is a feature quantity generated from first acoustic data, which is data of the operating sounds of a first circuit breaker, which is a plurality of said circuit breakers, as input data, and a first acoustic model, which is an acoustic model associated with the first feature quantity, as training data. A second learning unit outputs a second learning model by performing a second learning process, which takes a second feature quantity, which is generated based on second acoustic data, which is data of the operating sound of a second circuit breaker that is a different circuit breaker from the first circuit breaker, as input data, and uses the second acoustic model, which is the acoustic model associated with the second feature quantity, as training data. It has, The acoustic model is an activity level that is generated for each of the operating sounds emitted when the circuit breaker starts closing, when the circuit breaker finishes closing, when the circuit breaker starts opening, and when the circuit breaker finishes opening, and is represented by a triangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively, and then decays with a predetermined slope, or a rectangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively. In the second learning process, the first learning model output as a result of the first learning process is used as the initial value for the second learning model in the second learning process. A circuit breaker operation model generation device characterized by the following:

11. Regarding a circuit breaker that interrupts power, the first learning step involves outputting a first learning model through a first learning process, in which a first learning model is output by using a first feature quantity, which is a feature quantity generated from first acoustic data, which is data of the operating sounds of a first circuit breaker, which is a plurality of said circuit breakers, as input data, and a first acoustic model, which is an acoustic model associated with the first feature quantity, as training data. A second learning step involves outputting a second learning model through a second learning process, which uses a second feature, which is generated based on second acoustic data that is data of the operating sound of a second circuit breaker different from the first circuit breaker, as input data, and a second acoustic model, which is the acoustic model associated with the second feature, as training data. It has, The acoustic model is an activity level that is generated for each of the operating sounds emitted when the circuit breaker starts closing, when the circuit breaker finishes closing, when the circuit breaker starts opening, and when the circuit breaker finishes opening, and is represented by a triangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively, and then decays with a predetermined slope, or a rectangular wave that rises perpendicularly to the time axis starting from the start time and end time of the circuit breaker's closing or opening operation, respectively. In the second learning process, the first learning model output as a result of the first learning process is used as the initial value for the second learning model in the second learning process. A method for generating a circuit breaker operation model, characterized by the following features.

Images