The basic principle of the convolutional neural network [12] [13]: Input the recognized target image, convolution operation to obtain the feature image; pooling operation to achieve dimensionality reduction of the image; repeated operations for multiple convolution and pooling to fully extract image features; The classification of the target image is completed in the fully connected layer. The convolution operation is mainly operated on the original image of X to obtain a featured image, using H i to represent the feature image of the i layer of the convolutional neural network. H i can be expressed by the following formula:

H i = f ( H i − 1 * W i + b i ) .

In the formula, W_i represents the weight vector of the i layer convolution kernel, the symbol of * represents the convolution operation of the image or feature map of the layer i − 1, b_i represents the offset vector of the i layer convolution kernel, and f ( x ) represents the activation function.

In 1998, Yann Lecun proposed the classic LeNet-5 network model [14], which recognizes handwritten digits, and the input image size is 32 × 32. The network structure is configured as follows: the first and second convolutional layers are connected to the pooling layer respectively, and the third convolutional layer is connected to the two fully connected layers. As the first-generation classic convolutional neural network model, there is room for improvement and optimization of its network structure, convolution kernel size, activation function, and pooling kernel size in the diagnosis of insufficient fluid supply in oil wells.

AlexNet was created by Krizhevsky [15] [16] and others and won the ImageNet competition in 2012, which improved double the recognition accuracy of ImageNet. The model consists of an 8-layer structure and has outstanding advantages in image classification. The network structure of AlexNet is shown in Figure 1.

The network structure is matched as follows: the first, second, and fifth convolutional layers are connected to the pooling layer, the third and fourth convolutional layers are continuously convolved, finally, and the three fully connected

layers are connected. The AlexNet model is compared with the LeNet-5 model: A local response normalization mechanism (LRN) [17] is proposed. LRN creates a weight setting for the activity of local neurons in the model, which makes the image recognition process more effective Purposeful. The convolutional layer uses the ReLu activation function to replace the traditional Tanh activation function [18], which improves the accuracy and time of recognizing images. The pooling layer uses maximum pooling instead of traditional average pooling, which avoids the problem of blurring image feature extraction. The Dropout layer [19] is added to the fully connected layer to interrupt all connection parameters in proportion to prevent the occurrence of over-fitting. However, its relatively large network structure and large-size convolution kernel and pooling kernel still spend a lot of time in identifying the indicator diagram and the recognition accuracy is low.

Aiming at the problem of diagnosing the degree of insufficient fluid supply in oil wells, the traditional Alexnet model has low recognition accuracy for similar indicator diagrams and takes more time. Therefore, by improving the AlexNet model [9] [20], it can not only meet the requirements of image recognition accuracy but also reduce the image recognition time. Based on the traditional AlexNet, there are mainly the following improvements:

1) The input size of the improved AlexNet model is 227 × 227 × 3 pixels changed to 100 × 200 × 1 pixels, and the 1000 neurons in the output layer are changed to 10 neurons;

2) The improved AlexNet model is composed of 4 layers of convolutional layers, 3 layers of pooling, and 2 layers of fully connected, which reduces training time and improves model convergence speed;

3) To improve the learning ability of the AlexNet model, a small convolution kernel with a size of 3 × 3 is used to replace the 5 × 5 and 11 × 11 convolution kernels in the original AlexNet network model, and a 2 × 2 pooling kernel is used Replace the 3 × 3 pooling core in the original AlexNet network model, increase the number of nonlinear operations, and enhance the ability of image feature learning. The improved AlexNet network structure is shown in Figure 2.

In the training process of the neural network, the indicator diagram needs to be preprocessed first. The quality of the preprocessing determines the level of the sample set and the training effect of the network. The specific steps are as follows:

1) Drawing and marking of the indicator diagram

This paper collects the suspension point load and displacement data of a rod pump in an oil field and draws it into a standardized indicator diagram: The size is 100 × 200 pixels, the DPI is 50 pixels/inches, the curve line width is 3 pixels, and the color black. Then classify and filter the indicator diagrams, and obtain 10 groups of indicator diagrams representing the degree of insufficient fluid supply in oil wells. As shown in Table 1.

2) Pixel normalization [21] [22]

The input original indicator diagram has pixels between 0 and 255, and the difference of the pixels will interfere with the learning ability of the neural network, creating an error invisibly. So further, this article uses the maximum and minimum normalization method, the formula is as follows.

norm = x i − min ( x ) max ( x ) − min ( x )

x i is the image pixel value in the formula; min ( x ) and max ( x ) are the maximum and minimum values of image pixels.

The image is processed by the normalization method, and the pixels are returned to between 0 and 1, without changing the image characteristics, and eliminating the difference between the wells. Make the required information more prominent, hide noise information, and save training time.

The flow chart for identifying the degree of insufficient fluid supply in oil wells designed in this paper is shown in Figure 3.

Step 1: First, the load and displacement data collected from the oil field production site are drawn into an indicator diagram. The indicator diagrams that do not meet the conditions are screened, and a sample set of indicator diagrams containing 10 types of insufficient liquid supply is established through the 5000 images obtained through the screening.

Step 2: Disrupt the sample set data, and divide the sample set into a training set and a validation set at a ratio of 8:2. According to the characteristics of the indicator diagram of insufficient liquid supply, a convolutional neural network structure is created.

Step 3: Input the training set and validation set data into the convolutional neural network model. Use the backpropagation algorithm to continuously update the weight parameters of the convolutional neural network to obtain the appropriate accuracy and recognition time, and save the trained model.

Step 4: Use the trained model to identify the indicator diagrams collected on site.

Step 5: For the wrong indicator diagram, correct the wrong recognition result. Put the corrected indicator diagram into the corresponding classification sample set, training the convolutional neural network again, and continuously optimizes the recognition model of the convolutional neural network.

In this experiment, the fast training method of multi-layer neural network based on GPU [23] [24], the training time needs 485 s in the case of GPU training, and 2501 s in the case of no GPU training. The number of iteration steps in the experiment is set to 200 steps, the learning rate is 0.0001, and 200 images are learned each iteration. The total sample is divided into a training set and a verification set at a ratio of 8:2. The training set contains 4000 samples, and the verification set contains 1000 samples. The training accuracy and verification accuracy are shown in Figure 4 and Figure 5. We will standardize the confusion matrix by row before drawing, that is, the probability value is obtained, and the sum of all the probabilities in each row is 1, so the diagonal line represents the recall rate (recall rate) of each category. The Confusion matrix is shown in Figure 6.

The experiment was conducted with the sample set of the indicator diagram of insufficient liquid supply, and the learning was carried out 200 times. The network model converged about 75 times. It can be seen from Figure 4 that the initial training accuracy is on the rise, and after a period of time, the training accuracy remains above 99%; from Figure 5, it can be seen that the initial training loss results have a downward trend. After a period of time, training the loss result is basically maintained at about 0.06; it can be seen that the training model is relatively successful. Using the validation set to verify the training model, it can be seen that after a period of time, the verification accuracy rate is basically maintained above 98%, and the training loss result basically keeps it around 0.08, which can meet actual functional requirements.

It can be seen from the confusion matrix that the recognition recall rate for 10 different degrees of insufficient liquid supply exceeds 90%, which can meet the actual application requirements.

In order to compare the performance of the convolutional neural network in this paper to identify the degree of insufficient liquid supply, the experiment verified the recognition performance of the other two methods on the same indicator diagram sample set, including the different iteration times and network structure of the convolutional neural network.