Oil well pressure anomaly detection method and system based on deep learning

Abstract

The invention relates to the technical field of pressure detection, in particular to an oil well pressure anomaly detection method and system based on deep learning, and the method comprises the following steps: collecting a control signal and pressure data, calculating a slope difference and fluctuation to mark a sudden change point, extracting a multi-dimensional feature group to generate a training sample, and constructing a deep learning model to recognize pressure difference anomaly. The oil well pressure abnormal time period is predicted in real time. According to the method, accurate attribution of abnormal fluctuations in the oil well operation process is achieved by constructing the time sequence corresponding relation between control signals and pressure difference sudden change, the target pertinence of pressure change recognition is improved, a sliding structure extracts multi-dimensional dynamic characteristics, the capacity of capturing wellhead or underground pressure micro-amplitude abnormity is enhanced, and the training sample is combined with a control behavior label, so that the accuracy of the abnormal fluctuations in the oil well operation process is improved. The distinguishing capability of the model to different abnormal forms is optimized, the adaptability under complex working conditions is improved, a real-time prediction mechanism ensures that oil well pressure state recognition has continuity, rapidness and high confidence, and risk early warning and operation guarantee in the oil extraction process are effectively supported.

Description

Technical Field

[0001] The present invention relates to the technical field of pressure detection, and in particular, to a method and system for detecting abnormal oil well pressure based on deep learning. Background Art

[0002] The technical field of pressure detection includes the technical content of real-time monitoring, recording, and analysis of fluid pressure in a closed or open system, and is widely used in many industries such as petrochemical, machinery manufacturing, and energy transportation. The core content is to obtain the physical pressure data of the target medium under specific time and space conditions through a pressure sensor, and combine links such as electrical signal conversion, data acquisition, and data analysis to achieve an accurate description and tracking of the pressure state. The entire pressure detection technology system includes the implementation methods of pressure measurement principles, such as strain-type, capacitive, and piezoresistive pressure sensing mechanisms, as well as signal conditioning circuits, data acquisition terminals, and supporting pressure analysis methods, forming a complete system covering physical acquisition, numerical conversion, and data interpretation, aiming to serve the needs of industrial safety, process control, and equipment operation and maintenance.

[0003] Among them, the method for detecting abnormal oil well pressure based on deep learning refers to using an artificial neural network model to model and analyze the pressure data sequence during the oil production process of an oil well to identify potential abnormal fluctuations therein, covering the historical data collection of the pressure value at the wellhead or downhole of the oil well, the time-series arrangement and numerical normalization processing of the pressure data, the construction of a feature extraction structure based on a convolutional neural network, and the use of a long short-term memory network for state judgment and classification annotation in the time dimension. The parameters of the network model are adjusted through the error optimization process of the training set and the validation set to achieve the state discrimination of the input pressure data.

[0004] In the existing process of detecting abnormal oil well pressure, the pressure value itself is used as the analysis object, and no participation item of external behavior events is introduced, lacking the modeling of the linkage relationship between the control signal and the pressure response, resulting in a fuzzy attribution of abnormal fluctuations and difficulty in distinguishing the expected fluctuations caused by intervention from abnormal states. In the process of sample generation, the overall sliding window processing strategy is adopted, and the actual section location of the abnormal behavior itself is not combined, making the annotation process lack accuracy and affecting the training effect. At the feature representation level, it relies on limited basic statistical indicators and has insufficient response ability to non-linear fluctuation trends and periodic disturbances, and cannot effectively identify microscopic abnormal structures in complex sequences. During the model training process, it is mostly based on static labels and a single classification target, lacking label refinement for control behavior differences, resulting in incomplete coverage of abnormal forms by the model and limited generalization ability. In the actual operating environment, due to the prediction output depending on a static structure and a fixed time window scale, there are phenomena of response delay and mismatch of the recognition interval, and it is impossible to support the rapid judgment and accurate positioning of mutation events, especially in the continuous high-frequency sampling scenario, it is more likely to produce misidentifications and missed detections. Summary of the Invention

[0005] The object of the present invention is to solve the drawbacks existing in the prior art, and a method for detecting abnormal oil well pressure based on deep learning is proposed.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A method for detecting abnormal oil well pressure based on deep learning, comprising the following steps:

[0007] S1: Obtain the pump control start-stop signal and the valve control instruction timestamp, calculate the slope difference and fluctuation in combination with the pressure sequence, mark the mutation points, and generate a control-related pressure difference annotation set in combination with the control behavior;

[0008] S2: Based on the control-related pressure difference annotation set, extract the pressure sequence window, calculate the first-order difference, trend slope, sliding standard deviation and average value of the fluctuation interval, and construct a multi-dimensional time series feature group;

[0009] S3: Based on the multi-dimensional time series feature group, according to the control type information of the corresponding sample in the control-related pressure difference annotation set, assign labels respectively and pair the feature vectors with the corresponding labels to construct a control behavior training sample set;

[0010] S4: Based on the feature vectors and class labels of the control behavior training sample set, input them into the neural units in the deep learning network structure in turn, extract the internal change patterns and establish a classification boundary mapping, select the network weight parameters, and generate a pressure difference recognition model parameter group;

[0011] S5: Obtain the real-time pressure sampling data during the operation of the oil well, extract the feature vectors from the real-time data, use the pressure difference recognition model parameter group for prediction calculation, identify the time period of the abnormal pressure category, and generate the real-time detection result of the oil well pressure abnormality.

[0012] As a further solution of the present invention, the control-related pressure difference annotation set includes a pressure difference mutation section, a mutation position label, a control response type, and a control time index. The multi-dimensional time series feature group includes a time series structure feature, a statistical change index, a differential feature vector, and a local trend expression record. The control behavior training sample set includes a behavior category label, a feature representation combination, and a classification sample unit. The pressure difference recognition model parameter group includes a deep feature mapping parameter, a classification boundary function, and an optimization objective function. The real-time detection result of the oil well pressure abnormality includes an abnormal recognition interval, a corresponding category identifier, an abnormal time positioning, and a pressure difference abnormality amplitude.

[0013] As a further solution of the present invention, the specific steps for obtaining the control-related pressure difference annotation set are as follows:

[0014] S111: Based on the pump control start-stop signal and the valve control instruction timestamp, extract the corresponding pressure sampling sequence time points, calculate the change sequence of the pressure difference between adjacent sampling points within a continuous time period, determine whether the difference amplitude between adjacent pressure differences falls within the pump control response threshold interval, mark the continuous intervals that meet the threshold interval constraint conditions, and obtain the pressure difference response interval segments;

[0015] S112: According to the pressure difference response interval segments, extract the corresponding time points based on the extreme value positions of the pressure difference fluctuations therein, perform pressure difference jump discrimination processing on the continuous pressure difference sequence, and extract the jump points according to the sign directionality and numerical volatility calculation method of the pressure difference sequence mutation, generating a set of pressure difference mutation time points;

[0016] S113: According to the set of pressure difference mutation time points, find the pressure difference response interval segment to which each jump point belongs, and use the formula:

[0017]

[0018] Calculate the control-related pressure difference intensity value L of the interval corresponding to the pressure difference mutation point. Based on the control-related pressure difference intensity value as the indexing basis, bind the jump point and the interval information to obtain the control-related pressure difference annotation set, where ΔP k represents the pressure difference value at the k-th time point, T k is the sampling timestamp corresponding to the pressure difference value, and K represents the total number of pressure difference jump segments participating in the calculation in the current pressure difference interval segment.

