CN117036771A - Indicator diagram feature extraction method and system based on full convolution neural network - Google Patents

Abstract

The application discloses a method and a system for extracting characteristics of an indicator diagram based on a full convolution neural network, which relate to the technical field of oil exploitation fault diagnosis and comprise the steps of constructing an indicator diagram image data set; constructing a full convolution neural network model, wherein the input and output of the model are images of an indicator diagram; pre-training the model, wherein the trained evaluation index is used for judging the similarity between the output and input indicator diagram images of the model; performing super-parameter optimization on the model by adopting grid search, and optimizing the extraction capacity of the model; the convolution stage of the truncated model acts as a graph feature extractor. The method for extracting the characteristics of the indicator diagram, disclosed by the application, solves the problems of poor performance, low speed and high model optimization difficulty of direct learning of a large number of samples. In the subsequent establishment of the problem of oil well fault diagnosis according to the indicator diagram, the requirement of model learning can be met by only relying on a small quantity of marked samples, the certain labor and time cost during sample marking and model optimization is reduced, and the accuracy of indicator diagram identification is improved.

Description

A method and system for dynamometer feature extraction based on fully convolutional neural network

Technical field

The present invention relates to the technical field of oil production fault diagnosis, and in particular to a method and system for extracting features of dynamometer diagrams based on a fully convolutional neural network.

Background technique

Pumping units are the main equipment in oil field development. During the oil extraction process, they are prone to failure due to factors such as temperature and formation. Therefore, fault diagnosis of pumping unit wells is a key issue in the field of oil extraction. The dynamometer diagram is an important basis for diagnosing pumping unit faults. With the rapid development of data mining and storage technology, sensors collect various data from oil wells. These data are transmitted to the oilfield data center for storage and analysis, forming the formation of oil well production fault diagnosis. of big data. The new generation of artificial intelligence technology based on "big data + deep learning" continues to break through the limitations of existing technology and lead oil well fault diagnosis technology to a new stage.

The diagnostic accuracy of the pumping unit fault diagnosis model depends on the quantity and quality of the dynamometer sample set. On the one hand, fault diagnosis requires training a large number of labeled samples, and the process of training a large number of samples is slow in iteration and convergence, making model optimization time-consuming and difficult. On the other hand, current fault diagnosis research samples are small in size and cover few fault types. , the problem of sample imbalance is prominent.

The purpose of the present invention is to provide a dynamometer feature extraction method based on a fully convolutional neural network, which has a high recognition accuracy and reduces the model's dependence on a large amount of labeled data and the time spent on training.

Contents of the invention

The purpose of this section is to outline some aspects of embodiments of the invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, the abstract and the title of the invention to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the invention.

In view of the above and/or existing problems in a dynamometer feature extraction method based on a fully convolutional neural network, the present invention is proposed.

Therefore, the problems to be solved by this invention are: fault diagnosis requires training a large number of labeled samples, and the process of training a large number of samples has slow iteration and slow convergence speed, model optimization is time-consuming and difficult, and the sample size of current fault diagnosis research is small, There are few coverage fault types and sample imbalance is a prominent problem.

In order to solve the above technical problems, the present invention provides the following technical solution: a dynamometer diagram feature extraction method based on a fully convolutional neural network, which includes: constructing a dynamometer diagram image data set; constructing a fully convolutional neural network model, and model input and output are both dynamometer images; use the constructed dynamometer image data set to pre-train the fully convolutional neural network model. The training evaluation index is the judgment between the dynamometer image output by the model and the input dynamometer image. similarity between the models; use grid search to optimize the hyperparameters of the fully convolutional neural network model to optimize the extraction capability of the model; intercept the convolution stage of the fully convolutional neural network model as a feature extractor for the dynamometer diagram.

