CN115222738B - Online learning abnormity detection method and system based on feature migration - Google Patents

Abstract

The invention discloses an online learning abnormity detection method and system based on feature migration, which belong to the technical field of product abnormity detection and comprise the following steps: step 1: collecting sample data; step 2: feature migration of the feature extractor; and 3, step 3: sample data feature normalization; and 4, step 4: and (4) online learning of the anomaly detection model. The anomaly detection model is obtained through online learning, and the method can be well adapted to the condition that the customer requirements are not clearly defined or the requirements change, and meets the actual production requirement; the normal sample feature normalization method based on online learning is provided, so that the sample features can be well expressed, and the detection effect is improved; and the abnormal samples and the suspected abnormal samples are fed back to the client, and the client determines whether the samples are really abnormal or not, so that the requirements of the client can be met in time.

Description

Online learning abnormity detection method and system based on feature migration

Technical Field

The invention relates to the technical field of product abnormity detection, in particular to a method and a system for detecting online learning abnormity based on feature migration.

Background

In actual industrial production, a factory has a plurality of production lines for continuously producing industrial products (such as lead storage batteries), and each production line produces a large amount of products every day, and even reaches hundreds of thousands of products in a peak time. Due to the production quality variation, a certain proportion of defective products are produced every day. To find these blemishes, the factory is often configured with specialized quality inspectors to manually inspect the products. With the increase of the number of products, the manual inspection method needs to consume huge manpower and time, and the economical efficiency is poor. An automatic detection method is required to solve the above problems.

In recent years, computer vision technology is developed at a rapid speed, various technologies represented by deep learning are rapidly applied to the field of industrial detection, for example, chinese patent application with the publication number of CN113177924A discloses an industrial production line product flaw detection method, and when the method is used for detecting product flaws, a batch of data sets are collected in advance, and an algorithm model is obtained after the data sets are trained; therefore, the algorithm can only aim at fixed flaw types after actual deployment, and the algorithm can be invalid when new flaw types appear; if the method needs to adapt to new types of flaws again, new sample data needs to be collected again for model training, so that manpower and material resources are consumed greatly, and the requirements of customers are changed along with production requirements in real situations; this causes the previously deployed algorithms to be less effective or nearly unusable; the algorithm needs to be retrained if the requirements are met again, which is very time consuming. Chinese patent application publication No. CN106872487A, which discloses a vision-based surface flaw detection method, uses a support vector machine to determine whether flaws occur; the method has a great disadvantage that various new types of defects cannot be detected through an algorithm model established by collecting defect samples in advance because the types and the forms of the defects which can appear are various and people cannot exhaust all the types of the defects in actual production.

In order to solve the above problems, the present invention provides a method and a system for detecting abnormal online learning based on feature migration.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the method can conveniently extract some uncertain data samples after actual deployment, return the data samples to field customers, judge the abnormal conditions of the data by the customers, and update the parameters of the network model in real time according to the judgment conditions of the customers, thereby better adapting to the field environment and meeting the judgment requirements of the customers on the abnormal samples; moreover, when a customer's anomaly determination needs are adjusted, the network model may learn of such changes, making adjustments.

The invention solves the technical problems through the following technical scheme, and the invention comprises the following steps:

step 1: collecting sample data

Collecting product sample data in reality to obtain a sample data set;

step 2: feature migration for feature extractor

Migrating sample data collected in a real environment by adopting a residual network model, training and updating parameters of the residual network model on the whole sample data set, and selecting the last layer of a backbone network in the residual network model as a feature extractor;

and 3, step 3: sample data feature normalization

Performing feature extraction on the image sample through the feature extractor in the step 2 to obtain features of corresponding dimensions, and then performing normalization processing on the features;

and 4, step 4: online learning of anomaly detection models

Defining an online learning anomaly detection model, then initializing the whole online learning anomaly detection model by using randomized parameters, taking the characteristics after normalization in the step 3 as the input of the online learning anomaly detection model, carrying out forward propagation on the whole model to obtain the anomaly probability of the sample (namely the anomaly score of the sample, namely the prediction result of the model), feeding back the suspected anomaly sample and the sample falling into the set anomaly probability value range to a client, labeling the suspected anomaly sample and the sample falling into the set anomaly probability value range by the client to obtain the suspected anomaly sample and the label data of the sample falling into the set anomaly probability value range, calculating the loss of the sample according to the label data (namely calculating the error between the label data and the model prediction result by a loss function), calculating the gradient of each parameter by a backward propagation chain rule after obtaining the sample loss, updating the parameters, and continuously judging the subsequent sample data by the online learning anomaly detection model after the parameters are updated.

