Embedded artificial intelligence acoustic vibration detection system and model construction method

Abstract

The present application discloses an embedded artificial intelligence sound and vibration detection system and a model construction method, which relate to the field of AI technology. The system configuration center is configured with processor type and architecture information, sensor type information, a feature extraction library, a detection algorithm library, and application scenario information; the AI model verification and training module is used to perform model training and verification according to the selected feature extraction method, detection algorithm, and the input motor sound and vibration data; the embedded application export module is used to automatically compile according to the trained AI model and its model parameters to generate a model static library corresponding to the target AI model; the embedded application verification module is used to import the compiled model static library and perform model accuracy verification according to the input embedded verification data. By setting up the system configuration center, the most applicable feature extraction method and detection algorithm are extracted from the feature extraction library and the detection algorithm library according to the instance and weight score to construct the AI model, improving the detection accuracy and detection efficiency.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of AI technology, and particularly to an embedded artificial intelligence acoustic vibration detection system and a model construction method. Background Art

[0002] In industry, it is common to detect abnormalities of devices such as motors based on acoustic vibration signals. This kind of detection is often manually judged by experienced workers' so-called "golden ears". This judgment result is related to the experience intensity of the workers, and there are defects of instability, high labor cost, and low efficiency. In order to save labor, improve efficiency and accuracy, some advanced manufacturers seek external cooperation to develop an intelligent judgment system. However, for each new product added, the algorithm module in the intelligent judgment system needs to be customized. With the increase and decrease of production lines, the cost of external cooperation development and system maintenance continue to increase, which discourages many manufacturers.

[0003] In related technologies, some manufacturers use an acoustic vibration detection model based on artificial intelligence to replace the "golden ears". However, the motors produced may be applied to different scenarios, and the performance of the test platform or test equipment varies. If the same standardized detection model is used, it may lead to limited universality and reduced detection accuracy. Summary of the Invention

[0004] The embodiments of the present application provide an embedded artificial intelligence acoustic vibration detection system and a model construction method, which solve the problems of limited universality of the model test platform and the test model and low detection accuracy.

[0005] On the one hand, the present application provides an embedded artificial intelligence acoustic vibration detection system, which includes a system configuration center, an AI model verification and training module, an embedded application export module, and an embedded application verification module;

[0006] The system configuration center is configured with processor type and architecture information, sensor type information, a feature extraction library, a detection algorithm library, and application scenario information; wherein, the feature extraction library stores several feature extraction methods for extracting the characteristics of motor acoustic vibration data, and the detection algorithm library stores feature detection algorithms for model training;

[0007] The AI model verification and training module is used to perform model training and verification according to the selected feature extraction method, detection algorithm, and the input motor acoustic vibration data; wherein, the feature extraction method and the detection algorithm are determined according to the model running environment configuration parameters and the motor acoustic vibration data;

[0008] The embedded application export module is used to automatically compile according to the trained AI model and its model parameters to generate a model static library corresponding to the target AI model; the embedded application verification module is used to import the generated model static library and verify the model accuracy according to the input embedded verification data.

[0009] Specifically, the model running environment configuration parameters include at least one of the processor type structure, the occupied size of the chip resource space, and the computing power of the processor.

[0010] The application scenario information includes at least one of the automotive field, wind power generation, hydroelectric power generation, and industrial and agricultural production; the sensor type information includes at least one of sound data, vibration data, rotational speed data, voltage data, and pressure data, and the corresponding model configuration parameters include at least one of the number of sensor signals, signal types, and acquisition environments, and the acquisition environment includes a strong noise environment, a normal environment, and a low noise environment.

[0011] Specifically, the detection algorithm library includes a data balancing algorithm, a supervised classification learning algorithm, an outlier detection algorithm, a novelty point detection algorithm, and a deep anomaly detection algorithm; the feature extraction library includes feature extraction and feature selection methods; the feature extraction methods therein include at least one of the order analysis feature extraction method, the short-time Fourier feature extraction method, and the power feature extraction method; the feature selection method includes at least the correlation coefficient method and the information gain method, and is used to select all the extracted feature types.

[0012] Specifically, the AI model verification and training module further includes an AI model algorithm library; the AI model therein is constructed based on the selected feature extraction method, feature selection method in the feature extraction library, and the detection algorithm selected in the detection algorithm library, and the AI model algorithm library includes AI models composed of any feature extraction method, feature selection method, and detection algorithm and their corresponding model parameters, and is used to select the target AI model according to the model running environment configuration parameters and the motor acoustic and vibration data.

