KR102480899B1 - Method and apparatus for diagnosis of motor using multi-channel signals - Google Patents

Abstract

A method for diagnosing motor failure using multi-channel signals according to the present invention includes the steps of pre-processing a plurality of measurement signals of different types for a motor to generate a plurality of learning data respectively corresponding to a plurality of channels; learning a neural network model for classifying a failure mode of a motor using the plurality of learning data; and diagnosing a failure of a motor through the neural network model from a plurality of input data corresponding to the plurality of channels, wherein the neural network model takes the plurality of learning data as inputs and outputs a feature vector, respectively. multiple convolutional neural networks; a weight calculation block calculating weights respectively corresponding to the plurality of channels based on a plurality of feature vectors output from the plurality of convolutional neural networks; a weight sum calculation block calculating a weight sum vector of the plurality of feature vectors based on at least the weights; and a failure mode classification block for classifying failure modes of the motor by taking the weight sum vector as an input.

Description

Method and apparatus for diagnosis of motor using multi-channel signals {Method and apparatus for diagnosis of motor using multi-channel signals}

The present invention relates to a method and apparatus for diagnosing a motor failure, and more particularly, to a method and apparatus for diagnosing a motor failure capable of identifying a failure mode of a motor using multi-channel signals.

Since the motor is a rotating body, signals such as current, voltage, and vibration that can be acquired from the motor include periodic components, and the failure of the motor can be determined by analyzing these periodic components. Existing motor failure diagnosis technology mainly uses a method of tracking the failure characteristic frequency according to each failure mode through various frequency analysis techniques. However, in order to check the fault characteristic frequency, motor-related parameters such as the number of poles and slip in the case of an induction motor and the number of poles in the case of a synchronous motor must be checked, and a considerable level of expert knowledge is required to understand the characteristics of each signal. do. In addition, when the load torque of the motor changes, the magnitude and frequency of the fault characteristic frequency itself change in the case of an induction motor, and the fault characteristic frequency changes in the case of a synchronous motor, so it is very difficult to apply these conventional methods in real time in the actual field. follow

In addition, monitoring various types of individual measurement signals is inefficient because several types of signal conversion and calculation processes have to be performed, and since there are too many values to be observed, it is complicated for actual use and even signals with insufficient information related to failures have to be monitored. .

Meanwhile, recently, a technique for diagnosing a failure of a machine using deep learning has been researched. When learning by inputting various types of signals as inputs to the deep learning model using the existing deep learning method, signals that do not contain much failure information are used for learning, and unnecessary information unrelated to failures is included in failure diagnosis information or overloaded. Fault diagnosis performance may be degraded due to a sum phenomenon or the like. In addition, since it is difficult to track the calculation process that occurs inside the deep learning model, it is difficult to determine which signal is used for fault diagnosis with how important it is.

The technical problem to be achieved by the present invention is to use various types of signals measured in the motor together, and do not require expert knowledge such as complex signal processing technology, and weight each signal according to the amount of information related to the failure. It is to provide a method and apparatus for diagnosing motor failures capable of learning.

A method for diagnosing motor failure using multi-channel signals according to the present invention for solving the above technical problem is to preprocess a plurality of measurement signals of different types for a motor to generate a plurality of learning data corresponding to a plurality of channels, respectively step; learning a neural network model for classifying a failure mode of a motor using the plurality of learning data; and diagnosing a failure of a motor through the neural network model from a plurality of input data corresponding to the plurality of channels, wherein the neural network model takes the plurality of learning data as inputs and outputs a feature vector, respectively. multiple convolutional neural networks; a weight calculation block calculating weights respectively corresponding to the plurality of channels based on a plurality of feature vectors output from the plurality of convolutional neural networks; a weight sum calculation block calculating a weight sum vector of the plurality of feature vectors based on at least the weights; and a failure mode classification block for classifying failure modes of the motor by taking the weight sum vector as an input.

A concatenation layer for concatenating a plurality of feature vectors output from the plurality of convolutional neural networks and inputting them to the weight calculation block may be further included.

Each of the plurality of convolutional neural networks may include a plurality of sequentially connected convolution blocks and a GAP layer that outputs the feature vector by performing global average pooling (GAP) on an output of a last convolution block among them.

