Method and device for monitoring smelting process of vacuum induction furnace based on machine vision

Abstract

The invention discloses a method and device for monitoring the smelting process of a vacuum induction furnace based on machine vision, and relates to the technical field of vacuum induction furnace monitoring, a visual monitoring technology is adopted to obtain a smelting image of a key position around the vacuum induction furnace, and whether a lens is shielded or not is recognized through a lens shielding judgment model based on a convolutional neural network; the smelting state of a steel ingot is recognized through a smelting state recognition model based on a convolutional neural network, and when a smelting period is entered, the smelting state recognition model based on target detection is adopted to carry out visual mode recognition on a smelting image so as to recognize abnormal types such as dross, furnace wall scrap steel and molten steel splashing; and information such as the number of abnormal working conditions, starting time, ending data and duration time is recorded. The smelting condition of the vacuum induction furnace can be automatically monitored and accurately recognized, the reaction is rapid, tracing is convenient, and a foundation is laid for improving the smelting efficiency and precision.

Description

Technical Field

[0001] The invention belongs to the technical field of vacuum induction furnace monitoring, and in particular relates to a method and device for monitoring a smelting process of a vacuum induction furnace based on machine vision. Background Art

[0002] Vacuum induction furnace (VIM) is an important equipment for melting metal materials. It is mainly composed of furnace shell, inductor, crucible, etc. It uses electromagnetic induction to generate eddy current in metal conductor to heat the furnace charge for melting under vacuum conditions. The high-temperature alloys and ultra-high-strength steels smelted by it are widely used in key processes with high quality requirements for special key parts such as aviation, aerospace, nuclear power and weapons and equipment.

[0003] Vacuum induction furnace smelting is done by evacuating the air to form a low-pressure vacuum environment with almost no air. The vacuum is equivalent to a "black box", and operations such as adding materials, melting, measuring temperature, sampling, and adjusting composition need to be performed without destroying the vacuum. Parameters such as slag, scrap steel on the furnace wall, and splashing generated during the smelting process of the vacuum induction furnace have an important impact on the quality of the final product. Traditional furnace smelting process monitoring methods usually rely on manual observation, and there are problems such as inaccurate monitoring, slow response, and inability to trace the source. The smelting process is affected by multiple factors such as light and background differences, so it is difficult to use automatic monitoring and identification. In addition, the smelting process of the vacuum induction furnace is complex. A smelting cycle involves multiple smelting processes such as loading, melting, and sampling. The smelting cycle is long, and operators cannot focus on monitoring for a long time, and abnormalities cannot be identified and handled in time.

[0004] Due to the advancement of science and technology and industrial transformation and upgrading, the traditional smelting production model is gradually developing towards automation and intelligence. Especially in the field of vacuum induction furnace smelting, facing the rapidly developing and changing market demand and complex and changeable smelting conditions, the traditional process detection method can no longer meet the refined requirements of production, and has obvious limitations in smelting parameter control and real-time data feedback. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method and device for monitoring the smelting process of a vacuum induction furnace based on machine vision, which can monitor and identify the smelting process of a vacuum induction furnace, accurately and quickly identify abnormal situations, and facilitate staff to deal with abnormalities in time, thus laying a foundation for improving smelting efficiency and smelting quality.

[0006] In a first aspect, the present invention provides a method for monitoring a smelting process of a vacuum induction furnace based on machine vision, comprising:

[0007] Get the current smelting image of the vacuum induction furnace;

[0008] Preprocessing the smelting image to obtain a preprocessed smelting image;

[0009] Using the trained lens occlusion judgment model to perform a first detection on the preprocessed smelting image to determine whether the lens is blocked;

[0010] If not, a second detection is performed on the pre-processed smelting image using a trained smelting state recognition model to determine the smelting state of the steel ingot; wherein the smelting state of the steel ingot includes a smelting period;

[0011] If the smelting state of the steel ingot is in the smelting period, the trained smelting anomaly recognition model is used to perform anomaly type recognition on the pre-processed smelting image to obtain an anomaly type recognition result;

[0012] Displaying the abnormality type identification result;

[0013] Among them, the lens occlusion judgment model and the smelting state recognition model are trained based on convolutional neural network algorithms respectively, and the smelting anomaly recognition model is trained based on a target detection algorithm.

