WO2022052510A1

WO2022052510A1 - Anomaly detection system and method for sterile filling production line

Info

Publication number: WO2022052510A1
Application number: PCT/CN2021/095632
Authority: WO
Inventors: 彭力; 李贝贝; 李稳; 朱凤增; 张连富; 何子琎
Original assignee: 江南大学
Priority date: 2020-09-09
Filing date: 2021-05-25
Publication date: 2022-03-17
Also published as: CN112015153A; CN112015153B

Abstract

Provided are an anomaly detection system and method for a sterile filling production line. The detection system comprises a front-end device for collecting and uploading information; a front-end device management platform for receiving the information, realizing preliminary filtering and processing of the information, and eliminating problem information; and an information service platform for determining the information by means of an autoencoder network, and finding anomalous information. The detection method comprises Internet of Things data collection (S1), data pre-processing (S2), autoencoder network construction (S4), and final differentiation of normal data and anomalous data (S5). The accuracy, stability and robustness of the system are good, and a more excellent fault early-warning effect is achieved.

Description

A system and method for abnormality detection of aseptic filling production line

technical field

The invention relates to a detection system and method, in particular to a system and method for abnormality detection of an aseptic filling production line.

Background technique

Abnormal detection in the production process of aseptic filling production line has always been an important task in the working of aseptic filling production line. With the rapid development of Internet of Things technology and the improvement of various sensor technologies, more and more accurate data can be collected in real time on aseptic filling production equipment. Anomaly detection through various data on the production line has always been an important research and application. direction. Various fault detection and prediction methods are mostly based on the data collected on aseptic filling production equipment through the Internet of Things technology, and how to use these data to improve the accuracy of abnormality detection has become an important research topic.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a system and method for abnormality detection of aseptic filling production line, which combines neural network and Internet of Things data, greatly improves the accuracy of abnormality detection, and has good stability and robustness. A more excellent fault warning effect is achieved.

In order to solve the above technical problems, the present invention provides an abnormality detection system for an aseptic filling production line, including front-end equipment for collecting and uploading information; Front-end equipment management platform; an information service platform for judging information travel through the self-coding network and finding abnormal information.

In a preferred embodiment of the present invention, the front-end equipment further includes a plurality of sensors distributed on the aseptic filling production line and set for the production environment.

In a preferred embodiment of the present invention, it further includes that the sensors are connected in series in the form of an RS485 bus.

In a preferred embodiment of the present invention, the front-end equipment management platform further includes a sensor information processing module, a sensor association module, a sensor fault self-diagnosis module and a network communication module; the sensor information processing module is used to help the sensor to follow the correct Frequency acquisition data, filter the data and eliminate blank data; the sensor association module is used to integrate the data collected by each sensor, add time and space tags to it, and form the same set of data; the sensor fault self-diagnosis module is used for When the sensor itself fails, an alarm signal is generated, and the information service platform is notified to give an alarm; the network communication module is used for sending data to the information service platform.

In a preferred embodiment of the present invention, it further includes enterprise customers for displaying data collected by sensors, querying and displaying historical data, viewing sensor operating status, viewing the operating status of various functions of the aseptic filling production line, and alerting abnormal data. end.

In a preferred embodiment of the present invention, it further includes that the enterprise client is a web display page, the information service platform is deployed on a cloud server, and the web display page is communicatively connected to the information service platform.

