CN109840900B - A fault online detection system and detection method applied to intelligent manufacturing workshops - Google Patents

Abstract

The invention relates to a fault online detection system and a fault online detection method applied to an intelligent manufacturing workshop, which construct a manufacturing defect prediction model based on a deep neural network, effectively utilize a large number of typical manufacturing defects obtained in the manufacturing process of the actual intelligent manufacturing workshop, and perform training and learning on a standard sample image library formed by typical manufacturing defect images by combining image acquisition and image processing technology, so that the deep neural network model can be used for identifying and classifying the manufacturing defects in real time, dynamically obtain manufacturing information of the product through real-time image acquisition, processing and analysis in the manufacturing process of the product, and provide and automatically execute maintenance strategies of the manufacturing defects by combining a PLC (programmable logic controller) and a historical maintenance database, and can provide more accurate reference information for product manufacturing precision and manufacturing information collection and analysis of the intelligent manufacturing workshop.

Description

A fault online detection system and detection method applied to intelligent manufacturing workshops

Technical field

The invention belongs to the field of intelligent manufacturing technology, and specifically relates to an online fault detection system and method applied in an intelligent manufacturing workshop.

Background technique

The current intelligent manufacturing workshops use many high-speed and high-precision CNC machine tools, the allowable errors of processed and manufactured products are small, and the intelligent manufacturing workshops have high requirements for the comprehensive quality of personnel and the environment, making the products in the manufacturing workshops Manufacturing defects in the manufacturing process are sometimes difficult to detect manually in time and can usually only be discovered through manual inspection in subsequent processes. This causes problems such as an increase in waste products and a decrease in productivity in intelligent manufacturing. In severe cases, it may even cause subsequent manufacturing of machine tools. Scrapped, which has already occurred in the intelligent manufacturing assembly area. Therefore, dynamic manufacturing defect inspection and diagnosis of products in the manufacturing workshop during the manufacturing process is one of the hindering factors affecting the further development of intelligent manufacturing. It has also been Extensive research by scholars in the industry.

Contents of the invention

The purpose of the present invention is to provide an online fault detection system and detection method applied in intelligent manufacturing workshops.

According to the above object of the present invention, an online fault detection system is proposed, including a workpiece detection platform, an image acquisition unit, an image processing unit, a feature vector extraction unit, a deep neural network unit and a computer control unit, wherein,

The workpiece detection platform includes a detection station, and the area array camera in the image acquisition unit collects images of the workpiece detection platform on the detection station and sends them to the image processing unit;

The image processing unit performs a resolution scan on the received image to obtain the sensitive area image of the current detection station, denoises the sensitive area image, and then sends the denoised sensitive area image to the feature vector extraction unit;

The feature vector extraction unit performs edge detection on the sensitive area image to form a target area, and calculates the edge area, edge shape factor and average radius of the target area through formulas (1) to (3) respectively, plus the first 3 dimensions The Hu invariant moments form a feature vector of the sensitive area with four feature variables to reflect the workpiece quality information of the current workpiece detection platform. The feature vector is sent to the deep neural network unit as an input layer;

In the above formula, the parameters M and N are the number of edge points in the target area, Among them, t(x,y) is the gray value of each edge point; the parameter L is the perimeter of the target area, which can be calculated using the chain code method in image processing technology. The reference K is the number of edge points on the boundary of the target area. Number, (x _k ,y _k ) represents the pixel coordinates located on the boundary of the target area,/> Represents the centroid coordinates of the target area, which can be calculated by the following formula:

Among them, parameter A represents the area of the sensitive area, and is suitable for obtaining its size when a sensitive area is identified in image processing;

The deep neural network unit builds a manufacturing defect prediction model based on the neural network algorithm, trains, learns and classifies the image feature vectors of the workpiece detection platform, identifies the manufacturing defect type of the workpiece to be tested on the current workpiece detection platform, and feeds the classification results back to Computer control unit.