Further, in the step 1, the sample data includes a large number of normal samples and a small number of abnormal samples, and the types of the abnormal samples are various.

Furthermore, in the step 1, the camera with the set angle is set by the set positioning frame on the factory assembly line, and the product sample data is collected by the camera.

Further, in the step 2, the residual network model includes a backbone network Resnet18 and 1 full connection layer, the normal samples collected in the step 1 are labeled as 0, the abnormal samples are labeled as 1, the loss function is binary cross entropy, the learning rate is 0.01, and 32 samples are trained in a single time.

Furthermore, in the step 2, a training process of the residual network model on the sample data set is a feature migration process, and network parameters of the feature extractor are fixed and invariant.

Furthermore, in the step 3, a first-in first-out queue is defined, where the queue is used to store the feature distribution condition of the most recently set number of normal samples, and the set number of normal samples are collected and placed into the queue in the real-time updating process, and assuming that the mean value of the feature vector dimension of a certain normal sample is m and the variance is r at a certain moment, the normalization calculation formula of the current image sample after passing through the feature extractor is:

f’=(f-m)/r

the vector f is a feature vector which is not normalized, and f' is a feature vector which is calculated and normalized.

Further, in the step 4, the online learning anomaly detection model includes six fully-connected layers and one active layer, and the six fully-connected layers are connected with the one active layer in sequence, where each fully-connected layer parameter is a matrix, and the matrix size is: the input dimension and the output dimension, the activation function of the activation layer uses a sigmoid function, and the sigmoid function is defined as:

for sigmoid functionsIn mapping the prediction result to [0,1]]。

Further, in the step 4, if the abnormal probability is less than 0.25, the sample is a large possible normal sample, the sample is a suspected abnormal sample in the range of 0.25 to 0.5, and the sample with the abnormal probability value greater than 0.5 falls within the range of the set abnormal probability value.

Further, in step 4, assuming that a certain parameter θ is propagated backward to obtain a gradient grad, and the learning rate lr is 0.01, the update formula of the parameter is:

θ’=θ-lr*grad

where θ is the value before updating the parameter, and θ' is the value after updating.

The invention also provides an online learning anomaly detection system based on feature migration, which is used for carrying out anomaly detection work on products by adopting the method and comprises the following steps:

the sample collection module is used for collecting product sample data in reality to obtain a sample data set;

the characteristic migration module is used for migrating sample data collected in a real environment by adopting a residual network model, training and updating parameters of the residual network model on the whole sample data set, and selecting the last layer of a backbone network in the residual network model as a characteristic extractor;

the normalization module is used for extracting the features of the image sample through the feature extractor to obtain the features of corresponding dimensions, and then normalizing the features;

the online learning module is used for defining an online learning abnormity detection model, initializing the whole online learning abnormity detection model by using a randomized parameter, taking the characteristics normalized in the step 3 as the input of the online learning abnormity detection model, carrying out forward propagation on the whole model to obtain the abnormity probability of the sample, feeding back the suspected abnormal sample and the sample falling into the set abnormity probability value range to a client, labeling the suspected abnormal sample and the sample falling into the set abnormity probability value range by the client to obtain the suspected abnormal sample and the label data of the sample falling into the set abnormity probability value range, calculating the loss of the sample according to the label data, calculating the gradient of each parameter by a reverse propagation chain rule after obtaining the sample loss, updating the parameter, and continuously judging the subsequent sample data by the online learning abnormity detection model after updating the parameter;

the control processing module is used for sending instructions to other modules to complete related actions;

the sample collection module, the feature migration module, the normalization module and the online learning module are all electrically connected with the control processing module.

Compared with the prior art, the invention has the following advantages: according to the online learning anomaly detection method based on feature migration, the anomaly detection model is obtained through online learning, and the method can well adapt to the condition that the customer demand is unclear or the demand changes, and meets the actual production requirement; the normal sample feature normalization method based on online learning is provided, so that the sample features can be well expressed, and the detection effect is improved; the abnormal samples and the suspected abnormal samples are fed back to the client, and the client determines whether the samples are really abnormal or not, so that the requirements of the client can be met in time.

