阅读说明：本技术 一种基于深度信念网络的自动送钻方法 (Automatic bit feeding method based on deep belief network ) 是由 张译戈 于 2020-11-30 设计创作，主要内容包括：本发明公开了一种基于深度信念网络的自动送钻方法,首先建立DBN网络模型,构建DBN-PID控制器,将在钻井过程与井底钻压有关联的4个重要参数,钻头直径、井深、转盘转速、泥浆排量以及钻压,进行预处理后输入DBN网络进行训练,DBN根据当前系统运行状态,在线计算使控制系统达到性能要求的PID参数,PID控制器直接作用于主控制器,主控制器输出给定的速度控制信号送给变频器,然后控制钻头钻压,最终使钻头的钻压会达到希望的钻压值。从而提高整个网络的鲁棒性和适应性,达到提高钻速,减小钻头磨损,降低钻井成本的目的。(The invention discloses an automatic bit feeding method based on a depth belief network, which comprises the steps of firstly establishing a DBN network model, constructing a DBN-PID controller, preprocessing 4 important parameters related to the drilling pressure at the bottom of a well in the drilling process, namely the diameter of a drill bit, the depth of the well, the rotating speed of a rotary table, the slurry discharge capacity and the drilling pressure, inputting the preprocessed parameters into the DBN network for training, calculating PID parameters enabling a control system to meet performance requirements on line by the DBN according to the current system running state, directly acting on a main controller by the PID controller, outputting a given speed control signal to a frequency converter by the main controller, controlling the drilling pressure of the drill bit, and finally enabling the drilling pressure of the drill bit to reach a desired drilling pressure value. Therefore, the robustness and adaptability of the whole network are improved, and the purposes of improving the drilling speed, reducing the abrasion of the drill bit and reducing the drilling cost are achieved.)

1. An automatic bit feeding method based on a deep belief network is characterized by comprising the following steps:

s1: 4 important parameters, namely the diameter of a drill bit, the well depth, the rotating speed of a rotary table, the slurry discharge capacity and the drilling pressure, which are related to the drilling pressure at the bottom of a well in the drilling process are taken as a sample data set after pretreatment, and the sample data set is taken as an input variable of a DBN (direct bonded network), namely an input layer node of the network;

s2: constructing a deep belief network model, wherein the deep belief network model is composed of two layers of limited Boltzmann machines (RBMs) and one layer of BP neural network;

sending the sample data set in the S1 into a deep belief network model for training, thereby adjusting the parameters of the PID;

outputting a rotating speed control signal;

s3: the main controller outputs a given speed control signal to the frequency converter to control the rotating speed of the motor so as to control the bit pressure of the drill bit;

finally, the purpose of automatic bit feeding with constant bit pressure is achieved.

2. The automatic bit feeding method based on the deep belief network as claimed in claim 1, characterized in that: the input data in step S1 includes the bit diameter, well depth, rotary table rotation speed, mud displacement, and weight on bit, and is normalized.

3. The automatic bit feeding method based on the deep belief network as claimed in claim 1, characterized in that: the DBN network consists of two layers of limited Boltzmann machines (RBMs) and one layer of BP neural network, and the training process of the DBN model can be summarized into the following two steps:

firstly, performing unsupervised pre-training on the RBM, wherein the RBM is trained by adopting a contrast divergence method;

secondly, the weight is supervised and optimized, and the BP neural network adopts a gradient descent method to carry out fine adjustment and optimization.

4. The automatic bit feeding method based on the deep belief network as claimed in claim 1, characterized in that: the training of the DBN is divided into the following three steps:

inputting preprocessed data serving as a training sample into a visual layer of a first layer RBM, and training the first layer RBM by using a contrast divergence algorithm until an energy function is converged;

after the first layer of RBM is trained, network parameters are fixed, and the hidden layer of the first layer of RBM is taken as the visual layer of the second RBM; extracting preliminary characteristic data from the sample data set through a first-layer RBM for inputting a second-layer RBM, and training a second-layer RBM network by using a contrast divergence algorithm until an energy function is converged;

and thirdly, calculating the error between the actual output and the theoretical output, and carrying out fine adjustment and optimization by using a BP neural network and adopting a gradient descent method.