On the other hand, the present invention also provides an online fault detection method applied in intelligent manufacturing workshops, including:

Step 1: Construct a manufacturing defect prediction model based on a deep neural network, and train the deep neural network through sample images;

Step 2: Under the control instructions of the computer control module, the workpiece detection platform drives the workpiece to be tested to move on the detection station to reach the preset detection position, triggers the proximity switch, and sends the trigger signal to the computer control unit;

Step 3: The computer control unit sends instructions to the LED surface light source matrix and the image acquisition unit respectively. The LED surface light source matrix turns on the illumination. The image acquisition unit photographs the workpiece to be tested and sends the generated image to the image processing unit;

Step 4: The image processing unit identifies and segments the sensitive area of the current detection station, performs image denoising on the local area, and sends the image of the sensitive area to the feature vector extraction unit after denoising;

Step 5: The feature vector extraction unit performs edge detection on the sensitive area to form a target area, and calculates the edge area, edge shape factor and average radius of the target area. Combined with the Hu invariant moments of the first three dimensions, it forms a four-dimensional Feature vector of the sensitive area of the feature variable;

Step 6: Diagnose the manufacturing information based on the trained deep neural network on the feature vector, predict and classify the manufacturing defects of the workpiece to be tested, and feed the classification results back to the computer control unit.

The beneficial effects of the present invention are:

(1) For the first time, this invention applies image acquisition and image processing technology to the identification and diagnosis of product manufacturing defect information in an intelligent manufacturing workshop. Through a manufacturing defect prediction model built based on a deep neural network, real-time image acquisition of products in the intelligent manufacturing workshop is performed. , processing and feature extraction to obtain the feature vector representation of product manufacturing information, thereby realizing that product manufacturing information can be displayed, defects can be diagnosed, and maintenance strategies can be automatically generated in the intelligent manufacturing process, forming real-time monitoring of product manufacturing and defect conclusions. An organic integrated system with real-time analysis and automatic implementation of maintenance strategies, thus effectively solving one of the existing factors that restrict the widespread application of intelligent manufacturing technology;

(2) In order to obtain accurate image processing results, this invention effectively solves the tiny vibration bands that may occur in machine tools and inspection stations in intelligent manufacturing workshops by applying several optical fiber slip rings and resolution scanning technology in the image processing process. The improved image acquisition accuracy and the lens deflection problem of the area array camera improve the accuracy of online fault defect diagnosis and identification of the product manufacturing of the present invention;

(3) In order to reduce the time-consuming image processing, the present invention specifically sets the marking point position of the background mark based on the typical manufacturing defects that may occur in the workpiece to be tested at the current inspection station, so that the subsequent image processing can be directly The sensitive area image of the current detection station is located, and image processing and feature variable extraction and calculation are performed on this local area, which greatly improves the image processing efficiency and makes the product online fault prediction system and method of the present invention more efficient than traditional ones. The entire image processing and feature variable extraction are more real-time and efficient, which greatly improves the application value of the present invention in actual intelligent manufacturing workshops.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and examples.

Figure 1 is a schematic diagram of the fault online detection system of the present invention.

Detailed ways

The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams that only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the structures related to the present invention.

The present invention provides an online fault detection system and method applied to intelligent manufacturing workshops. It builds a manufacturing defect prediction model based on a deep neural network, effectively utilizes a large number of typical manufacturing defects obtained in the manufacturing process of actual intelligent manufacturing workshops, and combines image acquisition with With image processing technology, a standard sample image library composed of typical manufacturing defect images is trained and learned, so that the deep neural network model can be used for real-time manufacturing defect identification and classification, and can be used in the product manufacturing process through real-time image collection, Processing and analysis dynamically obtain the manufacturing information of the product, and combine it with the PLC controller and historical maintenance database to provide and automatically execute the maintenance strategy for manufacturing defects, which can provide better product manufacturing accuracy, manufacturing information collection and analysis in the intelligent manufacturing workshop. Accurate reference information.

The implementation process of the present invention is discussed in detail through the following examples.

As shown in Figure 1, with reference to Figure 1, the present invention provides an online fault detection system applied to an intelligent manufacturing workshop, including a product detection platform 1, an image acquisition unit 2, an image processing unit 3, and feature vector extraction. Unit 4, deep neural network unit 5 and computer control unit 6.