[0013] Specifically, the label types of the motor acoustic and vibration data include an abnormal label, an OK label, and an unknown label, which are used to represent the abnormal, normal, and unclear states of the motor respectively; and the abnormal label is divided into a first-class abnormal label, a second-class abnormal label to an n-class abnormal label, which respectively correspond to different functional abnormalities of the motor.

[0014] Specifically, the data balancing algorithm includes at least one of the K-nearest neighbor algorithm AllKNN, the close point elimination algorithm NearMiss, the neighborhood deletion rule algorithm NeighbourhoodCleaningRule, and the one-sided selection algorithm OneSidedSelection.

[0015] The supervised classification learning algorithm includes at least one of K-nearest neighbor algorithm, linear regression algorithm, support vector machine, decision tree, and random forest; the outlier detection algorithm includes at least one of KMeans algorithm, LOF algorithm, and Isolation Forest algorithm;

[0016] The novelty point detection algorithm includes at least one of OneClassSVM algorithm, LocalOutlierFactor algorithm, and autoencoder algorithm;

[0017] The deep anomaly detection algorithm includes at least one of deep convolutional neural network (CNN) algorithm, recurrent neural network (RNN) algorithm, and long short-term memory neural network (LSTM) algorithm.

[0018] On the other hand, the present application provides a method for constructing an embedded motor acoustic vibration detection model, the method comprising:

[0019] Obtaining a training sample of motor acoustic vibration data, and determining the sample data volume, sensor data type, model application scenario information, and model running environment configuration parameters; the model running environment configuration parameters are used to determine the processor resources and computing power, and calculate the weight score, and the weight score is used to determine the target detection algorithm and / or target feature extraction method of the target AI model;

[0020] When the sample data volume is less than the lowest data volume threshold, or the processor resources and computing power are less than the lowest threshold, determine to extract the sample data features using a signal processing-based feature extraction method, and determine the target detection algorithm according to the weight score, and combine and construct to generate a first target AI model;

[0021] When the sample data volume is greater than the highest data volume threshold, and the processor resources and computing power are greater than the highest threshold, determine the deep anomaly detection algorithm as the target detection algorithm, and determine the target feature extraction method according to the weight score, and combine and construct to produce a second target AI model;

[0022] When the sample data volume is between the lowest and highest data volume thresholds, or the processor resources and computing power are between the lowest and highest thresholds, determine the target detection algorithm and the target feature extraction method according to the label type of the sample data set and the weight score, and combine and construct to generate a third target AI model.

[0023] Specifically, the label types of the sample data include unknown label, anomaly label, and OK label; among them, the anomaly label and OK label are recognizable labels, and the unknown label is an unrecognizable label;

[0024] When the sample data volume is lower than the minimum data volume threshold, all sample data sets are identifiable labels;

[0025] When the sample data volume is between the minimum and maximum data volume thresholds, the sample data set includes identifiable labels and / or unidentifiable labels; among them, the identifiable labels are used to supervise the classification learning algorithm for supervised learning;

[0026] When the sample data volume is greater than the maximum data volume threshold, all sample data sets are identifiable labels, which are used for deep learning by the deep anomaly detection algorithm.

[0027] Specifically, when determining the first target AI model, determine the acquisition environment of the sample data set; when the acquisition environment is a low-noise environment or an ordinary environment, directly use the feature extraction method of signal processing and the weight score to determine the target detection algorithm for construction;

[0028] When the acquisition environment is a strong-noise environment, perform denoising filtering on the sample data set, and then use the feature extraction method of signal processing and the weight score to determine the target detection algorithm for construction; among them, the feature extraction method of signal processing includes the short-time Fourier and power spectrum feature extraction methods;

[0029] When determining the third target AI model, first determine the target feature extraction method according to the weight score, and then determine the label type in the sample data set; when all are known labels containing anomaly labels and the data types of the anomaly labels are balanced, use the supervised classification learning algorithm to construct the AI model; when all are known labels containing anomaly labels but the data types of the anomaly labels are unbalanced, first use the data balancing algorithm for data preprocessing, and then use the classification learning algorithm to construct the AI model; <L

[0030] When all are unknown labels and / or OK labels, determine the identifiable label type, and when all the identifiable label types are OK labels, determine the novelty point detection algorithm as the target detection algorithm, otherwise determine the outlier detection algorithm as the target detection algorithm and construct the AI model.

