A method for determining the life cycle of a spare part of a mobile device based on deep learning

Abstract

This invention belongs to the field of artificial intelligence technology in petroleum engineering, and particularly relates to a method for determining the lifecycle of spare parts for moving equipment based on deep learning. The method includes the steps of stabilizing real-time operating data, performing backpropagation training using sequence-to-sequence recurrent neural units combined with a deep autoencoder, and constructing a spare parts lifecycle discrimination model. This method extracts the sequential correlation characteristics of real-time oil well operating data using bidirectional sequence-to-sequence recurrent units, uses a deep autoencoder to encode and decode the time series data, and employs complementary set empirical mode decomposition to stabilize the real-time operating data. This reduces the impact of abnormal operating conditions, enabling the trained model to better identify the lifecycle status of spare parts for moving equipment, thus improving the accuracy of lifecycle discrimination.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology in petroleum engineering, and in particular relates to a method for determining the life cycle of spare parts for dynamic equipment based on deep learning. Background Technology

[0002] Pumping unit equipment is a crucial component of real-time oilfield production. Taking conveyor belts as an example: As a vital component of the pumping unit, the condition of the conveyor belt throughout its lifecycle directly affects the unit's operational status; belt damage leads to pumping unit shutdown. In practical applications, conveyor belts, being resilient spare parts, experience wear, depletion, and breakage during pumping unit operation, impacting unit performance and oilfield economic efficiency. Therefore, timely analysis of pumping unit production data, analysis of the lifecycle status of conveyor belts, and proactive prevention and resolution of belt breakage and other faults are critical issues that urgently need to be addressed in current oilfield production.

[0003] However, further research revealed that the current determination of the life cycle status of pumping unit equipment still relies on manual inspections. This not only consumes a lot of manpower and resources, but also suffers from defects such as low accuracy and unreliable inspection results, which is detrimental to optimizing oilfield production processes. Therefore, it is urgent for those skilled in the art to provide a new approach to solve the aforementioned problems in the life cycle management of pumping unit spare parts. Summary of the Invention

[0004] This invention provides a method for determining the lifecycle of spare parts for moving equipment based on deep learning. This method extracts the correlation characteristics of real-time operating data of oil wells by using bidirectional sequence-to-sequence cyclic units, uses a deep autoencoder to encode and decode the time series and backpropagates to train the model, and uses complementary set empirical mode decomposition to stabilize the real-time operating data, thereby reducing the impact of abnormal operating conditions. This enables the trained model to better identify the lifecycle status of spare parts for moving equipment and improves the accuracy of determining the lifecycle of spare parts for moving equipment.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A method for determining the lifecycle of spare parts for moving equipment based on deep learning includes the following steps:

[0007] Step 1: Stabilize the real-time operating data;

[0008] Step 2: Use sequence-to-sequence recurrent neural units in conjunction with a deep autoencoder for reverse training;

[0009] Step 3: Construct a lifecycle discrimination model for spare parts.

[0010] Preferably, step 1 can be specifically described as follows:

[0011] Add positive and negative auxiliary noise to the real-time operating data of oil wells;

[0012] EMD decomposition was performed on the data after adding positive and negative auxiliary noise:

[0013] The decomposition results are obtained by combining multiple components.

[0014] Preferably, the decomposition result obtained in step 1 satisfies:

[0015] Equation (1);

[0016] In its formula (1), the decomposition result The final result of CEEMD decomposition is the first... One IMF component, For the first The first signal Each IMF component represents a unit.

[0017] Preferably, step 2 can be specifically described as follows:

[0018] Initialize the bidirectional sequence to the initial model of the cyclic unit of the sequence:

[0019] Train a bidirectional sequence-to-sequence cyclic unit model;

[0020] Time series are encoded using a deep autoencoder.

[0021] Preferably, the process of training the initial model of the bidirectional sequence-to-sequence cyclic unit in step 2 can be specifically described as follows:

[0022] Equation (2);

[0023] In formula (2), express Activation function Represents the parameter matrix, , These represent the deviation vectors of the parameter matrix, respectively. , They represent function; This represents the element-wise product operation;

[0024] yes The state vector in the time-memory unit; yes The state vector output during training; Indicates the input gate. Represents the Gate of Oblivion This indicates the output gate.

[0025] Preferably, the time series encoded in step 2 satisfies:

[0026] Equation (3);

[0027] In formula (3), This represents the time series formed by the encoding. , These represent the full and bias connections between the visible and hidden layers, respectively. express Activation function.

[0028] A more preferred approach also includes the following steps:

[0029] The time series encoded in step 2 is decoded, and the reconstruction error is calculated.

[0030] Preferably, the autoencoder output error cost function used for reconstruction error calculation satisfies:

[0031] Equation (4);

[0032] In formula (4), Indicates the number of samples.

[0033] This invention provides a method for determining the lifecycle of spare parts for moving equipment based on deep learning. This method includes the steps of stabilizing real-time operating data, performing back-training using sequence-to-sequence recurrent neural units combined with a deep autoencoder, and constructing a spare parts lifecycle discrimination model. This method for determining the lifecycle of spare parts for moving equipment, with the above-described steps, has at least the following technical advantages compared to existing technologies:

[0034] 1. By adopting an improved complementary set empirical mode decomposition, the stabilization of operating condition data was achieved, the model's ability to learn extreme value features was enhanced, and the impact of abnormal peak data caused by human operation or changes in well status was reduced.