Drawings

Fig. 1 is a schematic flow chart of an online learning anomaly detection method based on feature migration in an embodiment of the present invention.

Detailed Description

The following examples are given for the detailed implementation and specific operation of the present invention, but the scope of the present invention is not limited to the following examples.

As shown in fig. 1, the present embodiment provides a technical solution: an online learning anomaly detection method based on feature migration comprises the following steps:

step 1: collecting sample data

Collecting a large amount of product sample data in reality, wherein the sample data comprises a large amount of normal samples and a small amount of abnormal samples, and obtaining a sample data set; collecting various abnormal samples of different types as far as possible, wherein the samples serve as training data of an abnormal detection classifier;

it should be noted that the present invention treats the anomaly detection problem as a classification problem in the field of machine learning. A feature extractor is trained offline in advance to acquire features of an input image, and then an abnormality detection classifier is deployed online and used for judging whether the acquired features are abnormal samples or not.

It should be noted that, in step 1, the acquisition of product sample data is to erect a camera at a fixed position and a fixed angle on a factory assembly line, and the camera can clearly acquire product image data to be detected and transmit an image stream to a processing terminal (i.e., an anomaly detection system).

Step 2: feature migration for feature extractor

In this step, the invention uses a residual network model to migrate the product sample data collected in the real environment. The structure of the residual network model is as follows: backbone network Resnet18 plus 1 full connection layer; and marking the normal sample collected in the last step as 0, and marking the abnormal sample as 1. And training and updating residual error network model parameters on the whole sample data set, wherein the loss function is binary cross entropy, the learning rate is 0.01, and 32 samples are trained at one time.

The training process of the residual network model on the sample data set is a feature migration process, and the last layer of the backbone network is selected as a feature extractor (512-dimensional features). To prevent overfitting, training was stopped by training only 15 rounds on the sample dataset. After training is completed, a feature extractor of the sample image can be obtained, and network parameters of the feature extractor are fixed and unchangeable in the following process.

And step 3: sample data feature normalization

For an acquired image sample, a 512-dimensional feature vector can be obtained by the feature extractor in step 2. The 512-dimensional features may better represent the sample data since the feature extractor has performed a feature migration on the sample data set.

Furthermore, the feature distribution of the normal sample is further described. The invention also normalizes the extracted features. The invention aims to search abnormal samples in a large number of normal samples, and only the characteristic distribution condition of the normal samples is counted, which is different from other methods for counting in the whole data set. In addition, in order to consider the situation of online learning real-time update, a first-in first-out queue is defined, and the queue is only used for storing the normal sample feature distribution situation of the latest stage (the part of feature vectors stored in the queue recently, the quantity is 50); wherein the queue length is 50.

In the real-time updating process, the algorithm collects 50 normal samples and puts the samples into the queue, and the mean value of a certain dimension at a certain moment is assumed to be m, and the variance is assumed to be r; the normalization calculation method after the current image sample passes through the feature extractor is as follows:

f’=(f-m)/r

wherein, the vector f is a feature vector which is not normalized, and f' is a feature vector which is calculated and normalized;

it should be noted that, assuming that the queue length is 50, 50 eigenvectors can be stored, and the shapes are { f1, f2, f3, \8230;, f50}; wherein each feature vector is n-dimensional, and has a form of f1= (x 1, x2, x3 \8230; xn); all the feature vectors in the queue can be averaged in respective dimensions, and the mean value of the feature vectors in the queue can be obtained as m = (m 1, m2, m3, \8230;, mn); the variance of the respective dimensions of the feature vectors in the queue can also be calculated as r = (r 1, r2, r3, \8230;, rn).

Because only normal samples are counted, when an abnormal sample is input, the characteristics of the sample deviate from the distribution of the normal samples; this anomaly is amplified by the normalization calculation.