[0031] Specifically, the computational complexity and storage capacity of the target feature extraction method and the target detection algorithm are positively correlated with the weight score; the weight score is determined based on the sample data volume, the type and quantity of sensor data, the processor resources and computing power, and the generated target AI model and built-in parameters are classified and stored according to the application scenario information of the sample data set.

[0032] The beneficial effects brought by the technical solutions provided in the embodiments of the present application at least include: providing a universal motor acoustic vibration detection system, configuring processor type and architecture information, sensor type information, feature extraction library, detection algorithm library, and application scenario information, etc. in the system configuration center, which is convenient to extract the most applicable feature extraction method and detection algorithm from the feature extraction library and detection algorithm library according to specific instances, and use them to construct a target AI model. During model training, the target AI model can be selected for training and verification according to various input motor acoustic vibration data. The trained AI model can generate a model static library for the actual detection device through the embedded application export module, and perform recall verification through the embedded application verification module to confirm the scene detection accuracy of the model.

[0033] The method for constructing a detection model provided in the present application can determine the weight score for constructing an AI model according to the hardware conditions of the actual detection device and configuration parameters of the model operating environment, and determine the specific target feature extraction method and target detection algorithm according to the number of training samples, so as to construct a target AI model that is most suitable for the hardware device and application scenario, improving the detection accuracy and detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a structural block diagram of an embedded artificial intelligence acoustic vibration detection system provided in an embodiment of the present application;

[0035] Figure 2 is a flowchart of a method for constructing an embedded motor acoustic vibration detection model provided in another embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the objectives, technical solutions, and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0037] As used herein, "a plurality of" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.

[0038] Such as Figure 1Figure 2 shows a block diagram of the embedded artificial intelligence acoustic and vibration detection system provided by an embodiment of the present application. The system includes a system configuration center, an AI model verification and training module, an embedded application export module, and an embedded application verification module. The system configuration center is configured with information about the processor type and architecture, sensor type, a feature extraction library, a detection algorithm library, and application scenario information. The feature extraction library stores several feature extraction methods for extracting features from motor acoustic and vibration data, while the detection algorithm library stores feature detection algorithms used for model training.

[0039] The AI model verification and training module is used to train and verify the model based on the selected feature extraction method, detection algorithm, and input motor acoustic and vibration data. The feature extraction method and detection algorithm are determined based on the model's operating environment configuration parameters and the motor acoustic and vibration data. During actual operation after training, personnel can import motor test data based on the system's existing models and parameters to verify accuracy. They can also import a large number of data paths for model retraining.

[0040] The embedded application export module is used to automatically compile the trained AI model and its model parameters to generate a model static library corresponding to the target AI model. This module has built-in embedded implementations of all algorithm models and feature extraction methods corresponding to the selected configuration. Based on the trained parameters, it can automatically compile and export a static library of the selected model plus parameters, which the user can then call according to the examples. The embedded application verification module is used to import the compiled model static library and verify the model accuracy based on the input embedded verification data. On the embedded simulator, import the compiled static library and write simple embedded code to verify it on a small amount of verification data.

[0041] The interactions between each of the above steps are linked within the system. For example, the model parameters trained by the AI model verification and training module will be automatically applied to the embedded application export module.

[0042] The system's primary process involves building an AI model based on the configuration and motor data in the system configuration center. Because the system requires greater intelligence and the generation of scenario-specific AI models based on specific instances, the system configuration center includes a rich set of optional parameter standards, allowing the system to freely match the optimal AI model.

[0043] The configuration parameters of the model running environment include at least the processor type structure, the occupied size of the chip resource space, and the computing power of the processor. Since the performance of different processor architectures is different, the occupied chip resource space will also affect the operation of the later AI model. In addition, the computing power of the processor will also affect the operation and detection efficiency of the AI model. The model configuration parameters for building the model include the number of sensor signals, signal types, and acquisition environment information, etc. Because when the sensor types are different, the collected motor acoustic vibration data is also different, including sound data, vibration data, rotation speed data, voltage data, etc., and the processing methods and acquisition frequencies of different data are also different. For example, the input acoustic vibration data includes one or more of them, etc. The application scenario information includes the automotive field, wind power generation, hydropower generation, industrial and agricultural production, etc. Factors such as the power consumption and frequency of the motor in different scenarios will also cause the acoustic vibration data to be different. Therefore, the AI model needs to be established according to the application scenario, and the target AI model corresponding to the scenario needs to be selected during the later acoustic vibration test.