[0019] As a further solution of the present invention, the steps for obtaining the multi-dimensional time series feature group are specifically as follows:

[0020] S211: According to the pressure difference mutation sections marked in the control-related pressure difference annotation set, obtain the original pressure sampling values within the corresponding time range, perform positioning extraction according to the time interval of the mutation section, and divide the extracted sequence into multiple overlapping segments along the time axis to obtain a sliding time segment sequence;

[0021] S212: Based on the sliding time segment sequence, for the original pressure sampling values within each time segment, perform first-order difference processing, trend slope calculation, sliding standard deviation calculation, and fluctuation interval mean calculation respectively, generating a set of pressure fluctuation characteristic values corresponding to each time segment;

[0022] S213: According to the set of pressure fluctuation characteristic values, using the time series order corresponding to each group of characteristics as the index, use the formula:

[0023]

[0024] Calculate the multi-dimensional time series characteristic value F, splice all the characteristic values in sequence into a vector structure, and establish a multi-dimensional time series feature group, where ΔSq denotes the q-th first-order difference, M q denotes the q-th trend slope, V q denotes the q-th moving standard deviation, A q denotes the average value of the q-th fluctuation interval, and Q represents the number of feature items in each moving time segment.

[0025] As a further solution of the present invention, the steps for obtaining the control behavior training sample set are specifically as follows:

[0026] S311: Based on the multi-dimensional time series feature group, combined with the sampling interval index corresponding to each time series feature, compare the control type annotation information in the control-related pressure difference annotation set, identify the corresponding control behavior attributes, obtain the label type as normal control or abnormal mutation, and generate a control behavior label sequence;

[0027] S312: According to the multi-dimensional time series feature group and the control behavior label sequence, extract the feature vectors and the corresponding positions of the labels item by item, pair them in order to establish sample units, and use the formula:

[0028]

[0029] Calculate the paired deviation combination value Z of the i-th sample i , characterize the structural differences and deviation relationships between feature dimensions, and generate a feature-label pairing group, where X1, X2, X3, X4 represent the numerical values of the first four dimensions in the feature vector, and Y1, Y2, Y3 represent the label quantization values on the corresponding previous dimensions;

[0030] S313: Based on the feature-label pairing group, combine all feature vectors and the corresponding control behavior labels in a set manner according to the pairing results, establish a data structure for training tasks, and obtain a control behavior training sample set.

[0031] As a further solution of the present invention, the steps for obtaining the pressure difference recognition model parameter group are specifically as follows:

[0032] S411: Based on the control behavior training sample set, sequentially extract the feature vectors and the corresponding class labels in the samples, input them into the neural units at all levels in the neural network structure in order, construct a data structure for step-by-step conduction, and obtain a neuron input vector set;

[0033] S412: According to the neuron input vector set, compare the output results of the input features after activation in each layer of the network item by item, analyze the change trend of the neural responses under different class labels, extract the internal structural patterns of the sequence, and obtain the classification boundary response features;

[0034] S413: According to the classification boundary response characteristics, compare the error value between the network output and the true label, screen the network weight combination corresponding to the minimum error in each sample training round, extract the weight structure with the highest matching degree, and obtain the parameter group of the differential pressure recognition model.

[0035] As a further solution of the present invention, the steps for obtaining the real-time detection result of the oil well pressure anomaly are specifically as follows:

[0036] S511: Obtain the real-time pressure sampling data during the operation of the oil well. According to the set sliding window structure in the control-related differential pressure annotation set, slide and segment the real-time data in the same window manner to generate continuous window segments, and construct a time index structure for each segment to obtain a set of real-time sliding subsequences;

[0037] S512: Based on the set of real-time sliding subsequences, extract the pressure characteristics of each subsequence in chronological order, construct a feature vector group with the same structure as that in the training stage, input the parameter group of the differential pressure recognition model for forward prediction, identify the abnormal category time period in the prediction result, establish an abnormal time point index set, and calculate the amplitude of the extreme value fluctuation of the differential pressure within the window where the abnormal sample is located to obtain the abnormal predicted differential pressure amplitude value;

[0038] S513: According to the abnormal predicted differential pressure amplitude value and the corresponding abnormal time point index set, summarize the abnormal segment range in chronological order and classify and organize it to generate a list structure of abnormal sections, and obtain the real-time detection result of the oil well pressure anomaly.

[0039] An oil well pressure anomaly detection system based on deep learning includes:

[0040] The data processing module obtains the pump control start / stop signal and the valve control instruction timestamp, combines the time points in the pressure sampling sequence, calculates the differential pressure fluctuation value between adjacent sampling points, determines whether the differential pressure fluctuation is within the pump control response threshold interval, identifies the mutation position and marks the differential pressure mutation section to obtain the control-related differential pressure annotation set;

[0041] The feature extraction module obtains the original pressure sampling values within the corresponding time range according to the differential pressure mutation section in the control-related differential pressure annotation set, constructs a sliding analysis window and calculates the first-order difference, trend slope, sliding standard deviation and average value of the fluctuation interval to generate a multi-dimensional time series feature group;

[0042] The sample construction module, based on the multi-dimensional time series feature set and combined with the control type information, assigns labels to normal control and abnormal mutation behaviors respectively, pairs the feature vectors with the labels to obtain a set of control behavior training samples;

[0043] Based on the control behavior training sample set, the deep learning training module inputs the feature vectors into the neural units in the deep learning network, extracts the internal change patterns of the sequence, performs classification boundary mapping, and selects the network weight parameters according to the minimum loss value to obtain the differential pressure recognition model parameter set;

[0044] The real-time detection module obtains the real-time pressure sampling data during the operation of the oil well, combines the sliding window structure in the control-related differential pressure annotation set, extracts the real-time data feature vectors, performs forward prediction calculations using the differential pressure recognition model parameter set, marks the abnormal category time period, and generates the real-time detection result of the oil well pressure anomaly.

[0045] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0046] In the present invention, by introducing the pump control start-stop signal and the valve control instruction timestamp into the analysis of the pressure change, the direct correspondence between the control behavior and the differential pressure fluctuation under the operation state of the oil well is realized, enabling the abnormal detection to have the working condition perception ability, effectively identifying the pressure change characteristics driven by the operation, strengthening the identification of the source of the abnormal fluctuation, introducing the dynamic annotation of the differential pressure mutation section and the timing binding of the mutation position, improving the label accuracy and the semantic integrity of the sample, avoiding the interference of the noise label in the training stage, adopting the sliding window structure in the feature construction stage to integrate the first-order difference, trend slope and local fluctuation index, comprehensively depicting the short-term change trend and local abnormal signals of the pressure sequence in the oil well oil production process, effectively capturing the hidden abnormal starting points in the complex downhole environment, setting multiple types of labels in the training sample organization in combination with the control response type, enhancing the ability of the model to distinguish the fluctuations caused by the operation and the equipment abnormalities, enabling the separation and identification of the high-frequency disturbances and unplanned events in the oil well working conditions, through the hierarchical extraction of the sequence features and the target optimization of the model parameters, enabling the neural network to form a high-sensitivity expression ability for the downhole pressure dynamic mode, improving the stability and discrimination accuracy of the abnormal detection, combining the unified sliding structure in the real-time sampling process to complete the feature generation and prediction output, constructing a continuous abnormal detection mechanism for the dynamic pressure change of the oil well, improving the immediacy, coverage and positioning accuracy of the abnormal response, and making the pressure safety monitoring have higher reliability and practicality in the actual oil production process. Brief Description of the Drawings

[0047] Figure 1 is the main step flow chart of the present invention;

[0048] Figure 2 is the flow chart for obtaining the control-related differential pressure annotation set of the present invention;

[0049] Figure 3 is the flow chart for obtaining the multi-dimensional time series feature group of the present invention;

[0050] Figure 4This is the flowchart for obtaining the control behavior training sample set of the present invention;

[0051] Figure 5 This is the flowchart for obtaining the parameter group of the differential pressure recognition model of the present invention;

[0052] Figure 6 This is the flowchart for obtaining the real-time detection result of abnormal oil well pressure of the present invention. Detailed implementation manners

[0053] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present 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 only used to explain the present invention and are not used to limit the present invention.

