WO2023151488A1

WO2023151488A1 - Model training method, training device, electronic device and computer-readable medium

Info

Publication number: WO2023151488A1
Application number: PCT/CN2023/074026
Authority: WO
Inventors: 陈曦; 王巍巍; 葛成; 王明
Original assignee: 阿里巴巴（中国）有限公司
Priority date: 2022-02-11
Filing date: 2023-01-31
Publication date: 2023-08-17
Also published as: CN114169539A

Abstract

Provided are a model training method, a training device, an electronic device and a computer readable medium. The method comprises: constructing an initial dataset of high-frequency vibration signals; screening the initial dataset to obtain a key dataset; using the key dataset to perform model training and obtain a fault identification model, wherein the initial dataset is represented by three dimensions, namely fault category, sample and feature. Selection steps comprise: for each feature, performing average distance calculation respectively on two dimensions, namely the sample and the fault category, and obtaining, on the basis of the average distance, a representation value of the importance of each feature; and forming a key dataset with the features of which representation values are greater than a first threshold value. According to the embodiments, the representation value of the importance of each feature is calculated on the basis of the average distance of the two dimensions, namely the sample and the fault category, and the initial dataset is screened according to the representation value of the importance, and model training is carried out by using the screened features, thereby improving the model training efficiency.

Description

Model training method, training device, electronic device and computer readable medium

This application claims the priority of the Chinese patent application with the application number 202210127454.X and the application name "model training method, training device, electronic equipment and computer readable medium" submitted to the China Patent Office on February 11, 2022, which The entire contents are incorporated by reference in this application.

technical field

The present disclosure relates to the field of artificial intelligence, and in particular to a model training method, a training device, electronic equipment and a computer-readable medium.

Background technique

Due to structural and processing installation defects of rotors, bearings, housings, seals and foundations, or due to external effects, a large number of industrial equipment vibrates during operation, and excessive vibration is often the main cause of equipment damage. According to statistics, for the large-scale and wide-ranging rotating machinery and reciprocating machinery in the industry, the failure of equipment due to vibration accounts for more than 60% of the total failure rate. Therefore, the vibration monitoring and analysis of mechanical equipment is very important. Compared with other state parameters, such as the temperature, pressure, flow rate of lubricating oil or internal fluid, or the current of the motor, vibration parameters can often reflect the operation of the unit more directly, quickly and accurately. The vibration signal is usually one of the main bases for diagnosing the equipment status.

With the development of my country's industrial modernization, the application of large-scale rotating equipment is becoming more and more extensive. From steel, coal, electric power, cement, to subways, airplanes, trains, ships, etc., rotating equipment is inseparable from the figure of these rotating equipment. Stable operation is more and more important to the development of national economy. During the long-term operation of the equipment, various failures will inevitably occur. If the early failure symptoms are not found in time, with its development and expansion, when it reaches a certain critical point, the equipment is prone to sudden and serious failures, resulting in a large number of abnormal Planned maintenance work. These failures may cause certain economic losses in the slightest, and cause casualties in severe cases.

Minor faults such as wear and degradation occur in a certain part of industrial rotating equipment during work. Due to the weak macroscopic representation, it is often impossible to effectively monitor only by manual identification, and it is time-consuming and laborious. The vibration signal is generated and continues with the operation of the machine. Even if the machine is in good operating condition, it will vibrate due to slight excitation. For mechanical equipment, there are usually two types of vibration sources with different properties: one is caused by the mass imbalance of mechanical moving parts, misalignment of geometric axes, poor gear kneading, improper coordination of transmission parts, and excessive journal bearing clearance. Mechanical forced vibration, including periodic vibration, impact vibration, random vibration, etc., also causes noise; another type of vibration is due to structural response, vibration response caused by self-excited vibration or environmental vibration, such as: fluid surge vibration, bearing oil film vibration, response vibration of the component itself, local vibration of the structure, etc. Once an early failure occurs, the corresponding vibration and noise will undergo a series of changes. Therefore, using scientific methods, the monitoring and diagnosis of vibration signals plays an important role in improving the stable operation of rotating equipment. The monitoring and diagnosis system based on modern fault diagnosis technology can monitor the operating status of equipment in real time. The processing and analysis of data can discover the cause of equipment failure and predict the possible failure of equipment, in order to prevent accidents, Scientific arrangements for maintenance provide a scientific basis, thereby saving maintenance costs and improving equipment reliability and safety.

