Drilling correction prediction method and system based on LSTM (Long Short Term Memory)

Abstract

The invention discloses a drilling deviation rectification prediction method and system based on LSTM, and belongs to the technical field of data recognizing.A drilling posture deviation prediction model is constructed through the LSTM, training data are adopted to train the drilling posture deviation prediction model, the trained drilling posture deviation prediction model is obtained, then real-time multi-source data are collected, and the real-time multi-source data are obtained; real-time multi-source data is acquired, normalization processing is carried out on the real-time multi-source data, the normalized real-time multi-source data is input into the trained drilling posture deviation prediction model to be recognized, a drilling posture deviation prediction result is acquired, and finally the drilling posture deviation prediction result serves as a drilling deviation correction prediction result. According to the method and the device, the drilling deviation rectification prediction result is transmitted to equipment specified by a worker, so that the worker performs drilling deviation rectification, the time sequence characteristics in the drilling process can be more accurately captured by utilizing the strong memory ability and the nonlinear fitting ability of the LSTM, and the drilling deviation rectification efficiency and accuracy are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data recognition, and particularly relates to a drilling deviation correction prediction method and system based on LSTM. Background Art

[0002] In geological exploration and drilling engineering, the accuracy and stability of drilling are of great significance for resource exploitation and geological research. However, due to the complexity and uncertainty of geological conditions, deviations often occur during the drilling process, resulting in exploration failures or resource waste. Although the mainstream solutions in the prior art introduce inclination sensors and drilling pressure monitoring, multi-source data (ground penetrating radar, vibration, gyroscope) are only independently processed, and spatio-temporal feature fusion is not achieved. For example, the correlation between the inclination change rate and fracture development is not quantified, resulting in the loss of key warning signals, and the deflection response delay when the drill passes through the fractured zone exceeds 200 ms. The prediction models based on traditional machine learning (such as support vector machines, random forests, etc.) are limited by the short-term memory characteristics and cannot effectively capture the coupling effect of the gradual change of formation inclination and drill tool wear. The cumulative deviation in the middle and late stages of drilling increases by 2-3 times compared with the initial stage, seriously restricting the construction accuracy of deep holes (>50 m). Summary of the Invention

[0003] The present invention provides a drilling deviation correction prediction method and system based on LSTM to solve the technical problem of poor drilling deviation correction accuracy in the prior art.

[0004] A drilling deviation correction prediction method based on LSTM includes: Obtaining historical multi-source data related to drilling and the corresponding historical attitude deviation of the historical multi-source data, and preprocessing the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data; Constructing a drilling attitude deviation prediction model using LSTM, and training the drilling attitude deviation prediction model with the training data to obtain a trained drilling attitude deviation prediction model; Collecting real-time multi-source data, normalizing the real-time multi-source data, and inputting the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for identification to obtain a drilling attitude deviation prediction result; Taking the drilling attitude deviation prediction result as the drilling deviation correction prediction result, and transmitting the drilling deviation correction prediction result to a device designated by the staff to enable the staff to perform drilling deviation correction.

[0005] Further, preprocessing the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data, including: Normalize the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain the historical multi-source data after normalization and the corresponding historical attitude deviation of the historical multi-source data; Construct training samples using the historical multi-source data after normalization, construct training labels using the corresponding historical attitude deviation of the historical multi-source data after normalization, and form training data by combining the training samples and the training labels.

[0006] Furthermore, construct a drilling attitude deviation prediction model using LSTM, and train the drilling attitude deviation prediction model using the training data to obtain the trained drilling attitude deviation prediction model, including: Construct a drilling attitude deviation prediction model using LSTM, and initialize the encoding of the model parameters of the drilling attitude deviation prediction model using a sequence initialization mechanism to obtain multiple different parameter encodings; Obtain the fitness corresponding to each parameter encoding using the training data, and determine the parameter encoding with the maximum fitness as the optimal parameter encoding; Based on the optimal parameter encoding, adaptively select the search space corresponding to the parameter encoding, and determine the parameter encoding after selecting the search space; Perform information fusion on the parameter encoding after selecting the search space using a multi-segment variable-speed information fusion method to determine the parameter encoding after information fusion; Perform joint search on the parameter encoding after information fusion using a multi-direction joint search method to determine the parameter encoding after joint search; Perform global search on the parameter encoding after joint search using a position crossover transformation method to determine the parameter encoding after global search; Judge whether the current training times have reached the maximum training times. If so, re-determine the optimal parameter encoding according to the parameter encoding after global search. Otherwise, return to the step of obtaining the optimal parameter encoding; Use the model parameters in the re-determined optimal parameter encoding as the final model parameters of the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model.

