WO2024014061A1

WO2024014061A1 - Isolator model generating system, isolator model generating device, and isolator model generating method

Info

Publication number: WO2024014061A1
Application number: PCT/JP2023/010966
Authority: WO
Inventors: 隆遠藤
Original assignee: 株式会社日立製作所
Priority date: 2022-07-11
Filing date: 2023-03-20
Publication date: 2024-01-18
Also published as: JP2024009691A

Abstract

In order to efficiently generate an isolator operation model, this isolator model generating system is characterized by comprising a common model learning unit (114) for outputting a common model (134) by means of common model learning (S1), which is learning that takes feature quantity tensors generated from acoustic data acquired from first isolators, being a plurality of installed isolators (201a, 201b), as input data, and takes degrees of activity associated with the feature quantity tensors (411) as teacher data, and an individual model learning unit (115) for outputting an individual model (136) by means of individual model learning (S2), which is learning that takes a feature quantity tensor generated on the basis of acoustic data acquired from an isolator (201c) different from the isolators (201a, 201b) as input data, and takes a degree of activity associated with the feature quantity tensor as teacher data, wherein, in the individual model learning (S2), the common model (134) output as a result of the common model learning (S1) is utilized as an initial value.

Description

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

The present invention relates to a circuit breaker model generation system, a circuit breaker model generation device, and a circuit breaker model generation method.

There is a growing need to remotely monitor the status of circuit breakers and perform maintenance without dispatching workers. In response to such needs, a method has been proposed in which various sensors are attached to the circuit breaker to grasp the operating status of the circuit breaker, and the state is determined based on the difference from the normal state. In such diagnosis, attempts have been made to monitor the state of the circuit breaker by installing a current sensor inside the circuit breaker or attaching an acceleration sensor to the casing of the circuit breaker.

There is a technique described in Patent Document 1 as a technique for attaching a sensor and making a judgment based on the acquired data. Patent Document 1 describes, "a power conversion circuit that converts input DC power into AC power, an output wiring that transmits the power converted by the power conversion circuit, and a voltage converter provided in the middle of the output wiring that converts the input DC power into AC power. an AC breaker constructed to be switched on and off based on a value; A detector, and a diagnostic device that diagnoses whether or not the AC circuit breaker that has emitted the sound is abnormal based on the sound detected by the sound detector and predetermined criterion information. ” A power conversion system and circuit breaker diagnostic device are disclosed.

JP2019-184341A

However, when applying the technique described in Patent Document 1 to a newly installed circuit breaker, it is necessary to actually operate the circuit breaker and collect a large amount of data on its normal state. However, as the number of operations of the circuit breaker increases, problems arise such as the deterioration of the circuit breaker progressing and the number of man-hours required for collection increasing.

The present invention was made in view of this background, and an object of the present invention is to efficiently generate an operational model of a circuit breaker.

In order to solve the above-mentioned problems, the present invention includes a sound collection device that collects the operating sound of a circuit breaker that cuts off power and outputs it as acoustic data, and a first sound collection device that is one of the plurality of circuit breakers installed. The acoustic model is an acoustic model that uses as input data a first feature amount that is a feature amount generated from first acoustic data that is the acoustic data obtained from the circuit breaker, and is associated with the first feature amount. A first learning unit that outputs a first learning model through first learning that is learning using a first acoustic model as teacher data, and a second learning unit that is the circuit breaker different from the first circuit breaker. As input data, a second feature quantity, which is a feature quantity, generated based on the second acoustic data, which is the acoustic data acquired from the circuit breaker, is used as input data, and the a second learning unit that outputs a second learning model through second learning that is learning using a second acoustic model that is an acoustic model as teacher data; In the second learning, the first learning model output as a result of the first learning is used as the initial learning model of the second learning model in the second learning. It is characterized by being used as a value.
Other solutions will be described as appropriate in the embodiments.

According to the present invention, an operational model of a circuit breaker can be efficiently generated.

FIG. 2 is a diagram showing an outline of circuit breaker diagnosis by the circuit breaker diagnosis system. FIG. 3 is a diagram showing details of collection of acoustic data and stroke position data used in this embodiment. 1 is a diagram showing a configuration example of a diagnostic device according to the present embodiment. It is a figure which shows each timing at the time of the closing operation of a circuit breaker. It is a figure which shows each timing at the time of the closing operation of a circuit breaker. FIG. 3 is a diagram illustrating an example of a feature tensor generation procedure in common model learning and individual model learning. FIG. 3 is a diagram showing a correspondence relationship between a feature amount tensor and an activation degree. FIG. 3 is a diagram illustrating processing in a common model learning unit performed in this embodiment. FIG. 2 is a diagram showing processing in an individual model learning unit performed in this embodiment. It is a figure showing the processing procedure of an acoustic diagnosis section performed in this embodiment. FIG. 2 is a diagram showing the configuration of common learning data and individual learning data. FIG. 3 is a diagram showing a data structure of teacher data in a common model. FIG. 3 is a diagram showing a specific example of the data structure of teacher data. It is a figure showing an example of a diagnosis result display screen outputted to an output device in this embodiment. FIG. 7 is a diagram showing another example of the degree of activity.

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

<Summary>
FIG. 1 is a diagram showing an outline of circuit breaker diagnosis by a circuit breaker diagnosis system (circuit breaker 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 section 114, an individual model learning section 115, and an acoustic diagnosis section 118.
Further, the diagnostic processing performed in the circuit breaker diagnostic system Z is performed in three stages: common model learning (S1), individual model learning (S2), and diagnosis (S3).

(Common model learning (S1: first learning, first learning step))
The circuit breaker 201 cuts off power.

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

Then, the stroke position measuring device 203 measures the position of the contact point during the opening/closing operation using a laser or the like. Specifically, the stroke position measuring device 203 measures the distance of an opening/closing rod (not shown) of the circuit breaker 201a. In this way, the stroke position measuring device 203 measures the position of the contact of the circuit breaker and outputs it as stroke position data 302 (contact position data, see FIGS. 4 and 5).

Note that the microphone 202 collects the sound when the circuit breaker 201a opens and closes, and the stroke position measuring device 203 measures the stroke position in synchronization. The acoustic data (first acoustic data) 301 that is the acoustic data collected by the microphone 202 and the stroke position data 302 that is the stroke position data are sent to the common model learning unit 114 as common learning data 133. Sent.

Any acquisition method may be used as long as the acoustic data 301 and the 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 constantly performed. Alternatively, when the open signal or the close signal of the circuit breaker 201 is detected, sound collection by the microphone 202 and measurement by the stroke position measuring device 203 may be started. Alternatively, when a person signals the opening or closing operation of the circuit breaker 201, the microphone 202 and the stroke position measuring device 203 may be turned on manually.

Similarly, from a circuit breaker 201b (201: first circuit breaker) different from the circuit breaker 201a, the sound of the opening/closing operation of the circuit breaker 201b by the microphone 202 and the stroke position by the stroke position measuring device 203 are synchronized. It is measured by The measured acoustic data (first acoustic data) 301 and stroke position data 302 of the circuit breaker 201b are sent to the common model learning unit 114 as common learning data 133. Note that the acoustic data 301 and stroke position data 302 of circuit breakers 201a of the same model having different operating conditions may be measured. The different operating conditions include, for example, different oils used.

Note that the circuit breakers 201a and 201b from which the common learning data 133 is collected during common model learning (S1) are circuit breakers 201 that are known to have no abnormality.

Note that 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.

Note that in the common model learning (S1) in FIG. 1, two sets of circuit breaker 201, microphone 202, and stroke position measuring device 203 are provided. However, the present invention is not limited to this, and the acoustic data 301 and the stroke position data 302 may be measured for three or more sets of circuit breakers 201, microphones 202, and stroke position measuring devices 203. In this way, in common model learning (S1), acoustic data 301 and stroke position data 302 are acquired from a plurality of circuit breakers 201 installed.

Then, the common model learning unit (first learning unit) 114 uses the common learning data 133 acquired from each of the circuit breaker 201a and the circuit breaker 201b to generate a common model (first learning model) 134. Learn. Specifically, the common model 134 is the connection strength between neurons of the deep neural network (neural network, first neural network) N1 (see FIG. 8). Learning of the common model 134 will be described later. The common model 134 is a learning result by the common model learning unit 114.