5. The automatic bit feeding method based on the deep belief network as claimed in claim 1, characterized in that: in a traditional PID control system, a DBN-PID controller capable of self-learning the 3 parameters of proportion, integration and differentiation is constructed and used for intelligent control of the bit pressure of automatic bit feeding of the oil drilling machine; the DBN network calculates PID parameters enabling the control system to meet performance requirements on line according to the current system running state, the PID controller directly acts on the main controller, the main controller outputs a given speed control signal to the frequency converter, then the bit pressure of the drill bit is controlled, and finally the bit pressure of the drill bit can reach a given value;

u(k)＝u(k-1)+K_p[e(k)-e(k-1)]+K_ie(k)+K_d[e(k)-2e(k-1)+e(k-2)]

wherein u (K) is a PID control response output signal, K_pIs a proportionality coefficient, K_iIs the integral coefficient, K_dIs a differential coefficient, e (k) is a velocityAnd (5) degree deviation.

6. The automatic bit feeding method based on the deep belief network as claimed in claim 1, characterized in that: and comparing the expected bit pressure value with the actually acquired bit pressure value, and then carrying out feedback correction to finally enable the actual bit pressure to be equal to or close to the expected bit pressure so as to achieve the purpose of constant bit pressure bit feeding.

Technical Field

The invention relates to the technical field of petroleum drilling, in particular to an automatic bit feeding method based on a deep belief network.

Background

The constant bit pressure automatic bit feeding technology is that the pressure of a drill bit to the bottom of a well is kept to be a constant value according to the requirements of a drilling process in the normal drilling process of a drilling machine. The expected control effect can be achieved by adopting a PID (proportion Integration differentiation) controller with appropriate parameters. However, the control model of the system is not constant during the drilling process, so in order to achieve better control effect, the parameters of the PID need to be corrected. However, it is difficult to properly calibrate the PID parameters, which is very demanding for the field operator, and results in less than optimal use of automatic bit delivery.

During drilling, the weight-on-bit can be influenced by various factors including geological structure, drilling fluid properties, shaft friction and the like, and the bit feeding motor has the characteristics of multivariable, strong coupling and nonlinearity. The drilling process is a time-varying process, and the modeling work is very difficult, so that the control effect achieved by only adopting the conventional PID control is not ideal.

In recent years, the development of deep belief network learning is very rapid, and the deep belief network learning has great development and progress in many fields.

Disclosure of Invention

The invention aims to provide an automatic bit feeding method based on a depth Belief network, and provides a method for adjusting PID parameters by using a DBN (deep Belief nets) network aiming at the problem that the control effect of the traditional PID control is not ideal, wherein the DBN calculates the PID parameters which enable a control system to meet the performance requirements on line according to the current system running state, the PID controller directly acts on a main controller, the main controller outputs a given speed control signal to a frequency converter, then the bit pressure of a bit is controlled, and finally the bit pressure of the bit can reach a given value, so that the purpose of constant bit pressure drilling is realized.

In order to achieve the purpose, the invention provides the following technical scheme: an automatic bit feeding method based on a deep belief network comprises the following steps:

s1: 4 important parameters, namely the diameter of a drill bit, the well depth, the rotating speed of a rotary table, the slurry discharge capacity and the drilling pressure, which are related to the drilling pressure at the bottom of a well in the drilling process are taken as a sample data set after pretreatment, and the sample data set is taken as an input variable of a DBN (direct bonded network), namely an input layer node of the network;

s2: constructing a deep belief network model, wherein the deep belief network model is composed of two layers of limited Boltzmann machines (RBMs) and one layer of BP neural network;

sending the sample data set in the S1 into a deep belief network model for training, thereby adjusting the parameters of the PID;

outputting a rotating speed control signal;

s3: the main controller outputs a given speed control signal to the frequency converter to control the rotating speed of the motor so as to control the bit pressure of the drill bit;

finally, the purpose of automatic bit feeding with constant bit pressure is achieved.

Preferably, in S1, the input data is preprocessed by using a scaling method. Namely, normalization processing, which limits the input data and output data of the network to the interval of [0,1 ].

Preferably, the maximum value and the minimum value of each component in the whole sample data set are determined, and normalization processing is performed, wherein the variation of normalization is as follows:wherein x is_iRepresenting input or output data, x_minMinimum value, x, representing variation of data_maxRepresenting the maximum value of the data.

Preferably, in S1, the complexity of the input-output nonlinear mapping relationship determines the selection of the training samples, and 5 to 10 times of the total number of the network connection weights are selected as the training samples.