As a preferred solution of the dynamometer feature extraction method based on a fully convolutional neural network according to the present invention, the dynamometer image data set includes, according to the massive dynamometer displacement load data in the database, With the abscissa as the displacement and the ordinate as the load, standardized dynamometer images are drawn in batches to construct a dynamometer data set.

As a preferred solution of the invention for extracting features of dynamometer diagrams based on a fully convolutional neural network, the fully convolutional neural network model includes convolution, pooling, deconvolution, and anti-pooling. Operation steps: The input of the model is a dynamometer image. The deconvolution stage is the reverse process of the convolution stage. The output of the convolution layer is decoded. The output of the model is a dynamometer image. The input and output of the model are The same, it is the same power diagram image.

As a preferred solution of the dynamometer feature extraction method based on a fully convolutional neural network according to the present invention, the deconvolution stage includes filling, convolution, and cropping.

As a preferred solution of the dynamometer feature extraction method based on a fully convolutional neural network according to the present invention, the pre-training includes using the dynamometer image data set to perform the full convolution on the dynamometer image data set. The neural network is pre-trained, and the characteristics of the dynamometer diagram output by the model convolution operation are restored to the dimensions of the input through the deconvolution operation. After each round of training, the characteristics of the dynamometer diagram when input into the convolution layer are calculated and the characteristics of the dynamometer diagram after the deconvolution operation are calculated. Feature similarity of the restored features; the feature similarity includes: the measurement method of feature similarity is the Jaccard similarity coefficient; for the two feature matrices A and B, calculate the Jaccard similarity coefficient of each vector, average it, and the Jaccard similarity The calculation formula of the coefficient is expressed as,

J(A,B)＝(J(A ₁ ,B ₁ )+J(A ₂ ,B ₂ )+...+J(A _n ,B _n ))/n

For two vectors A _n and B _n , as two sets, the Jaccard similarity coefficient is expressed as,

J(A _n ,B _n )=|S|/|U|

Among them, S is the intersection of sets A _n and B _n , and U is the union of sets A _n and B _n ; if the feature similarity exceeds 90%, the convolution stage of the fully convolutional neural network model is intercepted. If the feature similarity is Below 90%, perform hyperparameter optimization.

As a preferred solution of the dynamometer feature extraction method based on a fully convolutional neural network according to the present invention, the hyperparameter optimization includes performing hyperparameter optimization on the fully convolutional neural network model until When the feature similarity between the dynamometer diagram input by the convolution layer and the dynamometer diagram output by the deconvolution layer reaches more than 90%, the optimized fully convolutional neural network model is obtained; grid search is used as the algorithm for the fully convolutional neural network model. Hyperparameter optimization method pre-defines the value range of each hyperparameter to be optimized.

As a preferred solution of the dynamometer feature extraction method based on a fully convolutional neural network according to the present invention, the dynamometer feature extractor includes removing the deconvolution layer of the fully convolutional neural network model. , a dynamometer feature extractor based on the convolution stage of the fully convolutional neural network model is obtained.

Another object of the present invention is to provide a system based on a fully convolutional neural network-based dynamometer feature extraction method, which can reduce the model's dependence on a large amount of labeled data and training costs by constructing a dynamometer feature extraction system. time.

As a preferred solution of the dynamometer chart feature extraction system based on a fully convolutional neural network according to the present invention, it includes: a dynamometer chart input module, a pre-training module, a hyperparameter optimization module and a feature extraction module; The dynamometer input module is used to input dynamometer data of features to be extracted; the pre-training module uses the dynamometer image data set to pre-train the fully convolutional neural network, and convolves the model through deconvolution operations. The dynamometer features output by the operation are restored to the input dimensions, and the similarity between the features when the dynamometer graph is input into the convolution layer and the features restored after the deconvolution operation is calculated, and the Jaccard similarity coefficient is used as a measure of feature similarity. ; The hyperparameter optimization module performs hyperparameter optimization on the trained fully convolutional neural network model and finds the best hyperparameter combination through grid search until the power diagram input by the convolution layer and the output of the deconvolution layer When the dynamometer feature similarity reaches more than 90%, an optimized fully convolutional neural network model is obtained; the feature extraction module uses a dynamometer feature extractor to extract features from the dynamometer data.

