An oilfield site violation behavior identification method based on a multi-layer perception network

Abstract

The application discloses a kind of oilfield field violation behavior identification method based on multilayer perceptron network, belong to computer vision field, including the following steps: making oilfield violation behavior identification dataset;Build the oilfield field violation behavior identification model based on multilayer perceptron network;Training and optimization are carried out to model;Through the camera of oil production operation field, the video data to be detected at current time is obtained, it is input to the oilfield field violation behavior identification model of training completion, obtains violation behavior detection result data, and carries out violation behavior judgment;The real-time analysis result of oilfield field violation behavior is stored and visualized, to be viewed and handled by regulatory personnel.The application has the advantages of accurate detection behavior category, can intelligently judge whether behavior is illegal, etc., aims to solve the real-time detection problem of unsafe illegal behavior in oilfield operation site, and provides intelligent solution for production management of oilfield operation site.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, specifically relating to a method for identifying violations in oilfields based on a multilayer perceptron network. Background Technology

[0002] Safety is the cornerstone of development. The petroleum industry presents complex scenarios and high risks for personnel. 24 / 7 monitoring is costly in terms of manpower and resources, and inefficient. Therefore, utilizing automated intelligent technologies to intelligently analyze potential safety hazards at work sites is of great significance.

[0003] In recent years, with the development of computer hardware, its data processing capabilities have significantly improved. Artificial intelligence technologies such as machine learning and deep learning have become increasingly widespread. Simultaneously, the development of software infrastructure, including neural networks, and the comprehensive deployment of cameras at oilfield operation sites have made it feasible to use computers for large-scale automated processing of real-time monitoring videos, enabling intelligent detection at work sites. Deep neural networks represent the forefront of machine learning, capable of automatically learning deep-level feature information. Therefore, there is an urgent need for a method to automatically identify and issue early warnings for violations at oilfield sites using deep learning models. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method for identifying violations at oilfield sites based on a multilayer perceptron network. By using a behavior recognition model, this method automatically identifies and issues warnings for violations at oilfield sites. The method analyzes and issues warnings based on automatically extracted features, which can reduce manpower and material costs.

[0005] The technical solution of the present invention is as follows:

[0006] A method for identifying violations in oilfield operations based on a multilayer perceptron network includes the following steps:

[0007] Step 1: Create a dataset for recognizing violations in oil fields;

[0008] Step 2: Construct a model for identifying violations at oilfield sites based on a multilayer perceptron network;

[0009] Step 3: Train and optimize the oilfield on-site violation recognition model to obtain the trained oilfield on-site violation recognition model;

[0010] Step 4: Obtain the video data to be detected at the current moment through the camera at the oil production site, input it into the trained oilfield violation behavior recognition model, obtain the oilfield violation behavior detection result data, and judge the violation behavior based on the obtained detection result data.

[0011] Step 5: Store and visualize the real-time analysis results of violations at the oilfield site for regulatory personnel to view and handle.

[0012] Furthermore, the specific process of step 1 is as follows:

[0013] Step 1.1: Collect video data from the oilfield site. Based on the start and end times of various violations in the video data, segment the video data into multiple segments. Randomly select several segments from the segmented video data to form the original dataset.

[0014] Step 1.2: Classify and place the original dataset containing several segments, placing segments of different behavior categories in different folders, and placing video segments of the same category in the same folder;

[0015] Step 1.3: Randomly divide the classified dataset into training and validation sets according to the proportions.

[0016] Furthermore, in step 2, the oilfield on-site violation recognition model includes three parts: an encoding network, a feature extraction network, and a classification network. The encoding network is used to segment video data into frames, dividing each frame into a sequence of patch blocks composed of several patches of the same size. Then, the patch block sequence is encoded and mapped into a token sequence at fixed frame intervals. The feature extraction network utilizes three multilayer perceptron networks to achieve information interaction between the token sequences, thereby extracting the spatiotemporal information features of the entire video data. The classification network classifies the video based on the spatiotemporal information features obtained from the feature extraction network, simultaneously outputting category and confidence information during classification.