A method for detecting abnormality in an aseptic filling production line, comprising the following steps:

S1, collect IoT data through front-end equipment;

S2, preprocess the collected historical data, divide the collected historical data into normal data and obvious abnormal data according to the actual production line operation status, and complete normalization and feature construction for the normal data;

S3, the judgment threshold of the Mahalanobis distance is obtained from normal data. If the Mahalanobis distance is not within the threshold interval, the data is determined to be abnormal data; if the Mahalanobis distance is within the threshold interval, the data is added with its Mahalanobis distance feature and recorded as Uncertain data is input to the auto-encoder, and the sparsity limit is added for unsupervised training, and the parameters and dimensions of each layer are adjusted to obtain the optimal hidden layer expression;

S4, combine the optimal hidden layer expression with the Sigmoid classifier to construct an auto-encoding network, and perform supervised fine-tuning after inputting the labeled data into the self-encoding network to obtain the optimal parameters and complete the construction of the self-encoding network;

S5, for the uncertain data whose Mahalanobis distance is within the threshold interval, input it into the constructed auto-encoding network, and the auto-encoding network determines whether it is normal data or abnormal data.

In a preferred embodiment of the present invention, in step S3, the method for obtaining the discrimination threshold of the Mahalanobis distance is:

Calculate the Mahalanobis distance of n data sets X=(X ₁ , X ₂ , X ₃ ,..., X _n ) with each data dimension m, where the mean is μ=(μ ₁ , μ ₂ , μ ₃ ,...,μ _m ) ^T , and the covariance matrix is Σ, then for any data x=(x ₁ ,x ₂ ,x ₃ ,...,x _m ) ^T , the Mahalanobis distance is as follows shown:

where Σ ^-1 is the inverse of the covariance matrix;

For the collected normal data set X _N =(X ₁ ,X ₂ ,X ₃ ,...,X _l ), the normal data mean is μ _N =(μ ₁ ,μ ₂ ,μ ₃ ,..., μ _m ) ^T , the covariance matrix is Σ _N , and the Mahalanobis distance of each data in the normal data set calculated according to the above formula is recorded as:

M _N =(M ₁ ,M ₂ ,...M _q ,...,M _l )

Among them, M _q represents the Mahalanobis distance of the qth normal data in the normal data set;

The mean and covariance matrix used in calculating the Mahalanobis distance of all data are still μ _N and Σ _N in the normal data set, then the _Mahalanobis distance of the data Xi is recorded as:

_Denote the mean of the dataset MN as

Standard deviation is recorded as

According to the 3σ criterion in statistics, the Mahalanobis distance of most normal data is distributed in

In the interval, the Mahalanobis distance of some abnormal data is not in the above interval, so the initial detection of abnormal data can be carried out through the Mahalanobis distance of the data, and the Mahalanobis distance discrimination thresholds T _up and T _low are set, and the Mahalanobis distance is not in ( The data in the interval T _low , T _up ) is determined as abnormal data, and the expressions of T _up and T _low are as follows:

In a preferred embodiment of the present invention, it further includes that in step S3, the working process of the self-encoder includes an encoding process and a decoding process. For the data set X=(X ₁ , X ₂ , X ₃ ,...,X _n ) , n is the number of data, the dimension of each data is m, and each data X _i is expressed in the hidden layer through the encoding process. The encoding process can be described as:

h _i =σ _e (WX _i +b)

Among them, W and b are the coding weights and biases, and σ _e is the activation function of the coding layer;

The hidden layer expresses the reconstructed data X _i ' through the decoding process, and the decoding process can be described as:

X _i '=σ _d ( _W'hi +b')

Among them, W' and b' are decoding weights and biases, W'=W ^T , and σ _d is the activation function of the decoding layer.

In a preferred embodiment of the present invention, the method for obtaining the hidden layer expression in step S3 is further included:

First, the weights and biases are adjusted by a layer-by-layer greedy algorithm to minimize the reconstruction error. The cost function for the entire training dataset is:

Among them, L is the loss function of a single data, and L is the mean square error loss function;

Then, an L2 regularization weight decay term is added to the cost function, λ is the penalty factor, KL divergence is added as a constraint, and a sparse penalty term is added to the cost function to form a sparse autoencoder, and the final loss function is obtained. for:

in,

is the regularization term,

is the constraint condition of KL divergence, and k is the number of neurons in the hidden layer;

Finally, after identifying some abnormal data through Mahalanobis distance, the uncertain data is used as the input layer data of the autoencoder for training, so that the value of the formula loss function is the smallest, and then the best hidden layer expression is obtained.