In the preferred embodiment, the number of product detection platforms 1 can be several, each product detection platform 1 corresponds to an image acquisition unit 2, the image processing unit 3, the feature vector extraction unit 4 and the deep neural network unit 5 can be used as software The system is integrated in the computer control unit 6.

Among them, the product detection platform 1 includes a detection station 11, a proximity switch 12 and an intelligent positioning and release mechanism 13. In the preferred embodiment, the proximity switch 12 and the intelligent positioning and release mechanism 13 are all communicatively connected with the computer control unit 6, where The intelligent positioning and release mechanism 13 receives the control instructions from the computer control unit 6, drives the workpiece to be tested to move on the detection station 11 to reach the preset detection position, triggers the proximity switch 12 and sends a trigger signal to the computer control unit 6. The computer control unit 6 starts the image acquisition unit 2 to collect images.

On the other hand, a background mark 14 arranged continuously along the flexible production line is provided on one side of the detection station 11. A plurality of marking points corresponding to the current detection station 11 are provided on the background mark 14. In the preferred embodiment , based on the typical manufacturing defect information that may occur in the workpiece to be tested at the current inspection station, the position of the mark point on the background target plate 14 is set to facilitate the identification and segmentation of sensitive areas in subsequent image processing. In addition, an LED surface light source matrix 15 is provided on the other side of the detection station 11. The LED surface light source matrix 15 is communicatively connected to the computer control unit 6. The computer control unit 6 can realize intelligent control of the LED surface light source matrix 15 on the entire production line. On and off, when the workpiece to be tested moves to the preset position of the current detection station 11, the computer control unit 6 sends an instruction to the LED surface light source matrix 15 corresponding to the detection station 11 to make it light up. When the workpiece is transported to the next inspection station, the LED surface light source matrix 15 of the current inspection station 11 is controlled to be extinguished. The background target plate 14 and the LED surface light source matrix 15 together form a backlight lighting environment, which is conducive to the image of the workpiece to be tested and the The grayscale contrast of the background target plate 14 reduces the time-consuming edge detection during image processing.

The image acquisition unit 2 includes an area array camera 21 , a first photoelectric conversion element 22 , a fiber slip ring 23 and a second photoelectric conversion element 24 that are connected by communication in sequence. When the workpiece to be tested moves to a preset position on the detection station 11 When the detection position is close to the center of the shooting field of view of the area array camera 21, the proximity switch 12 is triggered and a trigger signal is sent to the computer control unit 6; the axis of the photosensitive lens of the area array camera 21 is set perpendicular to the flow direction of the detection station, taking into account the intelligent manufacturing workshop In the process, the operation of the equipment may cause certain vibrations at the detection station, thereby affecting the shooting accuracy of the area array camera 21 fixedly connected to the detection station. In order to obtain clear and reliable shooting effects, the area array camera 21 of the present invention shoots Before the image is sent to the image processing unit 3, it undergoes photoelectric conversion through the first photoelectric conversion element 22, the optical fiber slip ring 23 and the second photoelectric conversion element 24, so that the image electrical signal received by the image processing unit 3 tends to be complete and reliable.

Specifically, the first photoelectric conversion element 22 and the second photoelectric conversion element 24 are both provided with the same number of input terminals and multiple output terminals, and the number of optical fiber slip rings corresponds thereto. Each input terminal of the first photoelectric conversion element 22 They are respectively connected to the area scan camera, each output end is connected to the input end of the second photoelectric conversion element 24 through an optical fiber slip ring 23, and each output end of the second photoelectric conversion element 24 is connected to the image processing unit 3.

After the image processing unit 3 receives the photoelectrically converted image, in order to improve the image processing efficiency, the present invention first obtains the sensitive area image of the current detection station through resolution scanning and performs image segmentation, and then only denoises the obtained sensitive area. As well as edge detection, traditional image processing is to denoise the entire image and then perform image segmentation processing. The present invention first segments sensitive areas and then performs local denoising, which can effectively improve image processing efficiency.