[0017] Furthermore, the specific process of step 2 is as follows:

[0018] Step 2.1: The input video data first passes through an encoding network, which performs a linear mapping on the input video in both time and space dimensions, mapping it into a token sequence;

[0019] Assuming the original input video sequence has F frames, each frame has a resolution of H×W, and each image patch has a resolution of P1×P2, then the number of image patches is... Where H and W represent the height and width of the frame resolution, respectively, and P1 and P2 represent the height and width of the image patch resolution, respectively; using the coding network structure, a linear mapping is first performed on all image patches, so that the video V∈R F×W×H×C Mapping to the labeled sequence G∈R f×k×d The obtained tag sequence G is the token sequence; where C is the number of channels, f is the size of the time dimension in the coding network structure, k is the hidden dimension that is set in advance, and d = S is the number of image patches;

[0020] Step 2.2: The token sequence then passes through a feature extraction network containing three types of multilayer perceptrons to extract the spatiotemporal information features contained in the video;

[0021] The feature extraction network consists of three multilayer perceptron structures: a token information mixing layer, a channel information mixing layer, and an inter-frame information mixing layer. The token information mixing layer operates on the token dimension of the input feature information, mixing information between different tokens. The channel information mixing layer operates on the channel dimension of the feature information, mixing information between different channels. The inter-frame information mixing layer operates on the frame dimension of the input feature information, mixing information between different frames. In the feature extraction network, the token sequence first passes through the token information mixing layer, mixing information between different tokens in the sequence, while a residual structure is used to stabilize the information extraction process. The entire process involves one... The feature obtained after the token information mixing layer is the sum of the features between the token information mixing layers and the features after the token information mixing layer. This feature is then passed through the channel information mixing layer, which also uses a residual structure. After passing through one token information mixing layer and one channel information mixing layer, the spatial information feature extraction is achieved. The token information mixing layer and the channel information mixing layer are superimposed N1 times. Then, the spatial information extracted by superimposing N1 times is sent to the inter-frame information mixing layer superimposed N2 times. Each superposition uses a residual structure. The completion of the inter-frame information mixing layer extraction is equivalent to the completion of the spatiotemporal information feature extraction of the entire video.

[0022] Step 2.3: The extracted spatiotemporal information features are processed by a classification network to calculate the category of the behavior, and the detection confidence is obtained simultaneously; specifically:

[0023] The classification network consists of an input layer and an output layer. The input layer comprises U input nodes corresponding to the spatiotemporal information features; the output layer comprises V categories. After the spatiotemporal information features pass through the classification network, the confidence score for each category is obtained, calculated using the following formula:

[0024] Y v =x l w lv +x2w 2v +…+x u w uv +…+x U w Uv +b v (1);

[0025] Among them, Y v Let x represent the confidence level of the v-th category. u Let w represent the u-th input node. uv b represents the weight that maps the u-th input node to the v-th class. v This represents the bias of the v-th category;

[0026] The Softmax function is used to convert confidence scores into a probability distribution with positive numbers and a sum of confidence scores of 1. The formula for the Softmax function is:

[0027]

[0028] Where e is the natural constant and m is the category number. This represents the total confidence level across M categories.

[0029] Furthermore, the specific process of step 3 is as follows:

[0030] The training set and validation set from the dataset in step 1 are used to train and optimize the model. The model is fitted to the training set, the model parameters are set, and the model is trained. After multiple models are trained, each model is used to predict the validation set data, the model accuracy is recorded, and the parameters corresponding to the model with the highest accuracy are selected as the model parameters of the finally trained oilfield on-site violation recognition model.

[0031] The network training parameters set during training include: maximum number of iterations, learning rate, momentum, weight decay, and the number of categories detected.

[0032] Furthermore, the specific process of step 4 is as follows:

[0033] Step 4.1: Pre-set the detection and discrimination threshold;

[0034] Step 4.2: The detection confidence level included in the detection result data. When the detection confidence level is greater than or equal to the detection discrimination threshold, it is determined that there is a violation at the current oilfield site, and a real-time warning is issued; otherwise, it is determined that there is no violation at the current oilfield site.

[0035] Furthermore, in step 5, a violation recognition software system was developed based on the oilfield on-site violation recognition model. During development, the oilfield on-site violation recognition model was built under the deep learning PyTorch framework. The software system was installed on the computer in the monitoring room, which was connected to a server. The server received and stored video data from the oilfield on-site cameras, and also stored the detected oilfield violation recognition data. Judgment information was generated based on the stored detection data. When a violation occurred, a real-time reminder was given through the alarm bar of the software system on the monitoring room computer to achieve a visual display.