(Individual model learning (S2: second learning))
Subsequently, when a new circuit breaker 201c (201: a second circuit breaker that is a circuit breaker 201 different from the first circuit breaker) is additionally installed, the individual model learning unit (second learning unit) 115 Individual learning (S2) is performed.
First, an individual model (second learning model) 136 is created for diagnosing the opening/closing operation of the newly installed circuit breaker 201c. The circuit breaker 201c is the circuit breaker 201 to be diagnosed later. Therefore, the microphone 202 installed in the circuit breaker 201c collects the sound when the circuit breaker 201c is opened and closed. Further, a stroke position measuring device 203 installed in the circuit breaker 201c measures the stroke position. The sound collection in the circuit breaker 201c and the measurement of the stroke position are performed synchronously. Acoustic data (second acoustic data) 301 of the sound collected by the microphone 202 and stroke position data 302 measured by the 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 the stroke position data 302 collected as the individual learning data 135 may be smaller in number than the common learning data 133.

Note that the circuit breaker 201c from which the individual learning data 135 is collected during the individual model learning (S2) is a circuit breaker 201 that is known to have no abnormality at the time the individual model learning (S2) is performed. It is.

The individual model learning unit 115 uses a deep neural network (neural network, second neural network) N2 (see FIG. 9) using the already learned model (common model 134) as an initial value by the common model learning unit 114. , the individual model 136 is learned by inputting the individual learning data 135. The learning of the individual model 136 will be described later. The individual model 136 is a learning result by the individual model learning unit 115, and specifically, is the connection strength of neurons in the deep neural network N2.

(Diagnosis (S3))
After learning the individual model 136, the acoustic diagnosis section (diagnosis section) 118 diagnoses the circuit breaker 201c (S3). Note that the acoustic diagnosis section 118 corresponds to the activity estimation processing section 116 and the state diagnosis processing section 117 in FIG.
When diagnosing the circuit breaker 201c, the acoustic diagnosis unit 118 collects acoustic data (third acoustic data that is different from the second acoustic data) that is the data of the sound collected by the installed microphone 202. ) 301 and the individual model 136 to diagnose the circuit breaker 201c. First, the acoustic diagnosis unit 118 inputs the acquired acoustic data 301 to the individual model 136. The acoustic diagnosis unit 118 diagnoses the operation of the circuit breaker 201c based on the diagnosis result 211 that is the output result from the individual model 136.

<Details of collecting acoustic data 301 and stroke position data 302>
FIG. 2 is a diagram showing details of collection of acoustic data 301 and stroke position data 302 used in this embodiment. Reference is made to FIG. 1 as appropriate.
The collection of acoustic data 301 and stroke position data 302 in FIG. 2 is the same as the collection of common learning data 133 in common model learning (S1) in FIG. 1 and the collection of individual learning data 135 in individual model learning (S2). This is a common process.

As described above, the microphone 202 that collects the sound of the circuit breaker 201 and the stroke position measuring device 203 are installed near the circuit breaker 201 (distance L1). The logger 221 also synchronizes the acoustic data 301 (see FIGS. 4 and 5) output from the microphone 202 and the stroke position data 302 (see FIGS. 4 and 5) output from the stroke position measuring device 203. and obtain it. 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 section 114, an individual model learning section 115, and an acoustic diagnosis section 118 shown in FIG. Further, the diagnostic device 1 can store common learning data 133, common model 134, individual learning data 135, and individual model 136 in the auxiliary storage device 130 (see FIG. 3).

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 the common learning data 133 and the individual learning data 135 shown in FIG.

In this way, the system shown in FIG. 2 can be used when collecting the common learning data 133 when generating the common model 134. Furthermore, the system shown in FIG. 2 can also be used when collecting a small amount of individual learning data 135 when generating the individual model 136 of the new circuit breaker 201c. Further, for the diagnosis (S3 in FIG. 1), the system shown in FIG. 2 can be used, except that the stroke position measuring device 203 is not used.

<Diagnostic device 1>
FIG. 3 is a diagram showing a configuration example of the diagnostic device (breaker model generation device) 1 according to the present embodiment.
The diagnostic device 1 includes a main storage device 100 configured with a RAM (Random Access Memory) or the like, and an auxiliary storage device 130 configured with an HDD (Hard Disk Drive), an SDD (Solid State Drive), or the like. Furthermore, the diagnostic device 1 includes a central processing unit 101 including a CPU (Central Processing Unit) and a GPU (Graphic Processing Unit). The diagnostic device 1 also includes an input device 102 including a keyboard and a mouse, an output device (display device) 103 including a display, and a communication device 104 that communicates with a logger 221 and the like. Note that in FIG. 3, the logger 221 is indicated by a broken line to indicate that the logger 221 is not a component of the diagnostic device 1 in this embodiment.

(Auxiliary storage device 130)
The auxiliary storage device 130 stores a 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. There is.
The diagnostic software program 131 is a program for executing the diagnostic software 110.
The microphone position information 132 is information regarding the positional relationship (distance L1 in FIG. 2) between the microphone 202 and the circuit breaker 201.
The common learning data 133 is data used for common model learning (S1 in FIG. 1) by the common model learning unit 114.
The common model 134 is a model output by common model learning (S1 in FIG. 1) by the common model learning unit 114. As described above, the common model 134 is specifically the connection strength of each neuron in the neural network that constitutes the common model 134.

The individual learning data 135 is data used in individual model learning (S2 in FIG. 1) by the individual model learning unit 115.
The individual model 136 is a model output by individual model learning by the individual model learning unit 115 (S2 in FIG. 1). As described above, the individual model 136 is specifically the connection strength of each neuron in the neural network that constitutes the individual model 136.

Diagnostic parameters 137 are parameters used for diagnosis. For example, the diagnostic parameter 137 may be a threshold value indicating by what percentage deviation from the rated average of the estimated opening operation start timing and opening operation end timing is determined to be abnormal.
The diagnosis history information 138 is the diagnosis results 211 (see FIG. 1) performed by the state diagnosis processing unit 117 up to now.

The main storage device 100 is provided with diagnostic software 110 and a work area 120.
(Diagnostic software 110)
The diagnostic software 110 is realized by loading the diagnostic software program 131 stored in the auxiliary storage device 130 into the main storage device 100 and executing it by the central processing unit 101. The diagnostic software 110 includes an overall control processing section 111, a spectrogram generation section (feature generation section) 112, a feature extraction processing section (feature setting section, feature generation section) 113, a common model learning section 114, and an individual model learning section. 115, an activity estimation processing section (diagnosis section) 116, and a state diagnosis processing section (diagnosis section) 117.

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

The spectrogram generation unit 112 generates a spectrogram 401 (see FIG. 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 quantity extraction processing unit 113 extracts a feature quantity tensor (feature quantity) 411 (FIGS. 6 and 7) from the spectrogram 401 generated by the spectrogram generation unit 112. The feature amount tensor 411 is obtained by cutting out the spectrogram 401 at a predetermined time width. Note that details of the feature amount tensor 411 will be described later. Incidentally, a tensor is a multidimensional array that includes vectors and matrices.

The common model learning unit 114 inputs a feature tensor 411 obtained from the acoustic data 301 obtained from multiple models and circuit breakers 201 having different operating conditions into a neural network, and learns the common model 134. generate.
The individual model learning unit 115 inputs the feature quantity tensor 411 obtained from the acoustic data 301 when the circuit breaker 201 to be diagnosed is normal to the common model 134, and generates the individual model 136 by learning.

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

Incidentally, the activity estimation processing section 116 and the state diagnosis processing section 117 constitute the acoustic diagnosis section 118 in FIG.

(Work area 120)
Further, the work area 120 is an area used as a temporary storage unit when the diagnostic software 110 is being executed.
The work area 120 is provided with an acoustic data storage area 121, a spectrogram storage area 122, a feature storage area 123, a model storage area 124, an estimated activity storage area 125, a diagnostic parameter storage area 126, and a diagnostic result storage area 127. ing.
The acoustic data storage area 121 is an area where acoustic data 301 acquired from the microphone 202 is temporarily stored.
The spectrogram 401 is temporarily stored in the spectrogram storage area 122.
The feature amount tensor 411 is temporarily stored in the feature amount storage area 123.
A common model 134 and individual models 136 are temporarily stored in the model storage area 124.

The estimated activity level 681 estimated by the individual model 136 is temporarily stored in the estimated activity level storage area 125 .
In the diagnostic parameter storage area 126, diagnostic parameters 137 used when diagnosing the circuit breaker 201 are temporarily stored.
The diagnosis result storage area 127 temporarily stores the diagnosis result 211 obtained by the state diagnosis processing section 117.