Beneficial effects of the present invention:

In the abnormal detection system of the aseptic filling production line of the present invention, the front-end equipment can collect the production environment information of each node of the aseptic filling production line, and upload it to the front-end equipment management platform. , screening, and finally find out the abnormal information by the self-encoding network constructed in the information service platform. As a special neural network structure that makes the output equal to the input, the auto-encoding network can reconstruct the data in the process of image processing, anomaly detection, fault prediction, data classification, etc., and perform feature extraction to obtain the hidden layer. It has the characteristics of better representative data when it is expressed, which can greatly improve the accuracy, stability and robustness of abnormal detection of aseptic filling production line, and achieve more excellent fault warning effect. This method solves the linkage problem between neural network and IoT data. First, a sparse auto-encoder is constructed to obtain a better hidden layer expression, and then combined with the Sigmoid classifier, an auto-encoder network is constructed for supervised fine-tuning, and a complete anomaly detection model is obtained. . After inputting the data into the self-encoding network anomaly detection model, accurate detection results are obtained. Compared with the traditional anomaly detection method, this method combines the self-encoding network and the Internet of Things data for intelligent analysis and detection, which can achieve the characteristics of strong comprehensiveness, high accuracy, and strong robustness.

Description of drawings

1 is a schematic structural diagram of an abnormality detection system for an aseptic filling production line in a preferred embodiment of the present invention;

Fig. 2 is the schematic flow chart of the abnormality detection method of aseptic filling production line;

3 is a schematic structural diagram of an autoencoder;

Fig. 4 is the structural representation of self-encoding network;

Figure 5 is a schematic diagram of the fine-tuning of the autoencoder network.

detailed description

The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

Example

The invention discloses an abnormality detection system for an aseptic filling production line, comprising front-end equipment for collecting and uploading information; a front-end equipment management platform for receiving information, realizing preliminary filtering and processing of information, and eliminating problem information; An information service platform that judges the information travel through the self-encoding network and finds out abnormal information. With the above optimized structure, the front-end equipment can collect the production environment information of each node of the aseptic filling production line and upload it to the front-end equipment management platform. The self-encoding network built in the platform finds out abnormal information. As a special neural network structure that makes the output equal to the input, the auto-encoding network can reconstruct the data in the process of image processing, anomaly detection, fault prediction, data classification, etc., and perform feature extraction to obtain the hidden layer. It has the characteristics of better representative data when it is expressed, which can greatly improve the accuracy, stability and robustness of abnormal detection of aseptic filling production line, and achieve more excellent fault warning effect.

Specifically, as shown in FIG. 1 , the above-mentioned front-end equipment includes several sensors distributed on the aseptic filling production line and set for the production environment. The sensors include at least one or more of vibration sensors, voltage and current sensors, temperature sensors, humidity sensors and noise sensors. For a normally operating aseptic filling production line, its production environment is stable, that is, vibration, voltage, current, temperature, humidity, noise and other data are kept within a normal range, and once the production environment is abnormal, these data must also be Corresponding changes will occur.

In the aseptic filling production workshop, there are many factors that affect the network communication. In order to reduce the influence of these factors, the above sensors are connected in series in the form of RS485 bus.

The front-end device management platform is connected in communication with the front-end device. The front-end equipment management platform includes a sensor information processing module, a sensor association module, a sensor fault self-diagnosis module and a network communication module. The sensor information processing module can help the sensor to collect data according to the correct frequency, filter the data and eliminate blank data, and enhance the anti-interference ability. The sensor association module can integrate the data collected by each sensor, add time and space tags to it, and form the same set of data. The sensor fault self-diagnosis module can generate an alarm signal when the sensor itself fails, and notify the information service platform to give an alarm. The network communication module can send data to the information service platform.