[0054] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, in the description of the present invention, "a plurality of" means two or more, unless otherwise specifically defined.

[0055] Please refer to Figure 1 , an oil well pressure anomaly detection method based on deep learning, comprising the following steps:

[0056] S1: Obtain the pump control start / stop signal and the valve control instruction timestamp, combine the time points corresponding to the control instructions in the pressure sampling sequence, generate continuous pressure segments and calculate the differential pressure fluctuation values between adjacent sampling points, determine whether the differential pressure fluctuation is within the pump control response threshold interval, identify the mutation positions and mark the differential pressure mutation sections, bind the sections that meet the conditions with the mutation positions as training labels, and generate a control-related differential pressure annotation set;

[0057] S2: According to the differential pressure mutation sections marked in the control-related differential pressure annotation set, obtain the original pressure sampling values within the corresponding time range, construct a sliding analysis window and calculate the first-order difference, trend slope, sliding standard deviation and average value of the fluctuation interval, and splice each group of features into a vector structure in chronological order to generate a multi-dimensional time series feature group;

[0058] S3: Based on the multi-dimensional time series feature group, according to the control type information of the corresponding samples in the control-related differential pressure annotation set, respectively assign labels for normal control-induced changes and abnormal mutation behaviors, and pair the feature vectors with the corresponding labels to construct a control behavior training sample set;

[0059] S4: Based on the feature vectors and class labels of the control behavior training sample set, input them into the neural units in the deep learning network structure in sequence, extract the internal change patterns of the sequence and establish the classification boundary mapping relationship, select the network weight parameters with the minimum loss value as the criterion, and generate the differential pressure recognition model parameter group;

[0060] S5: Obtain the real-time pressure sampling data during the operation of the oil well, create sliding subsequences for the real-time data in combination with the same window structure in the control-related differential pressure annotation set, extract the feature vectors, perform forward prediction calculations using the differential pressure recognition model parameter group, mark the time periods identified as abnormal categories, and establish an abnormal list in combination with the actual sampling time index and differential pressure amplitude results to generate the real-time detection result of oil well pressure abnormality.

[0061] The control-related differential pressure annotation set includes differential pressure mutation sections, mutation position labels, control response types, and control time indexes. The multi-dimensional time series feature group includes time series structure features, statistical change indicators, differential feature vectors, and local trend expression records. The control behavior training sample set includes behavior category labels, feature representation combinations, and classification sample units. The differential pressure recognition model parameter group includes deep feature mapping parameters, classification boundary functions, and optimization objective functions. The real-time detection result of oil well pressure abnormality includes abnormal recognition intervals, corresponding category identifiers, abnormal time positioning, and differential pressure abnormality amplitudes.

[0062] Please refer to Figure 2 , step S1 is as follows:

[0063] S111: Based on the pump control start / stop signal and the valve control instruction timestamp, extract the corresponding pressure sampling sequence time points, calculate the differential pressure change sequence between adjacent sampling points within a continuous time period, determine whether the difference amplitude between adjacent differential pressures falls within the pump control response threshold interval, and mark the continuous intervals that meet the threshold interval constraint conditions to obtain the differential pressure response interval segments;

[0064] Based on the pump control start / stop signal and the valve control instruction timestamp, it is necessary to first extract the corresponding time points of the control instructions from the data acquisition device. For example, the pump start signal is recorded at time point 12.3 s, the pump stop signal is recorded at 37.6 s, and the valve control instructions are recorded at 21.5 s and 44.2 s. Corresponding these to the pressure sampling sequence, with the pressure sampling period set to once every 0.1 s, the index positions of the corresponding signal time points in the pressure sequence are 123, 215, 376, and 442 respectively. Further, based on these index positions, extract the sampling paragraphs of 5 s continuously before and after each control signal in the pressure sequence, and a total of 101 sampling point data segments are extracted. Set this segment of data as the control pressure segment, calculate the pressure difference change between adjacent sampling points in the control pressure segment, that is, calculate the difference between the pressure at each moment and the pressure at the previous moment. If a certain segment of the pressure sequence is [1.23 MPa, 1.27 MPa, 1.34 MPa, 1.33 MPa], then its pressure difference sequence is [0.04 MPa, 0.07 MPa, -0.01 MPa]. By judging whether the difference between every two consecutive pressure differences is within the set pump control response threshold range to identify whether there is a response. This response threshold range is measured by experiments. For a typical pneumatic pump, the acceptable range of its pressure difference change response is ±0.02 MPa. If the absolute value of the difference exceeds this range, it is not regarded as an effective response. The points with pressure difference exceeding 0.02 MPa will be marked and excluded, so as to screen out the stable change paragraphs. Perform a loop traversal on this process, and finally screen out the pressure difference response interval segments composed of multiple continuous paragraphs, among which the control signal is located, which can be used as the basis for subsequent jump identification.

[0065] The set threshold of ±0.02 MPa is based on the normal pressure difference fluctuation range of the equipment under controlled state in the experimental data. See the following table.

[0066] Table 1 Reference Table for Setting Pressure Difference Response Threshold

[0067] Control Status Minimum Differential Pressure Change (MPa) Maximum Differential Pressure Change (MPa) Set Threshold (MPa) Start 0.015 0.038 ±0.02 Stop 0.010 0.029 ±0.02 Valve Open 0.013 0.031 ±0.02 Valve Closed 0.011 0.027 ±0.02

[0068] As shown in Table 1, through the minimum and maximum pressure difference fluctuation values monitored during the actual operation of the equipment, it is determined that the fluctuation within ±0.02 MPa belongs to the response change interval of pump control start / stop or valve action. This threshold is used as the judgment basis. During the operation, calculate the pressure difference change between its front and rear points for each sampling point and judge whether it falls into this interval. The finally obtained pressure difference response interval segments are such as: from the 123rd to the 132nd sampling points, from the 215th to the 226th sampling points, etc.