<Closing operation>
FIG. 4 is a diagram showing each timing during the closing operation of the circuit breaker 201 (see FIGS. 1 and 2).
In FIG. 4, acoustic data 301a (301) and stroke position data 302b (302) are shown in order from the top. Furthermore, in FIG. 4, the activity level 310 at the start of the closing operation (activity level 311 at the start of the closing operation), the activity level 310 at the end of the closing operation (activity level 312 at the end of the closing operation), and the activity level 310 at the start of the opening operation. (Activity level 313 at the start of the opening operation) and activity level 310 at the end of the opening operation (Activity level 314 at the end of the opening operation) are shown. Further, in FIG. 4, the respective graphs are arranged along the time axis. Note that the activity level 310 is an acoustic model that models the sound generated when the circuit breaker 201 operates.

Note that the activation degree 311 at the start of the closing operation is the activation degree 310 generated for the operation sound emitted when the closing operation of the circuit breaker 201 starts. The activation degree 312 at the end of the closing operation is the activation degree 310 generated for the operation sound emitted when the closing operation of the circuit breaker ends. Furthermore, the opening operation start activation degree 313 is the activation degree 310 generated regarding the operation sound emitted when the opening operation of the circuit breaker 201 starts. The activation degree 314 at the end of the opening operation is the activation degree 310 generated regarding the operation sound emitted when the opening operation of the circuit breaker ends.

Note that, as described above, the activity level 310 is an acoustic model that models the sound generated when the circuit breaker 201 operates, such as when the circuit breaker 201 closes or opens.
As the activity level 310 related to the closing operation of the circuit breaker 201, two types of activity level 310 are set: an activity level 311 at the start of the closing operation, and an activity level 312 at the end of the closing operation. The reason for this is that two sounds are generated when the circuit breaker 201 is closed in the example of this embodiment.

Furthermore, two types of activation degrees 310 are set as the activation degrees 310 regarding the opening operation of the circuit breaker 201: an activation degree 313 at the start of the opening operation and an activation degree 314 at the end of the opening operation. The reason for this is also that two sounds are generated when the circuit breaker 201 is opened.

Therefore, the activity level 310 may be set according to the sound generated by the circuit breaker 201 during the closing and opening operations. For example, if one sound is generated in each of the closing operation and the opening operation, one type of activation level 310 may be prepared for each type, and if three or more sounds are generated, Three types of activation levels 310 may be set for each of the closing operation and the opening operation. Further, if the number of sounds generated is different for each of the closing operation and the opening operation, different numbers of activation degrees 310 may be set. For example, if three sounds are generated during the closing operation and two sounds are generated during the opening operation, three types of activation levels 310 are set during the closing operation, and two types of activation levels 310 are set during the opening operation. It only needs to be set.

Furthermore, since FIG. 4 shows the closing operation, the activation level 313 at the start of the opening operation and the activation level 314 at the end of the opening operation are inactive. Note that the broken lines in the graphs of the activity at the start of closing operation 311, the activity at the end of closing operation 312, the activity at the start of opening operation 313, and the activity at the end of opening operation 314 indicate the peak values of the respective activities 310. ing.

As described above, the acoustic data 301a is stored in synchronization with the stroke position data 302a. The true values of the start time and end time of the closing operation can be determined from the change start time 321 and change end time 322 of the stroke position data 302a. Thereby, the closing operation start activation level 311 is set so that it rises at the start time of the closing operation (change start time 321 of the stroke position data 302a) and decreases at a constant rate. The activation degree 311 at the start of the closing operation set in this way becomes a model of the acoustic intensity at the starting point of the closing operation. Similarly, the activation level 312 at the end of the closing operation is set so that it rises at the end time (change end time 322 of the stroke position data 302a) and decreases at a constant rate. The activation degree 312 at the end of the closing operation set in this way becomes a model of the sound intensity at the end point of the closing operation.

As shown in FIG. 4, each of the degrees of activity 310 rises perpendicularly to the time axis with the start time and end time of the closing or opening action as the starting point (action start time), and then rises at a predetermined slope. It is expressed in the form of a triangular wave that is attenuated by .

<Opening operation>
FIG. 5 is a diagram showing each timing during the opening operation of the circuit breaker 201 (see FIGS. 1 and 2).
Similarly to FIG. 4, in FIG. 5, acoustic data 301b and stroke position data 302b are shown in order from the top. Furthermore, in FIG. 5, the activity level 310 at the start of the closing operation (activity level 311 at the start of the closing operation), the activity level 310 at the end of the closing operation (activity level 312 at the end of the closing operation), and the activity level 310 at the start of the opening operation. (Activity level 313 at the start of the opening operation) and activity level 310 at the end of the opening operation (Activity level 314 at the end of the opening operation) are shown. Note that in FIG. 5, the respective graphs are aligned on the time axis.

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

As described above, in this embodiment, the opening operation and the closing operation of the circuit breaker 201 are used as the operation of the circuit breaker 201.

Further, in this embodiment, the activity level 310 is composed of an activity level at the start of the closing operation 311, an activity level at the end of the closing operation 312, an activity level at the start of the opening operation 313, and an activity level at the end of the opening operation 314. . By configuring the activity level 310 as described above, it can be applied to the circuit breaker 201 that makes two sounds during each of the opening operation and the closing operation.

(Generation of feature tensor 411)
FIG. 6 is a diagram illustrating an example of a procedure for generating the feature quantity tensor 411 in common model learning (S1 in FIG. 1) and individual model learning (S2 in FIG. 1). Refer to FIG. 3 as appropriate. Note that FIG. 6 explains in detail the generation process of the common learning data 133 in the common model learning (S1) of FIG. 1 and the individual learning data 135 in the individual model learning (S2).
The acoustic data 301 is subjected to frequency analysis (S101) by the spectrogram generation unit 112 and converted into a spectrogram 401. Hereinafter, the acoustic data 301 collected for use in learning will be appropriately referred to as learning acoustic data 301A (first acoustic data, second acoustic data). As shown in FIG. 6, in the spectrogram 401, the vertical axis represents frequency and the horizontal axis represents time. In other words, the spectrogram 401 is obtained by connecting the frequency analysis results of each waveform cut out in a fixed time window W1 with respect to the learning acoustic data 301A. In other words, the spectrogram 401 is obtained by converting an acoustic waveform into a spectral sequence at regular time intervals.

Next, the feature extraction processing unit 113 performs frame stacking (S102) on the generated spectrogram 401. That is, the feature amount extraction processing unit 113 extracts (stack) the spectrograms 401 in a constant time window W2 while shifting the time of the spectrograms 401 by a predetermined time. Each of the extracted spectrograms 401 is referred to as a feature amount tensor 411. In this way, the feature amount extraction processing unit 113 generates a plurality of feature amount tensors 411 by cutting out the spectrogram 401 at the predetermined time window W2. In this way, the feature quantity extraction processing unit 113 acquires the feature quantity tensor 411 while shifting its start time by a predetermined time width (obtains by shifting a predetermined time window by a predetermined time width). Thereby, the feature quantity extraction processing unit 113 generates a plurality of feature quantity tensors 411. In this way, the feature quantity tensor 411 is generated from the acoustic data 301 (in the example shown in FIG. 6, the learning acoustic data 301A). At this time, it is desirable that the time width for shifting is set smaller than the time width that the feature amount tensor 411 has. By doing so, the time resolution of the feature amount tensor 411 can be improved.

The feature tensor 411 shown in FIG. 6 has a time width of 17 msec and a frequency width of 77 frames. Note that the frequency width is divided into predetermined frequency units (for example, 10 Hz), and the frequency width of 77 frames means 77 frequency units (770 Hz). Further, the number given below the feature amount tensor 411 in FIG. 6 is the number of the feature amount tensor 411. In the example shown in FIG. 6, nfrm feature quantity tensors 411 are generated.