[0007] Furthermore, initialize the encoding of the model parameters of the drilling attitude deviation prediction model using a sequence initialization mechanism to obtain multiple different parameter encodings, including:

[0008] where, represents the i -th dimension model parameter of the d -th parameter encoding, d = 1, 2, …, D, D represents the total dimension of the model parameters in the parameter encoding, andi When = 1, the i th parameter is encoded as a randomly generated parameter encoding. It represents the i th dimension model parameter of the d +1th parameter encoding. represents the remainder function. represents pi, and sin represents the sine function.

[0009] Furthermore, training data is used to obtain the fitness corresponding to each parameter encoding, and the parameter encoding with the maximum fitness is determined as the optimal parameter encoding, including: For any parameter encoding, after applying the model parameters in the parameter encoding to the drilling attitude deviation prediction model, the training samples in the training data are used as inputs, and the training labels in the training data are used as the expected outputs to obtain the cross-entropy loss function value. After adding the cross-entropy loss function value to a preset constant, a non-zero intermediate parameter is obtained, and the reciprocal operation is performed on the non-zero intermediate parameter to obtain the fitness corresponding to the parameter encoding. Traverse all parameter encodings, obtain the fitness corresponding to each parameter encoding, and determine the parameter encoding with the maximum fitness as the optimal parameter encoding.

[0010] Furthermore, based on the optimal parameter encoding, the search space corresponding to the parameter encoding is adaptively selected, and the parameter encoding after the search space is selected is determined, including: According to the current training times, the adaptive search space selection factor is obtained as:

[0011] where represents the adaptive search space selection factor at the t th training, and t when = 1, it is set to a fixed initial value. represents the adaptive search space selection factor at the t -1th training. represents the natural constant, and T represents the maximum number of trainings. According to the adaptive search space selection factor and the optimal parameter encoding, the parameter encoding after the search space is selected is obtained as:

[0012]

[0013] where represents the t th parameter encoding at the m th training. Denotes the parameter encoding after the m th selection search space, m m = 1, 2, …, NP, where NP represents the total number of parameter encodings, Denotes the optimal parameter encoding, Denotes the first random number between (0, 1), Denotes the second random number between (0, 2 ). Denotes the third random number between (0, 1), and sin represents the sine function, Denotes the optimal parameter encoding, Denotes the adaptive position selection factor, Denotes the exponential function with the natural constant e as the base, Denotes the t th fitness corresponding to the m th parameter encoding during the Denotes the fitness corresponding to the optimal parameter encoding, Denotes another parameter encoding that is closest to the parameter encoding in terms of Euclidean distance, Denotes the fourth random number between (0, 1).

[0014] Furthermore, a multi-stage variable-speed information fusion method is adopted to fuse the parameter encodings after the selection search space, and the parameter encodings after the information fusion are determined, including: According to the current training times, the multi-stage variable-speed information fusion factor is obtained as:

[0015] where, Denotes the multi-stage variable-speed information fusion factor, Denotes the preset maximum value of the multi-stage variable-speed information fusion factor, Denotes the preset minimum value of the multi-stage variable-speed information fusion factor, Denotes the cosine function, Denotes the first training times threshold, Denotes the second training times threshold, and 1 < < < T, Denotes the multi-stage variable-speed information fusion factor when the training times is ; According to the multi-stage variable-speed information fusion factor, the parameter encodings after the selection search space are fused, and the parameter encodings after the information fusion are determined as:

[0016]

[0017]

[0018] Among them, represents the parameter encoding after the t th selection search space during the k th training, represents the parameter encoding after the k th information fusion, k n = 1, 2, …, NP, represents the first random parameter encoding except for the parameter encoding , represents the fifth random number between (0, ), represents the sixth random number between (0, 2 ), represents the first information fusion coefficient, represents the second information fusion coefficient, represents the first constant term and is set to - ; represents the second constant term and is set to ; represents the third constant term and is set to .

[0019] Furthermore, a multi-directional joint search method is adopted to jointly search the parameter encoding after information fusion to determine the parameter encoding after joint search, including: According to the current training times, obtain the joint search control factor as:

[0020] According to the joint search control factor, obtain the adaptive joint search speed as:

[0021]

[0022]

[0023] Among them, represents the t th parameter encoding after information fusion during the n th training, n n = 1, 2, …, NP, represents the search speed corresponding to the parameter encoding t during the th training, represents the adaptive joint search speed corresponding to the parameter encoding , Denote the sign function, Denote the parameter encoding The corresponding first neighborhood position encoding, Denote the parameter encoding The corresponding second neighborhood position encoding, Denote the fitness corresponding to the first neighborhood position encoding, Denote the fitness corresponding to the second neighborhood position encoding, Denote the seventh random number between (0, 1); Obtain the multi-direction search speed by combining the historical optimal direction and the current optimal direction as:

[0024] Wherein, Denote the t +(1)th training parameter encoding The corresponding search speed, Denote the preset first learning rate, Denote the preset second learning rate, Denote the eighth random number between (0, 1), Denote the ninth random number between (0, 1), Denote the parameter encoding The corresponding historical optimal position; According to the adaptive combined search speed and the multi-direction search speed, perform a combined search on the parameter encoding after information fusion, and determine the parameter encoding after the combined search as:

[0025] Wherein, Denote the n th parameter encoding after the combined search, Denote the tenth random number between (0, 1).