[0044] In addition, during the test and training stage, the motor acoustic vibration data is used as training samples and needs to cover a variety of noise environments, that is, the acquisition environment information covers strong noise environments, ordinary environments, and low noise environments (the specific decibel levels are determined according to the actual situation). The robustness and detection accuracy of the AI model obtained through training covering a variety of noise environments are relatively better. There are data balancing algorithms, supervised classification learning algorithms, outlier detection algorithms, novelty point detection algorithms, and deep anomaly detection algorithms built into the detection algorithm library. The model sizes of these detection algorithms and the requirements for the computer storage capacity and computing power are different, and the recognition accuracy and efficiency are also different.

[0045] In this application, the data balancing algorithms include at least one of the K-Nearest Neighbor algorithm (AllKNN), the Near Miss algorithm, the Neighbourhood Cleaning Rule algorithm, and the One Sided Selection algorithm. The supervised classification learning algorithms include at least one of the K-Nearest Neighbor algorithm, the linear regression algorithm, the support vector machine, the decision tree, and the random forest. The outlier detection algorithms include at least one of the KMeans algorithm, the LOF algorithm, and the Isolation Forest algorithm. The novelty point detection algorithms include at least one of the One-Class Support Vector Machine (OneClassSVM) algorithm, the Local Outlier Factor algorithm, and the autoencoder algorithm. The deep anomaly detection algorithms include at least one of the Deep Convolutional Neural Network (CNN) algorithm, the Recurrent Neural Network (RNN) algorithm, and the Long Short-Term Memory Neural Network (LSTM) algorithm.

[0046] Before feature recognition, it is necessary to extract necessary data features. After feature extraction, it is also necessary to screen all feature data according to specific circumstances and select several necessary features for data analysis. Therefore, a feature extraction method library and a feature selection method library are built into the feature extraction library. The feature extraction methods include at least one of order analysis feature extraction method, short-time Fourier feature extraction method, and power feature extraction method. The feature selection methods include at least the correlation coefficient method and the information gain method, which are used to select all extracted feature types for model construction. The feature selection method is based on filter multi-objective feature selection to select K features from the selection library. Taking the correlation coefficient method as an example, the feature correlation P1 between OK samples and NG samples, the feature correlation P2 between OK samples and OK samples, and the feature correlation P3 between NG samples and NG samples can be calculated. The first N features are arranged in descending order of P = P2 + P3 - p * P1, where p is a weight constant. Specific sample types are described in detail below.

[0047] In addition, the system configuration center is also configured with an AI model algorithm library, which contains several pre-trained AI models and is constructed based on the feature extraction methods selected from the feature extraction library and the detection algorithms selected from the detection algorithm library. For the well-trained AI model library, it should include AI models composed of any feature extraction methods and detection algorithms and their corresponding model parameters, which are used to select the target AI model according to the model running environment configuration parameters and the motor acoustic vibration data for quick matching and selection.

[0048] For the motor acoustic vibration data used for model training, its label types include abnormal labels, OK labels, and unknown labels, which are used to represent the abnormal, normal, and unclear states (unlabeled) of the motor respectively. The abnormal labels are divided into type I abnormal label NG1, type II abnormal label NG2 to type n abnormal label NGn according to the motor functions, corresponding to different functional abnormalities of the motor. When selecting a specific target AI model, it is also necessary to select the best usage method according to the type of acoustic vibration data. For example, in the case of a small amount of data, the label types of each data can be recognized to construct an AI model. In the case of a moderate amount of data, a small amount of labeled data is used for training and the corresponding target AI model is constructed.

[0049] Using this system, ordinary engineers can configure information such as the processor type and sensor type, select the application scenario according to needs. After configuration and selection, the collected data is imported to automatically train the detection algorithm model. After training and verification are completed, the embedded static library can also be exported according to the processor type provided by the user for subsequent engineers to update and iterate. Each configuration combination corresponds to an AI model, and the model has been trained with some data. The recall rate of some products is 100%, and the accuracy rate can reach about 95%.

[0050] In summary, by providing a universal motor acoustic vibration detection system, the present solution configures processor type and architecture information, sensor type information, feature extraction library, detection algorithm library, and application scenario information, etc. in the system configuration center, which is convenient to extract the most applicable feature extraction method and detection algorithm from the feature extraction library and the detection algorithm library according to specific examples, and build a target AI model with this. During model training, various motor acoustic vibration data inputs can be used to select the target AI model for training and verification. The trained AI model can generate a model static library for the actual detection device through the embedded application export module, and conduct recall verification through the embedded application verification module to confirm the scene detection accuracy of the model.