[0035] 2. A bidirectional sequence-to-sequence recurrent unit neural network diagnostic model was built, which can handle the temporal characteristics of oil well operating data; at the same time, it considers the forward and backward information of real-time oil well operating data, improving the feature extraction capability.

[0036] 3. The sequence is encoded and decoded by a deep autoencoder, and the model is retrained during the backpropagation iteration process, which weakens the influence of irrelevant features on the results and improves the generalization ability of the model. Detailed Implementation

[0037] This invention provides a method for determining the lifecycle of spare parts for moving equipment based on deep learning. This method extracts the correlation characteristics of real-time operating data of oil wells by using bidirectional sequence-to-sequence cyclic units, uses a deep autoencoder to encode and decode the time series and backpropagates to train the model, and uses complementary set empirical mode decomposition to stabilize the real-time operating data, thereby reducing the impact of abnormal operating conditions. This enables the trained model to better identify the lifecycle status of spare parts for moving equipment and improves the accuracy of determining the lifecycle of spare parts for moving equipment.

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Example 1

[0039] A method for determining the lifecycle of spare parts for moving equipment based on deep learning includes the following steps:

[0040] Step 1: Stabilize the real-time operating data;

[0041] Step 2: Use sequence-to-sequence recurrent neural units in conjunction with a deep autoencoder for reverse training;

[0042] Step 3: Construct a lifecycle discrimination model for spare parts.

[0043] In addition, as a preferred embodiment of the present invention, the method for determining the lifecycle of spare parts for moving equipment based on deep learning further preferably includes the following steps:

[0044] That is, the time series encoded in step 2 is decoded and the reconstruction error is calculated.

[0045] The following embodiments will be used as examples to further explain in detail the features of each step, and the present invention will be described exemplarily based on these features. Example 2

[0046] Specifically, Embodiment 2 includes all the technical features of Embodiment 1. Furthermore, the features of step 1 in Embodiment 2 are further explained as follows.

[0047] Specifically, step 1 can be further described as follows:

[0048] Add positive and negative auxiliary noise to the real-time operating data of oil wells;

[0049] For example, using real-time operating data of oil wells For example, the data length is represented as The number of features is represented as Initialization is represented as The above data were processed with positive and negative auxiliary white noise, respectively, where the positive noise was represented as... Negative noise is represented as At this point, the set of operating data including the aforementioned positive and negative auxiliary noise can be represented as follows: , This is the dataset of operating conditions after adding positive noise; This is the working condition dataset after adding negative noise.

[0050] Then, EMD decomposition was performed on the data after adding positive and negative auxiliary noise:

[0051] Specifically, the above-mentioned operating condition data set with added positive and negative auxiliary noise is subjected to EMD decomposition to obtain two corresponding IMF components. as well as .

[0052] Then, the decomposition results are obtained by combining multiple components.

[0053] The decomposition result satisfies:

[0054] Equation (1);

[0055] In its formula (1), the decomposition result The final result of CEEMD decomposition is the first... One IMF component, For the first The first signal Each IMF component represents a unit.

[0056] It should be noted that the decomposition results By further combining these components, a stable real-time operating condition sequence can be obtained. . Example 3

[0057] Specifically, Embodiment 3 includes all the technical features of Embodiment 1. Furthermore, the features of step 2 are further explained in Embodiment 3 as follows.

[0058] Specifically, step 2 can be described as follows:

[0059] Initialize the bidirectional sequence to the initial model of the cyclic unit of the sequence:

[0060] For example, the real-time operating condition sequence after stabilization processing obtained according to Example 2. For example, its real-time operating condition sequence after stabilization can be represented as: ,in, Indicates the data length. Indicates the number of features. This indicates the size of a single training data batch for initializing the model. Indicate the duration of time; and input variables The value is set to the number of features. Output dimension Set to 1.

[0061] Then, the initial model of the bidirectional sequence-to-sequence cyclic unit is trained;

[0062] Specifically, the process of training the initial model of the bidirectional sequence-to-sequence cyclic unit can be described as follows:

[0063] Equation (2);

[0064] In formula (2), express Activation function Represents the parameter matrix, , These represent the deviation vectors of the parameter matrix, respectively. , They represent function; This represents the element-wise product operation;

[0065] yes The state vector in the time-memory unit; yes The state vector output during training; Indicates the input gate. Represents the Gate of Oblivion This indicates the output gate.

[0066] Then, a deep autoencoder is used to encode the time series.

[0067] It is worth noting that the time series formed by the encoding satisfies:

[0068] Equation (3);

[0069] In formula (3), This represents the time series formed by the encoding. , These represent the full and bias connections between the visible and hidden layers, respectively. express Activation function. Example 4

[0070] Specifically, Embodiment 4 includes all the technical features of Embodiment 1. Furthermore, Embodiment 4 further supplements the features of steps 2 and 3 as follows.