Each of the plurality of convolution blocks may include a one-dimensional convolution layer.

The weight calculation block may include a fully connected layer and a softmax layer that outputs the weights.

The failure mode classification block may include a fully connected layer and a softmax layer outputting the failure mode.

The weight sum calculation block may calculate the weight sum vector by summing the weight sum of the plurality of feature vectors according to the weight and the sum of the plurality of feature vectors.

The preprocessing may include a log scale fast Fourier transform (FFT).

The plurality of measurement signals may include at least one of current, voltage, vibration, and torque.

The method for diagnosing motor failure may further include calculating a statistic of the weight output from the weight calculation block for each failure mode based on the learning result.

The motor failure diagnosis apparatus using multi-channel signals according to the present invention for solving the above technical problem is to pre-process a plurality of measurement signals of different types for the motor to generate a plurality of learning data corresponding to a plurality of channels, respectively preprocessing unit; a learning unit learning a neural network model for classifying a failure mode of a motor using the plurality of learning data; and a diagnostic unit for diagnosing motor failure through the neural network model from a plurality of input data corresponding to the plurality of channels, wherein the neural network model receives the plurality of learning data as inputs and outputs a feature vector, respectively. multiple convolutional neural networks; a weight calculation block calculating weights respectively corresponding to the plurality of channels based on a plurality of feature vectors output from the plurality of convolutional neural networks; a weight sum calculation block calculating a weight sum vector of the plurality of feature vectors based on at least the weights; and a failure mode classification block for classifying failure modes of the motor by taking the weighted sum vector as an input.

The motor failure diagnosis apparatus may further include a weight statistic calculating unit configured to calculate a statistic of the weight output from the weight calculation block for each failure mode based on the learning result.

According to the present invention described above, while using various types of signals measured in the motor together, it is possible to learn by giving weights to each signal according to the amount of information related to failure without requiring expert knowledge such as complex signal processing technology. It is possible to provide a method and apparatus for diagnosing motor failure.

In addition, according to the present invention, it is possible to determine the relative importance of which signal contains a large number of failure-related components, thereby improving interpretability.

1 shows a block diagram of a motor failure diagnosis apparatus according to an embodiment of the present invention.
2 shows an example of a measurement signal of each channel and a log scale FFT result.
3 shows the structure of a neural network model for diagnosing motor failures according to an embodiment of the present invention.
4 shows a specific example of a weight sum calculation block.
5 shows a flow chart of a method for diagnosing a motor failure according to an embodiment of the present invention.
6 is a diagram showing experimental results of testing a neural network model according to an embodiment of the present invention as a confusion matrix.
7 is a diagram showing experimental results of statistics of weights for each channel for each failure mode.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings. Substantially the same elements in the following description and accompanying drawings are indicated by the same reference numerals, respectively, and redundant description will be omitted. In addition, in describing the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

1 shows a block diagram of a motor failure diagnosis apparatus according to an embodiment of the present invention. The motor failure diagnosis apparatus according to the present embodiment may include a pre-processing unit 110, a learning unit 120, a neural network model 130, a diagnosis unit 140, and a weight statistic calculation unit 150.

The preprocessing unit 110 preprocesses a plurality of measurement signals of different types (channels) for the motor to obtain signals corresponding to the plurality of channels (ch.1, ch.2, ..., ch.N), respectively. Generate training data or input data. Failure modes may be labeled in the learning data. Failure modes may include, for example, normal, one rotor bar broken, two rotor bars broken, three rotor bars broken, four rotor bars broken, and the like.

The plurality of measurement signals may include, for example, current (phase current), voltage (phase voltage), vibration, and torque of the motor. Phase current may be measured through an AC current probe, phase voltage through an oscilloscope, vibration through a single axis accelerometer, and torque through a torque sensor. Measurement signals can include current, voltage, vibration, and torque as well as other types of signals.

Since a sufficient amount of data is required for deep learning learning, the pre-processor 110 may cut the measurement signals into equal lengths for data augmentation. In this case, since the lengths of the time series acquired during the same time period are different when the sampling frequencies of the respective signals are different, the signals may be trimmed to a size proportional to the sampling frequency for each channel.