In recent years, with the rise of deep learning, researchers have made important progress in applying neural network models to machine fault monitoring and diagnosis. However, in practice, the data sets generated based on high-frequency vibration signals are huge and diverse. If these data sets are directly used for model training, it will undoubtedly consume a lot of resources and be inefficient.

Contents of the invention

In view of this, the present disclosure aims to provide a model training method, a training device, an electronic device and a computer-readable medium, so as to improve the efficiency of model training.

According to a first aspect of the present disclosure, a model training method is provided, including:

Construct an initial data set of high-frequency vibration signals;

Screening the initial data set to obtain key data sets;

Using the key data set to perform model training to obtain a fault identification model,

Wherein, the initial data set is characterized by three dimensions of fault category, sample and feature, and the selection step includes:

For each feature, the average distance calculation is performed in the two dimensions of the sample and the fault category, and the representative value of the importance of each feature is obtained based on the average distance calculation results of the two dimensions; and

The key data set is composed of features whose importance characteristic value is greater than the first threshold.

In some embodiments, the method further includes: in the key data set, calculating the correlation between each feature and other features and filtering the features of the key data set based on the correlation between features.

In some embodiments, in the key data set, calculating the correlation between each feature and other features and filtering the features of the key data set based on the correlation between features includes:

Sorting the features of the key data set in descending order of importance and taking out the features with the highest importance and putting them into the feature subset;

Sorting by importance, calculating the variance inflation factor of each feature in the key data set and the feature subset;

comparing the variance inflation factor of each feature in the key data set with a second threshold, and if the variance inflation factor of the feature is less than the second threshold, putting the feature into a feature subset,

repeating said calculating the variance inflation factor for each feature in said key dataset and said subset of features and comparing the variance inflation factor for each feature in said key dataset to a second threshold, if the variance of the feature If the expansion factor is smaller than the second threshold, then put the feature into the feature subset until the comparison of all the features of the key data set is completed.

In some embodiments, for each feature, calculating the average distance in the sample dimension includes:

Calculate the average distance between different samples of each feature under the same fault category;

Based on the average distance between different samples of each feature under the same fault category, calculate the average value of each feature in multiple fault categories, and record it as the first average distance;

For each feature, calculating the average distance in the fault category dimension includes:

Calculate the average value of each feature in all samples under the same fault category;

Based on the average value of each feature in all samples under the same fault category, the average distance of each feature average among different fault categories is calculated, and recorded as the second average distance.

In some embodiments, the characterization value of the importance of each feature obtained based on the calculation result of the average distance of the two dimensions includes:

Calculate the variance factor of the first mean distance for each feature,

Calculate the variance factor of the second mean distance for each feature;

Calculate compensation factors based on the two variance factors for each feature;

A representative value of the importance of each feature is obtained based on the compensation factor of each feature, the first average distance, and the second average distance.

In some embodiments, the calculation of the average distance between different samples of each feature under the same fault category adopts formula (1):

The average value of each feature in multiple fault categories is calculated using formula (2):

Wherein, the initial data set includes K kinds of fault categories, q _i,k,j , i=1,2,...,I _k ; k=1,2,...,K; j=1,2,...,J , where q _i,k,j represent the j-th eigenvalue of the i-th sample in the k-th fault category, I _k is the number of samples of the k-th fault category, K is the number of fault categories, and J is each sample number of features.

In some embodiments, the calculation of the average value of each feature in all samples under the same fault category adopts formula (3):

The average distance of each characteristic mean value between the described calculation different fault categories adopts formula (4):

In some embodiments, for each feature in each key data set, the step of calculating the variance inflation factor of the feature subset includes: substituting into formula (5) to obtain the variance inflation factor of the feature,

in, Represents the goodness of fit of the regression equation X _j =β ₀ +βX′, where X′ is a feature of the feature subset.

According to a second aspect of the present disclosure, there is provided a model training device, comprising:

A feature extraction unit is used to construct an initial dataset of high-frequency vibration signals;

A feature screening unit, configured to screen the initial data set to obtain key data sets;

A model training unit, configured to use the key data set for model training to obtain a fault recognition model, wherein, The initial data set includes the use of three dimensions of fault category, sample and feature representation, then the feature screening unit includes:

According to a third aspect of the present disclosure, there is provided an electronic device, including a memory and a processor, the memory further stores computer instructions executable by the processor, and when the computer instructions are executed, any one of the above-mentioned The model training method described above.