In order to reduce the impact of the production workshop on communication, the network communication module can transmit data to the information service platform through wired Ethernet data transmission.

The above-mentioned information service platform is in communication connection with the front-end equipment management platform. The information service platform includes an IoT data processing module, an abnormal data detection module and an abnormal alarm module. The IoT data processing module can receive and process the data sent by the front end, providing a data basis for the abnormal data detection module. The abnormal data detection module can judge the data through the self-encoding network model, and realize the information linkage abnormal detection function. The abnormal alarm module can issue an alarm when the abnormal data detection module determines abnormal data. The information service platform is mainly responsible for processing the IoT data collected by the sensors, and detecting abnormal data through the IoT data combined with the self-encoding network.

In order to facilitate the access of the information service platform, it can be deployed on the cloud server, which can realize access anytime and anywhere, and greatly enhance the convenience of the system.

In some preferred embodiments of the present invention, an abnormality detection system for an aseptic filling production line using self-encoding network and Internet of Things data further includes an enterprise client. The enterprise client communicates with the information service platform. The enterprise client is used to display the data collected by the sensor, query and display historical data, check the operation status of the sensor, check the operation status of each function of the aseptic filling production line, and alarm abnormal data. Once the information service platform detects abnormal data, it will send the alarm data to the enterprise client, so that the enterprise can know the abnormality in time and observe the data in real time.

Enterprise clients can display pages for the web. It communicates and connects with the information service platform on the cloud server. Using web components, functions such as data display and alarm are realized. At the same time, the web page does not require complicated installation and debugging, which further improves the convenience and intuitiveness of the system. In addition, the use of web pages for data display and alarm improves the robustness of the entire system, and is flexible and convenient to use, reflecting a user-friendly Internet thinking, which is the general trend of production line anomaly detection systems in the Internet era.

The enterprise client can be composed of data display interface and alarm mechanism.

The invention discloses a method for detecting abnormality of an aseptic filling production line, comprising the following steps:

S1, collect IoT data through front-end equipment. In order to collect more abnormal data and build a better self-encoding network, it is possible to collect data for a long time, or even artificially cause abnormality for a period of time.

S2: Preprocess the collected historical data, divide the collected historical data into normal data and obvious abnormal data according to the actual production line operation state, and complete normalization and feature construction for the normal data. Obvious outliers include blank data and obviously wrong data.

S3, the judgment threshold of the Mahalanobis distance is obtained from normal data. If the Mahalanobis distance is not within the threshold interval, the data is determined to be abnormal data; if the Mahalanobis distance is within the threshold interval, the data is added with its Mahalanobis distance feature and recorded as The uncertain data is input to the autoencoder, and the sparsity restriction is added for unsupervised training, and the parameters and the dimensions of each layer are adjusted to obtain the optimal hidden layer expression.

S4, the optimal hidden layer expression is combined with the Sigmoid classifier to construct an auto-encoding network. After inputting the labeled data into the auto-encoding network, supervised fine-tuning is performed to obtain the optimal parameters, and the construction of the self-encoding network is completed. It solves the problem of neural network parameter initialization, shortens the training times of the classifier, and improves the accuracy of anomaly detection.

S5, for the uncertain data whose Mahalanobis distance is within the threshold interval, input it into the constructed self-encoding network, and the self-encoding network determines whether it is normal data or abnormal data.

This method solves the linkage problem between neural networks and IoT data. First, a sparse auto-encoder is constructed to obtain a better hidden layer expression, and then combined with the Sigmoid classifier, an auto-encoder network is constructed for supervised fine-tuning, and a complete anomaly detection model is obtained. . After inputting the data into the self-encoding network anomaly detection model, accurate detection results are obtained. Compared with the traditional anomaly detection method, this method combines the self-encoding network and the Internet of Things data for intelligent analysis and detection, which can achieve the characteristics of strong comprehensiveness, high accuracy, and strong robustness.