[0026] Furthermore, adopt the position crossover transformation method to perform a global search on the parameter encoding after the combined search, and determine the parameter encoding after the global search as: Adopt the position crossover transformation method to perform a global search on the parameter encoding after the combined search, and obtain the global search encoding as:

[0027]

[0028] Wherein, Denote the t th training j th parameter encoding after the combined search, Denote the jThe global search code corresponding to the parameter coding after combined search, j = 1, 2, …, NP, denotes the second random parameter coding except for the parameter coding and beyond, denotes the eleventh random number between (0, 1), denotes the twelfth random number between (0, 1), denotes the thirteenth random number between (0, 1), denotes the position transformation coding, denotes the parameter upper limit coding, denotes the parameter lower limit coding; When the fitness of the global search code is greater than the fitness of the parameter coding after the corresponding combined search, the global search code is used as the parameter coding after global search; otherwise, the original parameter coding after combined search is directly used as the parameter coding after global search.

[0029] On the other hand, the present invention provides an LSTM-based drilling deviation correction prediction system, including: a training data acquisition module, a model training module, a data prediction module, and a data feedback module; The training data acquisition module is used to acquire historical multi-source data related to drilling and the corresponding historical attitude deviation of the historical multi-source data, and preprocess the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data; The model training module is used to construct a drilling attitude deviation prediction model using LSTM, and train the drilling attitude deviation prediction model using the training data to obtain the trained drilling attitude deviation prediction model; The data prediction module is used to collect real-time multi-source data, perform normalization processing on the real-time multi-source data, input the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for identification, and obtain the drilling attitude deviation prediction result; The data feedback module is used to use the drilling attitude deviation prediction result as the drilling deviation correction prediction result, and transmit the drilling deviation correction prediction result to the device designated by the staff, so that the staff can perform drilling deviation correction.

[0030] A drilling deviation correction prediction method based on LSTM provided by the present invention constructs a drilling attitude deviation prediction model through LSTM, and uses training data to train the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model. Then, real-time multi-source data is collected, and the real-time multi-source data is normalized. The normalized real-time multi-source data is input into the trained drilling attitude deviation prediction model for recognition to obtain the drilling attitude deviation prediction result. Finally, the drilling attitude deviation prediction result is used as the drilling deviation correction prediction result, and the drilling deviation correction prediction result is transmitted to the device designated by the staff to enable the staff to perform drilling deviation correction. Utilizing the powerful memory ability and non-linear fitting ability of LSTM, it can more accurately capture the time series characteristics during the drilling process, effectively improving the efficiency and accuracy of drilling deviation correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present invention, and are used in conjunction with the specification to explain the principles of the present invention.

[0032] Figure 1 It is a flowchart of a drilling deviation correction prediction method based on LSTM provided by an embodiment of the present invention.

[0033] Figure 2 It is a schematic structural diagram of a drilling deviation correction prediction system provided by an embodiment of the present invention.

[0034] Among them, 201 - training data acquisition module, 202 - model training module, 203 - data prediction module, 204 - data feedback module.

[0035] Through the above drawings, specific embodiments of the present invention have been shown, and there will be more detailed descriptions hereinafter. These drawings and text descriptions are not intended to limit the scope of the inventive concept in any way, but to illustrate the concept of the present invention to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0037] The embodiments of the present invention will be described in detail below with reference to the drawings.

[0038] As Figure 1As shown in the figure, an embodiment of the present invention provides a drilling deviation correction prediction method based on LSTM, including: S101. Obtain historical multi-source data related to drilling and the corresponding historical attitude deviation of the historical multi-source data, and preprocess the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data; The historical multi-source data related to drilling may include time series data such as drilling parameters (such as drilling pressure, rotation speed, vibration) and geological features (such as lithology changes). Further, it may also include data such as drilling attitude and working condition parameters to increase the prediction accuracy, and the historical attitude deviation may be the drilling azimuth deviation or the pitch angle deviation. These data can be collected based on a preset data sampling frequency and at a fixed data length to form historical multi-source data.

[0039] By normalizing the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data, the data complexity can be effectively reduced and the data recognition efficiency can be improved.

[0040] S102. Use LSTM (Long Short-Term Memory Network) to construct a drilling attitude deviation prediction model, and use the training data to train the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model; LSTM (Long Short-Term Memory Network) is a special recurrent neural network (RNN), specifically designed to solve the problem of gradient disappearance / explosion that occurs in traditional RNNs when processing long sequence data, and can effectively capture long-term dependencies in time series. Therefore, the embodiment of the present invention uses LSTM to construct a drilling attitude deviation prediction model, which can effectively improve the accuracy of drilling deviation correction.