[0051] Figure 2 FIG. is a flowchart of an embedded motor acoustic vibration detection model construction method provided by another embodiment of the application. This process is mainly used for the preliminary construction and training of the AI model. The specific steps are as follows:

[0052] S1. Obtain motor acoustic vibration data training samples, and determine the sample data volume, sensor data type, model application scenario information, and model operating environment configuration parameters.

[0053] The model operating environment configuration parameters therein are used to determine the processor resources and computing power, as well as calculate the weighted score, and the weighted score is used to determine the target detection algorithm and / or target feature extraction method of the target AI model.

[0054] The target AI model in this solution needs to be selected based on the weighted score, or the target detection algorithm and / or target feature extraction method are determined based on the weighted score. That is to say, the memory size of the target feature extraction method and the target detection algorithm is positively correlated with the weighted score. The higher the weighted score, the larger the storage capacity and computing volume of the selected feature extraction method and detection algorithm, and the higher the requirements for the computer's hardware and the processor's processing performance. Specifically, it can also be obtained by summing up the various weight scores. For example, the amount of sample data, the type and quantity of sensor data, the processor resources and computing power respectively indicate the first weight score, the second weight score, the third weight score, etc. During operation, the actual score is calculated according to the level of the current computer device itself and the weight of the corresponding item, and the sum of all scores is determined as the weighted score, and the corresponding target detection algorithm and / or target feature extraction method is determined according to its numerical size, that is, the weighted score and the target detection algorithm and / or target feature extraction method are in a one-to-one mapping relationship.

[0055] S2. When the sample data volume is less than the minimum data volume threshold, or the processor resources and computing power are less than the minimum threshold, determine to use the feature extraction method of signal processing to extract the sample data features, and determine the target detection algorithm according to the weighted score, and combine and construct the first target AI model.

[0056] The following takes Figure 2 the specific embodiments in it as an example. When the sample data volume is less than the lowest data volume threshold (less than 20 pieces), or the processor resources and computing power are less than the lowest threshold (low processor resources and computing power), it is determined to select the feature extraction method of signal processing to extract the sample data features. Because in the case of less data volume or hardware performance limitations, the device operation load is reduced as much as possible, the features in the signal time domain are used for feature extraction and selection, and the first target AI model is constructed by combining the thresholds and empirical values set according to industry experience data. The target feature extraction method therein can be the short-time Fourier feature extraction method or the power feature extraction method, which is specifically determined according to the weight score.

[0057] After extracting the feature parameters, it is necessary to select the target detection algorithm. However, the feature extraction based on low data needs to consider the acquisition environment when collecting the sample data set. In a low-noise and ordinary-noise environment, there is no need to filter the parameters, and the target detection algorithm can be directly determined according to the weight score and the recommended empirical values. However, when the acquisition environment is a strong-noise environment, it is necessary to filter and denoise the collected audio signals, pressure signals, and vibration signals, and then construct the first target AI model based on the target detection algorithm and the target feature extraction method (and its target feature selection method).

[0058] The anomaly detection method (target detection algorithm) based on signal processing is to use the absolute value, standard deviation, skewness, kurtosis, entropy, root mean square, peak-to-peak value, crest factor, Clearence Factor, shape factor, impulse degree, etc. in the signal time domain, and the octave band, stft, power spectrum, order analysis (rotational speed signal needs to be provided), MFCC, etc. in the frequency domain. After selecting the appropriate features through the feature selection algorithm using these analysis results, a threshold is constructed to judge the sample. The threshold is determined according to industry data, and there are different numbers of signal processing threshold judgment methods for each application field. The built-in threshold of the model is the empirical value of the system in this field before.

[0059] S3. When the sample data volume is greater than the highest data volume threshold and the processor resources and computing power are greater than the highest threshold, the deep anomaly detection algorithm is determined as the target detection algorithm, and the target feature extraction method is determined according to the weight score, and the second target AI model is constructed by combination.

[0060] Schematically, when the sample data volume is greater than the highest data volume threshold (greater than 10,000 records), and the processor resources and computing power are greater than the highest threshold (high processor resources and computing power), it is determined to select a deep anomaly detection algorithm (DAD technology based on deep learning) as the target detection algorithm. Because with the support of big data, an algorithm with higher recognition accuracy and depth needs to be adopted as the target detection algorithm. For example, CNN algorithm, RNN algorithm, LSTM algorithm, etc. are adopted, and it is specifically selected according to the weight score. Similarly, when the weight score is relatively large, a method with higher prediction effect can be selected as the recommendation method. Then the second target AI model is constructed by combination.