[0036] The beneficial technical effects of this invention are as follows:

[0037] In terms of time efficiency, the method of this invention can automatically and in real time determine whether workers at the oilfield site have violated regulations through computers and cameras at the oil production site. This is more accurate and faster than the process of manually checking cameras for confirmation and then providing feedback, thus reducing a significant amount of manpower and material costs.

[0038] In terms of production efficiency, the method of this invention not only reduces time costs, but also allows inspectors to conveniently monitor the safety status of workers at various oil production sites from the monitoring room. The alarm system for violations at the oilfield site makes it very convenient to monitor the production situation of workers on-site and to respond quickly based on the actual situation, thereby improving the safety of production operations.

[0039] In terms of application cost, no additional hardware equipment such as sensors is required, which reduces hardware costs. This invention uses computer vision and deep learning technology. It can be achieved by adding a server and installing the corresponding software on the server side through the existing cameras at the production site, which facilitates the rapid realization of intelligent detection at the oil production site.

[0040] Furthermore, this invention enables intelligent detection of worker violations at oilfield sites and records the data in multimedia formats such as images and videos, resulting in richer and more reliable archival information, which is helpful for inspectors in subsequent searches and archiving. This invention has advantages such as accurate detection of behavior categories and intelligent judgment of whether behavior constitutes a violation, aiming to solve the problem of real-time detection of unsafe violations at oilfield work sites and providing an intelligent solution for production management at oilfield work sites. Attached Figure Description

[0041] Figure 1 This is a flowchart of the oilfield violation identification method based on a multilayer perceptron network according to the present invention.

[0042] Figure 2 This is a schematic diagram of the structure of the oilfield violation recognition model based on a multilayer perceptron network according to the present invention.

[0043] Figure 3 for Figure 2 A schematic diagram of the feature extraction network in the model. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0045] like Figure 1 As shown, a method for identifying violations in oilfields based on multilayer perceptron networks includes the following steps:

[0046] Step 1: Create a dataset for recognizing violations in oil fields. The specific process is as follows:

[0047] Step 1.1: Collect video data from the oilfield site. Based on the start and end times of various violations in the video data, segment the video data into multiple segments. Randomly select several segments from the segmented video data to form the original dataset.

[0048] In this embodiment, oilfield on-site video data containing only violations is collected, the entire video data is processed into 1000 segments, and then 8000 segments are randomly selected as the original dataset.

[0049] Step 1.2: Classify and place the original dataset containing several segments, putting segments of different behavior categories in different folders, and video segments of the same category in the same folder.

[0050] In this embodiment, there are eight types of violations at the oilfield site: smoking in the production area, crossing guardrails, going up and down stairs without holding the railings, kicking gas cylinders, kicking clamp frames, gathering of people, walking on conveyor belts, and using mobile phones in the production area.

[0051] Step 1.3: Randomly divide the classified dataset into training and validation sets according to the proportions.

[0052] In this embodiment, the training set and the validation set are randomly divided in a ratio of 8:2.

[0053] Step 2: Construct an oilfield on-site violation recognition model based on a multilayer perceptron network.

[0054] Figure 2 This is a schematic diagram of the model structure in the oilfield violation identification method based on multilayer perceptron networks provided by the present invention. The model includes three parts: an encoding network, a feature extraction network, and a classification network. The encoding network is used to perform frame processing on video data segments and divide each frame image into a sequence of patch blocks composed of several patches of the same size. Then, the patch block sequence is encoded and mapped into a token sequence at fixed frame intervals. The feature extraction network uses the three multilayer perceptron networks to realize information interaction between the token sequences, thereby extracting the spatiotemporal information features of the entire video data. The classification network classifies the video according to the spatiotemporal information features obtained by the feature extraction network. The classification network outputs category and confidence information simultaneously during classification.

[0055] The specific working process of the oilfield violation recognition model based on multilayer perceptron networks is as follows:

[0056] Step 2.1: The input video data first passes through an encoding network. The encoding network performs a linear mapping on the input video in both time and space dimensions, mapping it into a token sequence.