<Correspondence between feature quantity tensor 411 and activation degree 310>
FIG. 7 is a diagram showing the correspondence between the feature amount tensor 411 and the activity level 310.
For one feature quantity tensor 411, the time at the corresponding activation 310 (activation at the start of closing operation 311, activation at the end of closing operation 312, activation at the start of opening operation 313, activation at the end of opening operation 314) Intervals are always mapped. For example, the feature amount tensor 411z shown in FIG. 7 is associated with the time interval T1 in each activity level 310. In other words, the feature quantity tensor 411z corresponds to a decay period from before the rise in the activation degree 313 at the start of the opening operation. In other words, the feature quantity tensor 411z indicates the sound at the start of the opening operation of the circuit breaker 201. As shown in FIGS. 4 and 5, the acoustic data 301 and the respective activation degrees 310 are associated with each other based on the acoustic data 301 and the stroke position data 302. Therefore, the feature quantity tensor 411 generated based on the acoustic data 301 and each activation level 310 are associated with each other.

Similarly, other feature quantity tensors 411 are associated with time intervals in their respective activation levels 310. This kind of association between the feature quantity tensor 411 and the activation level 310 is performed by the feature quantity extraction processing unit 113 shown in FIG. In this way, the feature amount extraction processing unit 113 calculates the activity level 310, which is a model of the sound generated when the circuit breaker 201 operates, based on the acoustic data 301 and the stroke position data 302, and the acoustic data 301. The generated feature amount tensor 411 is associated with the generated feature amount tensor 411. Incidentally, the feature quantity tensor (first feature quantity) 411 used for learning the common model (see FIG. 1) 134 is also the feature quantity tensor (second feature quantity) used for learning the individual model (see FIG. 1) 136. ) 411 is also generated using the same procedure.

In the common model 134 and the individual model 136, learning is performed using the feature quantity tensor 411 as input data and the corresponding activation degree 310 as teacher data.

(Common model learning unit 114)
FIG. 8 is a diagram showing the processing performed by the common model learning unit 114 in this embodiment. FIG. 8 explains in detail the processing in the common model learning unit 114 of common model learning (S1) in FIG. 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. Further, the common model learning unit 114 includes a first full-connection neural network layer 521 and a second full-connection neural network layer 522. In this way, the deep neural network (first neural network) N1 used in common model learning (S1: see FIG. 1) is composed of a plurality of stages of neural networks.

First, the common learning data 133 is generated based on learning acoustic data (first acoustic data) 301A (see FIG. 6) acquired from (multiple) circuit breakers 201 of various models and operating conditions. A feature quantity tensor (first feature quantity) 411 (411a, 411b: first feature quantity) is stored. In addition, in FIG. 8, the following information is written near the feature amount tensor 411, the first converted data 502, the second converted data 503, the third converted data 504, the fourth converted data 505, and the fifth converted data 506. The number shown indicates the data size.

First, one of the feature quantity tensors 411 is input to the convolutional neural network layer 511 as input data. As shown in FIG. 8, the feature amount tensor 411 is 17×77×1 data.

The feature quantity tensor 411 is converted by the convolutional neural network layer 511 into first conversion data 502 having a data size of 15×75×16. Note that among the numbers indicating the data size of the first converted data 502, “16” is the number of channels and indicates the number of features (filters) of the spectrogram 401. The number of channels is a value preset by the user.

Further, this first converted data 502 is compressed by the first pooling layer 512 into second converted data 503 having a data size of 6×36×16. Among the numbers indicating the data size of the second converted data 503, “16” indicates the number of features (filters) of the spectrogram 401, similarly to the first converted data 502. Further, the second transformed data 503 is further compressed by the second pooling layer 513 into third transformed data 504 having a data size of 2×17×16. Among the numbers indicating the data size of the third converted data 504, “16” indicates the number of features (filters) of the spectrogram 401, similarly to the first converted data 502 and the second converted data 503.

Subsequently, the common model learning unit 114 converts the third converted data 504 into fourth converted data 505 having a data size of 1×544. The fourth converted data 505 is simply the third converted data 504 rearranged into 1×544 data based on a predetermined arrangement rule.

Then, the first full-connection neural network layer 521 converts the fourth transformed data 505 into fifth transformed data 506 having a data size of 1×128. The fifth transformed data 506 is then transformed into common model output data 507 by the second full-connection neural network layer 522. The data size of the common model output data 507 is 4c. Here, c is the model of the circuit breaker 201 and the number of operating conditions. And, "4" is the above-described four activation 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. In other words, the feature tensor 411 input to the convolutional neural network layer 511 includes an activation level at the start of closing operation 311, an activation level at the end of closing operation 312, and an activation level at the start of opening operation, which are prepared for each model and operating condition. 313 and the activation level 314 at the end of the opening operation, the data is output as data corresponding to the state of any one of the activation levels 310 .

On the other hand, the common learning data 133 holds the state of the activation degree 310 (310a, 310b: first acoustic model) that is associated with each feature amount data. Therefore, the common model learning unit 114 calculates the difference between the common model output data 507 and the teacher data associated with the feature quantity tensor 411 input to the convolutional neural network layer 511 as an error (S111). The common model learning unit 114 then performs backpropagation processing (S112) to update the connection strength of neurons in each layer based on the error. Note that the initial value of the connection strength of neurons in each layer is set by a random number.

After performing the backpropagation process, the common model learning unit 114 inputs the next feature quantity tensor 411 to the convolutional neural network layer 511. Then, when all the feature quantity tensors 411 are input to the convolutional neural network layer 511, the feature quantity tensor 411 that was first input to the convolutional neural network layer 511 is input to the convolutional neural network layer 511 again.

The common model learning unit 114 continues the above process a predetermined number of times or until the error reaches a predetermined condition (for example, a predetermined value or less). The common model learning unit 114 outputs the connection strength of neurons in each layer calculated through the above processing as a common model 134.

In this way, the common model learning unit 114 generates a common model 134 based on the feature quantity tensor 411 related to a large number of models and operating conditions, and the activity level 310 related to the opening/closing operation sound of the circuit breaker 201 in a plurality of models and operating conditions. generate.

(Individual model learning unit 115)
FIG. 9 is a diagram showing the processing performed by the individual model learning unit 115 in this embodiment. Note that FIG. 9 explains in detail the processing of the individual model learning unit 115 of the individual model learning (S2) in FIG. 1.
Like the common model learning unit 114, the individual model learning unit 115 includes a convolutional neural network layer 511, a first pooling layer 512, a second pooling layer 513, a first full-connection neural network layer 521, and a second full-connection neural network layer. A neural network layer 522 is included. In this way, the configuration of the deep neural network (second neural network) N2 used in individual model learning (S2: see FIG. 1) is composed of a plurality of stages of neural networks. Further, the deep neural network N2 used in the individual model learning (S2) has the same structure as the deep neural network N1 (see FIG. 8) used in the common model learning (S2).

Further, for the convolutional neural network layer 511 to the first full-connection 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 neuron connection strength regarding the data indicated by diagonal lines in FIG. 9 is set to the value of the common model 134 stored in the common model 134 as an initial value. However, in FIG. 9, the initial value of the connection strength of neurons in the second full-connection neural network layer (neural network one stage before the output) 522 is set as a random number. In this way, in the individual model learning (S2), the common model 134 output as a result of the common model learning (S1) is used as the initial value of the individual model 136 (deep neural network N2) in the individual model learning (S2).

Furthermore, in the individual model learning unit 115, the final output of the individual model output data 531 is four activation degrees 310 for the circuit breaker 201c (see FIG. 1) to be diagnosed. The four activation levels 310 are the activation level 311 at the start of the closing operation, the activation level 312 at the end of the closing operation, the activation level 313 at the start of the opening operation, and the activation level 314 at the end of the opening operation shown in FIG. 4 and FIG. Therefore, the individual model output data 531 becomes 4×1 data. Unlike the common model learning unit 114, the individual model learning unit 115 targets only one circuit breaker 201, so the individual model output data 531 is 1×4 data.

Feature tensor (second feature) 411c extracted from learning acoustic data (second acoustic data) 301A (see FIG. 6) acquired by the newly installed circuit breaker 201c (to be diagnosed) (411) is stored in the individual learning data 135. Further, the state of the activation level (second acoustic model) 310c (310) corresponding to the feature amount tensor 411c is stored in the individual learning data 135 as teacher data.

Then, the individual model learning unit 115 inputs the first feature quantity tensor 411c to the convolutional neural network layer 511. Thereafter, processing is performed by the convolutional neural network layer 511, the first pooling layer 512, the second pooling layer 513, and the first full-connection neural network layer 521, thereby outputting the fifth converted data 506. The fifth converted data 506 is then converted into individual model output data 531 by the second full-connection neural network layer 522.