The preprocessor 110 may transform the cut signal from the time domain to the frequency domain through a log scale Fast Fourier Transform (FFT). In this case, since the length of the FFT result is different when the sampling frequency of each signal is different, high-frequency components after a specific length (eg, 5,000 points) may be removed from the FFT result to make the length of the deep learning input the same. Periodic components of the signal of each channel can be highlighted through FFT, and harmonic components and peripheral components that are buried due to a relatively small value compared to the running component in the signal of each channel can be highlighted by log scale conversion.

In addition, since the scale of the frequency coefficient is different for each channel, it may be difficult to learn deep learning, so the preprocessor 110 may normalize the distribution of data so that the average is 0 and the variance is 1 based on frequency for each channel.

2 shows an example of a measurement signal of each channel and a log scale FFT result. Referring to FIG. 2 , measurement signals of phase current, phase voltage, and vibration are shown from top to bottom on the left, and log scale FFT results of phase current, phase voltage, and vibration are shown from top to bottom on the right.

The learning unit 120 uses learning data of a plurality of channels (ch.1, ch.2, ..., ch.N) to learn the neural network model 130 for classifying the failure mode of the motor.

3 shows the structure of a neural network model 130 for diagnosing motor failures according to an embodiment of the present invention.

Referring to FIG. 3, the neural network model 130 according to the present embodiment takes the learning data of each channel (ch.1, ch.2, ..., ch.N) as an input and uses the feature vector of each channel as input. A plurality of output convolutional neural networks (210_1, 210_2, ..., 210_N) and a plurality of feature vectors output from the plurality of convolutional neural networks (210_1, 210_2, ..., 210_N) are concatenated (concatenation) A weight calculation block 230 that calculates a weight for each channel based on feature vectors concatenated through a concatenation layer 220 and a concatenation layer 220, and a feature of each channel based on the weight for each channel. It may include a weighted sum calculation block 240 that calculates a weighted sum vector, which is a weighted sum of vectors, and a failure mode classification block 250 that classifies failure modes of the motor by taking the weighted sum vector as an input.

The plurality of convolutional neural networks 210_1, 210_2, ..., 210_N are independently trained for each channel to extract features for each channel. Each of the plurality of convolutional neural networks 210_1, 210_2, ..., 210_N is configured as a one-dimensional convolutional neural network, and high-level features related to fault diagnosis can be extracted from input data of each channel. Specifically, the convolutional neural network 210 performs global average pooling (GAP) on a plurality of sequentially connected convolution blocks 211, 212, and 213, and the output of the last convolution block 213 among them to perform feature vector It may include a GAP layer 214 that outputs. The first and second convolution blocks 211 and 212 are respectively a one-dimensional convolution layer, a batch normalization layer that improves the learning rate by uniformizing the input distribution of the activation function, a ReLU layer as an activation function, and overfitting. A max pooling layer for prevention may be included. The third convolution block 213 may include a 1D convolution layer, a batch normalization layer, and a ReLU layer. A plurality of feature vectors output from the third convolution block 213 may be summarized into one feature vector through the GAP layer 214 .

Each physically different signal has a different amount of failure-related information for each type of failure. The weight calculation block 230 is intended to quantitatively reflect this difference, and weights are given to the feature vectors of each channel, and through this, after deep learning learning is completed, the relative value of which signal has more failure-related features is determined. importance can be assigned.

The weight calculation block 230 may include a fully connected layer and a softmax layer. The weight calculation block 230 receives as an input a vector in which feature vectors independently extracted for each channel are connected in one dimension through the concatenation layer 220, and a mutual relationship in which all information of each channel is connected through the fully connected layer is obtained. After extracting, the weight of each channel can be output through the softmax layer. Due to the nature of the softmax function, each output (weight) of the softmax layer may be all positive numbers and the sum may be 1. The weight calculation block 230 may give more weight to a channel with more failure information and less weight to a channel with little failure information so that efficient learning can be achieved and overfitting can be prevented. Failure-related components can be further highlighted through the weight calculation block 230, and after learning is completed, the output of the weight calculation block 230 is analyzed to analyze which channel (signal) was mainly used for failure diagnosis. This can improve interpretability.