[0057] Assuming the original input video sequence has F frames, each frame has a resolution of H×W, and each image patch has a resolution of P1×P2, then the number of image patches is... Where H and W represent the height and width of the frame resolution, respectively, and P1 and P2 represent the height and width of the image patch resolution, respectively. Directly inputting S image patches into the network for feature extraction would incur a huge computational cost. Therefore, an encoding network structure is used to first perform a linear mapping on all image patches, assigning video B∈R. F ×W×H×C Mapping to the labeled sequence G∈R f×k×d Where C is the number of channels, f is the size of the time dimension in the encoding network structure, k is the hidden dimension preset by the user, and d = S is the number of image patches. The token sequence G obtained through this encoding method is the token sequence, which simultaneously integrates spatiotemporal information and reduces the computational load of subsequent operations.

[0058] Step 2.2: The token sequence then passes through a feature extraction network containing three types of multilayer perceptrons to extract the spatiotemporal information features contained in the video;

[0059] The structure of the feature extraction network is as follows Figure 3 As shown, the network mainly consists of three multilayer perceptron structures: a token information mixing layer, which operates on the token dimension of the input feature information to mix information between different tokens; a channel information mixing layer, which operates on the channel dimension of the feature information to mix information between different channels; and an inter-frame information mixing layer, which operates on the frame dimension of the input feature information to mix information between different frames. In the feature extraction network, the token sequence first passes through the token information mixing layer, mixing information between different tokens in the sequence, while using a residual structure to stabilize the information extraction process. Therefore, the feature obtained after passing through a complete token information mixing layer is the sum of the features between the token information mixing layers and the features after passing through the token information mixing layer (without using a residual structure). This feature then passes through the channel information mixing layer, also using a residual structure. After passing through one token information mixing layer and one channel information mixing layer, spatial information extraction is achieved. To fully extract spatial information, the token information mixing layer and the channel information mixing layer are superimposed N1 times. The spatial information extracted by superimposing N1 times is then fed into the inter-frame information mixing layer superimposed N2 times, with a residual structure used for each superposition. Once the inter-frame information mixing layer is extracted, the spatiotemporal information features of the entire video are extracted. Figure 3In this example, “⊕” represents feature addition, which means the feature information before passing through each information mixing layer and the feature information after passing through each information mixing layer are added together. “N1×” means that the spatial information mixing layer (composed of a token information mixing layer and a channel information mixing layer) is superimposed N1 times, and “N2×” means that the inter-frame information mixing layer is superimposed N2 times.

[0060] Step 2.3: The extracted spatiotemporal information features are processed by a classification network to calculate the confidence level of the current behavior in each category. Specifically:

[0061] The classification network consists of an input layer and an output layer. The input layer comprises U input nodes corresponding to the spatiotemporal information features; the output layer comprises V categories. In this invention, there are a total of 8 categories of violations, therefore V = 8. After the spatiotemporal information features pass through the classification network, the confidence score for each category is obtained, calculated using the following formula:

[0062] Y v =x l w lv +x2w 2v +…+x u w uv +…+x U w Uv +b v (1);

[0063] Among them, Y v Let x represent the confidence level of the v-th category. u Let w represent the u-th input node. uv b represents the weight that maps the u-th input node to the v-th class. v This represents the bias of the v-th category.

[0064] The higher the confidence score, the more likely the corresponding label is to be the true label. The Softmax function is used to convert the confidence scores into a probability distribution with positive numbers and a sum of 1. The formula for the Softmax function is:

[0065]

[0066] Among them, Y v Let represent the confidence level of the v-th category, e be the natural constant, and m be the category number. This represents the total confidence level across M categories.

[0067] The Softmax function converts the confidence scores of each category into a probability distribution of positive numbers in the range [0,1], where the sum of the confidence scores of all categories is 1. For example, if there are three categories, the Softmax function yields a confidence score of 0.7 for the first category, 0.2 for the second, and 0.1 for the third. All three confidence scores are within the range [0,1], and 0.7 + 0.2 + 0.1 = 1, meaning the sum of the confidence scores of the three categories is 1.