Specifically, the image processing unit 3 automatically locates the center positions of multiple mark points on the image, determines the angle between the background target 14 and the horizontal direction, and thereby calculates the deflection angle between the area array camera 21 and the background target 14 , the image processing unit 3 controls the image to perform resolution scanning along the deflection angle, thereby completing the identification of the sensitive area of the current detection station. After identifying the sensitive area, the image processing unit 3 performs an image segmentation operation on the image, thereby Obtain a new image to be processed. In the preferred embodiment, considering the backlight lighting environment composed of the background target plate 14 and the LED surface light source matrix 15, the contrast between the sensitive area of the image and the background is high, and the gray value difference is large. The present invention uses a region-based image segmentation algorithm to segment the sensitive area image.

When denoising the segmented sensitive area image, in order to make the sensitive area image have a more natural smoothing effect and enhance the processing effect of random noise in the sensitive area image, the present invention uses the Gaussian filtering method to denoise the sensitive area image. Denoising; after obtaining a smooth sensitive area image, the feature vector extraction unit further processes the processed sensitive area image.

Specifically, the purpose of the feature vector extraction unit is to reduce high-dimensional image information characterized by pixel sets into low-dimensional image information characterized by vector sets, so as to facilitate computer processing and ensure the accuracy of deep neural network unit classification; In the present invention, based on the flexible production characteristics of intelligent manufacturing, it is more appropriate to use the core image feature of shape characteristics to obtain the intelligent manufacturing information of the workpiece to be tested. In order to comprehensively capture the manufacturing information of the sensitive area image of the workpiece to be tested, This invention simultaneously considers the external edge information and internal area information of the target area, and uses the edge area, edge shape factor, average radius of the target area and the Hu invariant moment of the first three dimensions as characteristic variables to characterize the sensitive area image, and uses This is used as the feature vector of the sensitive area image and is input to the deep neural network-based unit 5.

In a preferred embodiment, based on the edge attributes of the sensitive area image, the feature vector extraction unit first performs edge detection on the sensitive area image to obtain the target area, and calculates the edge area of the target area through formulas (1)-(3) respectively. , edge shape factor and average radius of the target area, plus the Hu invariant moments of the first three dimensions, form a feature vector of the sensitive area with four feature variables to reflect manufacturing quality information such as processing or assembly of the current product inspection platform. The feature vector is sent to the deep neural network unit 5 as an input layer;

In the above formula, the parameters M and N are the number of edge points in the target area, Among them, t(x,y) is the gray value of each edge point; the parameter L is the perimeter of the target area, which can be calculated using the chain code method in image processing technology. The reference K is the number of edge points on the boundary of the target area. Number, (x _k ,y _k ) represents the pixel coordinates located on the boundary of the target area,/> Represents the centroid coordinates of the target area, which can be calculated by the following formula:

Among them, parameter A represents the area of the sensitive area, and its size can be obtained when the sensitive area is identified in image processing.

In addition, Hu invariant moments, as an important global feature of images, are not affected by light and noise, have good geometric invariance, and can effectively describe images of objects with complex shapes. Taking into account the typical manufacturing defects and image properties of smart manufacturing, For processing efficiency, it is effective to select the first 3-dimensional Hu invariant moment as one of the characteristic variables of the sensitive area image of the workpiece to be tested. The specific calculation method can refer to the conventional practice in image processing technology, which is different here. A further explanation.