That is, in the individual model learning unit 115, the input feature quantity tensor 411 is output as a state of one of the four activation degrees 310. The four activation levels 310 are an activation level 311 at the start of the closing operation, an activation level 312 at the end of the closing operation, an activation level 313 at the start of the opening operation, and an activity level 314 at the end of the opening operation.

On the other hand, as described above, the state of the activity level 310 associated with the feature quantity tensor 411 of the circuit breaker 201c is held as teacher data in the individual learning data 135. Therefore, the individual model learning unit 115 calculates the difference between the individual model output data 531 and the teacher data (activity level 310) associated with the feature amount tensor 411 in the input convolutional neural network layer 511 as an error. (S211). The common model learning unit 114 then performs backpropagation processing (S212) to update the connection strength of neurons in each layer based on the error.

After performing the backpropagation process, the individual model learning unit 115 inputs the next feature quantity tensor 411 to the convolutional neural network layer 511. Then, when all the feature quantity tensors 411 are input to the convolutional neural network layer 511, the feature quantity tensor 411 that was first input to the convolutional neural network layer 511 is input to the convolutional neural network layer 511 again.

The individual model learning unit 115 continues the above process a predetermined number of times or until the error reaches a predetermined condition (for example, a predetermined value or less). The individual model learning unit 115 stores the connection strength of neurons in each layer calculated through the above processing as an individual model 136 in the auxiliary storage device 130 (see FIG. 3).

In this manner, in generating the individual model 136, the common model 134 is first set as the initial value of the first full-connection 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 feature amount tensor 411 of the newly installed circuit breaker 201c and the activation degree 310 of the circuit breaker 201c. At this time, the individual model learning unit 115 inputs the activation degree 310 (teacher data) of the circuit breaker 201c and the feature quantity tensor 411 of the circuit breaker 201c to the convolutional neural network layer 511, thereby creating a second full-connection neural network layer. Learning is performed so that the difference between the output of 522 (individual model output data 531) and the activity level 310 as teacher data becomes small. In this way, learning of the individual model 136 is performed.

In this way, the initial value of the connection strength in the neurons from the convolutional neural network layer 511 to the first full-connection neural network layer 521 configuring the individual model 136 is the same as the value of the common model 134 generated by the common model learning unit 114. is set. This eliminates the need for the individual model learning unit 115 to learn the connection strengths from the convolutional neural network layer 511 to the first full-connection neural network layer 521 from the state of random numbers. Therefore, even if the number of feature quantity tensors 411 acquired for learning from the newly installed circuit breaker 201c is small, the learning of the individual model 136 can be stably converged.

(Processing of acoustic diagnosis unit 118)
FIG. 10 is a diagram showing the processing procedure of the acoustic diagnosis section 118 performed in this embodiment. Note that FIG. 10 explains in detail the processing of the acoustic diagnosis unit 118 of the diagnosis (S3) in FIG. 1.
First, acoustic data (third acoustic data) 301 is acquired from the circuit breaker 201c to be diagnosed. Like the acoustic data 301 acquired in FIG. 10, the acoustic data 301 acquired when diagnosing the circuit breaker 201 is appropriately referred to as diagnostic acoustic data (third acoustic data) 301B. The circuit breaker 201c is a circuit breaker 201c that has been trained by the individual model learning section 115.

Subsequently, the activity estimation processing unit 116 performs frequency analysis (S301) to convert the diagnostic acoustic data 301B into a spectrogram 401 composed of spectra for each time. Further, the activity estimation processing unit 116 performs frame stacking (S302) to acquire a plurality of spectrograms 401 by extracting spectrograms 401 of a predetermined time width while shifting time. As a result, a plurality of feature quantity tensors (third feature quantities) 411, which are spectrograms 401 having a predetermined time width, are generated. The data size of the feature quantity tensor 411 is the same as the feature quantity tensor 411 generated in FIG. The processing up to this point is similar to the processing shown in FIG.

Subsequently, the activity estimation processing unit 116 sequentially inputs the feature quantity tensor 411 to the individual model 136 trained by the individual model learning unit 115. Each feature tensor 411 input to the individual model 136 is output as estimated activity levels (estimated acoustic model) 681 of the four opening/closing operation sounds (S303). The estimated activity level 681 is the activity level 310 estimated by the individual model 136. As shown in FIG. 10, the estimated activity 681 is similar to the activity 310: estimated activity 681A at the start of closing operation, estimated activity 681B at the end of closing operation, estimated activity 681C at the start of opening operation, and estimated activity 681C at the end of opening operation. It has an estimated activity of 681D.

When all the feature quantity tensors 411 are converted into estimated activities 681 and the converted estimated activities 681 are arranged in chronological order (in the input order of the feature quantity tensors 411), four types of estimated activities based on the diagnostic acoustic data 301B are obtained. A state time series of 681 degrees is obtained. Based on the peak value of the estimated activity 681, the activity estimation processing unit 116 can estimate the start time and end time of the opening or closing operation of the circuit breaker 201c (estimates the operation timing). In this way, the activity estimation processing unit 116 estimates the operation timing of the circuit breaker 201c based on the estimated activity 681 estimated from the diagnostic acoustic data 301B.

The state diagnosis processing unit 117 compares the time required for the opening/closing operation with the rated timing (reference operation timing) based on the estimated start time and end time (estimated operation timing), and performs the shutoff. The health of the device 201 is determined. The state diagnosis processing unit 117 determines whether or not an abnormality has occurred in the circuit breaker 201c based on the operation timing estimated from the result outputted by inputting the feature quantity tensor 411 to the individual model 136.

Specifically, the state diagnosis processing unit 117 calculates the degree of deviation between the estimated start time and end time and the rated timing, and if the degree of deviation is a predetermined value or more, It is determined that an abnormality has occurred.

(Correspondence between feature quantity tensor 411 and activity level 310)
FIG. 11 is a diagram showing the configuration of the common learning data 133 and the individual learning data 135.
FIG. 11 shows an example of a method of storing the feature quantity tensor 411 and the activity degree 310 (see FIG. 6) in the common learning data 133 and the individual learning data 135. However, the method of storing the feature amount tensor 411 and the activity degree 310 in the common learning data 133 and the individual learning data 135 is not limited to the example shown in FIG. 11. In FIG. 11, reference is made to FIG. 1 as appropriate.
As shown in FIG. 11, in the common learning data 133 and the individual learning data 135, ID (item 701), feature value tensor 411 (item 702), and activation degree 310 (item 703) are stored in correspondence with each other. has been done.
The ID is an ID assigned to a pair of the feature amount tensor 411 and the activity level 310.
As shown in FIG. 11, the feature quantity tensor 411 is stored in the format of "float x _i,j [nfrm _i,j , freq, nstack]".
Here, "float" indicates that the value to be saved is a floating point tensor. Further, i of x _i,j indicates the model or operating condition of the circuit breaker 201, and j indicates the number of the feature amount tensor 411. The numbers of the feature quantity tensor 411 indicated by j are assigned in order from oldest to newest. Hereinafter, in order to simplify the explanation, it is assumed that i indicates the model of the circuit breaker 201.

FIG. 11 shows that such a feature quantity tensor 411 has a size of nfrm _i,j x freq x 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". Furthermore, freq is the number of frames in the frequency direction. In the example shown in FIG. 6, freq is "77". Further, nstack is the number of feature quantity tensors 411 having a size of nfrmi,j×freq that are simultaneously input to the common model 134 or the individual models 136 (input in a bundle). In the neural network, n feature quantity tensors 411 having a size of nfrm _i,j ×freq may be input at once. In the example shown in FIG. 6, the feature amount tensor 411 is input one by one, so ntack is "1".

Furthermore, the activity level 310 is stored in the format of "float y _i,j [nfrm _i,j , 4]". Here, "float" indicates that the value to be saved is a floating point tensor. Further, i in y _i,j indicates the model or operating condition, and j indicates a number like j in the feature amount tensor 411. The activity level 310 is a tensor (matrix) having a size of nfrm _i,j ×4.

Further, nfrm _i,j indicates the number of frames in the time direction of the activity level 310, and contains the same number as the feature amount tensor 411 (“17” in the example shown in FIG. 6). The last "4" means that there are four types of activation levels 310 (activation level 311 at the start of closing operation, activation 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). It shows. The first column of the tensor indicating the activation degree 310 stores the activation degree 311 at the start of the closing operation, and the second column stores the activation degree 312 at the end of the closing operation. The third column stores the activation degree 313 at the start of the opening operation, and the fourth column stores the activation degree 314 at the end of the opening operation.

