A method for intelligent identification of drilling conditions based on multimodal self-supervised alignment and parallel temporal networks

Abstract

This invention discloses an intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks. The method includes: acquiring multi-source temporal data during drilling operations; each window sample contains a sequence of engineering parameters, a sequence of vibration features, and a state / log sequence over a period of time; generating engineering features, vibration features, and state features using a multimodal encoder; inputting the engineering features into a parallel temporal network, mapping the vibration features and state features to the same feature space as the fused temporal features, and then combining them with the fused temporal features to obtain multimodal fused features; and classifying drilling conditions based on the multimodal fused features. This reduces label dependency, specifically through mask reconstruction and cross-modal comparison alignment, achieving robust representation even with minimal labeling; and improves the accuracy of complex condition identification by using parallel CNNs for local mutations combined with long dependencies of Informer algorithms for complementarity.

Description

Technical Field

[0001] This application relates to the field of oil and gas drilling data processing and intelligent identification technology, and in particular to an intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks. Background Technology

[0002] During oil and gas drilling operations, the on-site monitoring and control system, equipment control system, and data acquisition link generate a large amount of multi-source time-series data. This data acquisition link, comprised of sensors, PLCs / controllers, acquisition gateways, communication networks, and data platforms, is typically used for operation monitoring, efficiency analysis, anomaly alarms, post-mortem evaluation, remote support, and intelligent applications. Drilling conditions are characterized by frequent switching, blurred boundaries, and significant influence from differences in well section / formation / equipment, such as rotary drilling, sliding drilling, single-hole connection, tripping, running in, forward / reverse reaming, circulation, and directional drilling.

[0003] To achieve operational condition identification, existing technologies mainly employ manual judgment / manual reporting. This method relies on the driller's experience or daily records, resulting in strong subjectivity, poor timeliness, difficulty in scaling, and inability to achieve precision down to the second. Alternatively, fixed threshold / rule methods can be used, such as setting thresholds or logical conditions based on changes in drilling pressure, rotation speed, pump pressure, and well depth. The advantage is its simplicity, but the disadvantages include difficulty in generalizing thresholds across well sections; serious misjudgments and omissions under complex operating conditions; and sensitivity to noise, missing data, and anomalies.

[0004] To address the problems inherent in traditional operating condition identification methods, some solutions employ supervised learning, primarily utilizing traditional machine learning or deep learning for classification. However, these methods typically require a large number of high-quality labels, while on-site labeling is costly, inconsistent in scope, and difficult to maintain continuously. Furthermore, the models lack sufficient transferability across wells, blocks, and equipment. They also cannot adapt to the unique scenarios of drilling operations, such as the difficulty in time alignment of multi-sampling rate and multi-source heterogeneous data; while auxiliary modalities such as vibration and equipment logs are valuable, they are difficult to effectively integrate with engineering parameters; single network structures struggle to balance local abrupt changes (transient operating condition transitions) with long-term dependencies (operational segment trends); output boundary jitter leads to distorted operating condition time statistics; and the lack of interpretable output makes it difficult to support engineer review and trust. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, one objective of this application is to propose an intelligent drilling condition identification method based on multimodal self-supervised alignment and parallel temporal networks, which can achieve multimodal alignment and fusion under minimal labeling conditions, take into account both local and global temporal features, and provide boundary stabilization and interpretable evidence for drilling condition identification.

[0006] A drilling condition intelligent identification method based on multimodal self-supervised alignment and parallel temporal network disclosed in an embodiment of this application includes: acquiring multi-source time-series data during drilling operations; segmenting the multi-source time-series data based on a sliding window to obtain multiple window samples, each window sample containing a sequence of engineering parameters, a sequence of vibration features, and a sequence of states / logs over a period of time; encoding the sequence of engineering parameters, the sequence of vibration features, and the sequence of states / logs in the window samples using a multimodal encoder to generate engineering features, vibration features, and state features; inputting the engineering features into a parallel temporal network, the parallel temporal network including a local branch of a convolutional neural network for capturing local mutation features and a global branch of an Informer for capturing global long-range dependency features; adaptively fusing the outputs of the local branch of the convolutional neural network and the outputs of the global branch of the Informer through a gating network step-by-step to obtain fused temporal features; mapping the vibration features and the state features to the same feature space as the fused temporal features, and then combining them with the fused temporal features to obtain multimodal fused features; and classifying drilling conditions based on the multimodal fused features.