Preferably, in S2, the deep belief network DBN training is performed according to the following steps:

determining the structure and parameters of the DBN network;

taking the sample data set after normalization processing as the input of a network, taking the rotating speed of a motor as the output of the network, and firstly carrying out unsupervised pre-training by using a RBM (simplified Boltzmann machine);

inputting the preprocessed data serving as a training sample into a visual layer of a first layer RBM, and training the first layer RBM network by using a contrast divergence algorithm until an energy function is converged;

after the first layer of RBM is trained, network parameters are fixed, and the hidden layer of the first layer of RBM is taken as the visual layer of the second RBM;

extracting preliminary characteristic data from the sample data set through a first-layer RBM for inputting a second-layer RBM, and training a second-layer RBM network by using a contrast divergence algorithm until an energy function is converged;

calculating the error between the actual output and the theoretical output, and carrying out fine adjustment and optimization by using a BP (Back propagation) neural network and adopting a gradient descent method; the energy function of the RBM can be expressed as:

wherein, ω is_ijRepresenting the weight between the ith node and the jth node, c_jAnd b_iFor offset, I and J are the number of visible and hidden layer elements, respectively.

Preferably, in S3, on the basis of a conventional PID controller, a DBN-PID controller capable of automatically setting PID parameters is designed, the PID controller directly acts on the main controller, the main controller outputs a given speed control signal to the frequency converter, and then controls the bit pressure of the drill bit, so that the bit pressure of the drill bit can finally reach a given value; the purpose of constant bit pressure drilling is achieved.

Preferably, in S3, the parameter K of the PID controller is controlled by the DBN network algorithm according to the operation status of the system by using incremental digital PID control_p、K_i、K_dAdjusting in real time; the following were used:

u(k)＝u(k-1)+K_p[e(k)-e(k-1)]+K_ie(k)+K_d[e(k)-2e(k-1)+e(k-2)]

wherein u (K) is a PID control response output signal, K_pIs a proportionality coefficient, K_iIs the integral coefficient, K_dIs the differential coefficient, and e (k) is the velocity deviation.

Compared with the prior art, the invention has the beneficial effects that:

the invention provides an automatic bit feeding method based on a depth belief network, which is characterized in that a current actual bit pressure value is detected and compared with a desired bit pressure value, and a comparison result is sent to a DBN-PID controller so as to adjust the speed of a winch, so that the actual bit pressure value is infinitely close to the desired bit pressure value, thereby achieving the purposes of improving the drilling speed, reducing the abrasion of a drill bit and reducing the drilling cost. And training the DBN network by combining a contrast divergence algorithm, and then performing reverse whole network model tuning by using BP (Back propagation) so as to improve the robustness and adaptability of the whole network.

Drawings

FIG. 1 is a flow chart of a method of automatic drill delivery based on deep learning;

FIG. 2 is a schematic diagram of a DBN network architecture;

FIG. 3 is a schematic diagram of a DBN-PID control architecture;

FIG. 4 is a schematic diagram of an automatic bit feed process based on DBN-PID control.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to fig. 1-4, the present invention provides a technical solution: an automatic bit feeding method based on a deep belief network mainly comprises the following steps:

s1: 4 important parameters, namely the diameter of a drill bit, the well depth, the rotating speed of a rotary table, the slurry discharge capacity and the drilling pressure, which are related to the drilling pressure at the bottom of a well in the drilling process are taken as a sample data set after pretreatment, and the sample data set is taken as an input variable of a DBN (direct bonded network), namely an input layer node of the network;

s2: and constructing a deep belief network model, wherein the deep belief network model is formed by two layers of limited Boltzmann machines (RBMs) and one layer of BP neural network. And feeding the sample data set in the S1 into a deep belief network model for training so as to adjust the parameters of the PID. Outputting a rotating speed control signal;

s3: the main controller outputs a given speed control signal to the frequency converter to control the rotating speed of the motor, and further control the bit weight of the drill bit. Finally, the purpose of automatic bit feeding with constant bit pressure is achieved.

In S1, the input data is preprocessed by scaling. Namely, normalization processing, which limits the input data and output data of the network to the interval of [0,1 ]. Determining the maximum value and the minimum value of each component in the whole sample data set, and carrying out normalization processing, wherein the variation of normalization is as follows:

wherein x is_iRepresenting input or output data, x_minMinimum value, x, representing variation of data_maxRepresenting the maximum value of the data.