In a preferred embodiment, the deep neural network unit 5 of the present invention is specifically a manufacturing defect prediction model based on a deep neural network, and includes a three-layer neural network, namely an input layer, a hidden layer and an output layer, where the input layer It has the same scale as the output layer. As the input interface of the manufacturing defect prediction model, the input layer receives the feature vector of the workpiece image to be tested, reaches the hidden layer through information encoding, and then transforms to the output layer through information decoding. The present invention adopts the classic The encoding and decoding formula models will not be described one by one here; before the deep neural network unit 5 can formally classify the manufacturing defects of the workpiece to be tested, learning and training need to be carried out first. Specifically, for the current inspection station, the workpiece to be tested may appear Typical manufacturing defects, establish a standard sample image library. In the preferred embodiment, the standard sample image library can include four sample library types: manufacturing qualified, manufacturing defect I, manufacturing defect II, manufacturing defect III, etc., as a deep neural network The training sample library of the network; similar to the aforementioned extraction of feature vectors of the workpiece to be tested, the present invention also performs edge detection on the images in the standard sample image library, and sequentially extracts the edge area, edge standard deviation, shape factor and Hu invariant of the image Feature variables such as moments constitute the feature vectors of the training sample library; finally, the input layer of the deep neural network unit 5 reads the feature vectors in the training sample library, and based on the encoding and decoding of the deep neural network unit 5, each standard sample image The manufacturing information corresponding to the images in the library is subjected to deep learning to obtain the manufacturing defect prediction model of the current inspection station.

After the deep neural network unit 5 learns and trains the training sample library, it can be used to classify and predict the manufacturing defect information of the workpiece to be tested at the current inspection station 11, thereby identifying the manufacturing defects of the workpiece to be tested on the current product inspection platform 1 type, and further sends the manufacturing defect information of the workpiece to be tested to the computer control unit 6, and the computer control unit 6 performs repair and processing of the manufacturing defects.

In a preferred embodiment, the computer control unit 6 includes a PLC controller 61 and is communicatively connected to the product detection platform 1, the image acquisition unit 2 and the deep neural network unit 5. The computer control unit 6 is classified according to the classification of the deep neural network unit 5. The manufacturing defect information indicated by the results is combined with the historical maintenance database to determine the resources required to repair the manufacturing defect, and automatically generate a maintenance strategy to repair the defect based on the defect type, defect location, defect degree, maintenance personnel selection, etc. Corresponding work instructions are formed and sent to the PLC controller 61, which performs specific maintenance operations, and updates and adds a maintenance record for the execution in the historical maintenance database.

Embodiment 2

The present invention further provides an online fault detection method applied to an intelligent manufacturing workshop, which adopts the aforementioned intelligent manufacturing workshop fault online detection system of the present invention and includes the following steps:

Step 1: Construct a manufacturing defect prediction model based on a deep neural network, and train the deep neural network through sample images;

Step 2: Under the control instructions of the computer control module, the intelligent positioning and release mechanism drives the workpiece to be tested to move on the detection station to reach the preset detection position, triggers the proximity switch, and sends the trigger signal to the computer control unit;

Step 3: The computer control unit sends instructions to the LED surface light source matrix and the image acquisition unit respectively. The LED surface light source matrix turns on the lighting. According to the preset program and parameters, the image acquisition unit shoots the workpiece to be tested. After photoelectric conversion, it will generate The image is sent to the image processing unit;

Step 4: The image processing unit identifies and segments the sensitive area of the current detection station, performs image denoising on the local area, and sends the image of the sensitive area to the feature vector extraction unit after denoising;

Step 5: The feature vector extraction unit performs edge detection on the sensitive area to form a target area, and calculates the edge area, edge shape factor and average radius of the target area through formulas (1)-(3) respectively, combined with the first three dimensions The Hu invariant moment constitutes the eigenvector of the sensitive area with four eigenvariables;

Step 6: Diagnose the manufacturing information based on the trained deep neural network on the feature vector, predict and classify the manufacturing defects of the workpiece to be tested, and feed the classification results back to the computer control unit;

Step 7: Based on the classification results and historical maintenance database information, the computer control unit determines the maintenance strategy to repair the manufacturing defect, sends work instructions to the PLC controller, and the PLC controller performs specific maintenance operations.