A computer device includes a memory and a processor. The memory stores a computer program. It is characterized in that when the processor executes the computer program, the steps of the above method are implemented.

A computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the above method are implemented.

The beneficial effect of the present invention is to provide a dynamometer feature extraction method that uses a large number of unlabeled samples to improve learning performance when there is less labeled data, and uses a "convolution + deconvolution" fully convolutional neural network as a trainer. , measure the feature extraction performance based on the similarity of input and output features, and obtain a feature extractor that can accurately identify power diagrams, avoiding the problems of poor performance, slow speed, and difficulty in model optimization caused by direct learning of a large number of samples. In the subsequent problem of establishing oil well fault diagnosis based on dynamometer diagrams, only a small number of labeled samples can meet the needs of model learning, which reduces certain manpower and time costs in sample labeling and model optimization and improves the efficiency of dynamometer diagram identification. Accuracy.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the drawings needed to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. Those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting any creative effort. in:

Figure 1 is an overall flow chart of a method for extracting features of dynamometer diagrams based on a fully convolutional neural network according to an embodiment of the present invention.

FIG. 2 is a feature extraction flow chart of a feature extraction method for dynamometer diagrams based on a fully convolutional neural network provided by the first embodiment of the present invention.

Figure 3 is a system structure diagram of a power diagram feature extraction system based on a fully convolutional neural network provided by the first embodiment of the present invention.

Figure 4 is a schematic diagram of some dynamometer diagrams in the data set of a dynamometer diagram feature extraction method based on a fully convolutional neural network provided by the first embodiment of the present invention.

FIG. 5 is a schematic diagram of a fully convolutional neural network of a dynamometer feature extraction method based on a fully convolutional neural network provided by the first embodiment of the present invention.

Detailed ways

In order to make the above objects, features and advantages of the present invention more obvious and understandable, the specific implementation modes of the present invention will be described in detail below with reference to the accompanying drawings.

Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Those skilled in the art can do so without departing from the connotation of the present invention. Similar generalizations are made, and therefore the present invention is not limited to the specific embodiments disclosed below.

Second, "one embodiment" or "an embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. "In one embodiment" appearing in different places in this specification does not all refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

Example 1

Referring to Figures 1 and 2, a first embodiment of the present invention is provided. This embodiment provides a dynamometer feature extraction method based on a fully convolutional neural network, including: constructing a dynamometer image data set; constructing a full volume Convolutional neural network model, the model input and output are both dynamometer diagram images; use the constructed dynamometer diagram image data set to pre-train the fully convolutional neural network model, and the training evaluation index is to judge the dynamometer diagram image output by the model and Similarity between the input dynamometer images; use grid search to optimize the hyperparameters of the fully convolutional neural network model to optimize the extraction capability of the model; intercept the convolution stage of the fully convolutional neural network model as the dynamometer feature Extractor.

S1. Construct a dynamometer data set. According to the massive dynamometer displacement load data in the database, with the abscissa as the displacement and the ordinate as the load, standardized dynamometer images are drawn in batches, as shown in Figure 2.

S2. Construct a fully convolutional neural network model. The input of the model is a power diagram image. The model includes operation steps such as convolution, pooling, deconvolution, and depooling. The fully convolutional neural network model is based on the convolution stage and the deconvolution stage. The deconvolution stage It is the reverse process of the convolution stage. It decodes the output of the convolution layer. The output of the model is a dynamometer image. The input and output of the model are the same and are the same dynamometer image. Among them, the convolution stage uses but does not Limited to VGG16 as a feature extractor, when the 200*100*3 dynamometer diagram is input, the first layer of the VGG16 model extracts the features of the dynamometer diagram into 100*50*32 dimensions. After a series of feature extractions, the last layer The dynamometer characteristics are compressed to 7*4*512.