[0007] Its effects are as follows: reducing label dependence, specifically through mask reconstruction and cross-modal comparison alignment, robust representations can still be obtained under conditions of few labels; improving the recognition accuracy of complex working conditions, through the complementary use of parallel CNNs for local mutations and Informer for long dependencies; enhancing the ability to distinguish hidden working conditions, by introducing auxiliary modalities such as vibration spectrum and logs / states; reducing boundary jitter and improving statistical confidence, achieved through confidence gating combined with shortest duration constraints and optional sequence smoothing; and improving engineering usability through interpretable output, by outputting the contribution of key time periods / key variables, which facilitates verification and closed-loop optimization.

[0008] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0009] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a simplified schematic diagram of a drilling condition intelligent identification method based on multimodal self-supervised alignment and parallel temporal network according to some embodiments of this application; Figure 2 This is a simplified schematic diagram of cross-modal self-supervised alignment pre-training according to some embodiments of this application; Figure 3 This is a simplified schematic diagram of a parallel timing network according to some embodiments of this application; Figure 4 This is a flowchart illustrating the pooling attention aggregation and interpretation process according to some embodiments of this application; Figure 5 This is a simplified flowchart illustrating the process of restoring the operating condition sequence and stabilizing the boundary according to some embodiments of this application. Detailed Implementation

[0010] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0011] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0012] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0013] A drilling condition intelligent identification method based on multimodal self-supervised alignment and parallel temporal network according to an embodiment of this application includes: acquiring multi-source time-series data during drilling operations; segmenting the multi-source time-series data based on a sliding window to obtain multiple window samples, each window sample containing a sequence of engineering parameters, a sequence of vibration features, and a sequence of states / logs over a period of time; encoding the sequence of engineering parameters, the sequence of vibration features, and the sequence of states / logs in the window samples using a multimodal encoder to generate engineering features, vibration features, and state features; inputting the engineering features into a parallel temporal network, the parallel temporal network including a convolutional neural network branch for capturing local mutation features and an Informer branch for capturing global long-range dependency features; adaptively fusing the outputs of the convolutional neural network branch and the Informer branch step-by-step through a gating network to obtain fused temporal features; mapping the vibration features and the state features to the same feature space as the fused temporal features, and then combining them with the fused temporal features to obtain multimodal fused features; and classifying drilling conditions based on the multimodal fused features.

[0014] Specifically, the steps include the following: Step S1: Perform data input and preprocessing to obtain multi-source time-series data during the drilling operation. In this embodiment, a multi-source time-series dataset of the drilling operation is obtained. It should include at least three types of data: engineering parameters, vibration signals, and status / logs; use a unified timestamp format and sort by time to remove duplicates; and make preliminary marks for obvious anomalies such as out-of-bounds, spikes, saturation, and long-term invariance, which will be used for gating and degradation later.

[0015] Step S2: Unify the time base and align multiple sampling rates. Determine the main time axis and the unified sampling step size. Resample the engineering parameters according to the unified sampling step size. Extract frequency domain or time-frequency domain features from the vibration signal by windowing and align it to the main time axis. Aggregate the status / log data by time period or keep the latest status and align it to the main time axis, and generate missing masks and anomaly markers. Thus, engineering parameters are input as time-domain sequences, vibration is input as spectrum or time-frequency features, and status / log is input using discrete encoding or embedded encoding, forming a unified input structure.

[0016] Specifically, determine the main timeline. With uniform step size Engineering parameters are as follows Resampling; vibration signals are truncated by window and frequency domain / time-frequency domain features are extracted and then aligned; logs / status are aggregated by time period or the latest status is maintained; missing masks and anomaly markers are generated.

[0017] Step S3: Then, the aligned multi-source time-series data is segmented based on a sliding window to obtain multiple window samples. Set the window length. With step size Divide the aligned data into windows Each window simultaneously captures slices to obtain samples. ={ , , };in This is a window segment containing engineering parameters. For vibration feature window segments or window-level features, This is a log / status window fragment or window-level feature; it also records the window center time. To facilitate output reconstruction, gating is performed based on the missing / abnormal ratio, and the window center time is recorded for subsequent reconstruction.