[0041] It should be noted that using LSTM to construct a drilling attitude deviation prediction model is only a preferred example of the embodiment of the present invention. Other neural networks can also be used to construct a drilling attitude deviation prediction model. For example, the number of LSTMs can be increased and an attention mechanism can be added to improve the data prediction accuracy.

[0042] S103. Collect real-time multi-source data, normalize the real-time multi-source data, and input the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for recognition to obtain the drilling attitude deviation prediction result; Since the historical multi-source data has been normalized and then the drilling attitude deviation prediction model is trained, normalizing the real-time multi-source data can ensure accurate data recognition.

[0043] S104. Use the predicted result of the drilling attitude deviation as the predicted result of drilling deviation correction, and transmit the predicted result of drilling deviation correction to the device designated by the staff, so that the staff can perform drilling deviation correction.

[0044] In the embodiment of the present invention, by using the powerful memory ability and non-linear fitting ability of LSTM, the time series characteristics in the drilling process can be captured more accurately, and the efficiency and accuracy of drilling deviation correction can be effectively improved.

[0045] In the embodiment of the present invention, preprocessing the historical multi-source data and the historical attitude deviation corresponding to the historical multi-source data to obtain training data, including: Perform normalization processing on the historical multi-source data and the historical attitude deviation corresponding to the historical multi-source data to obtain the historical multi-source data after normalization processing and the historical attitude deviation corresponding to the historical multi-source data after normalization processing; Construct training samples using the historical multi-source data after normalization processing, construct training labels using the historical attitude deviation corresponding to the historical multi-source data after normalization processing, and form training data by combining the training samples and the training labels.

[0046] In the embodiment of the present invention, an LSTM is used to construct a drilling attitude deviation prediction model, and the training data is used to train the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model, including: Use LSTM to construct a drilling attitude deviation prediction model, and use a sequence initialization mechanism to perform initialization coding on the model parameters of the drilling attitude deviation prediction model to obtain multiple different parameter codings; For example, the model parameters (i.e., hyperparameters) of LSTM generally have corresponding upper and lower limits, which can be randomly initialized between the upper and lower limits, and then the initialized model parameters are encoded into vectors to obtain parameter codings. After repeating multiple times, multiple different parameter codings can be obtained.

[0047] Use the training data to obtain the fitness corresponding to each parameter coding, and determine the parameter coding with the maximum fitness as the optimal parameter coding; Based on the optimal parameter coding, adaptively select the search space corresponding to the parameter coding, and determine the parameter coding after selecting the search space; Use the multi-segment variable-speed information fusion method to perform information fusion on the parameter coding after selecting the search space, and determine the parameter coding after information fusion; Use the multi-directional joint search method to perform joint search on the parameter coding after information fusion, and determine the parameter coding after joint search; Use the position crossover transformation method to perform global search on the parameter coding after joint search, and determine the parameter coding after global search; Determine whether the current number of training times has reached the maximum number of training times. If so, re-determine the optimal parameter encoding according to the parameter encoding after global search. Otherwise, return to the step of obtaining the optimal parameter encoding. Take the model parameters in the re-determined optimal parameter encoding as the final model parameters of the drilling attitude deviation prediction model, and obtain the trained drilling attitude deviation prediction model.

[0048] Optionally, after the model encoding changes, out-of-bounds processing can be performed on the model encoding to ensure that the model encoding is always between its corresponding upper and lower limits. For example, when a certain model parameter in the model encoding is out of bounds, the out-of-bounds model parameter can be set to the nearest limit. For example, if it exceeds the upper limit, it is set to the upper limit value. Or, the out-of-bounds model parameter can be randomly generated between its upper and lower limits.

[0049] The prior art generally uses genetic algorithms or particle swarm algorithms for hyperparameter training. Although certain training effects can be achieved, the poor global search ability and poor training effects will lead to inaccurate prediction of drilling attitude deviation. Therefore, the embodiments of the present invention provide a training algorithm to improve the technical problems existing in the prior art, enhance the training effect of the drilling attitude deviation prediction model, and thus improve the accuracy of drilling attitude deviation prediction.

[0050] In the embodiments of the present invention, a sequence initialization mechanism is used to initialize the encoding of the model parameters of the drilling attitude deviation prediction model, and multiple different parameter encodings are obtained, including:

[0051] Among them, represents the i -th dimensional model parameter of the d -th parameter encoding, d = 1, 2, …, D, where D represents the total dimension of the model parameters in the parameter encoding, and when i = 1, the i -th parameter encoding is a randomly generated parameter encoding, represents the i +1-th dimensional model parameter of the d -th parameter encoding, represents the remainder function, represents pi, and sin represents the sine function.