[0061] The above two conditions have restrictions on the sample label type during training. Generally, the label types of sample data are divided into unknown labels, anomaly labels, and OK labels; anomaly labels and OK labels are recognizable labels, and unknown labels are unrecognizable labels. More specifically, anomaly labels are further divided into NG1 label, NG2 label,..., NGn label, respectively representing different fault types. When the sample data volume is lower than the lowest data volume threshold, all sample data sets are recognizable labels. When the sample data volume is between the lowest and highest data volume thresholds, the sample data set includes recognizable labels and / or unrecognizable labels; among them, recognizable labels are used to supervise the classification learning algorithm for supervised learning. When the sample data volume is greater than the highest data volume threshold, all sample data sets are recognizable labels, which are used for deep learning by the deep anomaly detection algorithm. The input of the first target AI model and the second target AI model during the training process are both recognizable labels, which are used to improve the model training accuracy and detection effect. For example, 100 groups of fixed-speed 20s audio signals, OK:NG = 2:8, wav format and other constraint information are used as the training data set.

[0062] S4. When the sample data volume is between the lowest and highest data volume thresholds, or the processor resources and computing power are between the lowest and highest thresholds, determine the target detection algorithm and the target feature extraction method according to the label type and weight score of the sample data set, and construct and generate the third target AI model by combination.

[0063] Exemplarily, when the sample data volume is between 20 and 10,000, or the processor resources and computing power are between the lowest and highest thresholds, although a machine learning-based method can be selected to construct the third target AI model and determine the target feature extraction method. It should be noted that there are various supervised classification learning algorithms, including fully supervised, semi-supervised, and unsupervised learning algorithms, and there are also many optional algorithms under each type of supervision mode. Therefore, it is necessary to determine in combination with the functional characteristics of supervised learning and the label type of the training data set.

[0064] After initially determining the target feature extraction method based on weight scores, all data labels are classified, which specifically includes the following steps:

[0065] 1. When it is determined that all label types are known labels containing abnormal labels and the data types of the abnormal labels are balanced, a third target AI model is constructed using a supervised classification learning algorithm.

[0066] The balanced data types of the abnormal labels mean that the proportion of the number of abnormal labels such as NG1, NG2, NG3, etc. is relatively balanced, which is beneficial to the accuracy of classification supervision. After processing, a supervised classification learning algorithm is selected as the target detection algorithm.

[0067] 2. When all are known labels containing abnormal labels but the data types of the abnormal labels are unbalanced, first perform data preprocessing using a data balancing algorithm, and then construct a third target AI model using a classification learning algorithm.

[0068] The proportion of the number of abnormal labels such as NG1, NG2, NG3, etc. is relatively unbalanced. For example, when the overall proportion of NG1 is 45%, the overall proportion of NG2 is 15%, and the overall proportion of NG3 is 5%, the data is significantly unbalanced. At this time, a data balancing algorithm needs to be selected to process the data. Specifically, the K-nearest neighbor algorithm (AllKNN), the close point elimination algorithm (NearMiss), the neighbor deletion rule algorithm (NeighbourhoodCleaningRule), the one-sided selection algorithm (OneSidedSelection), etc. can be used. After processing, a supervised classification learning algorithm is selected as the target detection algorithm.

[0069] The above two methods both select a supervised classification learning algorithm when the label types are known.

[0070] 3. When all are unknown labels and / or OK labels, determine the recognizable label types. And when all the recognizable label types are OK labels, determine the novelty point detection algorithm as the target detection algorithm; otherwise, determine the outlier detection algorithm as the target detection algorithm and construct an AI model.

[0071] The reason for outlier detection is that the types of NG are not clear, so the specific types of NG of the labels need to be determined first. When the recognizable label types are determined and all the recognizable label types are OK labels, a semi-supervised learning algorithm should be selected, that is, the novelty point detection algorithm is used as the target detection algorithm; on the contrary, when all are unknown labels or contain unknown labels, an unsupervised learning algorithm is selected, that is, the outlier detection algorithm is used as the target detection algorithm. The method of classification according to the supervision method can make the best use of things and give full play to the maximum performance to detect the acoustic vibration data.

[0072] In a possible embodiment, the final features can be selected first using the correlation coefficient method for all features in the feature library. Then, according to the relationship between the scores and the algorithms, a suitable classification algorithm can be selected. For example, if all the data labels are positive samples, and the resources and computing power of the processor are close to the highest data volume threshold, then a clustering algorithm with a large computational overhead and requiring historical data for prediction, such as the Local Outlier Factor algorithm, can be selected. It should be noted that the selected algorithm is not unique and can also be expanded with the upgrade of the system, and the mapping relationship between the scores and the specific algorithms can be adjusted.