Specifically, referring to Fig. 2, when combining IoT data with an auto-encoding network, considering the generalized distance of the correlation between variables, the Mahalanobis distance can be represented by the covariance matrix between vectors.

For a _dataset X ₌ ₍ X ₁ , X ₂ , X ₃ , _. ...,μ _m ) ^T , the covariance matrix is Σ, and any data is x=(x ₁ ,x ₂ ,x ₃ ,...,x _m ) ^T , then its Mahalanobis distance is as follows:

Among them, Σ ^-1 is the inverse matrix of the covariance matrix, and the Mahalanobis distance can be regarded as the distance between the data and the mean of the overall data.

Since the calculation of Mahalanobis distance needs to use the covariance matrix of the data set, the biggest advantage of other distances such as Euclidean distance is that the Mahalanobis distance considers the correlation between data features. In a data set, if the Mahalanobis distance of a data is smaller, it means that it is more similar to the mean data in the data set. In the collected aseptic filling production line data, due to the technological process and collection equipment, there is a non-negligible correlation between each feature of the data, and the Mahalanobis distance is more suitable for the distance of the aseptic filling production line data. Express.

Considering that in the data anomaly detection, assuming that a certain data is normal data, use the mean and covariance matrix of the normal data set to calculate its Mahalanobis distance according to formula (1). If the Mahalanobis distance of the data is close to the Mahalanobis distance of the normal data, it means that the data is more similar to the normal data, and the data is probably normal data; if the Mahalanobis distance of the data is far from the Mahalanobis distance of the normal data, it means If the similarity between the data and normal data is small, the data is likely to be abnormal data. Therefore, the Mahalanobis distance of the data can be used to judge the possibility that the data is abnormal data.

Due to the huge amount of data, in the training process of the autoencoder and neural network, the excessively large amount of data makes the training efficiency low and the detection accuracy cannot be effectively improved. In order to reduce the amount of training data for the autoencoder and neural network, the judgment threshold can be obtained from the Mahalanobis distance of the normal data samples, the Mahalanobis distance between the data and the average value of the normal data is calculated, and the data that does not exceed the judgment threshold is recorded as uncertain data and This part of the data is used for the training of autoencoders and neural networks.

It is assumed that the collected data set is denoted as X=(X ₁ , X ₂ , X ₃ , . . . , X _n ), and the data dimension is m. The normal data set is denoted as X _N =(X ₁ ,X ₂ ,X ₃ ,...,X _l ), and the normal data mean is μ _N =(μ ₁ ,μ ₂ ,μ ₃ ,...,μ _m ) ^T , the covariance matrix is Σ _N , and the Mahalanobis distance of each data in the normal data set calculated according to formula (1) is recorded as:

M _N =(M ₁ ,M ₂ ,...M _q ,...,M _l ) (2)

Among them, M _q represents the Mahalanobis distance of the qth normal data in the normal data set.

Next, the Mahalanobis distance of all data in the entire data set is calculated according to formula (1). Since the change of a certain data will affect the change of the mean value of the data set, the Mahalanobis distance exaggerates the effect of the small change vector, thus affecting the calculation of the Mahalanobis distance of other data. In order to improve the shortcomings of the above Mahalanobis distance, the mean and covariance matrix used in calculating the Mahalanobis distance of all data are still μ _N and Σ _N in the normal data set, obviously independent μ _N and Σ _N are not affected by vector changes ; The Mahalanobis distance of a certain data can be regarded as the distance between the data and the mean of the normal data set. Then the _Mahalanobis distance of the data Xi is written as:

According to relevant statistical knowledge and subsequent experimental analysis of vibration data, it can be known that if the data X _i is normal data, its Mahalanobis distance D _M ( _X _i ) should conform to the statistical distribution of the normal data Mahalanobis distance data set, namely MN; If X _i is abnormal data, then D _M ( _X _i ) does not conform to the statistical distribution of MN.