4 shows a specific example of the weight sum calculation block 240 . 4 shows an example of feature vectors f1, f2, and f3 of three channels for convenience of description. A vector f', in which f1, f2, and f3 are connected in one dimension, is output through the concatenation layer 220, and the weight calculation block 230 receives f' as an input and outputs weights w1, w2, and w3 for each channel. . f1, f2, and f3 are input to the weighted sum calculation block 240, and the weighted sum of f1, f2, and f3 according to the weights w1, w2, and w3 and the sum of f1, f2, and f3 that are not multiplied by weights are connected as residuals. (Residual Connection), the final weighted sum vector f is output. Residual connection is intended to prevent such information loss since the loss of information extracted from each channel can be large, and information multiplied with a weight is additionally added to reflect the relative importance of each channel, thereby increasing failure diagnosis performance.

Referring back to FIG. 3 , the failure mode classification block 250 may classify the failure mode through the fully connected layer and the softmax layer when the weighted sum vector is delivered from the weighted sum calculation block 240 . The failure mode class may include Normal and several failure types (Fault 1-4). Failure types may include, for example, one rotor bar breakage, two rotor bar breakage, three rotor bar breakage, four rotor bar breakage, and the like. The number of classes in the softmax layer may be determined according to the number of failure modes.

Referring back to FIG. 1 , when testing the neural network model 130 or diagnosing a motor, the diagnosis unit 140 inputs a plurality of input data obtained by preprocessing signals of multiple channels to the learned neural network model 130. Thus, the failure of the motor is diagnosed through the output of the neural network model 130.

The weight statistic calculation unit 150 calculates and outputs a statistic (eg, average, variance, etc.) of weight output from the weight calculation block 230 for each failure mode based on the learning result of the neural network model 130 . Through this, in fault diagnosis, it is possible to analyze which input channel (signal) contains a lot of fault information and is used importantly, and furthermore, it helps to understand the aspects of faults and the types of signals that must be acquired. can

5 shows a flow chart of a method for diagnosing a motor failure according to an embodiment of the present invention. Since the method for diagnosing motor failure according to the present embodiment is composed of steps processed by the above-described motor failure diagnosis apparatus, even if omitted below, the information described above with respect to the motor failure diagnosis apparatus is the motor failure according to the present embodiment. It also applies to diagnostic methods.

When measurement signals of multiple channels labeled with failure modes are prepared (step 510), the preprocessor 110 preprocesses the measurement signals to generate learning data of multiple channels (step 515).

The learning unit 120 builds the neural network model 130 and sets various parameters (step 520). In the neural network model 130, the number of convolutional neural networks 210, the number of outputs of the weight calculation block 230, and the number of outputs of the failure mode classification block 250 depend on the number of channels (measurement signals) and the number of failure modes. etc. can be designed. As various parameters, neural network models such as the number of convolutional layers, kernel size, number of channels, number of strides, type of activation function, presence and type of pooling, number of fully connected layers and number of nodes (130 ), learning parameters such as maximum epoch, batch size, learning rate, etc. may be set.

The table below shows parameters according to an experimental example when the number of channels is 3 and the number of failure modes is 5 as an example of parameters constituting the structure of the neural network model 130 .

layer name Parameter information number of parameters Conv1d 1 kernel size=128 , channel : 18 2,322*3 BatchNormalization 1 momentum=0.1 36*3 ReLU 1 - 0 MaxPool 1 kernel size=2 0 Conv1d 2 kernel size=8 , channel : 36 5220*3 BatchNormalization 2 momentum=0.1 72*3 ReLU 2 - 0 MaxPool 2 kernel size=2 0 Conv1d 3 kernel size=8 , channel : 72 20808*3 BatchNormalization 3 momentum=0.1 144*3 ReLU 3 - 0 GlobalAvgPool - 0 Weight Computation node number : 3 651 Classifier Dense 1 node number : 36 2592 Classifier Dense 2 node number : 5 180 all 89229

In the experimental example, the learning parameters may be set to 50 for maximum epoch, 16 for batch size, and 0.0005 for learning rate.

The learning unit 120 trains the neural network model 130 using the training data (step 525). The learning process of the neural network model 130 may be performed by a mini-batch gradient descent method.

When the learning of the neural network model 130 is completed (for example, learning up to the maximum epoch) (step 530), the diagnosis unit 140 may test the neural network model 130 using test data prepared separately from the training data (535). step).