[0014] In an optional embodiment, the abnormal types include slag, scrap steel on the furnace wall and molten steel splashing.

[0015] In an optional embodiment, the smelting state of the steel ingot also includes a charging period, a melting period and an end of smelting.

[0016] In an optional implementation manner, preprocessing the smelting image to obtain a preprocessed smelting image includes:

[0017] The smelting image is subjected to normalization preprocessing to obtain a smelting image after normalization preprocessing.

[0018] In an optional embodiment, the convolutional neural network is a ResNet-50 network.

[0019] In an optional embodiment, the smelting state recognition model is trained by the following method:

[0020] Construct an initial melting state recognition model based on the ResNet-50 network;

[0021] Acquire a training sample data set; wherein the training sample data set includes historical smelting images of a vacuum induction furnace;

[0022] Preprocessing and feature extracting the training samples in the training sample data to obtain a training sample feature set;

[0023] The initial smelting state recognition model is trained using the training sample feature set to obtain a trained smelting state recognition model.

[0024] In an optional implementation, the target detection algorithm is a YoLoV5 algorithm.

[0025] In an optional embodiment, if the smelting state of the steel ingot is in the smelting period, the pre-processed smelting image is subjected to abnormal type recognition using a trained smelting abnormality recognition model, and after obtaining the abnormal type recognition result, the method further includes:

[0026] Determine the abnormal data and the number of each abnormal type; wherein the abnormal data includes the start time, end time, duration, smelting image and area ratio corresponding to each abnormal type; the area ratio includes the slag area ratio.

[0027] In an optional implementation, the displaying of the abnormality type identification result includes:

[0028] The exception type and the exception data are displayed, and the exception type and the exception data are sent to a production system.

[0029] In a second aspect, the present invention provides a device for monitoring a smelting process of a vacuum induction furnace based on machine vision, comprising:

[0030] An acquisition module is used to acquire the current smelting image of the vacuum induction furnace;

[0031] A preprocessing module, used for preprocessing the smelting image to obtain a preprocessed smelting image;

[0032] A first detection module, used to perform a first detection on the pre-processed smelting image using a trained lens occlusion judgment model to determine whether the lens is blocked;

[0033] A second detection module is used for, if the answer is no, performing a second detection on the preprocessed smelting image using a trained smelting state recognition model to determine the smelting state of the steel ingot; wherein the smelting state of the steel ingot includes a smelting period;

[0034] An abnormality recognition module is used to identify the abnormality type of the pre-processed smelting image using a trained smelting abnormality recognition model if the smelting state of the steel ingot is in the smelting period, and obtain an abnormality type recognition result;

[0035] A display module, used to display the abnormality type identification result;

[0036] Among them, the lens occlusion judgment model and the smelting state recognition model are trained based on convolutional neural network algorithms respectively, and the smelting anomaly recognition model is trained based on a target detection algorithm.

[0037] The beneficial effects of the technical solution provided by the embodiment of the present invention are as follows: the present invention provides a method and device for monitoring the smelting process of a vacuum induction furnace based on machine vision. After preprocessing the current smelting image of the vacuum induction furnace, the smelting image is input into a pre-trained lens occlusion judgment model based on a convolutional neural network to determine whether the lens is blocked. If the lens is not blocked, the smelting image is input into a pre-trained melting state recognition model based on a convolutional neural network to identify the smelting state. When it is identified that the smelting enters the melting period, the smelting image is input into a trained smelting anomaly recognition model based on target recognition to identify the anomaly type, and finally the recognition result is displayed. The present invention can monitor and identify the smelting process of the vacuum induction furnace in real time, and can accurately and quickly identify abnormal situations, which is convenient for staff to deal with abnormalities in a timely manner, laying a foundation for improving smelting efficiency and smelting quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a method for monitoring a smelting process of a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0039] Figure 2 Another schematic flow chart of a method for monitoring a smelting process of a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0040] Figure 3 Another flow chart of a method for monitoring a smelting process of a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of lens occlusion in a method for monitoring a smelting process of a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0042] Figure 5 A schematic diagram of a smelting state of a method for monitoring a smelting process of a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0043] Figure 6 A schematic diagram of slag and furnace wall scrap steel in a method for monitoring a smelting process in a vacuum induction furnace based on machine vision provided by an embodiment of the present invention;

[0044] Figure 7 A schematic diagram of the system principle of a vacuum induction furnace smelting process monitoring device based on machine vision provided in an embodiment of the present invention.