According to a fourth aspect of the present disclosure, a computer-readable medium is provided, the computer-readable medium stores computer instructions that can be executed by an electronic device, and when the computer instructions are executed, the model training described in any one of the above is implemented method.

According to the model training method provided by the embodiments of the present disclosure, the data set is characterized into three dimensions of fault category, sample and feature, and the average distance calculation is performed in the two dimensions of sample and fault category, and based on the two average distance calculation results, the The characterization value of the importance of each feature, so that the initial data set can be screened according to the characterization value of importance to obtain an optimized feature combination, and then the optimized feature combination is used for model training, which can achieve the purpose of improving training efficiency.

Description of drawings

The above and other objects, features and advantages of the present disclosure will be more clear by describing the embodiments of the present disclosure with reference to the following drawings, in which:

FIG. 1 is a schematic flowchart of a model training method provided by an embodiment of the present disclosure;

FIG. 2 is a flowchart of a method for screening initial data sets in combination with importance and relevance provided by an embodiment of the present disclosure;

Fig. 3 is a schematic structural diagram of a model training device provided by an embodiment of the present disclosure;

Fig. 4 is a schematic structural diagram of an exemplary electronic device.

Detailed ways

The present disclosure is described below based on examples, but the present disclosure is not limited only to these examples. In the following detailed description of the disclosure, some specific details are set forth in detail. The present disclosure can be fully understood by those skilled in the art without the description of these detailed parts. In order to avoid obscuring the essence of the present disclosure, well-known methods, procedures, and procedures are not described in detail. Additionally, the drawings are not necessarily drawn to scale.

The model training method provided by the embodiment of the present disclosure is shown in FIG. 1 .

Step S11 is to construct an initial data set of high-frequency vibration signals.

Step S12 is to filter the initial data set to obtain the key data set, that is, perform feature screening on the initial data set containing rich features, and filter out a set of features with high importance and low correlation;

Step S13 is to use key data sets for training to obtain a fault classification model. That is to say, build a fault recognition model (using ML models such as support vector machine SVM, random forest RF, etc.), provide the data set screened in step S12 to the fault recognition model, train the fault recognition model, and in the training Continuously optimize the weight parameters of the fault recognition model, and stop training until the loss function meets expectations.

Among them, the high-frequency vibration signal usually comes from a vibration sensor, and the high-frequency vibration signal of the analog signal is converted into a digital signal through the vibration sensor, and then various methods are used to extract the feature of the high-frequency vibration signal of the digital signal. There are various existing feature extraction techniques, which can be roughly divided into time-domain features, frequency-domain features, time-frequency domain features, and features extracted for some specific scenes. Time-domain features are the most intuitive, and the following indicators are usually calculated: average, root mean square, peak-to-peak, pulse, margin, kurtosis, kurtosis, etc. The frequency domain feature is mainly to perform Fourier transform on the signal, observe the spectrum of the signal from another perspective, and extract various features of the spectrum, such as: mean value, maximum value, centroid frequency, spectral kurtosis, spectral power, etc. The time-frequency domain feature is mainly used in the start-stop stage of the equipment. The time-frequency spectrum of the signal is obtained to observe the change of the signal frequency with time, and the time-varying feature is extracted. All the extracted features form a data set, which is used as the input for feature selection in the next step.

The screening step is the process of selecting some of the most effective features from the data pool to reduce the dimensionality of the data set. It is an important means to improve the performance of the learning algorithm and is also a key data preprocessing step in pattern recognition. Through the screening step, the system can obtain Feature combinations for specific metric optimization (such as model recognition accuracy). According to whether it is independent of the subsequent machine learning algorithm, the feature selection methods used in the screening step are mainly divided into three categories:

(a) Filtering type: It has nothing to do with subsequent machine learning algorithms. Generally, the statistical performance of all training data is used to evaluate features directly, and various evaluation criteria are used to enhance the correlation between features and classes and reduce the correlation between features. Common selection methods include: variance selection method, correlation coefficient method, chi-square test method, mutual information method, etc.

(b) Encapsulation: The feature selection algorithm is used as an integral part of the subsequent machine learning algorithm, and the classification performance is directly used as the evaluation criterion for the degree of feature importance. It is based on the fact that a selected subset is ultimately used to construct a classification model. Therefore, if the features that can achieve higher classification performance are directly used when constructing the classification model, a classification model with higher classification performance can be obtained. The model here can be various machine learning algorithms.