In a preferred embodiment, the above step 1 specifically includes:

Step 1.1: Construct a manufacturing defect prediction model based on a deep neural network. The deep neural network is a three-layer neural network, including an input layer, an output layer, and a hidden layer, where the input layer and the output layer have the same scale;

Step 1.2: Establish a standard sample image library for typical manufacturing defects that may occur at the current inspection station, including four sample library types: manufacturing qualified, manufacturing defect I, manufacturing defect II, and manufacturing defect III, as training samples for deep neural networks library;

Step 1.3: Perform edge detection on the images in the standard sample image library, and sequentially extract feature variables such as edge area, edge standard deviation, shape factor, and Hu invariant moment of the image to form the feature vector of the training sample library;

Step 1.4: The input layer of the deep learning network unit reads the feature vectors in the training sample library and performs deep learning on the manufacturing information corresponding to the images in each standard sample image library to obtain the manufacturing defects of the current inspection station. Predictive model.

Taking the above-mentioned ideal embodiments of the present invention as inspiration and through the above description, relevant workers can make various changes and modifications without departing from the scope of the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the description, and must be determined based on the scope of the claims.

Claims (7)

1. An online fault detection system, characterized by including a workpiece detection platform, an image acquisition unit, an image processing unit, a feature vector extraction unit, a deep neural network unit and a computer control unit, wherein, The workpiece detection platform includes a detection station, and the area array camera in the image acquisition unit collects images of the workpiece detection platform on the detection station and sends them to the image processing unit; The image processing unit performs a resolution scan on the received image to obtain the sensitive area image of the current detection station, denoises the sensitive area image, and then sends the denoised sensitive area image to the feature vector extraction unit; The feature vector extraction unit performs edge detection on the sensitive area image to form a target area, and calculates the edge area, edge shape factor and average radius of the target area through formulas (1) to (3) respectively, plus the first 3 dimensions The Hu invariant moments form a feature vector of the sensitive area with four feature variables to reflect the workpiece quality information of the current workpiece detection platform. The feature vector is sent to the deep neural network unit as an input layer; In the above formula, the parameters M and N are the number of edge points in the target area, Among them, t(x,y) is the gray value of each edge point; the parameter L is the perimeter of the target area, which is calculated using the chain code method in image processing technology. The reference K is the number of edge points on the boundary of the target area. , (x _k ,y _k ) represents the pixel coordinates located on the boundary of the target area, Represents the centroid coordinates of the target area, calculated by the following formula: Among them, parameter A represents the area of the sensitive area, and is suitable for obtaining its size when a sensitive area is identified in image processing; The deep neural network unit builds a manufacturing defect prediction model based on the neural network algorithm, trains, learns and classifies the image feature vectors of the workpiece detection platform, identifies the manufacturing defect type of the workpiece to be tested on the current workpiece detection platform, and feeds the classification results back to computer control unit; The workpiece inspection platform also includes a proximity switch, an intelligent positioning and release mechanism, and a background mark plate continuously arranged along the flexible production line on one side of the inspection station. A plurality of background markers corresponding to the current inspection station are provided on the background mark plate. At the marking point, an LED surface light source matrix is provided on the other side of the detection station. The LED surface light source matrix is communicatively connected to the computer control unit. The background mark plate and the LED surface light source matrix together form a backlight lighting environment; and The proximity switch, intelligent positioning and release mechanism of the workpiece detection platform are all connected by communication with the computer control unit. The intelligent positioning and release mechanism receives the control instructions from the computer control unit and drives the workpiece to be tested to move on the detection station. Arrive at the preset detection position and approach the center of the shooting field of view of the area scan camera, trigger the proximity switch and send a trigger signal to the computer control unit, and the computer control unit starts the image acquisition unit to collect images; The image processing unit automatically locates the center positions of multiple mark points on the image, determines the angle between the background target and the horizontal direction, and thereby calculates the distance between the area array camera and the background target. Deflection angle, the image processing unit controls the image to perform resolution scanning along the deflection angle, thereby completing the identification of the sensitive area of the current detection station. 