_Denote the mean of the dataset MN as

Standard deviation is recorded as

In the interval, the Mahalanobis distance of some abnormal data is not within the above interval. Therefore, the initial detection of abnormal data can be carried out through the Mahalanobis distance of the data. Set the Mahalanobis distance discrimination thresholds T _up and T _low , and determine the data whose Mahalanobis distance is not in the interval (T _low , T _up ) as abnormal data. Combined with actual data analysis, in order to ensure the accuracy of abnormal detection, T _up and T up The expression for T _low is as follows:

The data whose Mahalanobis distance is in the interval of (T _low , T _up ) is determined as uncertain data, which is denoted as X _U , and then the self-encoding network is trained with the uncertain data set X _U to complete the model construction.

Therefore, a part of abnormal data is detected by Mahalanobis distance, and the remaining uncertain data is input into the self-encoding network. For the aseptic filling production line data with a large amount of data, the training data for the self-encoding network is reduced, and a part of abnormal data can be quickly identified according to the discrimination threshold of the Mahalanobis distance, which improves the detection efficiency.

In addition, the Mahalanobis distance of the aseptic filling production line data can be used to judge the possibility that the data is abnormal data, so consider the Mahalanobis distance of the data as a feature of the data. Since the Mahalanobis distance between normal data and abnormal data is quite different, in the training process of the neural network, adding the Mahalanobis distance feature as an important feature is beneficial to improve the abnormal detection effect of the neural network. Therefore, the Mahalanobis distance feature of the data is added to the data feature for the training of the self-encoding network.

Referring to FIG. 3 , the autoencoder mainly includes encoding and decoding stages, and the structure is symmetrical, that is, if there are multiple hidden layers, the number and structure of hidden layers in the encoding and decoding stages are the same. The main structure consists of input layer, hidden layer and output layer. The hidden layer encodes the input layer data, and the output layer decodes the hidden layer expression to reconstruct the original data, minimizing the reconstruction error to obtain the best hidden layer expression. The goal is to fit an identity function such that each output value is as equal as possible to the corresponding input value.

For data set X=(X ₁ , X ₂ , X ₃ ,...,X _n ), n is the number of data, and the dimension of each data is m. Each data X _i is expressed in the hidden layer through the encoding process, and the encoding process can be described as:

h _i =σ _e (WX _i +b) (6)

Among them, W and b are coding weights and biases, and σ _e is the activation function of the coding layer, which can be Sigmoid, Tanh, Relu, etc. Then the hidden layer expresses the reconstructed data X _i ' through the decoding process, and the decoding process can be described as:

X _i '=σ _d ( _W'hi +b') (7)

Among them, W' and b' are decoding weights and biases, W'=W ^T , and σ _d is the activation function of the decoding layer. The weights and biases are adjusted by a layer-by-layer greedy algorithm to minimize the reconstruction error. The cost function for the entire training data set is:

Among them, L is the loss function of a single data, and in Equation (8), L is the mean square error loss function.

In order to prevent over-fitting, an L2 regularization weight decay term is added to the cost function, and λ is a penalty factor to control the degree to which the regularization term affects the weight decay. In the example, KL divergence is added as a constraint on the basis of the autoencoder, and a sparse penalty term is added to its cost function to form a sparse autoencoder. The final loss function is obtained as:

in,

is the regularization term,

is the constraint condition of KL divergence, and k is the number of neurons in the hidden layer.

After some abnormal data are identified by the Mahalanobis distance from the IoT data collected by the sensor, the uncertain data is used as the input layer data of the autoencoder for training, so that equation (9) is minimized and the best hidden layer expression is obtained.

Referring to Figure 4, next, by using the obtained hidden layer as the input layer of the Sigmoid classifier, an auto-encoding network is constructed, and the data is judged as normal data and abnormal data. In this way, the expression of the hidden layer is obtained through the training of the autoencoder, combined with the Sigmoid classifier, the anomaly detection accuracy can be improved, and since the parameters of the hidden layer have been obtained through the training of the autoencoder, the parameter initialization problem of the neural network is solved, so that the automatic The number of training times of the encoding network is reduced, and the construction efficiency of the model is improved.