[0073] The method for constructing the detection model provided by this application can determine the weight scores for constructing the AI model according to the hardware conditions of the actual detection device and the configuration parameters of the model operating environment, and determine the specific target feature extraction method and target detection algorithm according to the number of training samples, so as to construct the target AI model most suitable for the hardware device and application scenario, improving the detection accuracy and detection efficiency.

[0074] The above describes the preferred embodiments of the present invention; it should be understood that the present invention is not limited to the above specific embodiments, and the devices and structures not described in detail should be understood to be implemented in a common manner in the art; any person skilled in the art can make many possible changes and modifications without departing from the technical solution of the present invention, or modify it into an equivalent embodiment with equivalent changes, which does not affect the essence of the present invention; therefore, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An embedded motor acoustic and vibration detection system, characterized in that The system includes a system configuration center, an AI model verification and training module, an embedded application export module, and an embedded application verification module; The system configuration center is configured with processor type and architecture information, sensor type information, a feature extraction library, a detection algorithm library, and application scenario information; among them, the feature extraction library stores several feature extraction methods for extracting the features of motor acoustic and vibration data, and the detection algorithm library stores feature detection algorithms for model training; the AI model verification and training module is used to perform model training and verification according to the selected feature extraction method, detection algorithm, and the input motor acoustic and vibration data; among them, the feature extraction method and the detection algorithm are determined according to the model running environment configuration parameters and the motor acoustic and vibration data; The embedded application export module is used to automatically compile according to the trained AI model and its model parameters to generate a model static library corresponding to the target AI model; The embedded application verification module is used to import the compiled model static library and perform model accuracy verification according to the input embedded verification data; Among them, when the sample data volume of the motor acoustic and vibration data is less than the minimum data volume threshold, or the processor resources and computing power are less than the minimum threshold, it is determined to use the feature extraction method of signal processing to extract the sample data features, and the target detection algorithm is determined according to the weight score, and the first target AI model is generated by combination; the weight score is based on the sample data volume, sensor data type and quantity, processor resources and computing power; When the sample data volume is greater than the maximum data volume threshold and the processor resources and computing power are greater than the maximum threshold, the deep anomaly detection algorithm is determined as the target detection algorithm, and the target feature extraction method is determined according to the weight score, and the second target AI model is generated by combination; When the sample data volume is between the minimum and maximum data volume thresholds, or the processor resources and computing power are between the minimum and maximum thresholds, the target detection algorithm and the target feature extraction method are determined according to the label type of the sample data set and the weight score, and the third target AI model is generated by combination.

2. The system according to claim 1, wherein The model running environment configuration parameters include at least one of processor type structure, chip resource space occupancy size, and processor computing power; The application scenario information includes at least one of the automotive field, wind power generation, hydropower generation, and industrial and agricultural production; the sensor type information includes at least one of sound data, vibration data, rotation speed data, and voltage data, and the corresponding model configuration parameters include at least one of the number of sensor signals, signal types, and acquisition environments, and the acquisition environments include strong noise environments, ordinary environments, and low noise environments.

3. The system according to claim 2, wherein The detection algorithm library includes a data balancing algorithm, a supervised classification learning algorithm, an outlier detection algorithm, a novelty point detection algorithm, and a deep anomaly detection algorithm; the feature extraction library includes feature extraction and feature selection methods; the feature extraction methods therein include at least one of order analysis feature extraction method, short-time Fourier feature extraction method, and power feature extraction method; the feature selection method includes at least the correlation coefficient method and the information gain method, and is used to select all extracted feature types; The supervised classification learning algorithm includes at least one of the K-nearest neighbor algorithm, linear regression algorithm, support vector machine, decision tree, and random forest; the deep anomaly detection algorithm includes at least one of the deep convolutional neural network CNN algorithm and the recurrent neural network RNN algorithm; among them, the recurrent neural network RNN algorithm includes the long short-term memory neural network LSTM algorithm.

4. The system according to claim 3, characterized in that, The AI model verification and training module also includes an AI model algorithm library; the AI model is constructed based on the selected feature extraction method, feature selection method in the feature extraction library, and the detection algorithm selected in the detection algorithm library, and the AI model algorithm library includes AI models composed of any feature extraction method, feature selection method, and detection algorithm and their corresponding model parameters, and is used to select the target AI model according to the model running environment configuration parameters and the motor acoustic vibration data.