[0052] Through the above sequence initialization mechanism for initialization encoding, the initial solutions can be effectively evenly distributed in the solution space, which can effectively improve the training speed and global search ability of the algorithm, and thus improve the training effect of the algorithm.

[0053] In an embodiment of the present invention, obtaining the fitness corresponding to each parameter encoding using training data and determining the parameter encoding with the maximum fitness as the optimal parameter encoding includes: For any parameter encoding, after applying the model parameters in the parameter encoding to the drilling attitude deviation prediction model, using the training samples in the training data as the input and the training labels in the training data as the expected output, obtaining the cross-entropy loss function value; After adding the cross-entropy loss function value to a preset constant (the preset constant can be 0.0001), obtaining a non-zero intermediate parameter, and performing a reciprocal operation on the non-zero intermediate parameter to obtain the fitness corresponding to the parameter encoding; Traverse all parameter encodings, obtain the fitness corresponding to each parameter encoding, and determine the parameter encoding with the maximum fitness as the optimal parameter encoding.

[0054] In an embodiment of the present invention, based on the optimal parameter encoding, adaptively selecting the search space corresponding to the parameter encoding and determining the parameter encoding after selecting the search space includes: According to the current training times, obtaining the adaptive search space selection factor as:

[0055] Wherein, represents the adaptive search space selection factor at the t -th training, and when t = 1, is set to a fixed initial value, represents the adaptive search space selection factor at the t -1-th training, represents the natural constant, T represents the maximum number of training times; According to the adaptive search space selection factor and the optimal parameter encoding, obtaining the parameter encoding after selecting the search space as:

[0056]

[0057] Wherein, represents the t -th parameter encoding at the m -th training, represents the m -th parameter encoding after selecting the search space, m = 1, 2, …, NP, NP represents the total number of parameter encodings, represents the optimal parameter encoding, represents a first random number between (0, 1), represents (0, 2 ), a second random number between represents a third random number between (0, 1), and sin represents the sine function. represents the optimal parameter encoding. represents the adaptive position selection factor. represents the exponential function with the natural constant e as the base. represents the t th time of training, the m fitness corresponding to the th parameter encoding. represents the fitness corresponding to the optimal parameter encoding. represents another parameter encoding that is closest to the parameter encoding in terms of the Euclidean distance.

[0058] In the embodiment of the present invention, the search space corresponding to the parameter encoding is adaptively selected, and the search space can be selected according to the position quality degree of the parameter encoding itself. At the same time, the position information of another closest parameter encoding is fused for selection, so that the un-searched area can be more effectively selected, the global search ability is improved, and as the algorithm progresses, the convergence fineness gradually increases, which can effectively ensure the training effect of the algorithm.

[0059] In the embodiment of the present invention, a multi-stage variable-speed information fusion method is used to fuse the parameter encoding after the search space is selected to determine the parameter encoding after information fusion, including: According to the current training times, obtain the multi-stage variable-speed information fusion factor as:

[0060] wherein, represents the multi-stage variable-speed information fusion factor, represents the preset maximum value of the multi-stage variable-speed information fusion factor, represents the preset minimum value of the multi-stage variable-speed information fusion factor, represents the cosine function, represents the first training times threshold, represents the second training times threshold, and 1 < < < T, represents the multi-stage variable-speed information fusion factor when the training times is ; According to the multi-stage variable-speed information fusion factor, fuse the parameter encoding after the search space is selected to determine the parameter encoding after information fusion as:

[0061]

[0062]

[0063] Among them, represents the parameter encoding after the t th selection search space during the k th training, represents the parameter encoding after the k th information fusion, k = 1, 2, …, NP, represents the first random parameter encoding except for the parameter encoding , represents the fifth random number between (0, ), represents the sixth random number between (0, 2 ), represents the first information fusion coefficient, represents the second information fusion coefficient, represents the first constant term and is set to - ; represents the second constant term and is set to ; represents the third constant term and is set to .

[0064] The multi-segment variable-speed information fusion method provided by the embodiments of the present invention can provide a relatively appropriate information fusion speed in the early stage of the algorithm, a relatively large information fusion speed in the middle stage of the algorithm to increase the global search ability, and then gradually reduce the information fusion speed in the later stage of the algorithm, thereby ensuring the convergence of the algorithm. At the same time, the idea of co-evolution and the golden sine search route are applied, which can effectively ensure the global search ability of the algorithm in the middle stage and the fine search ability of the algorithm.