Referring to Figure 5, labeled data is then required to be input for the training process of the self-encoding network for supervised fine-tuning. During fine-tuning, the error is back-propagated into the hidden layer and the Sigmoid classifier. Since the autoencoder solves the problem of parameter initialization of the classifier, the training times of the autoencoder network are reduced and the training efficiency is improved.

Finally, the construction of the self-encoding network anomaly detection model is completed. The main steps are summarized as follows:

S31, calculate the Mahalanobis distance of the normal data, and obtain the Mahalanobis distance discrimination threshold according to formula (4) and formula (5).

S32, obtain the Mahalanobis distance of all the data according to formula (3), determine the data whose Mahalanobis distance exceeds the threshold as abnormal data, and record the data that does not exceed the threshold as uncertain data, and add the Mahalanobis distance of the data to the data feature .

S33, take the uncertain data set added with Mahalanobis distance feature as the input of the autoencoder, train the autoencoder in an unsupervised manner, and obtain the optimal output and parameters of the encoding layer.

S34, take the self-encoder coding layer as the input layer of the Sigmoid classifier, and use the labeled uncertain data as the input to perform supervised fine-tuning to obtain the optimal parameters of the entire network, and complete the construction of the self-encoding network.

S35, for the data whose Mahalanobis distance is within the threshold interval, put it into the trained self-encoding network to obtain a judgment result.

The above-mentioned embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims

An aseptic filling production line anomaly detection system is characterized in that: it includes front-end equipment for collecting and uploading information; a front-end equipment management platform for receiving information, realizing preliminary filtering and processing of information, and eliminating problem information; It is an information service platform that judges and finds abnormal information through the self-encoding network.
The abnormality detection system for an aseptic filling production line according to claim 1, wherein the front-end equipment includes a plurality of sensors distributed on the aseptic filling production line and set for the production environment.
The abnormality detection system and method of the aseptic filling production line according to claim 2, wherein the sensors are connected in series in the form of an RS485 bus.
The abnormality detection system for an aseptic filling production line according to claim 1, wherein the front-end equipment management platform comprises a sensor information processing module, a sensor correlation module, a sensor fault self-diagnosis module and a network communication module; the sensor information The processing module is used to help the sensor collect data according to the correct frequency, filter the data and remove blank data; the sensor association module is used to integrate the data collected by each sensor, add time and space tags to it, and form the same set of data The sensor fault self-diagnosis module is used to generate an alarm signal when the sensor itself fails, and notify the information service platform to give an alarm; the network communication module is used to send data to the information service platform.
The abnormality detection system of the aseptic filling production line according to claim 1, characterized in that it includes the functions for displaying the data collected by the sensor, querying and displaying historical data, checking the operation status of the sensor, and checking the operation status of each function of the aseptic filling production line. And enterprise clients that alert on abnormal data.
The abnormality detection system for an aseptic filling production line according to claim 5, wherein the enterprise client is a web display page, the information service platform is deployed on a cloud server, and the web display page and information service Platform communication connection.
A method for detecting abnormality in an aseptic filling production line, comprising the following steps:

S1, collect IoT data through front-end equipment;

S2, preprocess the collected historical data, divide the collected historical data into normal data and obvious abnormal data according to the actual production line operation status, and complete normalization and feature construction for the normal data;

S3, the judgment threshold of the Mahalanobis distance is obtained from normal data. If the Mahalanobis distance is not within the threshold interval, the data is determined to be abnormal data; if the Mahalanobis distance is within the threshold interval, the data is added with its Mahalanobis distance feature and recorded as Uncertain data is input to the auto-encoder, and the sparsity limit is added for unsupervised training, and the parameters and dimensions of each layer are adjusted to obtain the optimal hidden layer expression;