[0068] Step 3: Train and optimize the oilfield on-site violation recognition model to obtain the trained oilfield on-site violation recognition model. The specific process is as follows:

[0069] The training and validation sets from the dataset in Step 1 are used for model training and optimization, respectively. The training set is used to fit the model, and the model is trained by setting its parameters. After training multiple models, in order to find the best-performing model, each model is used to predict the validation set data, and the model accuracy is recorded to adjust the model parameters. The parameters corresponding to the model with the highest accuracy are selected as the final model parameters.

[0070] The network training parameters set during training include: maximum number of iterations, learning rate, momentum, weight decay, and number of categories to be detected; after training and optimization, the trained oilfield violation recognition model is output.

[0071] In this embodiment, the specific training parameters are as follows: the maximum number of network iterations is set to 200; the initial learning rate is set to 0.01, and then modified to 0.001 in the 10th iteration and to 0.0001 in the 100th iteration; the momentum is set to 0.9; the weight decay is set to 0.0001; the pre-trained model is pre-trained using the ImageNet-1K dataset, which is a public dataset for image classification; the number of detected categories is modified to the number of violations at the oilfield site, and the number of categories is set to 8.

[0072] Step 4: Acquire video data to be detected at the current moment using cameras at the oilfield operation site, input it into the trained oilfield violation recognition model, obtain the oilfield violation detection result data, and judge the violation based on the obtained detection result data. The specific process is as follows:

[0073] Step 4.1: Pre-set the detection and discrimination threshold;

[0074] Step 4.2: The detection confidence level included in the detection result data. When the detection confidence level is greater than or equal to the detection discrimination threshold, it is determined that there is a violation at the current oilfield site, and a real-time warning is issued; otherwise, it is determined that there is no violation at the current oilfield site.

[0075] Step 5: Store and visualize the real-time analysis results of violations at the oilfield site for regulatory personnel to view and handle.

[0076] In this embodiment of the invention, a violation recognition software system was developed based on an oilfield on-site violation recognition model. During development, the oilfield on-site violation recognition model was constructed using the deep learning PyTorch framework. The software system is installed on a computer in the monitoring room, which is connected to a server. The server receives and stores video data from oilfield cameras and also stores detected oilfield violation recognition data (including multimedia formats such as images and videos). Judgment information is generated based on the stored detection data. When a violation occurs, a real-time alert is issued through the alarm bar of the software system on the monitoring room computer, achieving a visual display. This invention utilizes a deep learning model to automatically and in real-time process oilfield operation site monitoring videos and return the visualized results, which can improve the efficiency of oilfield operation site management and greatly reduce manpower and material resources, optimizing the cost and quality of oilfield operation site production management.

[0077] To demonstrate the feasibility and superiority of this invention, the model was pre-trained on the ImageNet-1K dataset and then tested and compared on the public dataset Kinetics-400. The model of this invention was compared with other models, including R(2+1)D, SlowFast-R101, MViT-B, Video Swin-B, and TimeSformer-L. The R(2+1)D model, proposed in 2018, explicitly decomposes 3D convolution into separate and continuous operations: a 2D spatial convolution and a 1D temporal convolution. This effectively doubles the number of nonlinearities, enabling the model to represent more complex functions. The SlowFast-R101 model, proposed in 2019, uses two channels—a slow path and a fast path—to extract features. The slow path uses low sampling and a high number of channels to primarily extract spatiotemporal features, while the fast path uses high temporal sampling and a low number of channels (mainly to reduce computational cost) to extract temporal features. The MViT-B model, proposed in 2021, applies a multi-scale hierarchical modeling method from visual prior knowledge to a purely attention-based architecture. This means that as the model deepens, the feature resolution decreases while the number of channels increases. The Video Swin-B model, proposed in 2022, uses adaptively learned sparse attention masks to address the problem of information redundancy in consecutive video frames. The TimeSformer-L model, also proposed in 2021, uses a spatiotemporal self-attention mechanism instead of traditional convolutional networks. Compared to traditional convolutional networks, it has faster training speed, higher testing efficiency, and can handle video clips longer than one minute.

[0078] The detailed information of the comparative experimental results is as follows:

[0079] The computational complexity of the model in this invention is 366 billion floating-point operations, the parameter count is 186.4 megabytes, the accuracy rate of the predicted probability ranking first category matching the actual result is 81.1%, and the accuracy rate of the predicted probability ranking top five categories matching the actual result is 94.9%.