[0065] In the embodiments of the present invention, a multi-directional joint search method is used to jointly search the parameter encoding after information fusion to determine the parameter encoding after joint search, including: According to the current training times, obtain the joint search control factor as:

[0066] According to the joint search control factor, obtain the adaptive joint search speed as:

[0067]

[0068]

[0069] Among them, represents the t th training, n after the n -th information fusion, the parameter encoding, = 1, 2, …, NP, t represents the search speed corresponding to the parameter encoding at the th training, represents the sign function, represents the first neighborhood position encoding corresponding to the parameter encoding, represents the second neighborhood position encoding corresponding to the parameter encoding, represents the fitness corresponding to the first neighborhood position encoding, represents the fitness corresponding to the second neighborhood position encoding, represents the seventh random number between (0, 1); The multi-directional search speed obtained by combining the historical optimal direction and the current optimal direction is:

[0070] Among them, represents the t +1-th training, search speed corresponding to the parameter encoding, represents the preset first learning rate, represents the preset second learning rate, represents the eighth random number between (0, 1), represents the ninth random number between (0, 1), represents the historical optimal position corresponding to the parameter encoding; According to the adaptive joint search speed and the multi-directional search speed, perform a joint search on the parameter encoding after information fusion, and determine that the parameter encoding after the joint search is:

[0071] Among them, represents the n th parameter encoding after the joint search, represents the tenth random number between (0, 1).

[0072] The multi-directional joint search method provided by the embodiments of the present invention can effectively search in a more optimal direction. Meanwhile, it integrates adaptive joint search, which can integrate a certain local search ability in the middle and early stages of the algorithm, and gradually search in the optimal direction. In the later stage of the algorithm, it gradually searches towards the optimal position, improving the convergence fineness of the algorithm.

[0073] In the embodiments of the present invention, the position crossover transformation method is used to perform global search on the parameter encoding after joint search, and the parameter encoding after global search is determined as: The position crossover transformation method is used to perform global search on the parameter encoding after joint search, and the global search encoding is obtained as:

[0074]

[0075] Wherein, represents the t th parameter encoding after joint search during the j th training, represents the global search encoding corresponding to the j rd parameter encoding after joint search, j = 1, 2,..., NP, represents the second random parameter encoding except the parameter encoding , represents the eleventh random number between (0, 1), represents the twelfth random number between (0, 1), represents the thirteenth random number between (0, 1), represents the position transformation encoding, represents the parameter upper limit encoding, represents the parameter lower limit encoding; When the fitness of the global search encoding is greater than the fitness of the corresponding parameter encoding after joint search, the global search encoding is used as the parameter encoding after global search. Otherwise, the original parameter encoding after joint search is directly used as the parameter encoding after global search.

[0076] The position crossover transformation method provided by the embodiments of the present invention can provide relatively strong global search ability, thus assisting the algorithm to jump out of the local optimal solution, and introducing an adaptive selection strategy to ensure the search speed of the algorithm.

[0077] In summary, through the mutual cooperation of several different search strategies, the present invention can effectively improve the training effect of the algorithm and ensure the prediction accuracy of the drilling attitude deviation prediction model.

[0078] A drilling deviation correction prediction method based on LSTM provided by the present invention constructs a drilling attitude deviation prediction model through LSTM, and uses training data to train the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model. Then, real-time multi-source data is collected, and the real-time multi-source data is normalized. The normalized real-time multi-source data is input into the trained drilling attitude deviation prediction model for recognition to obtain a drilling attitude deviation prediction result. Finally, the drilling attitude deviation prediction result is used as a drilling deviation correction prediction result, and the drilling deviation correction prediction result is transmitted to a device designated by a staff member to enable the staff member to perform drilling deviation correction. By utilizing the powerful memory ability and non-linear fitting ability of LSTM, it is possible to more accurately capture the time series characteristics during the drilling process, effectively improving the efficiency and accuracy of drilling deviation correction.

[0079] As Figure 2 shown, an embodiment of the present invention provides a drilling deviation correction prediction system based on LSTM, including: a training data acquisition module 201, a model training module 202, a data prediction module 203, and a data feedback module 204; The training data acquisition module 201 is configured to acquire historical multi-source data related to drilling and historical attitude deviations corresponding to the historical multi-source data, and preprocess the historical multi-source data and the historical attitude deviations corresponding to the historical multi-source data to obtain training data; The model training module 202 is configured to construct a drilling attitude deviation prediction model using LSTM, and use the training data to train the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model; The data prediction module 203 is configured to collect real-time multi-source data, normalize the real-time multi-source data, and input the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for recognition to obtain a drilling attitude deviation prediction result; The data feedback module 204 is configured to use the drilling attitude deviation prediction result as a drilling deviation correction prediction result, and transmit the drilling deviation correction prediction result to a device designated by a staff member to enable the staff member to perform drilling deviation correction.

[0080] A drilling deviation correction prediction system based on LSTM provided by an embodiment of the present invention can execute the above method technical solution, and its principle and beneficial effects are similar, and will not be elaborated here.