[0069] S112: According to the pressure difference response interval segment, extract the corresponding time points according to the extreme value positions of the pressure difference value fluctuations therein, perform pressure difference jump discrimination processing on the continuous pressure difference sequence, and extract the jump points according to the sign directionality and numerical volatility calculation method of the pressure difference sequence mutation to generate a set of pressure difference mutation time points;

[0070] Combined with the pressure difference response interval, the jump recognition operation is performed on the pressure difference sequence in the interval. First, the pressure difference sequence in each response interval is scanned, and the points where the absolute value of the pressure difference in the sequence has a mutation are extracted. The mutation is defined as the difference between a certain pressure difference point and its two adjacent points is greater than 0.03MPa at the same time. For example, in the interval [0.06MPa, 0.09MPa, 0.14MPa, 0.12MPa], the difference between the third point (0.14MPa) and the second point (0.09MPa) is 0.05M Pa, the difference between the 3rd point and the 4th point (0.12MPa) is 0.02MPa. Only the former difference is greater than the threshold, which does not constitute a mutation. If it is [0.06, 0.10, 0.15], the change range between 0.10 and 0.15 is 0.05MPa, which meets the mutation standard and is identified as a jump point. If there are multiple sampling points that meet the mutation conditions in the same interval in the jump point sequence, the maximum jump value is taken as the representative mutation point, and its time index is further used as the time point of the pressure difference mutation in this section, forming a The process is executed in all pressure difference response intervals in sequence, and finally a plurality of jump time points are obtained, such as index positions 127, 218, 379, etc., forming an identification point set with sudden pressure difference recognition as the core. The set mutation threshold is greater than ±0.03MPa, which comes from the synchronous amplitude analysis between the pump-controlled pressure difference fluctuation and the actual mechanical response intensity. Specifically, the pump outlet pressure monitoring curve and the corresponding flow rate change rate linkage change data are referred to. If the pressure difference change rate exceeds 0.03MPa / s within 1s, the flow mutation rate will show a significant jump phenomenon. Combined with the fluid inertia characteristics and the trend window of the pressure regulation rate stability range between 0.02MPa / s and 0.03MPa / s under the valve control mechanism, ±0.03MPa is comprehensively defined as the basis for the mutation threshold, and its value will be adjusted with the change of the rated output flow of the pump. For a system with a flow value of 12m3 / h, the pressure difference mutation value is a reasonable critical setting, and finally the pressure difference mutation time point set is extracted and formed.

[0071] S113: According to the pressure difference mutation time point set, for each jump point, find the pressure difference response interval to which it belongs, using the formula:

[0072]

[0073] Calculate the control-associated pressure difference strength value L of the interval corresponding to the pressure difference mutation point, and use the control-associated pressure difference strength value as the index basis to bind the jump point with the interval information to obtain the control-associated pressure difference annotation set, where ΔP k represents the pressure difference value at the kth time point, T k is the sampling timestamp corresponding to the pressure difference value, and K represents the total number of pressure difference jump segments involved in the calculation in the current pressure difference interval;

[0074] According to the set of differential pressure mutation time points, for each mutation point, find the corresponding differential pressure response interval segment to which it belongs, and determine whether it belongs to the previously extracted differential pressure interval index range. If the jump point is between interval indices 123 and 132, then bind this interval segment to the mutation point to form a labeling unit. Further calculate its control-related differential pressure intensity value, and all differential pressure sequences within this interval segment need to be extracted for calculation.

[0075] Substitute the differential pressure sequence of the interval segment [0.06, 0.09, 0.14, 0.10] MPa and the time point sequence [12.3, 12.4, 12.5, 12.6] s into the formula for calculation as follows:

[0076] The first item:

[0077] The second item:

[0078] The third item:

[0079]

[0080] The result shows that the control-related differential pressure intensity value of the interval segment where the current jump point is located is 1.089, belonging to a control signal response segment with obvious characteristics. By calculating this value for all jump points and binding their original intervals, a control-related differential pressure annotation set is formed.

[0081] Please refer to Figure 3 , and the steps of S2 are as follows:

[0082] S211: According to the differential pressure mutation section calibrated in the control-related differential pressure annotation set, obtain the original pressure sampling values within the corresponding time range, locate and extract them according to the time interval of the mutation section, and divide the extracted sequence into multiple overlapping segments along the time axis to obtain a sliding time segment sequence;

[0083] Mark the pressure difference mutation section according to the control correlation pressure difference annotation set, and obtain the original pressure sampling values within the corresponding time range. In specific operations, first read the start and end timestamps included in each record in the annotation set. If the start time of the pressure difference mutation section corresponding to a certain record is 18.20 seconds and the end time is 20.80 seconds, then extract the original data points within the corresponding time interval from the original pressure data sequence. Assuming a sampling interval of 0.02 seconds, the extraction interval should include 130 data points, thus forming a complete time period pressure sequence. Then, set a sliding window with a fixed time window width for segmented processing. For example, set the sliding window length to 20 points and the step size to 5 points. Then, multiple consecutive overlapping window segments will be generated within each pressure difference section. For example, the first window is at the 120th point, the second window is at the 625th point, and so on, forming multiple overlapping windows. For these window sequences, by traversing the pressure difference mutation section, extract the sliding windows in turn, and align the start and end times of the windows with the timestamps of the pressure sampling data to ensure that the time period division corresponds to the control signal sequence. Each window segment will form multiple time segment sets, which are convenient for subsequent unified processing. Suppose the number of windows is 6, then 6 sliding window segment data with different start and end times can be obtained under each mutation section, forming a sliding time segment sequence.

[0084] S212: Based on the sliding time segment sequence, for the original pressure sampling values within each time segment, perform first-order difference processing, trend slope calculation, sliding standard deviation calculation, and fluctuation interval mean calculation respectively to generate a set of pressure fluctuation characteristic values corresponding to each time segment;

[0085] According to the sliding time segment sequence, process the original pressure sampling values within each time segment, and calculate their first-order differences, trend slopes, sliding standard deviations, and fluctuation interval averages respectively. In specific implementation, first perform first-order difference processing on the sampling values within each time segment, that is, calculate the pressure difference between adjacent data points. Suppose the pressure sampling sequence in the first window segment is {1.02, 1.04, 1.05, 1.07, 1.09} MPa, then the first-order difference sequence is {0.02, 0.01, 0.02, 0.02} MPa; then calculate the trend slope of the data in this window. The slope of linear fitting can be used as the trend feature. If this time period is {0s, 0.02s, 0.04s, 0.06s, 0.08s}, then the linear slope can be fitted by the least squares method to obtain a trend value of 0.175 MPa / s; the sliding standard deviation uses the standard formula, that is, the square root of the average of the squared differences between each pressure value and the mean of this segment. If the pressure mean is 1.054 MPa, then the standard deviation is

[0086]

[0087] The average value of the fluctuation range is the arithmetic mean of all pressure sampling values within the window, which is 1.054 MPa. The combination of all the above eigenvalue constitutes the pressure fluctuation eigenvalue group corresponding to this window segment. By traversing all window segments and performing the above operations in sequence, the feature set corresponding to all time segments is obtained.

[0088] S213: According to the pressure fluctuation eigenvalue group, using the time series order corresponding to each group of features as the index, the formula is adopted:

[0089]

[0090] Calculate the multi-dimensional time series eigenvalue F, splice all eigenvalues in order into a vector structure, and establish a multi-dimensional time series eigenvalue group. Among them, ΔS q represents the q-th first-order difference, M q represents the q-th trend slope, V q represents the q-th moving standard deviation, A q represents the q-th average value of the fluctuation range, and Q represents the number of feature items within each sliding time segment;

[0091] According to the pressure fluctuation eigenvalue group, splice each group of pressure features in chronological order to construct a multi-dimensional vector structure for subsequent analysis and processing. During the splicing process, unify the structure format according to the feature dimension, and generate the feature vector value through the average function.

[0092] Among them, each group of eigenvalues is input into this formula for conversion in sequence. Assume the parameters in the first sliding window segment are as follows:

[0093] ΔS = {0.02, 0.01, 0.02, 0.02};

[0094] M = {0.17, 0.16, 0.18, 0.15};

[0095] V = {0.025, 0.023, 0.026, 0.022};

[0096] A = {1.05, 1.06, 1.04, 1.05};

[0097] Then the step-by-step operations are as follows:

[0098] 1) Calculate a single feature item and take the first item:

[0099]

[0100] ≈0.02 + 0.00287 = 0.02287;

[0101] 2) Calculate each item in sequence and then find the mean value:

[0102]

[0103] Finally, the multi-dimensional time series eigenvalue corresponding to this window segment is 0.02242. The eigenvalue vectors of multiple window segments are concatenated and combined in chronological order to establish a multi-dimensional time series eigenvalue group.