[0071] First, based on Example 3, Example 4 further explains the subsequent steps after time series encoding as follows:

[0072] Example 1 mentions decoding the time series encoded in step 2 and calculating the reconstruction error. This process can be specifically described as follows:

[0073] The input values ​​are reconstructed and recovered by continuously training the previous sequence to the recurrent sequence network through backpropagation, ensuring that the network reconstruction error is minimized. The decoding process can be referenced as follows: ;

[0074] in , These represent the full and bias connections between the visible and hidden layers, respectively. express Activation function. Of course, this decoding process is only an example, and the inventor can certainly further process and adjust the specific decoding method based on personal experience.

[0075] After the encoding and decoding processes are completed, the content is further reconstructed to calculate the error (that is, the original encoded content and the reconstructed input content are respectively input into the recurrent neural network for error calculation).

[0076] Specifically, the autoencoder output error cost function used for reconstruction error calculation satisfies:

[0077] Formula (4); in formula (4), Indicates the number of samples.

[0078] The purpose of this is to repeatedly train the model on each set of training data, thereby minimizing the mean squared error of the trained model.

[0079] Regarding the construction of the spare parts lifecycle discrimination model shown in step 3, as illustrated by the examples above, those skilled in the art can understand that after training the network until the model converges or reaches the maximum number of training iterations (e.g., n times), the spare parts lifecycle discrimination (network) model can be obtained. By using the untrained oil well real-time operating condition dataset as the test set to train the model, parameters such as model accuracy, precision, and recall can be obtained, and these parameters can be used as performance metrics for the model.

[0080] Thus, this invention provides a method for determining the lifecycle of spare parts for moving equipment based on deep learning, and completes and constructs a lifecycle discrimination model for spare parts; this lifecycle discrimination model can be used to determine the lifecycle of spare parts for moving equipment, ultimately achieving the purpose of replacing manual inspection and determination.

[0081] This invention provides a method for determining the lifecycle of spare parts for moving equipment based on deep learning. This method includes the steps of stabilizing real-time operating data, performing back-training using sequence-to-sequence recurrent neural units combined with a deep autoencoder, and constructing a spare parts lifecycle discrimination model. This method for determining the lifecycle of spare parts for moving equipment, with the above-described steps, has at least the following technical advantages compared to existing technologies:

[0082] 1. By adopting an improved complementary set empirical mode decomposition, the stabilization of operating condition data was achieved, the model's ability to learn extreme value features was enhanced, and the impact of abnormal peak data caused by human operation or changes in well status was reduced.

[0083] 2. A bidirectional sequence-to-sequence recurrent unit neural network diagnostic model was built, which can handle the temporal characteristics of oil well operating data; at the same time, it considers the forward and backward information of real-time oil well operating data, improving the feature extraction capability.

[0084] 3. The sequence is encoded and decoded by a deep autoencoder, and the model is retrained during the backpropagation iteration process, which weakens the influence of irrelevant features on the results and improves the generalization ability of the model.

[0085] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for determining the lifecycle of spare parts for moving equipment based on deep learning, characterized in that, The steps include the following: Step 1: Stabilize the real-time operating data; Step 2: Use sequence-to-sequence recurrent neural units in conjunction with a deep autoencoder for reverse training; Step 3: Construct a lifecycle discrimination model for spare parts; Step 1 is specifically described as follows: Add positive and negative auxiliary noise to the real-time operating data of oil wells; EMD decomposition was performed on the data after adding positive and negative auxiliary noise: The decomposition results were obtained by combining multiple components. The decomposition result obtained in step 1 satisfies: Equation (1); In its formula (1), the decomposition result The final result of CEEMD decomposition is the first... One IMF component, For the first The first signal Each IMF component represents; Step 2 is specifically described as follows: Initialize the bidirectional sequence to the initial model of the cyclic unit of the sequence: Train a bidirectional sequence-to-sequence cyclic unit model; Encode time series data using a deep autoencoder; The process of training the initial model of the bidirectional sequence-to-sequence cyclic unit in step 2 is specifically described as follows: Equation (2); In formula (2), express Activation function Represents the parameter matrix, , These represent the deviation vectors of the parameter matrix, respectively. , They represent function; This represents the element-wise product operation; yes The state vector in the time-memory unit; yes The state vector output during training; Indicates the input gate. Represents the Gate of Oblivion This indicates the output gate.

2. The method for determining the lifecycle of spare parts for moving equipment based on deep learning according to claim 1, characterized in that, The time series encoded in step 2 satisfies: Equation (3); In formula (3), This represents the time series formed by the encoding. , These represent the full and bias connections between the visible and hidden layers, respectively. express Activation function.

3. The method for determining the lifecycle of spare parts for moving equipment based on deep learning according to claim 2, characterized in that, It also includes the following steps: The time series encoded in step 2 is decoded, and the reconstruction error is calculated.

4. The method for determining the lifecycle of spare parts for moving equipment based on deep learning according to claim 3, characterized in that, The autoencoder output error cost function used for reconstruction error calculation satisfies: Equation (4); In formula (4), Indicates the number of samples.

Images