[0018] Step S4: Multimodal Feature Construction and Encoding Preparation. Engineering parameters are organized into a time-series matrix; vibration signals are converted into frequency band energy vectors or time-frequency spectra; logs / states are discretely encoded or embedded; forming a unified input structure. In this way, engineering parameters are resampled along a unified time axis; vibrations are truncated within a window and their frequency domain / time-frequency domain features are extracted and aligned; logs / states are aggregated by time period or kept up-to-date, and missing masks and anomaly markers are generated. This ensures that engineering parameters are input as a time-domain sequence; vibrations are input as spectral / time-frequency features; and logs / states are input using discrete encoding or embedded encoding.

[0019] Step S5: Cross-modal self-supervised alignment pre-training. Under conditions of no / few annotations, the encoder is jointly trained using mask reconstruction and cross-modal contrastive alignment to ensure consistency between different modalities within the same window in the same representation space.

[0020] In one embodiment, step S5 includes: S50: Input window sample. Input the constructed window sample. , , These correspond to engineering parameters, vibration characteristics, and logs / status, respectively.

[0021] S51: Three-modal coding and projection. For three types of modes—engineering parameters, vibration characteristics, and log / state—encoders are set up to obtain window representations, which are then mapped to the same dimensional space via a projection layer to obtain a shared representation vector. This shared representation space is used to carry semantic consistency after cross-modal alignment.

[0022] The multimodal encoder described in S52 undergoes cross-modal self-supervised alignment pre-training, the pre-training including: A random mask is applied to the sequence of engineering parameters, and the masked part is reconstructed by the decoder in the multimodal encoder, and the reconstruction loss is calculated.

[0023] Specifically, perform a mask reconstruction branch and work on the project parameter window. A random mask is applied, which in some examples can be random time steps or continuous segments. The masked portion is then reconstructed by the encoder and decoder, and the reconstruction loss is calculated. (Using metrics such as mean squared error (MSE)) enables the model to learn the temporal structure and the coupling relationship between variables, thereby improving its robustness to missing data, anomalies, and noise.

[0024] In terms of masking strategy, considering the physical coupling characteristics of drilling engineering parameters, highly correlated parameters (such as drill pressure and torque, pump pressure and displacement) are correlated and masked. This means simultaneously masking a set of correlated variables within the same time period, forcing the model to reconstruct the masked portion using other unmasked engineering parameters and cross-modal information such as vibration characteristics and logs / status from the same window. This reveals the collaborative variation patterns and physical constraints among multiple parameters during the drilling process. The length of continuous mask segments is set based on the duration of typical drilling anomalies, making the reconstruction task closer to actual data loss patterns.

[0025] In this embodiment, a random mask is applied to the sequence of engineering parameters. Specifically, based on a predefined group of physical correlation variables, correlation masks are applied to at least two engineering parameters with physical coupling relationships within the same group during the same time period.

[0026] For example, let's set up an engineering parameter window. Where N is the number of time steps and C is the number of parameter variables. A predefined set of related variable groups. The physical correlation variable group includes at least a first correlation group consisting of drill pressure and torque and / or a second correlation group consisting of pump pressure and displacement. In this example, g1 = {drill pressure, torque}, g2 = {pump pressure, displacement}, etc., and the parameters within each group have a physical coupling relationship. For each training sample, according to probability... Randomly select a group Perform masking; masking time interval length Sampling was performed based on the duration distribution of drilling events. ,in , … All of these are events.

[0027] Set of masked locations , Starting time, reconstruction loss Only the masked position is calculated.

[0028] The S53 cross-modal alignment branch uses different modal representations within the same window as positive sample pairs and representations from different windows as negative sample pairs to calculate the cross-modal alignment loss. During training, positive samples are brought closer together and negative samples are made more distinguishable, thereby aligning multimodal information within a shared representation space. Positive sample representations are brought closer together, while negative sample representations are moved further apart.

[0029] S54 is used for joint training. In this embodiment, a joint optimization objective of mask reconstruction and cross-modal alignment is adopted: ; in, For weight parameters, The loss for mask reconstruction (e.g., mean squared error) is calculated only for the masked location, as mentioned above. This represents the cross-modal alignment loss; the cross-modal alignment term is expressed in the form of bringing in same-window components and distancing different-window components as follows: ; in, These are the representation vectors of different modalities within the same window. For different windows, This is the interval threshold.