Furthermore, the deconvolution stage can be regarded as the inverse process of the convolution stage, by decoding the output of the convolution layer to obtain the reconstruction of the original input. The detailed process is divided into the following steps:

Padding: A padding operation is performed around the output feature map to keep the output the same size as the original input. Among them, the filling method adopts the "zero filling" method, which adds a circle of pixels with a value of zero around the output feature map.

Convolution: Convolution operation is performed on the filled feature map, usually using the same convolution kernel as the convolution layer. The size and stride of the convolution kernel are also usually the same as those of the convolutional layer. Among them, the purpose of convolution is to convolve the filled pixels with the input feature information to obtain more refined feature information.

Cropping: Finally, the edges of the output feature map need to be cropped to remove the filled parts and obtain the final deconvolution result.

Furthermore, the first layer of the deconvolution stage reads the features of the dynamometer diagram as 7*4*512 dimensions, and the last layer restores the features to 200*100*3.

S3. Fully convolutional neural network model pre-training. The fully convolutional neural network is pre-trained using the dynamometer data set, the weight parameters are initialized, and the dynamometer features output by the model convolution operation are continuously restored to the input dimensions through deconvolution operations.

Furthermore, after each round of training, the similarity between the features when the power diagram is input into the convolution layer and the features restored after deconvolution is calculated. The feature dimensions are both 200*100*3.

Wherein, the measurement method of the feature similarity is the Jaccard similarity coefficient. When the Jaccard similarity coefficient is closer to 1, it means that the two feature matrices are more similar.

Further, for the two feature matrices A and B, calculate the Jaccard similarity coefficient of each vector, and finally average it. The calculation formula of the Jaccard similarity coefficient is:

J(A,B)＝(J(A ₁ ,B ₁ )+J(A ₂ ,B ₂ )+...+J(A _n ,B _n ))/n

Furthermore, for two vectors A _n and B _n , consider them as two sets, their intersection is S, and their union is U, then the Jaccard similarity coefficient can be expressed as:

J(A _n ,B _n )=|S|/|U|

S4. Hyperparameter optimization. Perform hyperparameter optimization on the trained fully convolutional neural network model until the feature similarity between the dynamometer input by the convolution layer and the dynamometer output by the deconvolution layer reaches more than 90%, and the optimized fully convolutional model is obtained. Neural network model.

Furthermore, "grid search" is used as the hyperparameter optimization method of the fully convolutional neural network model, and the possible value range of each hyperparameter to be optimized is predefined.

Further, the hyperparameters that need to be optimized and their value ranges are defined as follows:

Batch size: batch_size_list=[10, 20, 30, 40, 50]

Inactivation rate: dropout_list=[0.3, 0.4, 0.5, 0.6, 0.7]

Learning rate: learning_rate_list=[0.1, 0.01, 0.001, 0.0001]

Activation function: activation_list=[sigmoid, tanh, relu]

Optimizer: optimizer_list=[sgd, Adadelta, Adam, RMSProp]

Further, "grid search" searches and evaluates by enumerating all possible hyperparameter combinations.

S5. Remove the deconvolution layer of the model. When the similarity of input and output features reaches 90%, it means that the 7*4*512 features extracted in the convolution stage can basically restore all the connotative information of the dynamometer diagram. At this time, the convolution stage of the fully convolutional neural network is intercepted as the dynamometer diagram. Graph feature extractor.

Furthermore, by utilizing the strong feature extraction capability of this model for dynamometer diagrams, in the subsequent problem of establishing oil well fault diagnosis based on dynamometer diagrams, only a small number of labeled samples can be relied upon to meet the needs of model learning, reducing the need for sample labeling and model There is a certain cost of manpower and time in training and optimization, and it can effectively improve the accuracy of dynamometer diagram recognition.