[0104] To clarify the acquisition and setting methods of each parameter value, the following schematic table of pressure sampling eigenvalues is constructed:

[0105] Table 2 Schematic Table of Pressure Eigenvalues of Sliding Time Windows

[0106] Window Number First-Order Difference Mean (MPa) Trend Slope (MPa / s) Moving Standard Deviation (MPa) Fluctuation Mean (MPa) 1 0.017 0.175 0.0253 1.054 2 0.015 0.162 0.0231 1.061 3 0.018 0.180 0.0267 1.049 4 0.016 0.159 0.0228 1.057

[0107] As shown in Table 2, the pressure characteristic parameters corresponding to each window segment are maintained within a reasonable floating range, which can truly reflect the pressure dynamics within the time segment. Based on the parameter values of Window 2 in this table, the eigenvalue calculated by substituting into the formula is 0.02242. This result indicates that the characteristic changes in this window segment remain stable and can be used to construct a representative sequence eigenvalue vector to further support the setting of the eigenvalue input layer in subsequent behavior prediction or response analysis.

[0108] Please refer to Figure 4 , and the steps of S3 are as follows:

[0109] S311: Based on the multi-dimensional time series eigenvalue group, combined with the sampling interval index corresponding to each time series eigenvalue, compare the control type annotation information in the control correlation pressure difference annotation set, identify the corresponding control behavior attributes, obtain the label type as normal control or abnormal mutation, and generate a control behavior label sequence;

[0110] Based on the multi-dimensional time series feature group, it is necessary to call the timestamps and source indexes of each item of data in the feature group, and correspondingly search for the time interval in which the feature segment is located in the control-related differential pressure annotation set, and extract its annotation type information. Among them, the control type information is often a binary marker, which respectively represents the differential pressure fluctuation caused by normal control or abnormal mutation behavior. Specifically, the quantization label can be set through a mapping method. For example, normal control is set to 0, and abnormal mutation is set to 1. The determination basis for abnormal mutation is that the differential pressure mutation amplitude is greater than 0.35 MPa and the corresponding number of sampling points is less than 5. When setting the differential pressure mutation amplitude, referring to the sampling period of 0.05 seconds during the valve opening and closing process, a pressure mutation exceeding 12% of the rated operating pressure (referring to 3 MPa) can cause transient reflux disturbance. Therefore, the mutation threshold is set to 0.35 MPa. The number of sampling points less than 5 indicates that the mutation process is completed within 0.2 seconds, belonging to non-stationary and rapid disturbance behavior. If only 0.25 MPa mutates, it is judged as a process fluctuation rather than an abnormal event. If 0.5 MPa mutates but spans more than 6 points, it is classified as a slow change process. This threshold tends to decrease as the sampling period shortens, and the sampling point limit becomes stricter as the fluctuation rate increases. For example, if the pressure is detected to mutate from 1.25 MPa to 1.63 MPa within a certain time period and only spans 3 sampling points, it meets the above conditions, so it is marked with the abnormal mutation label 1. During the execution process, it is necessary to sequentially match the control state types of all time series feature segments in order and generate a label set. Among them, the label set should be consistent with the length of the feature sequence, and the label assignment process is completed through the mapping relationship. The sample data listed in Table 3 is used to mark the actual control type to form a one-to-one corresponding label output.

[0111] Table 3 Control Behavior Sample Label Table

[0112] Time Period Number Start Time (s) End Time (s) Sudden Change in Differential Pressure (MPa) Number of Sampling Points Control Type Label 1 15.20 16.00 0.12 6 0 2 21.10 21.25 0.38 3 1 3 30.50 31.00 0.26 7 0 4 45.00 45.15 0.42 2 1

[0113] As shown in Table 3, using this type of label structure, a mapping rule can be established between the original sampling segment and the annotation set, and finally a control behavior label sequence can be obtained.

[0114] S312: According to the multi-dimensional time series feature group and the control behavior label sequence, extract the corresponding positions of the feature vector and the label item by item, and establish a sample unit in order of pairing. Use the formula:

[0115]

[0116] Calculate the paired deviation combination value Z of the i-th sample i , to characterize the structural differences and deviation relationships between feature dimensions, and generate a feature label pairing group. Among them, X1, X2, X3, X4 represent the numerical values of the first four dimensions in the feature vector, and Y1, Y2, Y3 represent the label quantization values on the corresponding previous dimensions;

[0117] According to the multi-dimensional time series feature group and the control behavior label sequence, each feature vector is paired with the label in the corresponding order. During the pairing process, representative values need to be extracted from the feature vector, and a structural comparison operation is performed in combination with the label quantization value to characterize the pairing difference degree between the label and the feature. The deviation reference value is set to 0.6, which is derived from the clustering distribution results of normal control and abnormal control samples in the pairing deviation value interval in the sample set. The setting basis is that in the case of a feature dimension of 4, the pairing structures greater than 0.6 are mostly concentrated in the abnormal pairing area, and this threshold is slightly increased as the feature dimension increases. The feature value units are all normalized to facilitate the direct reflection of the structural differences in the amplitude level during pairing. If it is lower than this value, the combination structure of the feature and the label tends to be consistent, and those between 0.4 and 0.6 are the overlapping areas. Therefore, 0.6 is set as the deviation limit value in the current dimension. The actual input sample is set as follows: the feature vector is [1.3, 0.8, 1.1, 0.7], and the label is abnormal mutation. Therefore, the quantized label value is [1, 1, 1]. Substituting it into the formula for calculation is as follows:

[0118]

[0119] This value represents the structural deviation intensity between the current feature and label combination. The higher the value, the greater the deviation degree. By comparing it with the deviation reference value of 0.6, it can be determined that the pairing structure of this sample shows a medium-strong deviation state and tends to be classified into the abnormal structure samples. This pairing result is used as a difference marker input for subsequent screening and modeling. The advantage of the formula is that it constructs a non-linear measurement structure through the difference combination between asymmetric features and labels, avoids the average weakening of fluctuation features, and amplifies the expression ability of discrete distributions in the pairing relationship, and finally generates a feature-label pairing group.

[0120] S313: Based on the feature-label pairing group, all feature vectors and the corresponding control behavior labels are combined in a set manner according to the pairing results to establish a data structure available for the training task, and a control behavior training sample set is obtained;

[0121] According to the pairing result structure in the pairing group of feature tags, based on the feature vectors and label combinations in the pairing items, they are recombined in a sequential binding manner to form an ordered input structure, that is, a control behavior training sample set is constructed. During the operation, it is necessary to ensure that the feature dimensions are completely retained after pairing, and at the same time, there is no order misalignment or label drift in the label combination. The feature-label order consistency can be verified through the sample index verification mechanism. If the input indices are sample numbers 1, 2, and 3, the corresponding feature vectors are [1.2, 0.6, 0.9, 0.8], [1.5, 1.0, 1.3, 0.9], [0.9, 0.5, 0.8, 0.6], and the corresponding labels are 0, 1, 0 respectively, then the combined result can be written as three complete sample structures, stored in a data table format, with column fields being number, feature vector, and behavior label. Finally, the input sample set structure is encapsulated to obtain the control behavior training sample set.