(c) Embedded: Directly use a certain machine learning model for training to obtain the weight coefficients of each feature value, and select features according to the importance of the coefficients from large to small as features.

In this paper, it is mainly aimed at improving the filter-type feature selection. First, when constructing the initial data set, the initial data set is represented as data in three dimensions of fault category, sample and feature. Specifically, for a segment of digital high-frequency signal output by the sensor, the segment of high-frequency signal is processed into a sample including multiple features, and each feature in the multiple features can be a time-domain feature, a time-domain feature, or a time-domain feature. Domain features or time-frequency domain features, and the sample has an assigned fault category. The fault category to which the sample belongs can be obtained through manual labeling or a trained fault recognition model. Specifically, for a fault recognition model that has been trained, a sample can be input into the model to obtain the fault category corresponding to the sample.

Then the average distance is performed in the two dimensions of sample and fault category respectively to obtain the first average distance of each feature in the sample dimension and the second average distance in the fault category dimension. Specifically, for the first average distance, first calculate the average distance between multiple samples of each feature under the same fault category, thus obtaining the multiple average distances under the fault category, and then calculate the average value of the average distances of the feature under multiple fault categories based on the multiple average distances of each feature under multiple fault categories as the first average distance of the feature; For the second average distance, first calculate the average value of all samples under the same fault category for each feature, thereby obtaining multiple average values of multiple features under multiple fault categories, and then based on each feature in multiple fault categories Calculate the average distance of the average value of the feature under multiple fault categories as the second average distance of the feature, and then get the feature's average distance based on the first average distance and the second average distance of each feature The characteristic value of importance, thus obtaining the characteristic value of the importance of all features, and then comparing the characteristic values of the importance of all features and selecting features from the initial data set according to the comparison results to form a key data set (the key data set is also characterized as fault categories, samples, and features), for example, as long as the characteristic value of the importance of a feature is greater than the set threshold, the feature will be assigned to the key data set.

To sum up, for each feature, the average distance is calculated in the two dimensions of sample and fault category, and the two average distances of each feature in the two dimensions are obtained, and the evaluation of each feature is obtained based on these two average distances. The representation value of the importance is used to select a combination of features from the data set to train the fault recognition model.

However, although the features selected by the above distance evaluation method have high importance, they may also have a high correlation with each other, which will increase the complexity of the model and affect the final fault identification effect. Therefore, in some embodiments, the correlation between each feature and other features is continuously calculated, and the screening is continued using the correlation.

Fig. 2 is a flowchart of a method for screening an initial data set in combination with importance and relevance provided by an embodiment of the present disclosure, which specifically includes the following steps.

In step S121, create an empty feature subset.

In step S122, the first average distance and the second average distance of each feature are calculated, and the characteristic value of the importance of each feature is obtained accordingly.

In step S123, the features whose importance characteristic value is greater than the first threshold are formed into a first feature set.

In step S124, the features of the first feature set are sorted in descending order of importance, and the feature with the highest importance is taken out and put into the feature subset.

In step S125, each feature in the first feature set and the variance inflation factor in the feature subset are calculated one by one in order of importance.

In step S126, it is judged whether the variance inflation factor of the feature is smaller than the second threshold, if yes, execute step S127, otherwise execute step S125. Steps S125 to S127 form a loop, and the number of repetitions of this loop is (N-1) times, where N is the number of features included in the first feature set.

In step S127, put the feature into a feature subset.

In step S128, the feature subset is output.

In this embodiment, the first average distance and the second average distance of each feature are calculated according to the calculation method of the previous embodiment, and the characterization value of the importance of each feature is obtained accordingly, and then the characterization of the importance The value greater than the first threshold is used as the screening condition to obtain the first feature set, the features of the first feature set are sorted according to the characteristic value of importance from large to small, and the feature with the highest characteristic value of importance is taken out and put into the feature subset, and then Arranged in descending order of importance Sequence, calculate the variance inflation factor of each feature and feature subset in the first feature set, and put the features whose variance inflation factor is less than the second threshold into the feature subset, which means that when a feature and feature subset When the variance inflation factor of is smaller than the second threshold, it is considered that the correlation between the feature and the feature subset is relatively small, so the feature can be put into the feature subset as a feature subset. It should be understood that the number of features in the feature subset increases as the number of cycles increases, that is, the number of features used to calculate the variance inflation factor gradually increases, so the calculation of the variance inflation factor of the features in the key data set and the feature subset is computationally and complex. degree increases with increasing.