Example 2

Referring to Figure 3, a second embodiment of the present invention is shown. What is different from the previous embodiment is that a dynamometer diagram feature extraction system based on a fully convolutional neural network is provided, including: a dynamometer diagram input module, a preset Training module, hyperparameter optimization module, feature extraction module.

The dynamometer diagram input module is used to input the dynamometer diagram data of the features to be extracted.

The pre-training module uses the dynamometer image data set to pre-train the fully convolutional neural network, and restores the dynamometer features output by the model convolution operation to the input dimensions through the deconvolution operation, and calculates the dynamometer input volume. The Jaccard similarity coefficient is used as a measure of feature similarity between the features during stacking and the features restored after deconvolution.

The hyperparameter optimization module performs hyperparameter optimization on the trained fully convolutional neural network model and finds the best hyperparameter combination through grid search until the dynamism diagram input by the convolution layer and the dynamism diagram output by the deconvolution layer When the feature similarity reaches more than 90%, the optimized fully convolutional neural network model is obtained.

The feature extraction module uses the dynamometer feature extractor to extract features from the dynamometer data.

If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code.

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, may be considered a sequenced list of executable instructions for implementing the logical functions, and may be embodied in any computer-readable medium, For use by, or in combination with, instruction execution systems, devices or devices (such as computer-based systems, systems including processors or other systems that can fetch instructions from and execute instructions from the instruction execution system, device or device) or equipment. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

More specific examples (non-exhaustive list) of computer readable media include the following: electrical connections with one or more wires (electronic device), portable computer disk cartridges (magnetic device), random access memory (RAM), Read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium may even be paper or other suitable medium on which the program may be printed, as the paper or other medium may be optically scanned, for example, and subsequently edited, interpreted, or otherwise suitable as necessary. process to obtain the program electronically and then store it in computer memory.

It should be understood that various parts of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if it is implemented in hardware, as in another embodiment, it can be implemented by any one or a combination of the following technologies known in the art: a logic gate circuit having a logic gate circuit for implementing a logic function on a data signal. Discrete logic circuits, application specific integrated circuits with suitable combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

Example 3

Referring to Figures 4 and 5, the third embodiment of the present invention is different from the first two embodiments in that the technical effects used in the present invention are verified and explained to verify the real effects of the method.

The embodiment of the present invention uses sensors arranged at the wellhead to record the displacement and load of the sucker rod every 30 minutes, which continuously collects the original data of the power indicator diagram in real time and transmits it to the database.

Among them, the displacement and load recorded every 30 minutes each contain 200 sample points, and these 400 sample points constitute the vector of a dynamometer diagram. This embodiment obtained a total of 20,000 vectors in the database, as shown in Table 1:

Table 1 Displacement load data

Furthermore, using the first 200 sample points of the vector as the abscissa (displacement) and the last 200 sample points as the corresponding ordinate (load), standardized dynamometer images are drawn in batches. The image consists of three channels of red, green and blue. The size of each image is 200*100*3 (3 represents 3 channels), the aspect ratio is 2:1, there is no axis label, and it contains a variety of oil well fault types. , some images are shown in Figure 4.

As shown in Figure 5, the model contains two stages of "convolution + deconvolution". The convolution stage uses VGG16 as the feature extractor. When the 200*100*3 power diagram is input, the first layer of the VGG16 model The features of the dynamometer diagram are extracted into 100*50*32 dimensions. After a series of feature extractions, the features of the last layer of the dynamometer diagram are compressed into 7*4*512.

Furthermore, the first layer of the deconvolution stage reads the features of the dynamometer diagram as 7*4*512 dimensions, and the last layer restores the features to 200*100*3.

Furthermore, after each round of training, the similarity between the features when the power diagram is input into the convolution layer and the features restored after deconvolution is calculated. The feature dimensions are both 200*100*3.