[0080] The R(2+1)D model has a computational cost of 75 billion floating-point operations and a parameter count of 61.8 megabytes. The accuracy rate of the predicted probability ranking first class matching the actual results is 72.0%, and the accuracy rate of the predicted probability ranking five classes matching the actual results is 90.0%.

[0081] The SlowFast-R101 model requires 234 billion floating-point operations and has 59.9 megabytes of parameters. The accuracy rate of the predicted class with the highest probability is 77.8%, and the accuracy rate of the predicted five classes with the highest probability is 92.8%.

[0082] The MViT-B model requires 455 billion floating-point operations and has 36.6 megabytes of parameters. The accuracy rate of the predicted class with the highest probability is 81.2%, and the accuracy rate of the predicted five classes with the highest probability is 95.1%.

[0083] The Video Swin-B model requires 242 billion floating-point operations and has 88.1 megabytes of parameters. The accuracy of predicting the class with the highest probability is 80.6%, and the accuracy of predicting the top five classes is 94.6%.

[0084] The TimeSformer-L model requires 2.38 trillion floating-point operations and has 121.4 megabytes of parameters. The accuracy rate of the predicted class with the highest probability is 80.7%, and the accuracy rate of the predicted five classes with the highest probability is 94.7%.

[0085] The results above show that, using the same ImageNet-1K dataset for pre-training, the oilfield violation recognition model based on a multilayer perceptron network, as proposed in this invention, achieves an accuracy of 81.1% on the public dataset Kinetics-400, which is 0.5% higher than the previous best model, Video Swin-B. Furthermore, the computational cost of this invention is significantly lower than that of the TimeSformer-L and Video Swin-B models; the TimeSformer-L model's computational cost is almost seven times that of the method proposed in this invention.

[0086] This invention also measured the actual inference runtime of the Kinetics-400 dataset on 20K validation videos with cross-validation (using 8 Tesla V100 GPUs). The SlowFast-R101 model required 15 hours to complete inference, the TimeSFormer-L model required 2.6 hours, while the method of this invention only required 1.8 hours. Although the computational cost of the model of this invention is greater than that of the SlowFast-R101 model, the efficiency of the inference stage of this invention is 8 times that of the SlowFast-R101 model and 1.4 times that of the TimeSFormer-L model.

[0087] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying violations in oilfield operations based on a multilayer perceptron network, characterized in that, Includes the following steps: Step 1: Create a dataset for recognizing violations in oil fields; Step 2: Construct an oilfield on-site violation recognition model based on a multilayer perceptron network; the specific working process of the oilfield on-site violation recognition model is as follows: Step 2.1: The input video data first passes through an encoding network, which performs a linear mapping on the input video in both time and space dimensions, mapping it into a token sequence; Assume the original input video sequence has Frames, each frame has a resolution of And the resolution of each image patch is Then the number of image patches ;in, , These represent the height and width of the frame resolution, respectively. , These represent the height and width of the image patch resolution, respectively; a coding network structure is used to first perform a linear mapping on all image patches, converting the video... Mapping to a label sequence The obtained tag sequence That is, a token sequence; where, For the number of channels, To encode the size of the time dimension in the network structure, A hidden dimension that you set up in advance. , where is the number of image patches; Step 2.2: The token sequence then passes through a feature extraction network containing three types of multilayer perceptrons to extract the spatiotemporal information features contained in the video; The feature extraction network consists of three multilayer perceptron structures: a token information mixing layer, a channel information mixing layer, and an inter-frame information mixing layer. The token information mixing layer operates on the token dimension of the input feature information, mixing information between different tokens. The channel information mixing layer operates on the channel dimension of the feature information, mixing information between different channels. The inter-frame information mixing layer operates on the frame dimension of the input feature information, mixing information between different frames. In the feature extraction network, the token sequence first passes through the token information mixing layer, mixing information between different tokens in the sequence, while using a residual structure to stabilize the information extraction process. The feature obtained after passing through a complete token information mixing layer is the sum of the features obtained from passing through the token information mixing layer and the features obtained after passing through the token information mixing layer. This feature then passes through the channel information mixing layer, also using a residual structure. After passing through one token information mixing layer and one channel information mixing layer, spatial information feature extraction is achieved. The token information mixing layer and the channel information mixing layer are then superimposed. Next, it will be superimposed. The extracted spatial information is then fed into the overlay process. Each inter-frame information mixing layer uses a residual structure for each layer. Once the inter-frame information mixing layer is extracted, it is equivalent to extracting the spatiotemporal information features of the entire video. Step 2.3: The extracted spatiotemporal information features are processed by a classification network to calculate the category of the behavior, and the detection confidence is obtained simultaneously; specifically: The classification network consists of an input layer and an output layer. The input layer is the spatiotemporal information features corresponding to the input layer. There are 1 input node; the output layer is 1. Each category has spatiotemporal information features, which are then processed by a classification network to obtain the confidence score for each category. The calculation formula is as follows: (1)； in, Indicates the first Confidence level of each category Indicates the first One input node, Indicates the first The input node is mapped to the first... The weights of each category, Indicates the first Bias of each category; use The function converts the confidence scores into a probability distribution where the sum of the confidence scores is 1 and the scores are positive. The formula for the function is: (2)； in, It is a natural constant. For the category number, Indicates that it contains The sum of confidence levels for each category; Step 3: Train and optimize the oilfield on-site violation recognition model to obtain the trained oilfield on-site violation recognition model; Step 4: Obtain the video data to be detected at the current moment through the camera at the oil production site, input it into the trained oilfield violation behavior recognition model, obtain the oilfield violation behavior detection result data, and judge the violation behavior based on the obtained detection result data. Step 5: Store and visualize the real-time analysis results of violations at the oilfield site for regulatory personnel to view and handle.