[0045] In the figure: 100 - acquisition module; 200 - preprocessing module; 300 - first detection module; 400 - second detection module; 500 - abnormality identification module; 600 - display module. DETAILED DESCRIPTION

[0046] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0047] See also Figure 1 and Figure 2 An embodiment of the present invention provides a method for monitoring a smelting process of a vacuum induction furnace based on machine vision, comprising the following steps S100 to S600.

[0048] Step S100, obtaining the current smelting image of the vacuum induction furnace.

[0049] Specifically, video recording devices are installed at key locations around the vacuum induction furnace, and the layout and angles of the cameras are optimized to ensure that images of the vacuum induction furnace smelting process can be captured accurately, comprehensively and clearly, thereby achieving real-time image signal acquisition.

[0050] Step S200, preprocessing the smelting image to obtain a preprocessed smelting image.

[0051] In some possible embodiments, this step specifically includes: performing normalization preprocessing on the smelting image to obtain a smelting image after normalization preprocessing.

[0052] In the field of image recognition, image quality has a direct impact on the design and effect accuracy of the recognition algorithm. Although algorithm optimization is the key, preprocessing technology also occupies an important position in the project, but this is often overlooked. Preprocessing steps usually include digitization, geometric transformation, normalization, smoothing, restoration and enhancement. In this embodiment, sampling normalization is used to preprocess the smelting image. Normalization can be understood to a certain extent as scaling the pixel values ​​originally in the range of 0-256 to between 0-1, thereby reducing the distribution distance of the pixel values. The grayscale difference between the smelting furnace area (target area) and the rest of the background area in the smelting image in this embodiment is large. Using normalization for preprocessing can reduce the grayscale difference and make the target area more prominent, thereby improving the accuracy, stability and reliability of detection and recognition.

[0053] Step S300: Perform a first detection on the pre-processed smelting image using the trained lens occlusion judgment model to determine whether the lens is blocked.

[0054] Since the vacuum induction furnace is fully enclosed and the enclosed space is highly vacuumed, a glass window is provided to facilitate observation of the furnace. In order to prevent the glass window from being damaged due to long-term high temperature in the furnace, an isolation steel plate is provided between the glass window and the furnace for heat insulation. The isolation steel plate is opened when it is necessary to check the furnace. To this end, it is necessary to determine whether the lens is blocked, that is, to determine whether the captured smelting image is blocked. This embodiment determines the trained lens occlusion judgment model through the following steps (1-1) to (1-4).

[0055] Step (1-1), construct an initial lens occlusion judgment model based on the ResNet-50 network.

[0056] Step (1-2), obtaining a training sample data set; wherein the training sample data set includes historical smelting images of a vacuum induction furnace.

[0057] Step (1-3), preprocessing and feature extraction of training samples in the training sample data to obtain a training sample feature set.

[0058] Step (1-4), using the training sample feature set to train the initial lens occlusion judgment model to obtain a trained lens occlusion judgment model.

[0059] In some embodiments, the structure of the initial lens occlusion judgment model constructed based on the ResNet-50 network can be: input layer, preprocessing layer, ResNet-50 backbone layer, global average pooling layer, fully connected layer and output layer. The input layer is a historical smelting image of a fixed size, such as a 224×224×3 RGB image, which includes an image with an occluded lens and a normal smelting image. The preprocessing layer includes image normalization, which uses the mean and standard deviation to scale the pixel values ​​to the range of [0, 1]. The ResNet-50 backbone layer includes a convolutional layer, a maximum pooling layer, and multiple residual block layers. The global average pooling layer is used to compress the spatial dimension of each channel of the feature map into a single value to generate a global feature vector. The fully connected layer is used to map the global feature vector to a binary classification output, and the output layer uses the Softmax activation function to output the probabilities of the two categories.

[0060] In some embodiments, the ResNet-50 backbone layers may be as follows.