And 4, step 4: online learning of anomaly detection models

The invention defines an online learning anomaly detection model, and the network structure of the model is as the following table 1:

TABLE 1 Online learning anomaly detection model Structure Table

Wherein, the parameter of the full connection layer is a matrix, and the size of the matrix is as follows: input dimension and output dimension;

the activation function uses a sigmoid function, which is defined as:

the function is used for mapping the prediction result to the interval of [0,1], and the sample is more likely to be abnormal when the probability is higher;

then, initializing the whole online learning anomaly detection model by using the randomized parameters;

and (3) regarding the normalized features (512 dimensions) acquired in the step (3), taking the normalized features as the input of the online learning anomaly detection model, and carrying out forward propagation on the whole model. Obtaining the abnormal probability score of the sample after the last layer of the active layer; under normal conditions, if score is equal to or greater than 0.5, the sample can be determined as an abnormal sample, and if score is less than 0.5, the sample can be determined as a normal sample. Because the problem to be solved by the invention is an abnormal detection task, the actual abnormal sample is far smaller than the normal sample; therefore, the invention herein proposes to set the threshold value to be 0.25, if score is less than 0.25, the sample is more likely to be normal, the interval from 0.25 to 0.5 is a suspected abnormal sample, and more than 0.5 is a more likely abnormal sample, i.e. a sample falling within the range of the set abnormal probability value;

and then feeding back the suspected abnormal samples and the samples falling into the range of the set abnormal probability value to the on-site client, and confirming the samples again. After the customer confirms, the labels of the samples are obtained, wherein the normal label is 0, and the abnormal label is 1; the loss of the sample can be calculated after the label data exists;

the loss function is the bisection cross entropy:

；

wherein L represents the loss function of both; y represents a true tag, wherein normal is 0 and abnormal is 1;

a value representing the prediction of the algorithm model, typically some value between 0 and 1; log is a logarithmic mathematical expression.

After the sample loss is obtained, the gradient of each parameter is calculated through a back propagation chain rule, and if a certain parameter theta is subjected to back propagation to obtain a gradient grad, the learning rate lr is 0.01, then the updating formula of the parameter is as follows:

θ’=θ-lr*grad

where θ is a value before the parameter is updated, and θ' is an updated value.

After learning one sample, the parameters of the model are updated as described above. After the parameters are updated, the online learning anomaly detection model (anomaly detection classifier) can continuously judge the subsequent sample data.

It should be noted that many parameters θ to be learned exist in the online learning anomaly detection model, and after the model is propagated backward each time, each parameter θ has its own gradient grad. When the model parameters are updated, each parameter theta is updated according to the updating formula of the parameters.

The network model learning process:

input sample- > forward propagation- > computation loss- > backward propagation computation gradient- > parameter update.

Each parameter updating is equivalent to one learning, and after the learning is finished, the following samples can be continuously judged, and then the learning process is repeated.

In summary, in the online learning anomaly detection method based on feature migration in the embodiment, the anomaly detection model is obtained through online learning, and this way can be well adapted to the situation that the customer demand is unclear or the demand changes, and meets the actual production needs; the normal sample feature normalization method based on online learning is provided, so that the sample features can be well expressed, and the detection effect is improved; the abnormal samples and the suspected abnormal samples are fed back to the client, and the client determines whether the samples are really abnormal or not, so that the requirements of the client can be met in time, and the method is worthy of popularization and use.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims (8)

1. An online learning anomaly detection method based on feature migration is characterized by comprising the following steps:

step 1: collecting sample data

Collecting product sample data in reality to obtain a sample data set;

step 2: feature migration for feature extractor

Migrating sample data collected in a real environment by adopting a residual network model, training and updating parameters of the residual network model on the whole sample data set, and selecting the last layer of a backbone network in the residual network model as a feature extractor;

and step 3: sample data feature normalization

Performing feature extraction on the image sample through the feature extractor in the step 2 to obtain features of corresponding dimensions, and then performing normalization processing on the features;

in the step 3, a first-in first-out queue is defined, where the queue is used to store the feature distribution condition of the latest set number of normal samples, the set number of normal samples are collected and placed into the queue in the real-time updating process, and assuming that the mean value of the feature vector dimensions of a certain normal sample is m and the variance is r at a certain time, the normalization calculation formula of the current image sample after passing through the feature extractor is:

f’=(f-m)/r

wherein, the vector f is a feature vector which is not normalized, and f' is a feature vector which is calculated and normalized;