[0122] Please refer to Figure 5 , and step S4 is as follows:

[0123] S411: Based on the control behavior training sample set, sequentially extract the feature vectors and corresponding class labels in the samples, and input them into the neural units at all levels in the neural network structure in order to construct a gradually conductive data structure and obtain the neuron input vector set;

[0124] Based on the control behavior training sample set, it is necessary to extract the feature vectors in the samples and their corresponding class labels. The feature vectors are the multi-dimensional time series feature groups generated in the previous steps, including information such as the difference, slope, and fluctuation amplitude of the pressure signal in each time period. In a specific implementation, for example, a pressure data sequence {102.5, 101.8, 100.6, 100.1, 99.8} kPa within a control period is selected. Using a sliding window, the first-order difference is extracted as {-0.7, -1.2, -0.5, -0.3} kPa, the trend slope is -0.675, and the average value of the fluctuation interval is 100.36 kPa. When actually constructing the training sample, such numerical combinations are combined into a feature vector of length 4 [-0.7, -1.2, -0.5, -0.3]. If this section of data is determined to be "normal control-induced pressure difference change" in the control annotation set, its corresponding label is 0; conversely, if it is an "abnormal mutation behavior", its label is 1. After all samples are organized in chronological order to construct a data stream, each group of input data is aligned with the label and input into the neural network neurons to prepare for the training stage. During the process, it is necessary to perform standardization processing on the dimension of the feature vector. The normalization operation controls the value range within the interval [0, 1] to ensure the stability of the subsequent network input layer. For example, after normalization, the above data becomes [0.583, 1.0, 0.417, 0.25], and it is combined with the label "0" to form a sample input unit. The entire data set is organized into a two-dimensional input tensor by batches and then input into the network structure, and finally a set of neuron input vectors is obtained.

[0125] S412: According to the set of neuron input vectors, compare the output results of the activated input features layer by layer in the network, analyze the changing trends of neural responses under different class labels, extract the internal structural patterns of the sequence, and obtain the classification boundary response features;

[0126] According to the neuron input vector set, it is necessary to calculate the output value of each neuron layer by layer according to the network structure. The network generally adopts a three-layer structure, and each layer contains different numbers of neuron units, corresponding to the feature extraction layer, the feature transformation layer, and the output discrimination layer respectively. In practice, the input sample vector [0.583, 1.0, 0.417, 0.25] is input into the neurons of the first layer for calculation. The output of each neuron is activated after weighted summation. For example, the calculation result of the first neuron in the first layer is z = 0.583×w1 + 1.0×w2 + 0.417×w3 + 0.25×w4 + b. During actual training, the initial weights are set as w = [0.4, -0.3, 0.2, 0.5], and the bias b = 0.1. The basis for this setting is that after input feature normalization, the value range is limited between [0, 1]. Therefore, the corresponding initial weights need to be kept within the same numerical range as the input to avoid the activation value getting out of control during the first round of forward propagation. At the same time, positive and negative weights are selected and juxtaposed to simulate the pulling effect of different input dimensions on the output activation direction. Among them, 0.4, 0.2, and 0.5 correspond to the enhancement effects of difference, trend, and fluctuation indicators, and -0.3 is used to reflect the reverse guiding characteristic of a certain feature when the label is expected to be 0. The bias term is set to 0.1, and its setting is based on the zero-centered output adjustment logic, that is, to make the neuron output deviate slightly from 0 when not activated to facilitate the stable response of the activation function. This value fluctuates with the network depth and the sample mean of batch input, and needs to be fixed after testing multiple sample means. Usually, the standard deviation of the sample mean multiplied by 0.1 is used as the basic judgment interval. After substituting the above weights and biases, the calculation is as follows

[0127] z = 0.583×0.4 + 1.0×(-0.3) + 0.417×0.2 + 0.25×0.5 + 0.1 = 0.2332. After being processed by the activation function, the output value is used as the input for the next layer. Repeating this operation process forms multi-layer feature transformation, and a prediction output is generated at the last layer. Under the input of different category samples, the network activation paths are different, and the response boundaries shown at the output nodes also vary. It is necessary to calculate the deviation degree between the true label and the output. In practice, if the true label is 0 and the output value is 0.76, the error is 0.76. If the output is 0.21, the error is 0.21, indicating that the latter is better under the current network weight combination. Thus, the neural response state under each weight combination is extracted and the corresponding error change trend is recorded, and finally the classification boundary response feature is obtained.

[0128] S413: According to the classification boundary response feature, compare the error value between the network output and the true label, screen the network weight combination corresponding to the minimum error in each sample training round, extract the weight structure with the highest matching degree, and obtain the differential pressure recognition model parameter group;

[0129] According to the classification boundary response characteristics, in the set of error values between each sample input and the network output, select the combination of weight structures with the smallest error value as the preferred path. By setting an error tolerance threshold, a parameter set with an error lower than a given limit can be screened out. This threshold can be set by the lower quartile value of the error distribution of the sample training rounds. For example, in a training, the set of all sample errors is {0.76, 0.21, 0.53, 0.39, 0.11}, and after sorting, it becomes 0.11, 0.21, 0.39, 0.53, 0.76. Then the lower quartile is 0.21. Select the parameter combination with an error lower than this value, which can be used as the optimal parameter group in the current training stage. Record the corresponding weight matrix and bias value, establish the parameter group structure, form a training network for the subsequent recognition and inference stage, and finally obtain the parameter group of the differential pressure recognition model.

[0130] Please refer to Figure 6 , step S5 is as follows:

[0131] S511: Obtain the real-time pressure sampling data during the operation of the oil well. According to the set sliding window structure in the control correlation differential pressure annotation set, slide and segment the real-time data in the same window manner to generate continuous window segments, and construct a time index structure for each segment to obtain a set of real-time sliding subsequences;

[0132] Obtain the real-time pressure sampling data during the operation of the oil well, which is collected based on the pressure sensors arranged at the wellhead and the middle section of the tubing in actual production. The pressure value of a group is collected every second in real time, and the sample frequency is 1 Hz, forming a pressure sequence {101.2, 100.8, 101.5, 102.1, 100.6, ……}. Through timestamp calibration, the correlation between the data stream and the clock is established. Subsequently, according to the defined sliding window structure in the control correlation pressure difference annotation set, a sliding segmentation operation is performed. If the sliding window length is 10 seconds and the step size is set to 5 seconds, the sequence is slid and divided into multiple overlapping segments. For example, the first segment is from the 1st second to the 10th second, the second segment is from the 6th second to the 15th second, and so on, forming a set of real-time sliding subsequences containing start and end timestamps. For each subsequence, extract the internal time axis order, construct a data structure indexed by the window number and identified by the start and end time information, and represent it in an array as {subsequence 1: [101.2, 100.8, 101.5, 102.1, 100.6, 101.3, 101.8, 101.1, 100.9, 101.4]}. Then, match the sampling points corresponding to the time index for each period, and calibrate the coverage time range of each subsequence as "2025-03-31 10:00:00~10:00:09", "2025-03-31 10:00:05~10:00:14", etc. At the data interface layer, synchronously record the data packet number and the sampling source node. Verify the data continuity of the subsequence through the sampling label sent by the acquisition end in real time to ensure that the real-time data forms a continuous, complete, and sequential sliding structure. Combining the actual background of the oil well operation stage (such as the stable production stage and the liquid drainage stage), the data sub-segments are correspondingly bound to the production status. The setting of the "abnormal fluctuation lower limit value of 95 kPa" is determined based on the pressure range fluctuation monitored during the stable operation of the oil well liquid drainage stage. The minimum pressure under stable conditions is usually not less than 96.2 kPa. Considering the fluctuation tolerance, the threshold is set to a 1.5% decrease, that is, 95 kPa. This value has a linkage fluctuation with the change of the wellbore liquid level height and the pump speed stability within the stage. If the sampling value of the oil pressure sensor is continuously lower than 95 kPa for 2 seconds, it is marked as an abnormal fluctuation event segment. Finally, complete the set of real-time sliding subsequences constructed based on the actual sliding window rules.