[0061] Convolutional layer 1: 7×7 convolution, 64 output channels, stride 2, padding 3.

[0062] Max pooling layer: 3×3 max pooling, stride 2.

[0063] Convolutional layer 2: 3 residual blocks, each with 64 output channels (first reduced to 64 by 1x1 convolution, then by 3×3 convolution, and finally by 1x1 convolution to 256, but the actual number of output channels is 64 because of the use of residual connections).

[0064] Convolutional layer 3: 4 residual blocks, each with 128 output channels (first reduced to 128, then 3×3 convolution, and finally increased to 512, but the actual output is 128).

[0065] Convolutional layer 4: 6 residual blocks, each with 256 output channels (first reduced to 256, then 3×3 convolution, and finally increased to 1024, but the actual output is 256).

[0066] Convolutional layer 5: 3 residual blocks, each with 512 output channels (first reduced to 512, then 3×3 convolution, and finally increased to 2048, but the actual output is 512).

[0067] The output of convolution layer 5 is globally averaged and pooled to obtain a 2048-dimensional global feature vector. The fully connected layer outputs two neurons, occluded and unoccluded, and the output layer outputs the probabilities of the two categories. In addition, the cross entropy loss function is used to optimize the network parameters. Finally, the output results of the lens occlusion judgment model are thresholded to obtain the category of the image and determine whether the lens is occluded.

[0068] Table 1 shows the training data of the lens occlusion judgment model based on ResNet-50. It can be seen that the recognition accuracy of the model reaches 100%.

[0069] Table 1 Training data of lens occlusion judgment model

[0070] Iteration round Training set-occlusion Training set - normal Test Set - Occlusion Test set - normal Accuracy First round 1000 1000 500 500 100％ Second round 1011 1270 500 500 100％ Third round 1074 1271 500 500 100％ Fourth round 1147 1271 500 500 100％

[0071] like Figure 4 As shown in the figure, the two images above are when the lens is blocked, and the two images below are when the lens is normal. Figure 4 It can be seen that the smelting image when the lens is completely or partially blocked is significantly different from the smelting image when the lens is normal. This embodiment adopts a lens occlusion judgment model based on ResNet-50, which introduces the structure of residual blocks. The smelting image under lens occlusion has information missing or deformation problems. Using a deep neural network such as ResNet-50 for recognition can extract key features from limited image information with its powerful feature extraction ability, thereby improving the robustness of recognition. At the same time, there are various situations in which the lens is blocked, such as partial occlusion, complete occlusion, etc. The deep network structure and residual learning mechanism of ResNet-50 enable it to adapt to different types of occlusion situations, thereby improving the accuracy and stability of recognition. ResNet-50 itself has a faster recognition speed and lower computational complexity. In the video surveillance scenario of this embodiment, the use of ResNet-50 can improve the recognition efficiency of the entire algorithm.

[0072] In summary, in this step, in order not to affect subsequent recognition, it is necessary to identify whether the lens is blocked. This embodiment identifies blocked images and unblocked smelting images based on the ResNet-50 lens occlusion recognition model. This embodiment samples the deep neural network ResNet-50 for classification, and uses the powerful feature extraction and learning capabilities of ResNet to classify the image for occlusion, accurately determine whether the monitoring screen is blocked, and prepare for the subsequent recognition model. It should be noted that in some embodiments, some non-smelting state pictures also need to be identified, such as images outside the furnace.

[0073] Step S400, if the judgment result of the lens occlusion judgment model is no, then the trained smelting state recognition model is used to perform a second detection on the pre-processed smelting image to judge the smelting state of the steel ingot; wherein the smelting state of the steel ingot includes the feeding period, the melting period, the smelting period and the end of smelting. wherein the melting period refers to the period when the steel ingot is not completely melted, and when the steel ingot is heated to be completely melted, the smelting period is entered. Figure 5 As shown, the first image in the first row is the charging period (the smelting start time is recorded at this time), the second and third images in the first row are the melting period, the second image in the second row is the smelting period, and the first image in the second row is the end of smelting (the smelting end time is recorded at this time).

[0074] Specifically, this embodiment determines the trained smelting state recognition model through the following steps (2-1) to (2-4).