Both the feature quantity tensor 411 and the activity level 310 have data whose number of frames (number of frames at time: nfrm _i,j ) depends on the length of the acoustic waveform of the acoustic data 301 as the source. The feature quantity tensor 411 stores data obtained by bundling data for the number of freqs in the frequency direction into adjacent nstacks. As a result, the feature amount tensor 411 is a three-dimensional array (tensor) in which the number of elements (sizes) in each dimension are nfrm, freq, and nstack.

On the other hand, as described above, the activity level 310 consists of four types of activity level 310, which are a combination of two types of operation start and operation end for two types of operations, the closing operation and the opening operation. The activity level 310 is stored for each frame in correspondence with the feature amount tensor 411. As a result, the activation level 310 becomes a two-dimensional array (tensor), and the number of elements in the first dimension (row) is nfrm, and the second dimension (column) is stored in an array of elements corresponding to each of the four types of activation level 310. be done.

(Common model output data 507)
FIG. 12 is a diagram showing the data structure of the teacher data in the common model 134. In FIG. 12, the teacher data in the common model 134 is simply described as teacher data.
The teacher data ID (symbol 711) is an ID uniquely given to the teacher data.
As indicated by reference numeral 712, the teacher data includes data of "Y _i,j [nfrm _i,j , 4c]". The teacher data may be expressed as floating point type data, or may be expressed as a [1,0] type integer.

i in Y _i,j indicates the model or operating condition. Further, j indicates the number of teacher data. The numbers indicated by j are assigned in ascending order of time. Further, it is shown that the teacher data is a tensor (matrix) having a size of nfrm _i,j ×4c. nfrm _i,j is the number of frames (number of data) in the time direction for model i and number j, as in FIG. Specifically, nfrm _i,j is "17" as in FIG. 11. c is the number of models and operating conditions. In FIG. 12, only the model is considered.

That is, the teacher data has nfrm _i,j elements (data) in the row direction and 4c data in the column direction. In the column direction, four types of activity levels 310 are stored in line for each model.

Reference numeral 720 in FIG. 12 indicates a specific configuration of each teacher data. In "y _i,j [nfrm _i,j ,4]", i, j and nfrm _i,j are the same as "Y _i,j ". "4" means that the activation 310 having the size nfrm _i,j in the row direction is the activation 311 at the start of the closing operation, the activation 312 at the end of the closing operation, the activation 313 at the start of the opening operation, and the activation 313 at the end of the opening operation. It shows that they are arranged in the column direction in order of activity level 314. Note that "zeros[nfrm _{i, j} , 4]" indicates that all components are "0". In "Y _{i, j} [nfrm _{i, j} , 4c]", such activity 310 is calculated for each model ("Model A activity", "Model B activity", ... "Model C activity"). ” (codes 721 to 723)) are arranged in the column direction.

In other words, as shown in FIG. 13, one piece of teacher data (at a certain time) is "activity level at the start of closing operation of model A", "activity level at the end of closing operation of model A", etc. in the row direction. It has a configuration arranged in . In FIG. 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", . . . has the number of elements of 1×nfrm _i,j .

The common model 134 is learned as a model that predicts the value of the activity level 310 of multiple models. Therefore, as shown in FIG. 12, activation levels 310 for common model learning are prepared as teacher data, in which activation levels 310 corresponding to different models are padded with zero (zero[nfrm _{i, j} , 4]). Then, the common model 134 performs learning so that the error between the common model output data 507 and the teacher data becomes small. If c types of models are used when creating the common model 134, the teacher data will be nfrm _i,j ×4c-dimensional data.

(Diagnosis result display screen 800)
FIG. 14 is a diagram showing an example of a diagnosis result display screen 800 output to the output device 103 in this embodiment. Refer to FIGS. 1, 4, and 6 as appropriate.
The diagnosis result display screen 800 has 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 diagnosis result display area 823.
In the acoustic data display area 801, an acoustic waveform indicated by the diagnostic acoustic data 301B (see FIG. 10) acquired through the microphone 202 is displayed.
In the spectrogram display area 802, a spectrogram 401 (see FIG. 10), which is the result of frequency analysis of the acoustic waveform (diagnostic acoustic data 301B) displayed in the acoustic data display area 801, is displayed.

The estimated activity display area 810 displays the estimated activity (acoustic estimated from the third acoustic data) that is output as a result of inputting the diagnostic acoustic data 301B collected at the time of diagnosis into the individual model 136 by the activity estimation processing unit 116. model) 681 is displayed. In the example shown in FIG. 14, the estimated activity display area 810 includes an estimated activity display area 811 at the start of the closing operation, an estimated activity display area 812 at the end of the closing operation, an estimated activity display area 813 at the start of the opening operation, and an area 812 for displaying the estimated activity at the end of the closing operation. An estimated activity level 814 at the end of the operation is displayed. The estimated activity level at the start of the closing operation 681A shown in FIG. 10 is displayed in the estimated activity level at the start of the closing operation display area 811. The estimated activity level at the end of the closing operation 681B shown in FIG. 10 is displayed in the estimated activity level at the end of the closing operation display area 812. The estimated activity level at the start of the opening operation 681C shown in FIG. 10 is displayed in the estimated activity level at the start of the opening operation display area 813. In the estimated activity level 814 at the end of the opening operation, the estimated activity level 681D at the end of the opening operation shown in FIG. 10 is displayed.

In the estimated operation display area 821, the state diagnosis processing unit 117 determines whether the acoustic waveform (diagnostic acoustic data 301B) displayed in the acoustic data display area 801 is the one at the time of the closing operation of the circuit breaker 201 or the one at the time of the opening operation. The results are shown. In the example shown in FIG. 14, as shown in the estimated activity display area 810, the estimated activity 681 at the start of the opening operation and at the end of the opening operation is in the active state. In other words, the estimated activity at the start of the opening operation 681C displayed in the estimated activity at the start of the opening operation display area 813 and the estimated activity at the end of the opening operation 681D displayed in the estimated activity at the end of the opening operation 814 are activated. It has become Accordingly, the state diagnosis processing unit 117 determines that the acoustic waveform displayed in the acoustic data display area 801 is the one at the time of the opening operation (“Open”).

The rated deviation display area 822 displays a numerical value indicating how much the timing of the closing or opening operation estimated by the estimated activity level 681 deviates from the rated timing. The example shown in FIG. 14 shows that the average of the opening operation start timing and the opening operation end timing estimated from the estimated activity level 681 deviates from the rating by +0.1%. Note that "+" in the rated deviation display area 822 indicates that the timing is late, and "-" indicates that the timing is early.

In the diagnosis result display area 823, the diagnosis result by the condition diagnosis processing unit 117 is displayed. Here, the result of the state diagnosis processing unit 117 determining whether the operation of the circuit breaker 201 is normal or abnormal is displayed based on the deviation from the rating displayed in the rating deviation display area 822. The example shown in FIG. 14 shows that it is normal.

According to the present embodiment, it is possible to provide a diagnostic device 1 that diagnoses the opening/closing timing state of the circuit breaker 201 during operation based on the acoustic data 301. The diagnostic device 1 acquires the true value of the stroke position (stroke position data 302) obtained by measuring the position of the contact point on the opening/closing rod in advance with the stroke position measuring device 203, and the acoustic data 301 measured in synchronization with the true value. do. The diagnostic device 1 then estimates the start/end timing of the opening/closing operation of the circuit breaker 201 based on the stroke position data 302. Then, the diagnostic device 1 generates an activity level 310 expressed by a triangular wave that rises vertically for each opening/closing and start/end, based on the acquired start/end timings of the opening/closing operation of the circuit breaker 201. Furthermore, the diagnostic device 1 performs learning using the common model 134 as an estimation model that performs regression estimation from the feature quantity tensor 411 generated based on the acoustic data 301 to the activity degree 310. Furthermore, when the circuit breaker 201 to be diagnosed is newly installed, the diagnostic device 1 learns the individual model 136 of the newly installed circuit breaker 201 using the common model 134 as an initial value.