[0030] Through the joint training described above, the final output is an aligned encoder and a shared representation space, which are used for subsequent working condition recognition.

[0031] Step S6: CIPA-Informer parallel timing modeling.

[0032] First, the three modalities of the same window , , The input is taken together; the convolutional neural network extracts mutation / transient features from local branches. Specifically, it performs local modeling on the time-series input of engineering parameters, extracts fine-grained features near mutations, short-period fluctuations, and switching boundaries, and outputs local sequence features aligned with the time step. The Informer global branch extracts long dependencies and intervariate relationships, performs long-sequence modeling on the same engineering parameter time-series input, captures the trend of work segments and intervariate coupling relationships, and outputs global sequence features aligned with time steps. , The global sequence features output by the Informer global branch, with dimensions and same.

[0033] In this CNN local branch, a multi-scale one-dimensional convolutional kernel parallel structure is adopted. Convolutional kernels of different scales correspond to transient events of different durations during drilling (such as millisecond to second-level torque spikes caused by drill bit cutting tooth impact, and parameter fluctuations of several seconds to tens of seconds caused by single-bit connection). Multi-granularity local mutation features are extracted in parallel, and multi-scale features are spliced ​​and fused step by step. In the Informer global branch, in the probabilistic sparse self-attention calculation, in view of the sparsity of parameter changes within the long drilling sequence, attention weights are preferentially assigned to time steps with drastic feature changes (such as parameter inflection points and trend reversals), while attention sparsification is automatically performed on a large number of redundant time steps in the stable working condition section, thereby efficiently capturing the global parameter drift trend caused by encountering different formations or the gradual wear of the drill bit.

[0034] For example, a local branch of a CNN uses three sets of parallel one-dimensional convolutional kernels, with a kernel size of... , , This captures torque spikes in the millisecond to second range, parameter fluctuations in the several-second range, and single-root connection processes in the tens of seconds range, respectively. For the input... The outputs of each convolutional group are concatenated after padding for a holding time step N. Select the one with the largest M value The queries constitute a sparse query matrix. The output of the remaining queries is directly taken as the mean vector. Since drilling parameters change slowly during the steady drilling phase, the queries are sparsified, and attention is focused only on time steps with drastic changes in features such as parameter inflection points and formation interfaces encountered during drilling, thereby reducing the amount of computation and highlighting key trend turning points.

[0035] Then, gated fusion is performed: to balance local mutations and long dependencies, this invention employs a gated fusion mechanism to adaptively synthesize the features of the two branches step-by-step. First, a gate coefficient matrix is ​​generated through a gated network. in N is the number of time steps, and dF is the feature dimension; then, fusion is performed by time step: ; in, The gating coefficient matrix ( ), This indicates element-wise multiplication.

[0036] In addition to local and global features, the gating network also incorporates first-order differential components of engineering parameters (such as the rate of change of drilling pressure and the rate of change of rotational speed) as gating modulation signals. This enables the gating coefficients to sense the instantaneous drasticness of parameter changes. When the short-term rate of change of parameters such as drilling pressure or torque exceeds the adaptive threshold, the gating coefficients automatically tilt towards the local feature branch at that time step, enhancing the response to the boundary of working condition switching. In the continuous drilling section where parameter changes are gradual, the gating coefficients tilt towards the global feature branch, preserving the continuity of long-term trends.

[0037] Finally, multimodal enhancement injection is performed to distinguish between vibration and log / state enhancement hidden operating conditions; vibration features and log / state features are mapped to a feature space consistent with the fused feature Fk and then injected into Fk. For example, vibration features are... Vectors are obtained through mapping Log / status characteristics Vectors are obtained through mapping This allows them to be placed in the same feature space as the fused feature Fk, thereby enhancing the ability to distinguish between hidden working conditions and similar working conditions, while keeping the time dimension unchanged, which facilitates subsequent interpretation output and boundary stabilization. Finally, the fused feature Fk is output and sent to the subsequent aggregation module.

[0038] Step S7: Pooling Attention Aggregation and Interpretation of Output (CIPA-Informer-Pooling).

[0039] Input the fusion feature Fk obtained in step S6.

[0040] Attention pooling is performed on the fused feature Fk. First, the attention weights at each time step are calculated. It outputs a working condition discrimination vector and generates explanatory information, such as key time periods and contributions of key variables. Among these, it integrates features. Attention-weighted aggregation is performed to obtain a window-level discriminant vector. : ; Where N is the number of time steps within the window. The time step weights are used to measure the contribution of each time step to the condition classification. Classification is performed based on zk, outputting the condition category yk and confidence level pk.