[0075] Step (2-1), construct an initial smelting state recognition model based on the ResNet-50 network.

[0076] Step (2-2), obtaining a training sample data set; wherein the training sample data set includes historical smelting images of the vacuum induction furnace.

[0077] Step (2-3), preprocessing and feature extraction of training samples in the training sample data to obtain a training sample feature set.

[0078] Step (2-4), using the training sample feature set to train the initial smelting state recognition model to obtain a trained smelting state recognition model.

[0079] In this embodiment, the initial smelting state recognition model is constructed based on the convolutional neural network of ResNet-50. In some possible embodiments, the initial smelting state recognition model based on the ResNet-50 network can be: input layer, initial convolution layer, residual block layer, full average pooling layer and full connection layer. The input layer is a historical smelting image of a fixed size, such as an RGB image of 224×224×3. The initial convolution layer includes: the first convolution layer, the normalization layer, the ReLU activation function to introduce nonlinearity, and the maximum pooling layer. The normalization layer uses the mean and standard deviation to scale the pixel values ​​to the range of [0, 1]. The residual block layer uses 16 residual blocks, and solves the problems of gradient disappearance and degradation in the network through jump connections.

[0080] Specifically, the parameters of each layer are as follows.

[0081] Convolutional layer 1: 7×7 convolution, 64 output channels, stride 2, for preliminary feature extraction.

[0082] Normalization layer: used to improve the stability of training.

[0083] The ReLU activation function introduces non-linearity.

[0084] Max pooling layer: 3×3 max pooling with stride 2 is used to further reduce the size of the feature map.

[0085] The residual block layer uses 16 residual blocks divided into 4 stages. Each residual block consists of 3 convolutional layers: 1×1 convolution to reduce the dimension, 3×3 convolution to process the main features, and another 1×1 convolution to restore the dimension.

[0086] First stage: 3 residual blocks, each with 64 output channels.

[0087] First stage: 4 residual blocks, each with 128 output channels.

[0088] First stage: 6 residual blocks, each with 256 output channels.

[0089] First stage: 3 residual blocks, each with 512 output channels.

[0090] Global average pooling layer: A global average pooling layer is used after the residual block layer to compress the spatial dimension of each channel of the feature map into a single value to generate a global feature vector.

[0091] Fully connected layer: maps the global feature vector to four categories: charging period, melting period, smelting period, and smelting end. This layer is followed by a Softmax layer to calculate the probability of each category for the final classification decision.

[0092] The iteration of the vacuum induction furnace melting state model is shown in Table 2. It can be seen from Table 2 that after 8 iterations, the accuracy of the trained melting state recognition model is close to 100%.

[0093] Table 2 Training data of smelting state recognition model

[0094]

[0095] Finally, according to the output results of the smelting state recognition model, the category with the highest probability is used as the smelting state of the target smelting image. In specific applications, it is necessary to determine the start and end marks of smelting based on the recognition results, and record the start time and end time of each heat of steel smelting to facilitate the subsequent management of each heat of steel. In addition, it is sometimes necessary to record the start time, duration and end time of each smelting state. Among them, only when the ingot is in a fully molten state can subsequent abnormality recognition be performed, that is, it is necessary to identify the time when the melting period ends and the smelting period is about to begin, in preparation for subsequent abnormality recognition.

[0096] The smelting state recognition model of this embodiment adopts an advanced image processing algorithm, namely, a convolutional neural network based on the ResNet-50 network for classification, performs feature extraction and learning on the captured images, accurately identifies the smelting state, and records the start and end time of smelting, so as to prepare for subsequent abnormality recognition.

[0097] Step S500: If the smelting state of the steel ingot is in the smelting period, the trained smelting anomaly recognition model is used to perform anomaly type recognition on the pre-processed smelting image to obtain an anomaly type recognition result.

[0098] Specifically, the abnormal types include slag, furnace wall scrap and molten steel splashing, among which slag and furnace wall scrap such as Figure 6 As shown. Among them, the target detection algorithm is YoLoV5 algorithm. This embodiment determines the trained smelting anomaly recognition model through the following steps (3-1) to (3-4).

[0099] Step (3-1), construct an initial smelting anomaly recognition model based on the YoLoV5 algorithm.