When the learning of the neural network model 130 is completed, the weight statistic calculation unit 150 outputs the weight statistic (eg, average, variance, etc.) from the weight calculation block 230 for each failure mode from the learning result of the neural network model 130 can be calculated (step 540). That is, statistics for each output of the weight calculation block 230 may be calculated for each class of the failure mode classification block 250 . Through this, it is possible to analyze which channel (measurement signal) contains a lot of failure information and is used significantly.

When a multi-channel measurement signal for a motor to be diagnosed is input (Step 545), the pre-processing unit 110 pre-processes the measurement signal to generate input data of multiple channels (Step 550).

The diagnosis unit 140 inputs input data of multiple channels to the neural network model 130 and diagnoses motor failure through the output of the neural network model 130 (step 555).

6 is a diagram showing experimental results of testing the neural network model 130 according to an embodiment of the present invention as a confusion matrix. Referring to FIG. 6 , it can be seen that the experimental results show very high diagnostic performance with an accuracy of 99.25%.

7 is a diagram showing experimental results of statistics of weights for each channel for each failure mode. 7 (a) to (e) show 3 channels (current: Ia, voltage: Va) for normal, 1 rotor bar breakage, 2 rotor bar breakage, 3 rotor bar breakage, and 4 rotor bar breakage, respectively. , Vibration: Vib_acpe) represents the average of each weight. Common to all failure modes, the weight of voltage is almost zero, which is not given importance at all, which is consistent with the actual physical phenomenon in which failure-related characteristics are not well represented in voltage. Current was judged to be the most important signal for most failure modes, as rotor bar failure affects the magnetic field and ultimately reflects in the current. However, it was found that the weight of the vibration signal was higher than that of the current signal in the three rotor bar breakage failures (d). means different.

Devices according to embodiments of the present invention include a processor, a memory for storing and executing program data, a permanent storage unit such as a disk drive, a communication port for communicating with an external device, a touch panel, a key, and a button. It may include user interface devices such as the like. Methods implemented as software modules or algorithms may be stored on a computer-readable recording medium as computer-readable codes or program instructions executable on the processor. Here, the computer-readable recording medium includes magnetic storage media (e.g., read-only memory (ROM), random-access memory (RAM), floppy disk, hard disk, etc.) and optical reading media (e.g., CD-ROM) ), and DVD (Digital Versatile Disc). A computer-readable recording medium may be distributed among computer systems connected through a network, and computer-readable codes may be stored and executed in a distributed manner. The medium may be readable by a computer, stored in a memory, and executed by a processor.

Embodiments of the invention may be presented as functional block structures and various processing steps. These functional blocks may be implemented with any number of hardware or/and software components that perform specific functions. For example, an embodiment may include an integrated circuit configuration, such as memory, processing, logic, look-up tables, etc., capable of executing various functions by means of the control of one or more microprocessors or other control devices. can employ them. Similar to components of the present invention that may be implemented as software programming or software elements, embodiments may include various algorithms implemented as data structures, processes, routines, or combinations of other programming constructs, such as C, C++ , Java (Java), can be implemented in a programming or scripting language such as assembler (assembler). Functional aspects may be implemented in an algorithm running on one or more processors. In addition, the embodiment may employ conventional techniques for electronic environment setting, signal processing, and/or data processing. Terms such as "mechanism", "element", "means", and "composition" may be used broadly and are not limited to mechanical and physical components. The term may include a meaning of a series of software routines in association with a processor or the like.

Specific executions described in the embodiments are examples, and do not limit the scope of the embodiments in any way. For brevity of the specification, description of conventional electronic components, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connection of lines or connecting members between the components shown in the drawings are examples of functional connections and / or physical or circuit connections, which can be replaced in actual devices or additional various functional connections, physical connection, or circuit connections. In addition, if there is no specific reference such as "essential" or "important", it may not necessarily be a component necessary for the application of the present invention.

So far, the present invention has been looked at with respect to its preferred embodiments. Those skilled in the art to which the present invention pertains will be able to understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from an illustrative rather than a limiting point of view. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent scope will be construed as being included in the present invention.