[0041] Therefore, an explanatory output is performed. While the present invention outputs the output working condition category and confidence level, it also outputs explanatory evidence associated with the attention weight, including at least: key time period evidence, a set of time segments with higher weights for playback review; key variable evidence, a set of variables with higher discriminative contributions for engineering explanation and parameter adjustment; evidence window index, bound to the original data interval for audit retention and review. Finally, yk, pk, and the explanatory evidence are bound to the central time ck of the window and output.

[0042] Step S8: Restoration of the working condition sequence.

[0043] Input (ck, yk, pk, explanatory evidence) of each window. After mapping the working condition category and confidence level of each window sample back to the continuous time axis according to the central time of the window, that is, according to the central time of the window Map the window prediction back to the continuous time axis; the overlapping regions are merged by voting or confidence level weighting to obtain the initial output of the continuous working condition sequence. For example, according to the window length And the step size Construct a set of windows, synchronously organize three types of modal inputs within the window and record the central time . Thus, map the window output back to the continuous time axis, and the overlapping regions are merged by confidence level weighting or voting to obtain the initial working condition sequence, ensuring that the working condition output can form an expression on the continuous time axis. Make the window output map back to the continuous time axis according to the central time, and the overlapping regions form the initial working condition sequence through the merging strategy, providing input for stabilization.

[0044] Step S9: Boundary stabilization output. When the window confidence level is lower than the preset confidence level threshold, mark the working condition category of this window as uncertain or maintain the previous working condition state; when the duration of a continuous working condition segment is lower than the preset shortest duration threshold, merge this segment into the adjacent stable working condition; perform smoothing filtering on the working condition sequence to eliminate isolated jumps. Specifically, it includes: Perform gating and constraint on the continuous working condition sequence to reduce short-term jumps, and output the stable working condition sequence and the low-confidence interval. For example, confidence level gating: Set the confidence level threshold p0. If the confidence level pk of a certain window < p0, then mark the output at this place as "uncertain" or maintain the previous state to avoid short-term jumps caused by noise. And, the shortest duration constraint: For the continuous working condition segment interval Define the duration: ; when the segment duration is less than the shortest threshold , merge this short segment into the adjacent stable working condition or perform a rollback process.

[0045] Then, consistent smoothing is performed: the operating condition sequence is smoothed and filtered to further eliminate isolated jumps. Finally, the stable operating condition sequence Yt is output, along with the low-confidence intervals, and the differences before and after stabilization are recorded for auditing purposes.

[0046] It is understood that the method according to the embodiments of this application unifies multi-source, multi-sampling-rate data to a common time reference and constructs trainable samples; cross-modal feature space alignment is achieved through self-supervised learning, reducing reliance on manual annotation; parallel networks are used to fuse local and global temporal features, improving the accuracy and robustness of complex working condition recognition; attention pooling is introduced to output the contribution of key time periods / key variables, providing interpretable evidence; boundary stabilization of the recognition results is performed to reduce jitter and improve online availability and statistical accuracy; and a stable working condition time axis is output, providing a reliable foundation for subsequent efficiency analysis and automatic KPI statistics.

[0047] Therefore, the following measures are taken: reducing the dependence on annotations, specifically through mask reconstruction and cross-modal comparison alignment, to obtain robust representations even with fewer annotations; improving the accuracy of complex working condition recognition by combining local mutations in parallel CNNs with long dependencies in Informer; enhancing the ability to distinguish hidden working conditions by introducing auxiliary modalities such as vibration spectrum and logs / states; reducing boundary jitter and improving statistical confidence by combining confidence gating with shortest duration constraints and optional sequence smoothing; and improving engineering usability by outputting the contribution of key time periods / key variables, which facilitates verification and closed-loop optimization.

[0048] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. Those skilled in the art will understand that all or part of the steps carried out by the methods of the above embodiments can be implemented by a program instructing related hardware, the program being stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiments.

[0049] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0050] The storage medium in this application embodiment may be a read-only memory, a disk, or an optical disk, etc.