[0100] Step (3-2), obtaining a training sample data set; wherein the training sample data set includes smelting images in the smelting period.

[0101] Step (3-3), preprocessing and feature extraction of training samples in the training sample data to obtain a training sample feature set.

[0102] Step (3-4), using the training sample feature set to train the initial smelting anomaly recognition model to obtain a trained smelting anomaly recognition model.

[0103] This embodiment adopts an initial smelting anomaly recognition model based on the YoLoV5 algorithm. In some possible embodiments, the specific structure of the model can be: input layer, backbone network layer, neck network layer and output layer. The input layer scales the pixel value to the range of [0, 1] by normalization. In some embodiments, Mosaic enhancement, MixUp and other methods can be used to further enrich the training data set and improve the model's recognition ability for complex scenes. The backbone network includes the Focus structure and the CSP (CrossStage Partial Networks) structure. The Focus structure converts the input image into a higher-level feature representation through slicing operations, reducing the number of parameters and the amount of calculation. The CSP structure divides the gradient flow, so that the gradient flow propagates through different network paths and then converges, thereby ensuring the recognition accuracy of the model, while reducing the amount of calculation and improving the reasoning speed. The neck network layer uses a method combining FPN (Feature Pyramid Networks) and PAN (Path Aggregation Network) structures to fuse information from different feature map levels. The FPN structure is top-down, and high-level features are fused with low-level features through upsampling to obtain feature maps for prediction; the PAN structure is bottom-up, and complements FPN to further enhance the fusion effect of features. The output layer uses GIOU Loss as the loss function (which includes a combination of classification loss, positioning loss, and confidence loss), and adds DIOU_nms for prediction box screening, and uses NMS (non-maximum suppression) to eliminate redundant detection results to obtain the final detection results.

[0104] During the specific implementation, the initial smelting anomaly recognition model is trained and iterated through a large amount of experimental data, so that it can accurately extract key parameters and make adaptive adjustments according to actual conditions.

[0105] This embodiment uses visual pattern recognition technology to establish a model for key scenarios of vacuum induction furnace smelting, and uses pattern recognition technology to accurately judge and classify abnormal operating conditions, such as splashing, slag, and scrap steel on the furnace wall.

[0106] In some possible embodiments, after step S500, the following steps are further included: determining abnormal data and determining the number of each abnormal type; wherein the abnormal data includes the start time, end time, duration, smelting image and area proportion corresponding to each abnormal type. The area proportion here includes the slag area proportion.

[0107] Specifically, in addition to identifying abnormal conditions such as slag and scrap steel on the furnace wall during the smelting process when the ingot is completely melted, this embodiment also records information during the smelting process, including the start and end time of smelting, and the number of abnormal conditions, start time, end time, duration, smelting images, slag area, etc.

[0108] In some embodiments, Figure 3 As shown in the figure, during the recognition process, the lens occlusion recognition model starts recording the start time and end time when it recognizes that the lens is not blocked. The smelting state recognition model also records the start time and end time of each smelting stage. The abnormality type recognition model also records the abnormality category, abnormality number, start time and end time.

[0109] This embodiment uses a model to perform target detection and size calculation in the vacuum induction furnace abnormality recognition model. Its main task is to detect abnormal conditions such as slag and scrap steel on the furnace wall in the image, and calculate the area ratio, so as to realize automatic recognition and quantification of abnormal conditions in the smelting process.

[0110] Step S600, displaying the abnormality type identification result.

[0111] This step specifically includes: displaying the abnormal type and abnormal data, and sending the abnormal type and abnormal data to the production system. Displaying the abnormal type and abnormal data facilitates on-site staff to take timely measures to improve smelting quality. Sending the abnormal type and abnormal data to the production system facilitates quality monitoring and recording of the current smelting process, and facilitates the quality evaluation of the smelting process and subsequent quality traceability.