When diagnosing the circuit breaker 201, the diagnostic device 1 estimates the activation level 310 of the circuit breaker 201 to be diagnosed, based on the acoustic data 301 and using the already learned individual model 136 (estimated activation level). degree 681). Then, the diagnostic device 1 estimates the opening/closing operation timing of the circuit breaker 201 to be diagnosed from the peak value of the estimated activation level 310 (estimated activation level 681). Furthermore, the diagnostic device 1 calculates the deviation from the rating regarding the estimated opening/closing operation timing. Then, the diagnostic device 1 diagnoses the operating state of the circuit breaker 201 based on the deviation.

When learning about a new circuit breaker 201, it has been necessary to newly collect a large amount of sets of learning acoustic data 301 and contact stroke position data 302. For this reason, in the conventional technology, the number of man-hours for collecting data for learning with respect to the new circuit breaker 201 increases. Furthermore, in order to collect data for learning, it is necessary to actually operate the circuit breaker 201. Therefore, the deterioration of the circuit breaker 201 becomes excessive due to the opening/closing operation of the circuit breaker 201 accompanying the measurement.

In this embodiment, a common model 134 is created in advance from a large number of circuit breakers 201. When creating the individual model 136 for the new circuit breaker 201, the individual model 136 for the new circuit breaker 201 is learned using the learned common model 134 as an initial value. This makes it possible to significantly reduce the amount of learning data required for learning the individual model 136 regarding the new circuit breaker 201. Thereby, the number of times the circuit breaker 201 is opened and closed for collecting learning data can be reduced, and deterioration of the circuit breaker 201 can be prevented.

For example, a pair of the feature quantity tensor 411 extracted from the acoustic data 301 of the plurality of types of circuit breakers 201 and the true value of the activity degree 310 regarding the opening/closing operation sound is prepared. Then, the diagnostic device 1 converts the feature quantity tensor 411 into a latent space with a predetermined number of dimensions. A latent space with a certain number of dimensions has four types of activation degrees 310.

Then, the diagnostic device 1 classifies the timing of the opening/closing operations of the plurality of types of circuit breakers 201 based on the converted expressions (four types of activity levels 310). Furthermore, the diagnostic device 1 learns a model (common model 134) that regression predicts the activity level 310 of the opening/closing operation sound of the circuit breaker 201. Thereby, the diagnostic device 1 performs learning such that four types of activity levels 310 suitable for predicting the timing of opening/closing operation sounds can be determined.

In order to construct a model (individual model 136) for the new circuit breaker 201, the diagnostic device 1 generates the feature quantity 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 can regressively estimate the activation degree 310 of the opening/closing operation sound from the fifth converted data 506. (second full-connection neural network layer 522) to train the individual model 136. The individual model 136 only needs to be able to regression estimate the activity level 310 at the start and end timings of the opening/closing operation for the new circuit breaker 201 to be diagnosed. Therefore, the output of the individual model 136 corresponds to four nodes (outputs corresponding to the activation level at the start of closing operation 311, the activation level at the end of closing operation 312, the activation level at the start of opening operation 313, and the activation level at the end of opening operation 314). Become. By performing individual model learning (S1) using such a procedure, the convergence of learning of the individual model 136 can be improved even with a small number of learning samples.

When diagnosing the circuit breaker 201, only sound is collected by the microphone 202, and the learned individual model 136 is used to estimate the An activity level 681 is calculated. The diagnostic device 1 can determine the opening/closing operation timing of the circuit breaker 201 from the peak position of the estimated activity level 681. Thereafter, the diagnostic device 1 tests the health of the circuit breaker 201 by comparing the estimated opening/closing operation timing with the rated timing. By doing so, it becomes possible to quickly diagnose the circuit breaker 201 using the individual model 136 with improved learning convergence.

That is, 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/closing operation event of the circuit breaker 201 using only the microphone 202 that measures sound. Furthermore, the diagnostic device 1 of this embodiment diagnoses the state of the circuit breaker 201 based on changes in the timing of important events of opening/closing operations (opening/closing operations in this embodiment) estimated using the individual model 136. becomes possible.

By doing so, it becomes possible to diagnose the circuit breaker 201 using only the microphone 202 without using the stroke position measuring device 203. In addition, the operation timing is estimated based on the estimated activation level 310 (estimated activation level 681) as a result of inputting the diagnostic feature tensor 411 to the individual model 136. Further, the circuit breaker 201 is diagnosed based on the estimated operation timing, specifically, based on the degree of deviation between the estimated operation timing and the rated operation timing. Thereby, the diagnosis of the circuit breaker 201 can be performed quantitatively.

In this way, according to the present embodiment, the common model 134 that has been trained on a large number of circuit breakers 201 is prepared in advance, and when diagnosing a new circuit breaker 201, the common model 134 is prepared in advance, and the model is not created from scratch. Additional learning is performed based on the model 134. Thereby, the individual model 136 can be effectively created from a small number of learning samples. In particular, in the individual model learning (S2), only the second full-connection neural network 522 differs from the common model 134 in its configuration. Thereby, a model for diagnosing the new circuit breaker 201 can be constructed at low cost. In addition, in this embodiment, in the individual model learning (S2), the initial value of the connection strength of neurons is a random number only in the second full-connection neural network 522. This allows the individual model 136 to converge reliably and quickly.

In this embodiment, the operation start time and operation end time are estimated based on the stroke position data 302 of the stroke position measuring device 203. Then, the activity level 310 is associated with the estimated operation start time and operation end time. Thereby, the teacher data (activity level 310) and the input data (feature amount tensor 411) can be associated without any manual work by the user.

Furthermore, in this embodiment, the feature amount tensor 411 is generated by cutting out the spectrogram 401 at the time window W2. Thereby, learning is 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 unique to the circuit breaker 201, thereby improving diagnostic accuracy.

In this embodiment, when generating the common model 134, the distinction between circuit breakers 201 is not taken into consideration. When generating the common model 134, the common model 134 may be learned based on the acoustic data 301 and stroke position data 302 collected from the same type of circuit breaker 201 as the circuit breaker 201 to be diagnosed. The types of circuit breakers 201 include a circuit breaker 201 in a substation, a circuit breaker 201 in a power receiving facility, a circuit breaker 201 in a bullet train, a circuit breaker in a general household, and the like. Furthermore, the types of circuit breakers 201 may be distinguished by the time between opening and closing operations.

The deep neural network N1 shown in FIG. 8 is often used in the image field. When the deep neural network N1 is used in the image field, images are used as training data. In this embodiment, the activity level 310 is used, which is a triangular wave representing a model of the sound produced during the opening and closing operations of the circuit breaker 201. In this way, the deep neural network N1 used in this embodiment is used in a completely different manner from the deep neural network N1 used in the image field.

Further, in this embodiment, a triangular wave is used as the activity level 310, which rises vertically at the time when the sound is generated when the circuit breaker 201 is operated, and then decays at a predetermined slope, but the present invention is not limited to this. For example, as shown in FIG. 15, activation levels 310A and 310B based on rectangular waves may be used as the activation level 310. Activities 310A and 310B shown in FIG. 15 are expressed by rectangular waves that rise perpendicularly to the time axis at the time when the circuit breaker 201 starts operating. In FIG. 15, the activity level 310A is a model of the start sound of the opening or closing operation, and the activity level 310B is a model of the end sound of the opening or closing operation.

Further, in this embodiment, the common model 134, the individual model 136, and the circuit breaker 201 are diagnosed based on the sounds generated during the opening and closing operations of the circuit breaker 201, but the diagnosis is not limited to this. do not have. For example, the common model 134, individual model 136, and circuit breaker 201 may be diagnosed based on the sound of a fuse blowing.

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 explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to having all the configurations described.

Further, a part or all of the above-described configurations, functions, units 111 to 117, auxiliary storage device 130, etc. may be realized in hardware by designing, for example, an integrated circuit. Further, as shown in FIG. 3, each of the above-described configurations, functions, etc. may be realized by software by having the central processing unit 101 such as a CPU interpret and execute programs for realizing the respective functions. Information such as programs, tables, files, etc. that realize each function can be stored in memory, recording devices such as SSD, IC (Integrated Circuit) cards, SD (Secure Digital) cards, DVDs, etc., in addition to being stored on the HD. (Digital Versatile Disc) and other recording media.
Furthermore, in each embodiment, control lines and information lines are shown that are considered necessary for explanation, and not all control lines and information lines are necessarily shown in terms of the product. In reality, almost all configurations can be considered interconnected.