[0051] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent identification of drilling conditions based on multimodal self-supervised alignment and parallel temporal networks, characterized in that, include: Acquire multi-source time-series data during drilling operations; The multi-source time-series data is segmented based on a sliding window to obtain multiple window samples. Each window sample contains a sequence of engineering parameters, a sequence of vibration characteristics, and a sequence of status / logs for a certain period of time. The engineering parameter sequence, vibration feature sequence, and state / log sequence in the window sample are encoded using a multimodal encoder to generate engineering features, vibration features, and state features. The engineering features are then input into a parallel temporal network, which includes a local branch of a convolutional neural network for capturing local mutation features and a global branch of an Informer network for capturing global long-range dependency features. By using a gating network to adaptively fuse the outputs of the local branches of the convolutional neural network and the outputs of the global branches of the Informer network step by step, fused temporal features are obtained. After mapping the vibration features and the state features to the same feature space as the fused temporal features, they are combined with the fused temporal features to obtain multimodal fusion features; Drilling conditions are classified based on the multimodal fusion features.

2. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 1, characterized in that, The multimodal encoder undergoes cross-modal self-supervised alignment pre-training, which includes: A random mask is applied to the sequence of engineering parameters, and the masked part is reconstructed by the decoder in the multimodal encoder, and the reconstruction loss is calculated. Using different modal representations of samples within the same window as positive sample pairs and representations of samples from different windows as negative sample pairs, cross-modal alignment loss is calculated to make positive sample representations closer to each other and negative sample representations farther apart. Wherein, the cross-modal alignment loss Calculate using the following formula: In the formula, These are the representation vectors of different modalities within the same window. For different windows, This is the interval threshold.

3. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 2, characterized in that, The application of a random mask to the engineering parameter sequence specifically involves: Based on predefined groups of physically related variables, correlation masks are applied to at least two engineering parameters with physical coupling relationships within the same group during the same time period. These physically related variables include multiple correlation groups and constitute a set of correlation variable groups. ; The mask time interval length of the correlation mask Determined by sampling from a pre-defined drilling event duration distribution: in, , … All are events; The set of masked positions is , Starting time, reconstruction loss Only the masked position is calculated.

4. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 1, characterized in that, The local branches of the convolutional neural network employ multiple parallel one-dimensional convolutional kernels, each with a different scale, to extract local mutation features of different durations in parallel; wherein, at least a first convolutional kernel with a kernel size of 3, a second convolutional kernel with a kernel size of 7, and a third convolutional kernel with a kernel size of 15 are included.

5. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 4, characterized in that, Local sequence features of convolutional neural network branch outputs aligned with time steps ; The Informer global branch outputs global sequence features aligned with the time step. , The global sequence features output by the Informer global branch, with dimensions and same.

6. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 5, characterized in that, The gating network generates the gating coefficient matrix in the following manner. : in N is the number of time steps, and dF is the feature dimension; then, fusion is performed by time step: in, The gating coefficient matrix ( ), This indicates element-wise multiplication.

7. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 1, characterized in that, The method according to claim 1, characterized in that, the step of classifying drilling conditions based on multimodal fusion features includes: Fusion features Attention-weighted aggregation is performed to obtain a window-level discriminant vector. : Where N is the number of time steps within the window. The time step weights are used to measure the contribution of each time step to the condition classification; classification is performed based on zk, and the condition category yk and confidence level pk are output.

8. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 7, characterized in that, After mapping the operating condition category and confidence level of each window sample back to the continuous time axis according to the window center time, at least one of the following boundary stabilization processes is performed: When the confidence level of a window is lower than the preset confidence level threshold, the working condition category of the window is marked as uncertain or the previous working condition status is maintained. When the duration of a continuous operating condition segment is lower than the preset minimum duration threshold, the segment is merged into an adjacent stable operating condition. Smoothing filters are applied to the operating condition sequence to eliminate isolated jumps.

9. The intelligent identification method for drilling conditions based on multimodal self-supervised alignment and parallel temporal networks according to claim 1, characterized in that, Before the multi-source time-series data is segmented based on a sliding window, the following is also included: Determine the main time axis and a uniform sampling step size, and resample the engineering parameters according to the uniform sampling step size; The vibration signal is truncated by a window and its frequency domain or time-frequency domain features are extracted and then aligned to the main time axis; The status / log data is aggregated by time period or kept up-to-date and aligned to the main time axis, and missing masks and anomaly markers are generated.

Images