2. The method for identifying violations at oilfield sites based on multilayer perceptron networks according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Collect video data from the oilfield site. Based on the start and end times of various violations in the video data, segment the video data into multiple segments. Randomly select several segments from the segmented video data to form the original dataset. Step 1.2: Classify and place the original dataset containing several segments, placing segments of different behavior categories in different folders, and placing video segments of the same category in the same folder; Step 1.3: Randomly divide the classified dataset into training and validation sets according to the proportions.

3. The method for identifying violations at oilfield sites based on multilayer perceptron networks according to claim 1, characterized in that, In step 2, the oilfield violation identification model comprises three parts: an encoding network, a feature extraction network, and a classification network. The encoding network segments video data into frames, dividing each frame into a sequence of patches of equal size. These patches are then encoded and mapped into token sequences at fixed frame intervals. The feature extraction network uses three multilayer perceptron networks to enable information exchange between token sequences, thereby extracting the spatiotemporal information features of the entire video data. The classification network classifies the video based on the spatiotemporal information features obtained from the feature extraction network, simultaneously outputting category and confidence information during classification.

4. The method for identifying violations at oilfield sites based on multilayer perceptron networks according to claim 1, characterized in that, The specific process of step 3 is as follows: The training set and validation set from the dataset in step 1 are used to train and optimize the model. The model is fitted to the training set, the model parameters are set, and the model is trained. After multiple models are trained, each model is used to predict the validation set data, the model accuracy is recorded, and the parameters corresponding to the model with the highest accuracy are selected as the model parameters of the finally trained oilfield on-site violation recognition model. The network training parameters set during training include: maximum number of iterations, learning rate, momentum, weight decay, and the number of categories to be detected.

5. The method for identifying violations at oilfield sites based on multilayer perceptron networks according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Pre-set the detection and discrimination threshold; Step 4.2: The detection confidence level included in the detection result data. When the detection confidence level is greater than or equal to the detection discrimination threshold, it is determined that there is a violation at the current oilfield site, and a real-time warning is issued. Otherwise, it is determined that there are no violations at the current oilfield site.

6. The method for identifying violations at oilfield sites based on multilayer perceptron networks according to claim 1, characterized in that, In step 5, a violation identification software system was developed based on the oilfield on-site violation identification model. During development, the oilfield on-site violation identification model was built under the deep learning PyTorch framework. The software system was installed on the computer in the monitoring room. The computer was connected to the server. The server received and stored video data from the oilfield on-site cameras. The server also stored the detected oilfield violation identification data and generated judgment information based on the stored detection data. When violations occur, real-time alerts are issued through the alarm bar of the software system on the monitoring room computer to achieve visual display.

Images