5. The system according to claim 4, wherein The label types of the motor acoustic vibration data include an anomaly label, an OK label, and an unknown label, which are used to represent the abnormal, normal, and unclear states of the motor respectively; and the anomaly label is divided into a type-I anomaly label, a type-II anomaly label to an n-type anomaly label, which respectively correspond to different functional anomalies of the motor.

6. The system according to claim 3, wherein The data balancing algorithm includes at least one of the K-nearest neighbor algorithm AllKNN, the close point elimination algorithm NearMiss, the neighbor deletion rule algorithm NeighbourhoodCleaningRule, and the one-sided selection algorithm OneSidedSelection; The outlier detection algorithm includes at least one of the KMeans algorithm and the Isolation Forest algorithm; The novelty point detection algorithm includes at least one of the one-class support vector machine OneClassSVM algorithm, the local outlier factor algorithm LocalOutlierFactor algorithm, and the autoencoder algorithm; 7. A method for constructing an embedded motor acoustic vibration detection model, characterized in that, The method is used for the embedded motor acoustic vibration detection system according to any one of claims 1-6, and the method includes: Obtain the motor acoustic vibration data training samples, and determine the sample data volume, sensor data type, model application scenario information, and model running environment configuration parameters; the model running environment configuration parameters are used to determine the processor resources and computing capabilities, and calculate the weighted score, and the weighted score is used to determine the target detection algorithm and / or target feature extraction method of the target AI model; When the sample data volume is less than the minimum data volume threshold, or the processor resources and computing power are less than the minimum threshold, determine to use the feature extraction method of signal processing to extract the features of the sample data, and determine the target detection algorithm according to the weight score, and combine and construct to generate the first target AI model; When the sample data volume is greater than the maximum data volume threshold, and the processor resources and computing power are greater than the maximum threshold, determine the deep anomaly detection algorithm as the target detection algorithm, and determine the target feature extraction method according to the weight score, and combine and construct to generate the second target AI model; When the sample data volume is between the minimum and maximum data volume thresholds, or the processor resources and computing power are between the minimum and maximum thresholds, determine the target detection algorithm and the target feature extraction method according to the label type of the sample data set and the weight score, and combine and construct to generate the third target AI model.

8. The method according to claim 7, characterized in that, The label types of the sample data include unknown labels, anomaly labels, and OK labels; among them, the anomaly labels and OK labels are recognizable labels, and the unknown labels are unrecognizable labels; When the sample data volume is lower than the minimum data volume threshold, all the sample data sets are recognizable labels; When the sample data volume is between the minimum and maximum data volume thresholds, the sample data set includes recognizable labels and / or unrecognizable labels; among them, the recognizable labels are used to supervise the classification learning algorithm for supervised learning; When the sample data volume is greater than the maximum data volume threshold, all the sample data sets are recognizable labels and are used for deep learning by the deep anomaly detection algorithm.

9. The method according to claim 8, wherein When determining the first target AI model, determine the acquisition environment of the sample data set; when the acquisition environment is a low-noise environment or an ordinary environment, directly use the feature extraction method of signal processing and the weight score to determine the target detection algorithm for construction; When the acquisition environment is a strong-noise environment, denoise and filter the sample data set, and then use the feature extraction method of signal processing and the weight score to determine the target detection algorithm for construction; among them, the feature extraction method of signal processing includes the short-time Fourier and power spectrum feature extraction methods; When determining the third target AI model, first determine the target feature extraction method according to the weight score, and then determine the label type in the sample data set; when all are known labels containing anomaly labels and the anomaly label data types are balanced, use the supervised classification learning algorithm to construct the third target AI model; when all are known labels containing anomaly labels but the anomaly label data types are unbalanced, first use the data balancing algorithm for data preprocessing, and then use the classification learning algorithm to construct the AI model; When all are unknown labels and / or OK labels, determine the recognizable label type, and when all the recognizable label types are OK labels, determine the novelty point detection algorithm as the target detection algorithm, otherwise determine the outlier detection algorithm as the target detection algorithm, and construct the AI model.

10. The method according to any one of claims 7-9, characterized in that, The computational complexity and storage requirements of the target feature extraction method and the target detection algorithm are positively correlated with the weight score; the weight score is determined based on the sample data volume, the type and quantity of sensor data, the processor resources, and the computing power, and the generated target AI model and built-in parameters are classified and stored according to the application scenario information of the sample data set.

Images