Combined with the formula below, we will introduce in more detail how to calculate the average distance in the two dimensions of sample and fault category to obtain two average distances for each feature, and obtain the characterization value to evaluate the importance of each feature based on these two average distances.

Suppose the initial data set contains K kinds of fault categories: q _i,k,j , i=1,2,…,I _k ; k=1,2,…,K; j=1,2,…,J, where, q _i,k,j represent the jth eigenvalue of the i-th sample in the k-th fault category. I _k is the number of samples of the kth fault category, K is the number of fault categories, and J is the number of features for each sample.

Among them, the average distance is calculated in the two dimensions of sample and fault category, and the specific operation steps to obtain the two average distances of each feature are as follows:

(1) Calculate the average distance between different samples under the same fault category:

(2) Then calculate the average value of each feature in K fault categories:

(3) Calculate the average values of different characteristics of all samples under the same fault category:

(4) Then calculate the average distance of each feature average among different fault categories:

Then, based on these two average distances, the specific operation steps to obtain the characterization value for evaluating the importance of each feature are as follows:

(5) Define the average distance variance factor

(6) Define the average distance variance factor

(7) Considering the two variance factors comprehensively to define the compensation factor:

(8) Calculate the coefficient after considering the compensation factor:

(9) Normalized to obtain the feature importance index:

In some embodiments, it is proposed that the variance inflation factor for the jth feature be calculated as follows:

in, Represents the goodness of fit of the regression equation X _j = β ₀ + βX', where X' includes other features to be compared for correlation besides X _j . Corresponding to the embodiment in FIG. 2 , X' refers to the feature subset.

A brief description will be given below based on the embodiment shown in FIG. 2 . First, in the initial state, there is only one feature with the highest importance in the feature subset, and then the feature with the second highest importance is taken from the key data set, and between this feature and the feature with the highest importance in the feature subset Calculate the variance inflation factor, if the variance inflation factor is less than the second threshold, put the feature into the feature subset, then take out the feature with the third importance from the key data set, and calculate the importance of the feature and the feature subset is the variance inflation factor between the combination of the highest and second highest two features, if the variance inflation factor is less than the second threshold, put the feature into the feature subset, and so on.

Correspondingly, an embodiment of the present disclosure provides a fault detection device, including a feature extraction unit 301 , a feature screening unit 302 and a model training unit 303 .

Feature extraction unit 301 is used to construct the initial data set of high-frequency vibration signal;

The feature screening unit 302 is used to screen the initial data set to obtain key data sets;

The model training unit 303 is used to perform model training using key data sets to obtain a fault recognition model. Among them, the initial data set includes the three-dimensional representation of fault category, sample and feature, and the model training unit 303 includes: for each feature, the average distance calculation is performed in the two dimensions of sample and fault category, and based on the two dimensions The characterization value of the importance of each feature is obtained from the average distance calculation result, and the features whose characterization value of importance is greater than the first threshold are selected to form the key data set.

For a more detailed description of the feature extraction unit 301 , the feature screening unit 302 and the model training unit 303 , please refer to the above description.

FIG. 4 is a schematic structural diagram of an exemplary electronic device 400 . The electronic device can be used to execute the application system including the above-mentioned fault identification model, and at the same time, the electronic device can be used to train the fault identification model. As shown in the figure, the server 400 includes a scheduler 401 , a storage unit 403 , an I/O interface 404 , and multiple model acceleration units 402 coupled via a bus 405 .

The storage unit 403 may include readable media in the form of volatile storage units, such as random access memory units (RAM) and/or cache storage units. The storage unit 403 may also include a readable medium in the form of a non-volatile storage unit, for example, a read only memory unit (ROM), a flash memory, and various magnetic disk memories.

The storage unit 403 can store various program modules and data, and various program modules include an operating system, providing such as Applications for text processing, video playback, software editing and compiling, etc. The executable codes of these application programs are read from the storage unit 403 by the scheduler 401 and executed, so as to realize the predetermined operations of these program modules. The scheduler 401 is generally a processor (CPU). In particular, the storage unit 403 stores the fault detection system based on the above.

The bus 405 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local area using any bus structure in a variety of bus structures. bus.