Wherein, the measurement method of the feature similarity is the Jaccard similarity coefficient. When the Jaccard similarity coefficient is closer to 1, it means that the two feature matrices are more similar. After 10 rounds of training, the Jaccard similarity coefficient at the input and output ends of the model reached 0.6.

Further, for the two feature matrices A and B, calculate the Jaccard similarity coefficient of each vector, and finally average it. The calculation formula of the Jaccard similarity coefficient is:

J(A,B)＝(J(A ₁ ,B ₁ )+J(A ₂ ,B ₂ )+...+J(A _n ,B _n ))/n

Furthermore, for two vectors A _n and B _n , consider them as two sets, their intersection is S, and their union is U, then the Jaccard similarity coefficient can be expressed as:

J(A _n ,B _n )=|S|/|U|

Perform hyperparameter optimization on the trained fully convolutional neural network model until the feature similarity between the dynamometer input by the convolution layer and the dynamometer output by the deconvolution layer reaches more than 90%, and the optimized fully convolutional model is obtained. Neural network model.

Furthermore, "grid search" is used as the hyperparameter optimization method of the fully convolutional neural network model, and the possible value range of each hyperparameter to be optimized is predefined.

Further, the hyperparameters that need to be optimized and their value ranges are defined as follows:

Batch size: batch_size_list=[10, 20, 30, 40, 50]

Inactivation rate: dropout_list=[0.3, 0.4, 0.5, 0.6, 0.7]

Learning rate: learning_rate_list=[0.1, 0.01, 0.001, 0.0001]

Activation function: activation_list=[sigmoid, tanh, relu]

Optimizer: optimizer_list=[sgd, Adadelta, Adam, RMSProp]

Further, "grid search" searches and evaluates by enumerating all possible hyperparameter combinations. After 10 rounds of training, the best hyperparameter combination is: batch size: batch_size_list=20, inactivation rate: dropout_list=0.3, learning rate: learning_rate_list=0.01, activation function: activation_list=relu, optimizer: optimizer_list=sgd , at this time the Jaccard similarity coefficient of the input and output ends reaches 0.9.

When the similarity of input and output features reaches 90%, it means that the 7*4*512 features extracted in the convolution stage can basically restore all the connotative information of the dynamometer diagram. At this time, the convolution stage of the fully convolutional neural network is intercepted as the dynamometer diagram. Graph feature extractor.

Furthermore, by utilizing the strong feature extraction capability of this model for dynamometer diagrams, in the subsequent problem of establishing oil well fault diagnosis based on dynamometer diagrams, only a small number of labeled samples can be relied upon to meet the needs of model learning, reducing the need for sample labeling and model There is a certain cost of manpower and time in training and optimization, and it can effectively improve the accuracy of dynamometer diagram recognition.

It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solution of the present invention can be carried out. Modifications or equivalent substitutions without departing from the spirit and scope of the technical solution of the present invention shall be included in the scope of the claims of the present invention.

Claims (10)