Claims (20)

As a method for diagnosing motor failure using multi-channel signals,
generating a plurality of learning data respectively corresponding to a plurality of channels by pre-processing a plurality of measurement signals of different types for the motor;
learning a neural network model for classifying a failure mode of a motor using the plurality of learning data; and
Diagnosing a failure of a motor through the neural network model from a plurality of input data corresponding to the plurality of channels,
The neural network model,
a plurality of convolutional neural networks each outputting feature vectors by taking the plurality of learning data as inputs;
a weight calculation block calculating weights respectively corresponding to the plurality of channels based on a plurality of feature vectors output from the plurality of convolutional neural networks;
a weight sum calculation block calculating a weight sum vector of the plurality of feature vectors based on at least the weights; and
and a failure mode classification block for classifying failure modes of the motor by taking the weighted sum vector as an input. According to claim 1,
The neural network model,
The method of diagnosing motor failure, characterized in that it further comprises a concatenation layer for concatenating a plurality of feature vectors output from the plurality of convolutional neural networks and inputting them to the weight calculation block. According to claim 1,
Each of the plurality of convolutional neural networks,
A method for diagnosing motor failure, comprising a plurality of sequentially connected convolution blocks and a GAP layer that outputs the feature vector by performing global average pooling (GAP) on an output of the last convolution block among them. According to claim 3,
Each of the plurality of convolution blocks includes a one-dimensional convolution layer. According to claim 1,
The weight calculation block includes a fully connected layer and a softmax layer that outputs the weights. According to claim 1,
The motor failure diagnosis method of claim 1 , wherein the failure mode classification block includes a fully connected layer and a softmax layer outputting the failure mode. According to claim 1,
The weight sum calculation block calculates the weight sum vector by summing the weight sum of the plurality of feature vectors according to the weight and the sum of the plurality of feature vectors. According to claim 1,
The motor failure diagnosis method, characterized in that the pre-processing comprises a log scale FFT. According to claim 1,
The plurality of measurement signals include at least one of current, voltage, vibration, and torque. According to claim 1,
The motor failure diagnosis method of claim 1 , further comprising calculating a statistic of the weight output from the weight calculation block for each failure mode from the learning result. As a motor failure diagnosis device using multi-channel signals,
a pre-processing unit generating a plurality of learning data corresponding to a plurality of channels by pre-processing a plurality of measurement signals of different types for the motor;
a learning unit learning a neural network model for classifying a failure mode of a motor using the plurality of learning data; and
A diagnosis unit for diagnosing a failure of a motor through the neural network model from a plurality of input data corresponding to the plurality of channels;
The neural network model,
a plurality of convolutional neural networks each outputting feature vectors by taking the plurality of training data as inputs;
a weight calculation block calculating weights respectively corresponding to the plurality of channels based on a plurality of feature vectors output from the plurality of convolutional neural networks;
a weight sum calculation block calculating a weight sum vector of the plurality of feature vectors based on at least the weights; and
and a failure mode classification block for classifying failure modes of the motor by taking the weighted sum vector as an input. According to claim 11,
The neural network model,
and a concatenation layer for concatenating a plurality of feature vectors output from the plurality of convolutional neural networks and inputting the concatenated feature vectors to the weight calculation block. According to claim 11,
Each of the plurality of convolutional neural networks,
A motor fault diagnosis device comprising a plurality of sequentially connected convolution blocks and a GAP layer that outputs the feature vector by performing Global Average Pooling (GAP) on the output of the last convolution block among them. According to claim 13,
Each of the plurality of convolution blocks includes a one-dimensional convolution layer. According to claim 11,
The weight calculation block includes a fully connected layer and a softmax layer that outputs the weights. According to claim 11,
The motor failure diagnosis apparatus according to claim 1 , wherein the failure mode classification block includes a fully connected layer and a softmax layer outputting the failure mode. According to claim 11,
The weight sum calculation block calculates the weight sum vector by summing the weight sum of the plurality of feature vectors according to the weight and the sum of the plurality of feature vectors. According to claim 11,
The motor failure diagnosis apparatus, characterized in that the pre-processing comprises a log scale FFT. According to claim 11,
The plurality of measurement signals include at least one of current, voltage, vibration, and torque. According to claim 11,
The motor failure diagnosis apparatus of claim 1, further comprising a weight statistic calculation unit configured to calculate a statistic of the weight output from the weight calculation block for each failure mode from the learning result.

Images