The server 400 can communicate with one or more external devices (such as keyboards, pointing devices, Bluetooth devices, etc.), and can also communicate with one or more devices that enable users to interact with the server 400, and/or communicate with the server 400 to enable Any device (eg, router, modem, etc.) that communicates with one or more other computing devices communicates. Such communication may occur through input/output (I/O) interface 404 . Moreover, the server 400 can also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN) and/or a public network such as the Internet) through a network adapter (not shown). The terminal 103 in FIG. 1 can access the server 400, for example through a network adapter. It should be appreciated that although not shown in the figure, other hardware and/or software modules may be used based on the server 400, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and Data backup storage system, etc.

As shown in the figure, the server 400 includes a plurality of model acceleration units 402 . The traditional processor architecture design is very effective in logic control, but not efficient enough in large-scale parallel computing, so it is not efficient for model calculations. Therefore, a model acceleration unit is developed, and different models can be adapted to different model acceleration units. The model acceleration unit is, for example, a neural network acceleration unit (NPU). The NPU adopts a data-driven parallel computing architecture and is used as a processing unit for processing a large number of operations (such as convolution, pooling, etc.) of each neural network node. Or it is a graphics processing unit (GPU), which is used to do image and graphics-related calculations. Since the graphics processing unit uses a large number of computing units dedicated to graphics calculations, the graphics card reduces its dependence on the CPU and assumes the responsibility of the CPU. Originally undertaken some computationally intensive graphics and image processing work, so the processing efficiency of image data is greatly improved. Multiple model acceleration units 402 are controlled by the scheduler 401 , and under the control of the scheduler 401 , multiple model acceleration units 402 can work together.

Taking each embodiment of the present disclosure as an example, referring to FIG. 1 and FIG. 3 , the step S11 corresponding to the feature extraction unit 301 and the step S12 corresponding to the feature screening unit 302 can be executed by the CPU, and the step S13 corresponding to the model training unit 303 can be It is completed by the cooperation of the scheduler 401 and multiple model acceleration units 402. At this time, the scheduler 401 generally controls the process, puts the execution of the trained fault classification model on multiple model acceleration units 402 and summarizes the multiple model acceleration units 402 As a result of execution, the data interaction between multiple model acceleration units 402 can be realized through the scheduler 401 , or the data interaction can be directly performed by multiple model acceleration units 402 .

It should be understood that the model training method and model training device provided by the embodiments of the present disclosure use a supervised distance evaluation method to select an optimized feature combination from a data set to improve model training efficiency. The supervised distance evaluation method refers to the evaluation using the average distance of the dimension of the fault category (the data for the dimension needs to be marked or generated through a trained fault recognition model).

In addition, an embodiment of the present disclosure also provides a computer-readable medium for storing the above-mentioned model training method computer readable instructions.

Commercial value of this disclosure

Embodiments of the present disclosure provide a model training method. Compared with the prior art, the embodiment of the present disclosure adopts a supervised distance evaluation method to select an optimized feature combination from a data set for model training, so as to achieve the purpose of improving model training efficiency and saving computing resources. Therefore, the model training method has certain practical value and economic value.

It should be appreciated that the above descriptions are only preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. For those skilled in the art, there are many variations to the embodiments of the present disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

It should be understood that each embodiment in this specification is described in a progressive manner, the same or similar parts of each embodiment can be referred to each other, and each embodiment focuses on the difference from other embodiments .

It should be understood that the foregoing describes specific embodiments of the present specification. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. Multitasking and parallel processing are also possible or may be advantageous in certain embodiments.

It should be understood that describing an element herein in the singular or showing only one in a drawing does not mean limiting the number of that element to one. Furthermore, modules or elements described or illustrated herein as separate may be combined into a single module or element, and modules or elements described or illustrated herein as a single may be split into a plurality of modules or elements.

It should also be understood that the terms and expressions used herein are for description only, and one or more embodiments of this specification should not be limited to these terms and expressions. The use of these terms and expressions does not mean to exclude any equivalent features shown and described (or parts thereof), and it should be recognized that various modifications may also be included within the scope of the claims. Other modifications, changes and substitutions are also possible. Accordingly, the claims should be read to cover all such equivalents.