1. A dynamometer feature extraction method based on a fully convolutional neural network, which is characterized by: including, Construct dynamometer image data set; Construct a fully convolutional neural network model, and the model input and output are both power diagram images; Use the constructed dynamometer image data set to pre-train the fully convolutional neural network model. The evaluation index of the training is to judge the similarity between the dynamometer image output by the model and the input dynamometer image; Use grid search to optimize hyperparameters of the fully convolutional neural network model to optimize the extraction capability of the model; The convolution stage of the fully convolutional neural network model is intercepted as the dynamometer feature extractor. 2. A dynamometer diagram feature extraction method based on a fully convolutional neural network as claimed in claim 1, characterized in that: the dynamometer diagram image data set includes, according to a large amount of dynamometer diagram displacement load data in the database , with the abscissa as the displacement and the ordinate as the load, draw standardized dynamometer images in batches to construct a dynamometer data set. 3. A method for extracting features of dynamometer diagrams based on a fully convolutional neural network as claimed in claim 2, characterized in that: the fully convolutional neural network model includes convolution, pooling, deconvolution, and anti-pooling. Operation steps, the input of the model is a dynamometer image, the deconvolution stage is the reverse process of the convolution stage, the output of the convolution layer is decoded, the output of the model is a dynamometer image, the input of the model The output is the same, which is the same dynamometer image. 4. A dynamometer feature extraction method based on a fully convolutional neural network as claimed in claim 3, characterized in that: the deconvolution stage includes filling, convolution and cropping. 5. A dynamometer feature extraction method based on a fully convolutional neural network as claimed in claim 4, characterized in that: the pre-training includes using the dynamometer image data set to The convolutional neural network is pre-trained, and the dynamometer characteristics output by the model convolution operation are restored to the dimensions of the input through the deconvolution operation. After each round of training, the characteristics of the dynamometer diagram when input into the convolution layer are calculated and deconvolved. Feature similarity of features after operation reduction; The feature similarity includes: the measurement method of feature similarity is Jaccard similarity coefficient; For the two feature matrices A and B, calculate the Jaccard similarity coefficient of each vector and average it. The calculation formula of the Jaccard similarity coefficient is expressed as, J(A,B)＝(J(A ₁ ,B ₁ )+J(A ₂ ,B ₂ )+...+J(A _n ,B _n ))/n For two vectors A _n and B _n , as two sets, the Jaccard similarity coefficient is expressed as, J(A _n ,B _n )=|S|/|U| Among them, S is the intersection of sets A _n and B _n , and U is the union of sets A _n and B _n ; If the feature similarity exceeds 90%, the convolution stage of the fully convolutional neural network model is intercepted. If the feature similarity is less than 90%, hyperparameter optimization is performed. 6. A dynamometer feature extraction method based on a fully convolutional neural network as claimed in claim 5, characterized in that: the hyperparameter optimization includes performing hyperparameter optimization on the fully convolutional neural network model, Until the feature similarity between the dynamometer diagram input by the convolution layer and the dynamometer diagram output by the deconvolution layer reaches more than 90%, the optimized fully convolutional neural network model is obtained; Grid search is used as a hyperparameter optimization method for the fully convolutional neural network model, and the value range of each hyperparameter to be optimized is predefined. 7. A dynamometer feature extraction method based on a fully convolutional neural network as claimed in claim 6, characterized in that: the dynamometer feature extractor includes deconvolution that removes the fully convolutional neural network model. layer to obtain a dynamometer feature extractor based on the convolution stage of the fully convolutional neural network model. 8. A system using a dynamometer feature extraction method based on a fully convolutional neural network as described in any one of claims 1 to 7, characterized in that it includes: a dynamometer input module, a pre-training module, Hyperparameter optimization module and feature extraction module; The dynamometer diagram input module is used to input dynamometer diagram data of features to be extracted; The pre-training module uses the dynamometer image data set to pre-train the fully convolutional neural network, and restores the dynamometer features output by the model convolution operation to the input dimensions through the deconvolution operation, and calculates the dynamometer The similarity between the features input to the convolutional layer and the features restored after the deconvolution operation is measured using the Jaccard similarity coefficient as a measure of feature similarity; The hyperparameter optimization module performs hyperparameter optimization on the trained fully convolutional neural network model and finds the best hyperparameter combination through grid search until the performance diagram input by the convolution layer and the output diagram of the deconvolution layer are When the functional map feature similarity reaches more than 90%, the optimized fully convolutional neural network model is obtained; The feature extraction module uses a dynamometer feature extractor to extract features from the dynamometer data. 9. A computer device, comprising a memory and a processor, the memory stores a computer program, characterized in that: when the processor executes the computer program, the method of any one of claims 1 to 7 is implemented. step. 10. A computer-readable storage medium with a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.