[0112] like Figure 2 As shown, this embodiment has built a set of vacuum induction furnace intelligent monitoring system, and the overall system mainly includes lens occlusion recognition model, smelting state recognition model, and abnormal type recognition model. First, the smelting image is input into the system, and the lens occlusion recognition model outputs the recognition result; and it is determined whether the lens is blocked. If so, it enters the image recognition step of the new furnace. If not, the smelting image is input into the smelting state recognition model and the recognition result is output; if the recognition result is the feeding period or the melting period, it returns to the step of inputting the smelting image. If the recognition result is the end of smelting, it enters the image recognition step of the new furnace. If the recognition result is the smelting period, the smelting image is input into the abnormal type recognition model to identify whether there is an abnormality, and to identify abnormalities such as scrap steel on the furnace wall, splashing of molten steel and slag. This embodiment monitors the video stream on site through machine vision and pattern recognition technology, and automatically monitors the smelting process to achieve quantitative evaluation of the quality process and quality traceability.

[0113] This embodiment adopts an advanced image processing algorithm based on convolutional neural network, uses a lens occlusion judgment model based on ResNet-50 to extract and learn features of the captured image, classifies the image occlusion, accurately judges whether the monitoring screen is blocked, and prepares for the subsequent recognition module; uses a smelting state recognition model based on ResNet-50 to judge the smelting state of the ingot, and when entering the smelting period, uses a smelting anomaly recognition model based on YoLoV5 to identify the abnormal type, and accurately and real-timely judges, classifies and quantifies data for key scenes such as bubbles, slag and scrap steel in the molten pool. Information on smelting anomalies in a smelting process is recorded, including the start and end time of smelting, the number of abnormal conditions, duration, corresponding pictures, area ratio and other information, which facilitates the quality evaluation of the smelting process and subsequent quality tracing. This embodiment realizes the automated monitoring of the smelting process of the vacuum induction furnace, accurately and quickly identifies abnormal conditions, facilitates the staff to deal with abnormalities in a timely manner, improves the efficiency and accuracy of the vacuum induction furnace smelting process, and has great technical innovation significance and production application value.

[0114] See also Figure 7 The smelting process monitoring device of a vacuum induction furnace based on machine vision proposed in this embodiment includes an acquisition module 100, a preprocessing module 200, a first detection module 300, a second detection module 400, an abnormality recognition module 500 and a display module 600. The acquisition module 100 is used to acquire the current smelting image of the vacuum induction furnace. The preprocessing module 200 is used to preprocess the smelting image to obtain the preprocessed smelting image. The first detection module 300 is used to perform a first detection on the preprocessed smelting image using a trained lens occlusion judgment model to determine whether the lens is blocked. The second detection module 400 is used to perform a second detection on the preprocessed smelting image using a trained smelting state recognition model to determine the smelting state of the ingot if it is not, wherein the smelting state of the ingot includes the smelting period. The abnormality recognition module 500 is used to perform abnormality type recognition on the preprocessed smelting image using a trained smelting abnormality recognition model if the smelting state of the ingot is the smelting period to obtain the abnormality type recognition result. The display module 600 is used to display the abnormality type recognition result. Among them, the lens occlusion judgment model and the smelting state recognition model are trained based on the convolutional neural network algorithm respectively, and the smelting anomaly recognition model is trained based on the target detection algorithm.

[0115] In possible embodiments, the anomaly types include dross, furnace wall scrap, and steel splashing.

[0116] In a possible embodiment, the smelting state of the steel ingot further includes a charging period, a melting period and the end of smelting.

[0117] In a possible embodiment, the preprocessing module 200 is used to perform normalization preprocessing on the smelting image to obtain a smelting image after normalization preprocessing.

[0118] In a possible embodiment, the convolutional neural network is a ResNet-50 network.

[0119] In a possible embodiment, the second detection module 400 includes a model building module, a training sample module, a training sample preprocessing module and a training module. The model building module is used to build an initial smelting state recognition model based on the ResNet-50 network. The training sample module is used to obtain a training sample data set; wherein the training sample data set includes historical smelting images of a vacuum induction furnace. The training sample preprocessing module is used to preprocess and extract features from the training samples in the training sample data to obtain a training sample feature set. The training module is used to train the initial smelting state recognition model using the training sample feature set to obtain a trained smelting state recognition model.

[0120] In a possible embodiment, the target detection algorithm is the YoLoV5 algorithm.