2. The fault online detection system according to claim 1, characterized in that, The image acquisition unit includes a first photoelectric conversion element, an optical fiber slip ring, and a second photoelectric conversion element that are sequentially communicated with the area array camera. The axis of the photosensitive lens of the area array camera is perpendicular to the flow direction of the detection station. The area array camera After the collected image undergoes photoelectric conversion, the second photoelectric conversion element sends the electrical signal of the image to the image processing unit; The first photoelectric conversion element and the second photoelectric conversion element of the image acquisition unit are both provided with the same number of input terminals and multiple output terminals, and the number of the optical fiber slip rings corresponds to the number of the first photoelectric conversion element. Each input end is connected to the area array camera respectively, and each output end of the first photoelectric conversion element is connected to the input end of the second photoelectric conversion element through an optical fiber slip ring. The second photoelectric conversion element Each output terminal of the component is connected to the image processing unit. 3. The fault online detection system according to claim 1, characterized in that, The computer control unit includes a PLC controller and is communicated with the workpiece detection platform, the image acquisition unit and the deep neural network unit respectively. After receiving the manufacturing defect information of the workpiece detection platform, the computer control unit automatically generates a maintenance strategy and Send work instructions to the PLC controller, and the PLC controller will perform specific maintenance operations. 4. The fault online detection system according to any one of claims 1-3, characterized in that, The computer control unit determines the resources required to repair the manufacturing defect based on the manufacturing defect information indicated by the classification result of the deep neural network unit, combined with the historical maintenance database, and determines the defect type, defect location, defect degree, The maintenance personnel choose to automatically generate a maintenance strategy to repair the defect, and form a corresponding work instruction and send it to the PLC controller, which performs the specific maintenance operation and updates and adds an update to the historical maintenance database for the execution situation. Maintenance records. 5. An online fault detection method applied to an intelligent manufacturing workshop using a fault online detection system as claimed in claim 1, characterized in that it includes: Step 1: Construct a manufacturing defect prediction model based on a deep neural network, and train the deep neural network through sample images; Step 2: Under the control instructions of the computer control module, the workpiece detection platform drives the workpiece to be tested to move on the detection station to reach the preset detection position, triggers the proximity switch, and sends the trigger signal to the computer control unit; Step 3: The computer control unit sends instructions to the LED surface light source matrix and the image acquisition unit respectively. The LED surface light source matrix turns on the illumination. The image acquisition unit photographs the workpiece to be tested and sends the generated image to the image processing unit; Step 4: The image processing unit identifies and segments the sensitive area of the current detection station, performs image denoising on the sensitive area, and sends the image of the sensitive area to the feature vector extraction unit after denoising; Step 5: The feature vector extraction unit performs edge detection on the sensitive area to form a target area, and calculates the edge area, edge shape factor and average radius of the target area. Combined with the Hu invariant moments of the first three dimensions, it forms a four-dimensional Feature vector of the sensitive area of the feature variable; Step 6: Diagnose the manufacturing information based on the trained deep neural network on the feature vector, predict and classify the manufacturing defects of the workpiece to be tested, and feed the classification results back to the computer control unit. 6. The fault online detection method according to claim 5, characterized in that, The fault online detection method is suitable for detecting manufacturing defects of workpieces using the fault online detection system as claimed in claim 1. 7. The fault online detection method according to claim 6, characterized in that, Step 1: Construct a manufacturing defect prediction model based on a deep neural network. The method of training and learning the deep neural network through sample images includes: Step 1.1: Construct a manufacturing defect prediction model based on a deep neural network. The deep neural network is a three-layer neural network, including an input layer, an output layer, and a hidden layer, where the input layer and the output layer have the same scale; Step 1.2: Establish a standard sample image library for typical manufacturing defects that may occur at the current inspection station, including four sample library types: manufacturing qualified, manufacturing defect I, manufacturing defect II, and manufacturing defect III, as a training sample library for the deep neural network ; Step 1.3: Perform edge detection on the images in the standard sample image library, and sequentially extract the edge area, edge standard deviation, shape factor and Hu moment invariant feature variables of the image to form the feature vector of the training sample library; Step 1.4: The input layer of the deep learning network unit reads the feature vectors in the training sample library and performs deep learning on the manufacturing information corresponding to the images in each standard sample image library to obtain the manufacturing defects of the current inspection station. Predictive model.