1 Diagnostic device (breaker 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 section (first learning section)
115 Individual model learning section (second learning section)
116 Activity estimation processing unit (diagnosis unit)
117 Status diagnosis processing unit (diagnosis unit)
118 Acoustic Diagnosis Department (Diagnosis Department)
133 Common learning data 134 Common model (first learning model)
135 Individual learning data 136 Individual model (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 Acoustic data (first acoustic data, second acoustic data, third acoustic data)
301a acoustic data 301b acoustic data 301A learning acoustic data (first acoustic data, second acoustic 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 (first acoustic model, triangular wave)
310b Activity (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 Activation level at the start of closing operation (activity level generated regarding the operation sound emitted when the closing operation of the circuit breaker starts)
312 Activation level at the end of closing operation (activity level generated for the operation sound emitted when the closing operation of the circuit breaker ends)
313 Activation level at the start of opening operation (activity level generated regarding the operation sound emitted when the opening operation of the circuit breaker starts)
314 Activation degree at the end of opening operation (activation degree generated regarding the operation sound emitted when the opening operation of the circuit breaker ends)
401 Spectrogram 411 Feature tensor (feature, first feature, second feature)
411a Feature tensor (feature, first feature)
411b Feature tensor (feature, first feature)
411c Feature tensor (feature, second feature)
411z Feature tensor (feature)
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 layer 512 First pooling layer 513 Second conversion data Pooling layer 521 First full-connection neural network layer 522 Second full-connection neural network layer 531 Individual model output data 681 Estimated activity (estimated acoustic model)
681A Estimated activity at the start of closing operation 681B Estimated activity at the end of closing operation 681C Estimated activity at the start of opening operation 681D Estimated activity at the end of opening operation 801 Acoustic data display area 802 Spectrogram display area 810 Estimated activity display area (third Displays the acoustic model estimated from the acoustic data of
811 Estimated activity display area at the start of closing operation 812 Estimated activity display area at the end of closing operation 813 Estimated activity display area at the start of opening operation 814 Estimated activity at the end of opening operation 821 Estimated operation display area 822 Rated deviation display area 823 Diagnosis Result display area (displays diagnosis results)
N1 Deep neural network (neural network, first neural network)
N2 Deep neural network (neural network, second neural network)
W1 Time window W2 Time window Z Circuit breaker diagnosis system (breaker model generation system)
S1 Common model learning (first learning, first learning step)
S2 Individual model learning (second learning, second learning step)

Claims

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 feature quantity that is a feature quantity generated from the first acoustic data that is the acoustic data acquired from the first circuit breaker that is the plurality of circuit breakers installed is input data, and the first a first learning unit that outputs a first learning model through first learning that is learning using a first acoustic model that is an acoustic model that is associated with the feature amount as teacher data;
Input a second feature amount that is a feature amount generated based on second acoustic data that is the acoustic data obtained from the second circuit breaker that is the circuit breaker different from the first circuit breaker. data, and outputs a second learning model through second learning, which is learning using the second acoustic model, which is the acoustic model associated with the second feature amount, as teacher data. Learning department and
has
The acoustic model is a model of the operating sound of the circuit breaker,
In the second learning, the first learning model output as a result of the first learning is used as an initial value of the second learning model in the second learning. generation system.
The neural network used in the first learning carried out in the first learning unit and the second learning carried out 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 has the same structure as the first neural network used for the first learning,
The first learning section includes:
When performing the first learning, the learning of the first neural network is performed using the initial value of the connection strength of neurons constituting the first neural network as a random number, so that the connection strength of the neurons is output as the first learning model,
The second learning section includes:
Among the neural networks constituting the second neural network, the initial value of the connection strength of the neurons in the neural network one stage before the output is a random number, and the connection strength of the neurons in the other neural networks The circuit breaker model generation system according to claim 1, wherein an initial value of is set as the coupling strength by the first learning model.
The operation of the circuit breaker is an opening operation and a closing operation of the circuit breaker,
The acoustic model includes the operating sound emitted when the circuit breaker starts to close, the operating sound emitted when the circuit breaker finishes closing, and the circuit breaker starting to open. The circuit breaker model generation system according to claim 2, wherein the circuit breaker model generation system is generated for each of the operation sound emitted when opening the circuit breaker, and the operation sound emitted when the opening operation of the circuit breaker ends. .
a third feature quantity that is the feature quantity that is obtained from the second circuit breaker by the sound collection device and that is generated based on third acoustic data that is the acoustic data different from the second acoustic data; is input into the second learning model that is output as a result of the second learning, and based on the output result output from the second learning model, a diagnosis regarding the operation of the second circuit breaker is made. The circuit breaker model generation system according to claim 1, further comprising a diagnostic section for performing diagnosis.
The diagnostic department includes:
The third feature amount, which is the feature amount generated based on the third acoustic data, is input to the second learning model and the operation timing is estimated from the result outputted. The circuit breaker model generation system according to claim 4, further comprising determining whether or not an abnormality has occurred in the second circuit breaker.
The diagnostic department includes:
Calculating the degree of deviation between the estimated operation timing and the reference operation timing, and determining that an abnormality has occurred in the second circuit breaker if the degree of deviation is a predetermined value or more. The circuit breaker model generation system according to claim 5, characterized in that:
The operation of the circuit breaker is an opening operation and a closing operation of the circuit breaker,
The acoustic model includes the operating sound emitted when the opening operation of the circuit breaker starts, the operation sound emitted when the opening operation of the circuit breaker ends, and the operation sound starting when the circuit breaker closes. the operating sound emitted when the circuit breaker closes, and the operating sound emitted when the closing operation of the circuit breaker ends;
In the diagnosis unit, the output result is the acoustic model estimated by inputting the third acoustic data to the second learning model,
The diagnostic department includes:
The circuit breaker model generation system according to claim 5, wherein the operation timing is estimated based on the acoustic model estimated from the third acoustic data.
The circuit breaker according to claim 7, wherein the acoustic model estimated from the third acoustic data and the diagnosis result of the second circuit breaker by the diagnosis unit are displayed on at least a display device. Model generation system.
a contact position measuring device that measures the position of the contact of the circuit breaker and outputs it as contact position data;
a feature setting unit that associates the acoustic model with a feature 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 model generation system according to claim 1, characterized in that it has the following.
Claim characterized by comprising: a feature amount generation unit that generates a spectrogram from the acoustic data, and generates the feature amount by acquiring a predetermined time window of the spectrogram while shifting it by a predetermined time width. 1. The circuit breaker model generation system according to 1.
The acoustic model is
The circuit breaker model generation system according to claim 1, characterized in that the circuit breaker model generation system is expressed as a triangular wave that rises perpendicularly to the time axis at the start of operation of the circuit breaker and then attenuates at a predetermined slope.
The acoustic model is
The circuit breaker model generation system according to claim 2, wherein the circuit breaker model generation system is expressed by a rectangular wave that rises perpendicularly to the time axis at the time when the circuit breaker starts operating.
Regarding a circuit breaker that cuts off power, input data is a first feature amount that is a feature amount generated from first acoustic data that is data of the operating sound of a first circuit breaker that is a plurality of circuit breakers. , a first learning unit that outputs a first learning model through first learning that is learning using a first acoustic model that is an acoustic model associated with the first feature amount as teacher data; ,
Input data is a second feature amount that is a feature amount generated based on second acoustic data that is data of an operating sound of the second circuit breaker that is different from the first circuit breaker. , a second learning unit that outputs a second learning model through second learning that is learning using the second acoustic model that is the acoustic model associated with the second feature amount as training data; and,
has
The acoustic model is a model of the operation sound of the circuit breaker,
In the second learning, a first learning model output as a result of the first learning is used as an initial value of a second learning model in the second learning. .
Regarding a circuit breaker that cuts off power, input data is a first feature amount that is a feature amount generated from first acoustic data that is data of the operating sound of a first circuit breaker that is a plurality of circuit breakers. , a first learning step of outputting a first learning model through first learning that is learning using a first acoustic model that is an acoustic model associated with the first feature amount as training data; ,
Input data is a second feature amount that is a feature amount generated based on second acoustic data that is data of an operating sound of the second circuit breaker that is different from the first circuit breaker. , a second learning step of outputting a second learning model by second learning that is learning using the second acoustic model that is the acoustic model associated with the second feature amount as teacher data; and,
has
The acoustic model is a model of the operation sound of the circuit breaker,
A circuit breaker model generation method characterized in that, in the second learning, a first learning model output as a result of the first learning is used as an initial value of a second learning model in the second learning. .