Claims

A model training method, comprising:

Construct an initial data set of high-frequency vibration signals;

Screening the initial data set to obtain key data sets;

Using the key data set to perform model training to obtain a fault identification model,

Wherein, the initial data set is characterized by three dimensions of fault category, sample and feature, then the screening includes:

For each feature, the average distance calculation is performed in the two dimensions of the sample and the fault category, and the representative value of the importance of each feature is obtained based on the average distance calculation results of the two dimensions; and

The key data set is composed of features whose importance characteristic value is greater than the first threshold.
The model training method according to claim 1, further comprising: in the key data set, calculating the correlation between each feature and other features and filtering the features of the key data set based on the correlation between features.
The model training method according to claim 1, wherein, in the key data set, the correlation between each feature and other features is calculated and the features of the key data set are screened based on the correlation between features include:

Sorting the features of the key data set in descending order of importance and taking out the features with the highest importance and putting them into the feature subset;

Sorting by importance, calculating the variance inflation factor of each feature in the key data set and the feature subset;

comparing the variance inflation factor of each feature in the key data set with a second threshold, and if the variance inflation factor of the feature is less than the second threshold, putting the feature into a feature subset,

repeating said calculating the variance inflation factor for each feature in said key dataset and said subset of features and comparing the variance inflation factor for each feature in said key dataset to a second threshold, if the variance of the feature If the expansion factor is smaller than the second threshold, then put the feature into the feature subset until the comparison of all the features of the key data set is completed.
The model training method according to claim 1, wherein, for each feature, calculating the average distance in the sample dimension comprises:

Calculate the average distance between different samples of each feature under the same fault category;

Based on the average distance between different samples of each feature under the same fault category, calculate the average value of each feature in multiple fault categories, and record it as the first average distance;

For each feature, calculating the average distance in the fault category dimension includes:

Calculate the average value of each feature in all samples under the same fault category;

Based on the average value of each feature in all samples under the same fault category, the average distance of each feature average among different fault categories is calculated, and recorded as the second average distance.
The model training method according to claim 4, wherein said obtaining the characterization value of the importance of each feature based on the calculation result of the average distance of two dimensions comprises:

Calculate the variance factor of the first mean distance for each feature,

Calculate the variance factor of the second mean distance for each feature;

Calculate compensation factors based on the two variance factors for each feature;

A representative value of the importance of each feature is obtained based on the compensation factor of each feature, the first average distance, and the second average distance.
The model training method according to claim 4, wherein the calculation of the average distance between different samples of each feature under the same fault category adopts formula (1):

The average value of each feature in multiple fault categories is calculated using formula (2):

Wherein, the initial data set includes K kinds of fault categories, q i,k,j , i=1,2,...,I k ; k=1,2,...,K; j=1,2,...,J , where q i,k,j represent the j-th eigenvalue of the i-th sample in the k-th fault category, I k is the number of samples of the k-th fault category, K is the number of fault categories, and J is each sample number of features.
The model training method according to claim 4, wherein the calculation of the average value of each feature in all samples under the same fault category adopts formula (3):

The average distance of each characteristic mean value between the described calculation different fault categories adopts formula (4):

Wherein, the initial data set includes K kinds of fault categories, q i,k,j , i=1,2,...,I k ; k=1,2,...,K; j=1,2,...,J , where q i,k,j represent the j-th eigenvalue of the i-th sample in the k-th fault category, I k is the number of samples of the k-th fault category, K is the number of fault categories, and J is each sample number of features.
The model training method according to claim 3, wherein, for each feature in each key data set, the step of calculating the variance inflation factor with the feature subset comprises: substituting into formula (5) to obtain the variance of the feature expansion factor,

in, Represents the goodness of fit of the regression equation X j =β 0 +βX′, where X′ is a feature of the feature subset.
A model training device, comprising:

A feature extraction unit is used to construct an initial dataset of high-frequency vibration signals;

A feature screening unit, configured to screen the initial data set to obtain key data sets;

A model training unit, configured to use the key data set for model training to obtain a fault recognition model, wherein,

The initial data set is characterized by three dimensions of fault category, sample and feature, and the feature screening unit includes:

For each feature, the average distance is calculated in the two dimensions of the sample and the fault category, and the representative value of the importance of each feature is obtained based on the average distance calculation results of the two dimensions; and

The key data set is composed of features whose importance characteristic value is greater than the first threshold.
An electronic device comprising a memory and a processor, the memory further storing information executable by the processor computer instructions, when the computer instructions are executed, the model training method according to any one of claims 1 to 8 is realized.
A computer-readable medium, the computer-readable medium stores computer instructions executable by an electronic device, and when the computer instructions are executed, the model training method according to any one of claims 1 to 8 is implemented.