In S1, the complexity of the input-output nonlinear mapping relation determines the selection of training samples, and 5-10 times of the total number of network connection weights is selected as the training samples.

In S2, the structure and parameters of the DBN network are determined first, and the network structure diagram of the DBN is shown in fig. 2; and then inputting the sample data set after normalization processing into a DBN network, starting to train a first RBM, fixing the weight and offset of the first RBM after the training is finished, then taking the state of a recessive neuron as an input vector of a second RBM, starting to train the second RBM, and finally carrying out fine adjustment and optimization on the whole DBN network by using a BP neural network through a gradient descent method.

The training process of the DBN model can be summarized as the following two steps:

RBM unsupervised pre-training, wherein the RBM is trained by adopting a contrast divergence method, and the training process is as follows: according toTo calculate Q (h)_1i＝1|x₁) And sampling it h_1iWherein h is_1i∈(0，1)；

According toTo calculate P (x)_2j＝h₁) And sampling it x_2jWherein x is_2j∈(0，1)；

Calculate Q (h)_2i＝1|x₂) And updating the weight according to the following formula:

W←W+ω(h₁x₁-Q(h₂＝1|x₂)x₂)

b←b+ω(x₁-x₂)

c←c+ω(h₁-Q(h₂＝1|x₂))

wherein the content of the first and second substances,is a sigmoid logic function.

Wherein x is all samples of the training sample set; omega is the learning rate in the gradient descent method; h is a hidden layer unit vector; w is a weight matrix of the RBM; b is an input offset vector; c is the output offset vector.

Secondly, the supervised weight is optimized: calculating the error between the actual output and the theoretical output, and performing fine adjustment and optimization by using a BP neural network and adopting a gradient descent method; in the unsupervised layer-by-layer training process, the parameter theta of the RBM cannot be guaranteed to be { omega ═ omega_ij,a_i,b_jGlobal optimality can be achieved.

Therefore, data received by the BP neural network positioned at the top layer is used as input, then errors between theoretical output and actual output are transmitted layer by layer to the RBM along the opposite direction, each layer can conduct optimization on the parameter theta, and finally theta ← theta + delta theta is enabled to achieve global optimization.

In S3, a DBN-PID controller is constructed in which a deep belief network model and a PID controller are combined, a structure diagram of the DBN-PID control is shown in fig. 3, and a control algorithm is as follows:

determining a network structure of the DBN, namely determining that the number of nodes of an input layer is 5, the number of nodes of an output layer is 3 and the number of hidden layer layers is 2;

obtaining R by sampling_inAnd Y_out. Then, the time error e ═ R is calculated_in-Y_out；

Inputting the differentiated error into a DBN network, wherein the output of the DBN network can adjust three parameters of a PID controller;

finally, outputting u (k) which is the output of the PID controller;

training the DBN, and adjusting a weight coefficient on line;

and (c) making k equal to k +1, stopping training the network only when the precision of the system is the same as the set precision, and otherwise, returning to the first step for re-execution.

In S3, the PID controller acts directly on the main controller, the main controller outputs a given speed control signal to the frequency converter, and then the bit pressure of the drill bit is controlled, and finally the bit pressure of the drill bit reaches a given value. The purpose of constant bit pressure drilling is achieved.

In S3, incremental digital PID control is adopted, and according to the running state of the system, the parameter K of the PID controller is subjected to the DBN network algorithm_p、K_i、K_dAnd performing real-time adjustment. The following were used:

u(k)＝u(k-1)+K_p[e(k)-e(k-1)]+K_ie(k)+K_d[e(k)-2e(k-1)+e(k-2)]

wherein u (K) is a PID control response output signal, K_pIs a proportionality coefficient, K_iIs the integral coefficient, K_dIs the differential coefficient, and e (k) is the velocity deviation.

As shown in fig. 4, in S3, the DBN-PID-based automatic bit-feeding process compares the desired weight-on-bit value with the actually collected weight-on-bit value, and then performs feedback correction to finally make the actual weight-on-bit equal to or close to the desired weight-on-bit, thereby achieving the purpose of constant weight-on-bit feeding.

It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein, and any reference signs in the claims are not intended to be construed as limiting the claim concerned.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.