[0133] S512: Based on the set of real-time sliding subsequences, extract the pressure characteristics of each subsequence in chronological order, construct a feature vector group with the same structure as that in the training stage, input the parameter group of the differential pressure identification model for forward prediction, identify the abnormal category time period in the prediction result, establish an abnormal time point index set, and calculate the amplitude quantity by combining the extreme value fluctuation of the differential pressure within the window where the abnormal sample is located to obtain the abnormal predicted differential pressure amplitude value;

[0134] According to the real-time sliding subsequence set, for each pressure data point in the sequence, extract time series difference features, first-order trend, difference between maximum and minimum values, local average, range index and other feature items in sequence, and construct the feature vector format used in the training stage. For example, if the values in the subsequence are [101.2, 100.8, 101.5, 102.1, 100.6], then calculate its first-order difference as [-0.4, 0.7, 0.6, -1.5], the trend value is 0.225, the maximum value is 102.1, the minimum value is 100.6, the range is 1.5, and the average value is 101.24. Thus, construct the feature vector T1 = [0.225, 1.5, 101.24]. Concatenate all the extracted feature items to form a vector set with consistent structure. Perform forward prediction operations on each feature vector using the pressure difference recognition model parameter group, and extract the response values of various output nodes through the network feed-forward structure. During model training, the abnormal category probability threshold is set to 0.7. This value is based on the statistical distribution of the output probabilities of different category samples in the training set during forward prediction. The maximum output probability of the normal category mostly concentrates in the interval of 0.3 - 0.6, and the output probability of abnormal category samples during the training stage is mostly higher than 0.75. Therefore, set 0.7 as the cut-off boundary to ensure that most high-confidence abnormal samples are retained. This value is dynamically adjusted according to the proportion of abnormal categories in the sample set. If the probability score of the abnormal category in the model output exceeds 0.7, then mark this window as the abnormal category. For example, if the corresponding output result of a certain subsequence is: the normal category is 0.28, and the abnormal category is 0.72, then the time period "2025-03-31 10:00:00~10:00:09" is marked as an abnormal category window, and record the range value of the pressure value change amplitude in this sequence as the fluctuation intensity result, which is 102.1 - 100.6 = 1.5 kPa. Organize the subsequence number, start and end time, and pressure difference fluctuation range into three items of data. The example is as follows:

[0135] Table 4 Real-time Subsequence Prediction and Identification Table

[0136] Subsequence Number Start Time End Time Differential Pressure Fluctuation Value (kPa) 1 2025-03-3110:00:00 2025-03-3110:00:09 1.5 2 2025-03-3110:00:05 2025-03-3110:00:14 2.3 3 2025-03-3110:00:10 2025-03-3110:00:19 1.2

[0137] As shown in Table 4, extract the pressure difference fluctuation range of each subsequence and mark the abnormal category, which effectively supports the subsequent construction and output of the abnormal list.

[0138] S513: According to the abnormal predicted pressure difference amplitude value and the corresponding abnormal time point index set, summarize the abnormal segment range in chronological order and classify and organize it to generate the abnormal section list structure, and obtain the real-time detection result of the oil well pressure abnormality;

[0139] According to the abnormal predicted pressure difference amplitude value and its corresponding abnormal time point index set, all abnormal category subsequences are merged and classified in chronological order to construct a list of data entries based on the start and end times and the pressure difference amplitude. By judging whether there is time continuity between adjacent abnormal windows, if the interval between the two is less than half of the window length, they are merged into one abnormal record. For example, if there is no gap between the time periods of subsequence 1 and 2, they are merged into "2025-03-31 10:00:00~10:00:14", and the maximum pressure difference value between the two segments is selected as the representative pressure difference amplitude for this segment, that is, max(1.5, 2.3) = 2.3 kPa. At the same time, record the well number, data source node, and pressure sequence index information belonging to this segment. Finally, an abnormal list with a standard structure is formed. The list fields include "abnormal segment number, start time, end time, abnormal type, pressure difference amplitude value, data source number", which are organized according to the structure specifications and exported to a callable data structure to form the real-time detection result of the oil well pressure anomaly.

[0140] An oil well pressure anomaly detection system based on deep learning, comprising:

[0141] The data processing module obtains the pump control start / stop signal and the valve control instruction timestamp, combines the time points in the pressure sampling sequence, calculates the pressure difference fluctuation value between adjacent sampling points, judges whether the pressure difference fluctuation is within the pump control response threshold interval, identifies the mutation position and marks the pressure difference mutation section to obtain the control-related pressure difference annotation set;

[0142] The feature extraction module obtains the original pressure sampling values within the corresponding time range according to the pressure difference mutation section in the control-related pressure difference annotation set, constructs a sliding analysis window and calculates the first-order difference, trend slope, sliding standard deviation, and average value of the fluctuation interval to generate a multi-dimensional time series feature group;

[0143] The sample construction module assigns labels to normal control and abnormal mutation behaviors respectively based on the multi-dimensional time series feature set and combines the control type information, pairs the feature vectors with the labels to obtain the control behavior training sample set;

[0144] The deep learning training module inputs the feature vectors into the neural units in the deep learning network based on the control behavior training sample set, extracts the internal change patterns of the sequence, performs classification boundary mapping, and selects the network weight parameters according to the minimum loss value to obtain the pressure difference recognition model parameter set;

[0145] The real-time detection module obtains the real-time pressure sampling data during the operation of the oil well, combines the sliding window structure in the control-related pressure difference annotation set, extracts the real-time data feature vectors, performs forward prediction calculations using the pressure difference recognition model parameter set, marks the abnormal category time period, and generates the real-time detection result of the oil well pressure anomaly.

[0146] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the relevant art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting abnormal oil well pressure based on deep learning, characterized in that, Including the following steps: S1: Obtain the pump control start / stop signal and the valve control instruction timestamps, calculate the slope difference and fluctuation by combining with the pressure sequence, mark the mutation points, and generate a control-related differential pressure annotation set in combination with the control behavior; S2: Based on the control-related differential pressure annotation set, extract the pressure sequence window, calculate the first-order difference, trend slope, sliding standard deviation, and average value of the fluctuation interval, and construct a multi-dimensional time series feature group; S3: Based on the multi-dimensional time series feature group, according to the control type information of the corresponding samples in the control-related differential pressure annotation set, assign labels respectively and pair the feature vectors with the corresponding labels to construct a control behavior training sample set; S4: Based on the feature vectors and class labels of the control behavior training sample set, input them into the neural units in the deep learning network structure in sequence, extract the internal change patterns and establish the classification boundary mapping, select the network weight parameters, and generate a differential pressure recognition model parameter group; S5: Obtain the real-time pressure sampling data during the operation of the oil well, extract the feature vectors from the real-time data, use the differential pressure recognition model parameter group for prediction calculation, identify the time period of the abnormal pressure category, and generate the real-time detection result of the oil well pressure anomaly.

2. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, characterized in that The control-related differential pressure annotation set includes the differential pressure mutation section, mutation position label, control response type, and control time index. The multi-dimensional time series feature group includes the time series structure feature, statistical change index, differential feature vector, and local trend expression record. The control behavior training sample set includes the behavior category label, feature representation combination, and classification sample unit. The differential pressure recognition model parameter group includes the deep feature mapping parameter, classification boundary function, and optimization objective function. The real-time detection result of the oil well pressure anomaly includes the abnormal recognition interval, corresponding category identifier, abnormal time positioning, and differential pressure anomaly amplitude.

3. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, characterized in that, The specific steps for obtaining the control-related differential pressure annotation set are as follows: S111: Based on the pump control start / stop signal and the valve control instruction timestamps, extract the corresponding pressure sampling sequence time points, calculate the change sequence of the differential pressure between adjacent sampling points within a continuous time period, judge whether the difference amplitude between adjacent differential pressures falls within the pump control response threshold interval, mark the continuous intervals that meet the threshold interval constraint conditions, and obtain the differential pressure response interval segment; S112: According to the differential pressure response interval segment, extract the corresponding time points according to the extreme value positions of the differential pressure fluctuation therein, perform differential pressure jump discrimination processing on the continuous differential pressure sequence, and extract the jump points according to the sign directionality and numerical volatility calculation method of the differential pressure sequence mutation to generate a set of differential pressure mutation time points; S113: According to the set of differential pressure mutation time points, find the differential pressure response interval segment to which each jump point belongs, and use the formula: Calculate the control-related pressure difference intensity value L corresponding to the interval where the pressure difference mutation point is located. Based on the control-related pressure difference intensity value as the index, bind the jump point with the interval information to obtain the control-related pressure difference annotation set, where ΔP k represents the pressure difference value at the k-th time point, and T k is the sampling timestamp corresponding to the pressure difference value. K represents the total number of pressure difference jump segments participating in the calculation in the current pressure difference interval segment.

4. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, wherein The specific steps for obtaining the multi-dimensional time series feature group are as follows: S211: According to the differential pressure mutation section calibrated in the control-related differential pressure annotation set, obtain the original pressure sampling values within the corresponding time range, perform positioning extraction according to the time interval of the mutation section, and divide the extracted sequence into multiple overlapping segments along the time axis to obtain a sliding time segment sequence; S212: Based on the sliding time segment sequence, for the original pressure sampling values within each time segment, perform first-order difference processing, trend slope calculation, sliding standard deviation calculation, and fluctuation interval mean calculation respectively to generate a pressure fluctuation eigenvalue group corresponding to each time segment; S213: According to the pressure fluctuation eigenvalue group, using the time series order corresponding to each group of features as the index, adopt the formula: Calculate the multi-dimensional time series eigenvalue F, concatenate all the eigenvalues in order into a vector structure, and establish a multi-dimensional time series feature group, where ΔS q represents the q-th first-order difference, M q represents the q-th trend slope, V q represents the q-th moving standard deviation, A q represents the average value of the q-th fluctuation interval, and Q represents the number of feature items within each sliding time segment.

5. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, wherein The specific steps for obtaining the control behavior training sample set are as follows: S311: Based on the multi-dimensional time series feature group, combined with the sampling interval index corresponding to each time series feature, compare the control type annotation information in the control-related differential pressure annotation set, identify the corresponding control behavior attributes, obtain the label type as normal control or abnormal mutation, and generate a control behavior label sequence; S312: According to the multi-dimensional time series feature group and the control behavior label sequence, extract the feature vectors and the corresponding positions of the labels item by item, and establish sample units in sequence pairing using the formula: Calculate the paired deviation combination value Z of the i-th sample i , characterize the structural differences and deviation relationships between feature dimensions, and generate a paired group of feature labels, where X1, X2, X3, X4 represent the numerical values of the first four dimensions in the feature vector, and Y1, Y2, Y3 represent the label quantization values on the corresponding previous dimensions; S313: Based on the feature-label pairing group, combine all the feature vectors and the corresponding control behavior labels in a set manner according to the pairing results, establish a data structure for training tasks, and obtain the control behavior training sample set.

6. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, characterized in that The specific steps for obtaining the differential pressure recognition model parameter group are as follows: S411: Based on the control behavior training sample set, sequentially extract the feature vectors and the corresponding class labels in the samples, and input them into the neural units at all levels in the neural network structure in order to construct a data structure for step-by-step conduction, and obtain the neuron input vector set; S412: According to the neuron input vector set, compare the output results of the input features after activation in each layer of the network item by item, analyze the change trend of the neural response under different class labels, extract the structural pattern inside the sequence, and obtain the classification boundary response feature; S413: According to the classification boundary response feature, compare the error value between the network output and the true label, screen the network weight combination corresponding to the minimum error in each sample training round, extract the weight structure with the highest matching degree, and obtain the differential pressure recognition model parameter group.

7. The method for detecting abnormal oil well pressure based on deep learning according to claim 1, wherein The specific steps for obtaining the real-time detection result of abnormal oil well pressure are as follows: S511: Obtain the real-time pressure sampling data during the operation of the oil well. According to the set sliding window structure in the control-related differential pressure annotation set, slide and segment the real-time data in the same window manner to generate continuous window segments, and construct a time index structure for each segment to obtain the real-time sliding subsequence set; S512: Based on the real-time sliding subsequence set, extract the pressure features for each subsequence in chronological order, construct a feature vector group with the same structure as that in the training stage, input it into the differential pressure recognition model parameter group for forward prediction, identify the abnormal category time period in the prediction result, establish an abnormal time point index set, and calculate the amplitude of the differential pressure extreme value fluctuation within the window where the abnormal sample is located to obtain the abnormal predicted differential pressure amplitude value; S513: Summarize the abnormal section ranges in chronological order and classify and organize them according to the predicted abnormal pressure difference amplitude values and the corresponding abnormal time point index sets, generate an abnormal section list structure, and obtain the real-time detection results of oil well pressure anomalies.

8. An oil well pressure anomaly detection system based on deep learning, characterized in that, The system is used to execute the method according to any one of claims 1-7, including: The data processing module obtains the pump control start / stop signal and the valve control instruction timestamp, combines the time points in the pressure sampling sequence, calculates the pressure difference fluctuation value between adjacent sampling points, determines whether the pressure difference fluctuation is within the pump control response threshold interval, identifies the mutation position and marks the pressure difference mutation section to obtain the control-related pressure difference annotation set; The feature extraction module obtains the original pressure sampling values within the corresponding time range according to the pressure difference mutation sections in the control-related pressure difference annotation set, constructs a sliding analysis window and calculates the first-order difference, trend slope, sliding standard deviation and average value of the fluctuation interval to generate a multi-dimensional time series feature group; The sample construction module, based on the multi-dimensional time series feature set and combined with the control type information, assigns labels to normal control and abnormal mutation behaviors respectively, pairs the feature vectors with the labels to obtain a control behavior training sample set; The deep learning training module, based on the control behavior training sample set, inputs the feature vectors into the neural units in the deep learning network, extracts the internal change patterns of the sequence, performs classification boundary mapping, and selects the network weight parameters according to the minimum loss value to obtain the pressure difference recognition model parameter set; The real-time detection module obtains the real-time pressure sampling data during the operation of the oil well, combines the sliding window structure in the control-related pressure difference annotation set, extracts the real-time data feature vectors, performs forward prediction calculations using the pressure difference recognition model parameter set, marks the abnormal category time periods, and generates the real-time detection results of oil well pressure anomalies.

Images