S4, combine the optimal hidden layer expression with the Sigmoid classifier to construct an auto-encoding network, and perform supervised fine-tuning after inputting the labeled data into the self-encoding network to obtain the optimal parameters and complete the construction of the self-encoding network;

S5, for the uncertain data whose Mahalanobis distance is within the threshold interval, input it into the constructed auto-encoding network, and the auto-encoding network determines whether it is normal data or abnormal data.
The method for detecting abnormality of aseptic filling production line as claimed in claim 7, characterized in that: in step S3, the method for obtaining the discrimination threshold of the Mahalanobis distance is:

Calculate the Mahalanobis distance of n data sets X=(X 1 , X 2 , X 3 ,..., X n ) with each data dimension m, where the mean is μ=(μ 1 , μ 2 , μ 3 ,...,μ m ) T , and the covariance matrix is Σ, then for any data x=(x 1 ,x 2 ,x 3 ,...,x m ) T , the Mahalanobis distance is as follows shown:

where Σ -1 is the inverse of the covariance matrix;

For the collected normal data set X N =(X 1 ,X 2 ,X 3 ,...,X l ), the normal data mean is μ N =(μ 1 ,μ 2 ,μ 3 ,..., μ m ) T , the covariance matrix is Σ N , and the Mahalanobis distance of each data in the normal data set is calculated according to the above formula, which is recorded as:

M N =(M 1 ,M 2 ,...M q ,...,M l )

Among them, M q represents the Mahalanobis distance of the qth normal data in the normal data set;

The mean and covariance matrix used in calculating the Mahalanobis distance of all data are still μ N and Σ N in the normal data set, then the Mahalanobis distance of the data Xi is recorded as:

Denote the mean of the dataset MN as
Standard deviation is recorded as
According to the 3σ criterion in statistics, the Mahalanobis distance of most normal data is distributed in
In the interval, the Mahalanobis distance of some abnormal data is not in the above interval, so the initial detection of abnormal data can be carried out through the Mahalanobis distance of the data, and the Mahalanobis distance discrimination thresholds T up and T low are set, and the Mahalanobis distance is not in ( The data in the interval T low , T up ) is determined as abnormal data, and the expressions of T up and T low are as follows:
The abnormality detection method for an aseptic filling production line according to claim 7, characterized in that: in step S3, the working process of the self-encoder includes an encoding process and a decoding process, and for the data set X=(X 1 , X 2 , X 3 ,...,X n ), n is the number of data, the dimension of each data is m, and each data X i is expressed in the hidden layer through the encoding process, and the encoding process can be described as:

h i =σ e (WX i +b)

Among them, W and b are the coding weights and biases, and σ e is the activation function of the coding layer;

The hidden layer expresses the reconstructed data X′ i through the decoding process, and the decoding process can be described as:

X' i =σ d (W'h i +b')

Among them, W' and b' are decoding weights and biases, W'=W T , and σ d is the activation function of the decoding layer.
The abnormality detection method of aseptic filling production line as claimed in claim 9, wherein the method for obtaining the hidden layer expression in step S3 is:

First, the weights and biases are adjusted by a layer-by-layer greedy algorithm to minimize the reconstruction error. The cost function for the entire training dataset is:

Among them, L is the loss function of a single data, and L is the mean square error loss function;

Then, add an L2 regularization weight decay term to the cost function, λ is a penalty factor, add KL divergence as a constraint, add a sparse penalty term to its cost function to form a sparse autoencoder, and get the final loss function for:

in,
is the regularization term,
is the constraint condition of KL divergence, and k is the number of neurons in the hidden layer;

Finally, after identifying some abnormal data through Mahalanobis distance, the uncertain data is used as the input layer data of the autoencoder for training, so that the value of the formula loss function is the smallest, and then the best hidden layer expression is obtained.