[0121] In a possible embodiment, if the smelting state of the steel ingot is in the smelting period, a recording module is further connected between the abnormality identification module and the display module, and the recording module is used to determine the abnormal data and the number of each abnormality type; wherein the abnormal data includes the start time, end time, duration, smelting image and area ratio corresponding to each abnormality type; the area ratio includes the slag area ratio.

[0122] In a possible embodiment, the display module 600 is used to display the exception type and the exception data, and send the exception type and the exception data to the production system.

[0123] The device provided in the embodiment of the present application is adopted. Since the device adopts the same inventive concept as the above-mentioned method provided in the embodiment of the present application, on the premise that the method can solve the technical problem, the device can also solve the technical problem, which will not be repeated here.

[0124] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A method for monitoring the smelting process of a vacuum induction furnace based on machine vision, characterized in that: include: Get the current smelting image of the vacuum induction furnace; Preprocessing the smelting image to obtain a preprocessed smelting image; Using the trained lens occlusion judgment model to perform a first detection on the preprocessed smelting image to determine whether the lens is blocked; If not, a second detection is performed on the pre-processed smelting image using a trained smelting state recognition model to determine the smelting state of the steel ingot; wherein the smelting state of the steel ingot includes a smelting period; If the smelting state of the steel ingot is in the smelting period, the trained smelting anomaly recognition model is used to perform anomaly type recognition on the pre-processed smelting image to obtain an anomaly type recognition result; Displaying the abnormality type identification result; Among them, the lens occlusion judgment model and the smelting state recognition model are trained based on convolutional neural network algorithms respectively, and the smelting anomaly recognition model is trained based on a target detection algorithm.

2. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: The abnormal types include slag, scrap steel on the furnace wall and molten steel splashing.

3. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: The smelting state of the steel ingot also includes a charging period, a melting period and a smelting end.

4. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: Preprocessing the smelting image to obtain a preprocessed smelting image includes: The smelting image is subjected to normalization preprocessing to obtain a smelting image after normalization preprocessing.

5. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: The convolutional neural network is a ResNet-50 network.

6. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 5 is characterized in that: The smelting state recognition model is trained by the following method: Construct an initial melting state recognition model based on the ResNet-50 network; Acquire a training sample data set; wherein the training sample data set includes historical smelting images of a vacuum induction furnace; Preprocessing and feature extracting the training samples in the training sample data to obtain a training sample feature set; The initial smelting state recognition model is trained using the training sample feature set to obtain a trained smelting state recognition model.

7. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: The target detection algorithm is the YoLoV5 algorithm.

8. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 1 is characterized in that: If the smelting state of the steel ingot is in the smelting period, the trained smelting anomaly recognition model is used to perform anomaly type recognition on the pre-processed smelting image, and after obtaining the anomaly type recognition result, the method further includes: Determine the abnormal data and the number of each abnormal type; wherein the abnormal data includes the start time, end time, duration, smelting image and area ratio corresponding to each abnormal type; the area ratio includes the slag area ratio.

9. The method for monitoring the smelting process of a vacuum induction furnace based on machine vision according to claim 8 is characterized in that: The displaying of the abnormal type identification result includes: The exception type and the exception data are displayed, and the exception type and the exception data are sent to a production system.

10. A vacuum induction furnace smelting process monitoring device based on machine vision, characterized in that: include: An acquisition module is used to acquire the current smelting image of the vacuum induction furnace; A preprocessing module, used for preprocessing the smelting image to obtain a preprocessed smelting image; A first detection module, used to perform a first detection on the pre-processed smelting image using a trained lens occlusion judgment model to determine whether the lens is blocked; A second detection module is used for, if the answer is no, performing a second detection on the preprocessed smelting image using a trained smelting state recognition model to determine the smelting state of the steel ingot; wherein the smelting state of the steel ingot includes a smelting period; An abnormality recognition module is used to identify the abnormality type of the pre-processed smelting image using a trained smelting abnormality recognition model if the smelting state of the steel ingot is in the smelting period, and obtain an abnormality type recognition result; A display module, used to display the abnormality type identification result; Among them, the lens occlusion judgment model and the smelting state recognition model are trained based on convolutional neural network algorithms respectively, and the smelting anomaly recognition model is trained based on a target detection algorithm.

Images