and 4, step 4: online learning of anomaly detection models

Defining an online learning anomaly detection model, then initializing the whole online learning anomaly detection model by using randomized parameters, taking the characteristics after normalization in the step 3 as the input of the online learning anomaly detection model, carrying out forward propagation on the whole model to obtain the anomaly probability of the sample, feeding back the suspected anomaly sample and the sample falling into the set anomaly probability value range to a client, labeling the suspected anomaly sample and the sample falling into the set anomaly probability value range by the client to obtain label data of the suspected anomaly sample and the sample falling into the set anomaly probability value range, calculating the loss of the sample according to the label data, calculating the gradient of each parameter by a backward propagation chain rule after the sample loss is obtained, updating the parameters, and continuously judging the subsequent sample data by the online learning anomaly detection model after the parameters are updated.

2. The online learning anomaly detection method based on feature migration according to claim 1, characterized in that: in the step 1, the camera with the set angle is arranged on the positioning frame on the factory assembly line, and the camera is used for collecting product sample data.

3. The online learning anomaly detection method based on feature migration according to claim 1, characterized in that: in the step 2, the residual network model includes a backbone network Resnet18 and 1 full connection layer, the normal samples collected in the step 1 are labeled as 0, the abnormal samples are labeled as 1, the loss function is a dichotomous cross entropy, the learning rate is 0.01, and 32 samples are trained in a single time.

4. The online learning anomaly detection method based on feature migration according to claim 1, characterized in that: in step 2, the training process of the residual error network model on the sample data set is a feature migration process, and the network parameters of the feature extractor are fixed and invariant.

5. The online learning anomaly detection method based on feature migration according to claim 1, characterized in that: in the step 4, the online learning abnormality detection model includes sixThe full-connection layer and an active layer are connected in sequence, six full-connection layers and an active layer are connected in sequence, wherein parameters of each full-connection layer are a matrix, and the size of the matrix is as follows: the input dimension and the output dimension, the activation function of the activation layer uses a sigmoid function, and the sigmoid function is defined as:

sigmoid function is used to map the prediction result to [0,1]。

6. The feature migration-based online learning anomaly detection method according to claim 5, characterized in that: in the step 4, if the abnormal probability is less than 0.25, the sample is a large possible normal sample, the sample is a suspected abnormal sample in the interval of 0.25 to 0.5, and the sample falling within the range of the set abnormal probability value is greater than 0.5.

7. The feature migration-based online learning anomaly detection method according to claim 6, wherein: in step 4, assuming that a certain parameter θ is propagated backward to obtain a gradient grad, and the learning rate lr is 0.01, the update formula of the parameter is:

θ’=θ-lr*grad

where θ is a value before the parameter is updated, and θ' is an updated value.

8. An online learning anomaly detection system based on feature migration, which is used for carrying out anomaly detection work on products by adopting the method according to any one of claims 1 to 7, and comprises the following steps:

the sample collection module is used for collecting product sample data in reality to obtain a sample data set;

the characteristic migration module is used for migrating the sample data collected in the real environment by adopting a residual network model, training and updating parameters of the residual network model on the whole sample data set, and selecting the last layer of a backbone network in the residual network model as a characteristic extractor;

the normalization module is used for extracting the features of the image sample through the feature extractor to obtain the features of corresponding dimensions, and then normalizing the features;

the online learning module is used for defining an online learning abnormity detection model, initializing the whole online learning abnormity detection model by using a randomized parameter, taking the characteristics normalized in the step 3 as the input of the online learning abnormity detection model, carrying out forward propagation on the whole model to obtain the abnormity probability of the sample, feeding back the suspected abnormal sample and the sample falling into the set abnormity probability value range to a client, labeling the suspected abnormal sample and the sample falling into the set abnormity probability value range by the client to obtain the suspected abnormal sample and the label data of the sample falling into the set abnormity probability value range, calculating the loss of the sample according to the label data, calculating the gradient of each parameter by a reverse propagation chain rule after obtaining the sample loss, updating the parameter, and continuously judging the subsequent sample data by the online learning abnormity detection model after updating the parameter;

the control processing module is used for sending instructions to other modules to complete related actions;

the sample collection module, the feature migration module, the normalization module and the online learning module are all electrically connected with the control processing module.

Images