[0081] Other embodiments of the present invention will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the invention, which follow the general principles of the present invention and include common general knowledge or conventional technical means in the technical field not disclosed by the present invention. It should be understood that the present invention is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. A drilling deviation correction prediction method based on LSTM, characterized in that, Including: Obtain historical multi-source data related to drilling and the corresponding historical attitude deviation of the historical multi-source data, and preprocess the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data; Construct a drilling attitude deviation prediction model using LSTM, and train the drilling attitude deviation prediction model with the training data to obtain the trained drilling attitude deviation prediction model; Collect real-time multi-source data, perform normalization processing on the real-time multi-source data, and input the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for identification to obtain the drilling attitude deviation prediction result; Use the drilling attitude deviation prediction result as the drilling deviation correction prediction result, and transmit the drilling deviation correction prediction result to the device specified by the staff to enable the staff to perform drilling deviation correction.

2. The drilling deviation correction prediction method based on LSTM according to claim 1, characterized in that Preprocess the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data, including: Perform normalization processing on the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain the normalized historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data; Construct training samples using the normalized historical multi-source data, construct training labels using the corresponding historical attitude deviation of the normalized historical multi-source data, and form training data by combining the training samples and the training labels.

3. The drilling deviation correction prediction method based on LSTM according to claim 1, characterized in that Construct a drilling attitude deviation prediction model using LSTM, and train the drilling attitude deviation prediction model with the training data to obtain the trained drilling attitude deviation prediction model, including: Construct a drilling attitude deviation prediction model using LSTM, and perform initialization encoding on the model parameters of the drilling attitude deviation prediction model using a sequence initialization mechanism to obtain multiple different parameter encodings; Obtain the fitness corresponding to each parameter encoding using the training data, and determine the parameter encoding with the maximum fitness as the optimal parameter encoding; Based on the optimal parameter encoding, adaptively select the search space corresponding to the parameter encoding, and determine the parameter encoding after selecting the search space; Perform information fusion on the parameter encoding after selecting the search space using a multi-segment variable-speed information fusion method to determine the parameter encoding after information fusion; Perform joint search on the parameter encoding after information fusion using a multi-direction joint search method to determine the parameter encoding after joint search; Perform global search on the parameter encoding after joint search using a position crossover transformation method to determine the parameter encoding after global search; Judge whether the current training times have reached the maximum training times. If so, re-determine the optimal parameter encoding according to the parameter encoding after global search. Otherwise, return to the step of obtaining the optimal parameter encoding; Use the model parameters in the re-determined optimal parameter encoding as the final model parameters of the drilling attitude deviation prediction model to obtain the trained drilling attitude deviation prediction model.

4. The method for predicting borehole deviation correction based on LSTM according to claim 3, wherein, Perform initialization encoding on the model parameters of the drilling attitude deviation prediction model using a sequence initialization mechanism to obtain multiple different parameter encodings, including: Among them, represents the i -th d -dimensional model parameter of the d -th parameter encoding, where i = 1, 2, …, D, and D represents the total dimension of the model parameters in the parameter encoding. When i = 1, the -th parameter encoding is a randomly generated parameter encoding. i + 1-th d -dimensional model parameter of the represents the remainder function, represents pi, and sin represents the sine function.

5. The method for predicting borehole deviation correction based on LSTM according to claim 3, wherein Using the training data to obtain the fitness corresponding to each parameter encoding, and determining the parameter encoding with the maximum fitness as the optimal parameter encoding, including: For any parameter encoding, after applying the model parameters in the parameter encoding to the drilling attitude deviation prediction model, using the training samples in the training data as the input and the training labels in the training data as the expected output, to obtain the cross-entropy loss function value; After adding the cross-entropy loss function value to a preset constant, obtaining a non-zero intermediate parameter, and performing a reciprocal operation on the non-zero intermediate parameter to obtain the fitness corresponding to the parameter encoding; Traversing all parameter encodings, obtaining the fitness corresponding to each parameter encoding, and determining the parameter encoding with the maximum fitness as the optimal parameter encoding.

6. The method for predicting borehole deviation correction based on LSTM according to claim 3, wherein, Based on the optimal parameter encoding, adaptively selecting the search space corresponding to the parameter encoding, and determining the parameter encoding after selecting the search space, including: According to the current training times, obtaining the adaptive search space selection factor as: Among them, represents the adaptive search space selection factor during the t -th training, and when t = 1, it is set to a fixed initial value. represents the adaptive search space selection factor during the t -1-th training, where \(e\) represents the natural constant and \(T\) represents the maximum number of training times. According to the adaptive search space selection factor and the optimal parameter encoding, obtaining the parameter encoding after selecting the search space as: Among them, represents the t th parameter encoding during the m th training, represents the parameter encoding after selecting the m th search space, m = 1, 2, …, NP, where NP represents the total number of parameter encodings, represents the optimal parameter encoding, represents the first random number between (0, 1), represents the second random number between (0, 2 ), represents the third random number between (0, 1), and sin represents the sine function, represents the optimal parameter encoding, represents the adaptive position selection factor, represents the exponential function with the natural constant e as the base, represents the t th fitness corresponding to the m th parameter encoding during the th training, represents the fitness corresponding to the optimal parameter encoding, represents another parameter encoding that is closest to the parameter encoding in terms of Euclidean distance, and represents the fourth random number between (0, 1).

7. The LSTM-based drilling deviation correction prediction method according to claim 6, wherein Using the multi-segment variable speed information fusion method to perform information fusion on the parameter encoding after selecting the search space, and determining the parameter encoding after information fusion, including: According to the current training times, obtaining the multi-segment variable speed information fusion factor as: Among them, represents the multi-stage variable speed information fusion factor, represents the preset maximum value of the multi-stage variable speed information fusion factor, represents the preset minimum value of the multi-stage variable speed information fusion factor, represents the cosine function, represents the first training times threshold, represents the second training times threshold, and 1 < < < T, represents that the training times is when the multi-stage variable speed information fusion factor; According to the multi-segment variable speed information fusion factor, performing information fusion on the parameter encoding after selecting the search space, and determining the parameter encoding after information fusion as: Among them, represents the parameter encoding after the t th training and the k th selection search space, represents the parameter encoding after the k th information fusion, k = 1, 2, …, NP, represents the first random parameter encoding except for the parameter encoding , represents the fifth random number between (0, ), represents the sixth random number between (0, 2 ), represents the first information fusion coefficient, represents the second information fusion coefficient, represents the first constant term and is set to - ; represents the second constant term and is set to ; represents the third constant term and is set to .

8. The method for predicting borehole deviation correction based on LSTM according to claim 7, wherein, Using the multi-direction joint search method to perform joint search on the parameter encoding after information fusion, and determining the parameter encoding after joint search, including: According to the current training times, obtaining the joint search control factor as: According to the joint search control factor, obtaining the adaptive joint search speed as: Among them, represents the t parameter encoding after information fusion for the n th training, n where \(i = 1, 2, \ldots, N_P\), represents the t search speed corresponding to the parameter encoding during the th training, represents the adaptive joint search speed corresponding to the parameter encoding, represents the sign function, represents the first neighborhood position encoding corresponding to the parameter encoding, represents the second neighborhood position encoding corresponding to the parameter encoding, represents the fitness corresponding to the first neighborhood position encoding, represents the fitness corresponding to the second neighborhood position encoding, represents the seventh random number between (0, 1); Obtaining the multi-direction search speed by combining the historical optimal direction and the current optimal direction; Among them, represents the t parameter encoding during the +1 - th training, represents the preset first learning rate, represents the preset second learning rate, represents the eighth random number between (0, 1), represents the ninth random number between (0, 1), represents the corresponding historical optimal position; According to the adaptive joint search speed and the multi-direction search speed, performing joint search on the parameter encoding after information fusion, and determining the parameter encoding after joint search as: Among them, represents the parameter encoding after the n th combined search, represents the tenth random number between (0, 1).

9. The method for predicting borehole deviation correction based on LSTM according to claim 8, characterized in that Using the position crossover transformation method to perform global search on the parameter encoding after joint search, and determining the parameter encoding after global search as: Using the position crossover transformation method to perform global search on the parameter encoding after joint search, and obtaining the global search encoding as: Among them, represents the parameter encoding after the t -th training and the j -th joint search, represents the global search encoding corresponding to the parameter encoding after the j -th joint search, j = 1, 2, …, NP, represents the second random parameter encoding except the parameter encoding , represents the eleventh random number between (0, 1), represents the twelfth random number between (0, 1), represents the thirteenth random number between (0, 1), represents the position transformation encoding, represents the parameter upper limit encoding, represents the parameter lower limit encoding; When the fitness of the global search encoding is greater than the fitness of the corresponding parameter encoding after joint search, then using the global search encoding as the parameter encoding after global search, otherwise directly using the original parameter encoding after joint search as the parameter encoding after global search.

10. A drilling deviation correction prediction system based on LSTM, characterized in that Including: A training data acquisition module, a model training module, a data prediction module, and a data feedback module; The training data acquisition module is used to obtain the historical multi-source data related to drilling and the corresponding historical attitude deviation of the historical multi-source data, and perform preprocessing on the historical multi-source data and the corresponding historical attitude deviation of the historical multi-source data to obtain training data; The model training module is used to construct a drilling attitude deviation prediction model using LSTM, and train the drilling attitude deviation prediction model with training data to obtain the trained drilling attitude deviation prediction model; The data prediction module is used to collect real-time multi-source data, perform normalization processing on the real-time multi-source data, input the normalized real-time multi-source data into the trained drilling attitude deviation prediction model for identification, and obtain the drilling attitude deviation prediction result; The data feedback module is used to use the drilling attitude deviation prediction result as the drilling deviation correction prediction result, and transmit the drilling deviation correction prediction result to the device designated by the staff, so that the staff can perform drilling deviation correction.

Images