CCL pulse identification and positioning method based on attention mechanism and adaptive wavelet
By employing an attention-based and adaptive wavelet-based CCL pulse recognition method, the problem of poor CCL signal quality in complex downhole environments was solved, achieving high recall for coupling positioning and improving the accuracy and safety of perforation depth positioning.
Patent Information
- Application Number
- CN202511436735.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-09
- Publication Date
- 2026-01-13
AI Technical Summary
Existing technologies suffer from poor CCL signal quality under high temperature and high pressure conditions downhole, and are affected by cable noise and non-standard operation interference, resulting in a high rate of missed detection of coupling signals, which affects the accuracy and safety of perforation depth positioning.
A CCL pulse recognition method based on attention mechanism and adaptive wavelet is adopted. By learning-enhancing wavelet transform, channel attention mechanism, time-varying gating and cross-subband multi-head attention mechanism, multi-scale features of CCL signal are extracted. Combined with temporal backbone network for classifier processing, high recall of clamp localization is achieved.
It effectively reduces the missed detection rate and improves the F1 score under complex well conditions, ensuring the accuracy and safety of perforation depth positioning and robustness to adapt to complex downhole environments.
Smart Images
Figure CN121327596A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of oilfield geological exploration technology, specifically a CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet. Background Technology
[0002] The casing collar locator (CCL) is a key piece of equipment in oil well logging and perforation operations. Perforation refers to the process of creating a hole in the formation at a designated location using a specially designed shaped charge device, allowing fluids to flow smoothly into the wellbore. Perforation technology is widely used in the extraction of oil and gas fields and coal seams. During operation, the perforating gun is lowered to the target depth in the well, and the perforating charge penetrates the casing and cement sheath, forming a channel between the formation and the wellbore, providing a pathway for oil and gas resource extraction. In oil well logging operations, the casing collar locator (CCL) is needed to help determine the depth of the downhole tool. The CCL consists of a permanent magnet and an induction coil. When the tool passes through the casing collar, the sudden change in metal thickness causes a change in magnetic flux, inducing a voltage pulse in the coil. These pulse signals typically exhibit one primary peak and two secondary peaks. By detecting the appearance of these pulses, the number of collars can be counted, thereby estimating the tool's lowering depth. In logging engineering, collar signals are often used to locate the perforation depth, achieving precise perforation.
[0003] Currently, there are two main methods for identifying coupling signals (CCL pulse signals): manual identification and machine calculation of CCL signal thresholds. Both methods use CCL signals to determine whether a signal is a coupling and are used for perforation depth positioning. These characteristic CCL coupling pulse signals typically exhibit one primary peak and two secondary peaks. By detecting the appearance of pulses, the number of couplings can be counted, thereby estimating the tool's descent depth. However, in actual operating conditions, the quality of CCL signals is very poor: the high temperature and pressure environment downhole and long-term wear cause the coupling signal amplitude to attenuate, and cable noise and non-standard operations introduce a large amount of environmental interference. CCL signals are non-stationary and have a low signal-to-noise ratio. Changes in tool operating speed and tilt angle can cause pulse waveform distortion. For example, if the CCL sensor tilt is too large, during high-speed descent, the primary peak is almost submerged by the secondary peaks. Traditional methods relying on fixed thresholds or standard templates often miss detections in these situations. Since missing a single coupling can lead to cumulative errors in depth calculation, significantly impacting operational safety and cost, a more robust and highly recall automated detection method is needed. However, the downhole environment contains many interfering factors that can affect the quality of CCL signals and interfere with coupling signals.
[0004] In summary, there is an urgent need for a CCL pulse identification and localization method that can still achieve high recall, low false negatives, and engineering feasibility under conditions of strong noise and non-stationarity. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a CCL pulse recognition and localization method based on an attention mechanism and adaptive wavelet, which includes:
[0006] S1: Acquire the CCL voltage signal and preprocess it to obtain the preprocessed CCL voltage signal;
[0007] S2: Perform a learnable lifting wavelet transform on the preprocessed CCL voltage signal to obtain a multi-scale subband tensor;
[0008] S3: The channel attention mechanism is used to process the multi-scale subband tensor to obtain the channel scaling feature;
[0009] S4: Perform time-varying gating processing on the channel scaling features to obtain the gated output;
[0010] S5: A cross-subband multi-head attention mechanism is used to process the gated output to obtain cross-subband features;
[0011] S6: A temporal backbone network is used to process the cross-subband features to obtain time-point features;
[0012] S7: Input the time-point features into the classifier for processing to obtain the CCL pulse localization result.
[0013] Preferably, the process of performing a learnable lifting wavelet transform on the preprocessed CCL voltage signal includes:
[0014] A multi-layer learnable boosting wavelet network is used to process the CCL voltage signal, with the output of the previous layer serving as the input to the next layer. The detailed features of each preceding layer are concatenated with the output signal of the last layer to obtain a multi-scale subband tensor. The processing steps of a single-layer learnable boosting wavelet network include:
[0015] The input signal is split into odd and even sequences;
[0016] Detailed features are obtained by predicting even sequences based on odd sequences.
[0017] The detailed features and even sequences are updated to obtain the output signal of the learnable boosting wavelet network of this layer.
[0018] Furthermore, the prediction process based on odd sequences for even sequences can be expressed as follows:
[0019]
[0020] in, Indicates the first Layers can learn to enhance the detailed features of wavelet networks. Represents the first energy normalization factor. Indicates the first Layers can learn to improve the output of the wavelet network after splitting into odd sequences. Indicates the first Layers can learn to improve the even sequences output by the wavelet network after splitting. Indicates the first Layers can learn to improve the prediction steps of wavelet networks using one-dimensional deep separable convolutions.
[0021] Furthermore, the update processing of detailed features and even sequences is represented as follows:
[0022]
[0023] in, Indicates the first Layers can learn to improve the output signal of a wavelet network. This represents the second energy normalization factor. Indicates the first Layers can learn to improve the even sequences output by the wavelet network after splitting. Indicates the first Layers can learn to enhance the detailed features of wavelet networks. Indicates the first Layers can learn to improve the update step of wavelet networks and one-dimensional depthwise separable convolutions.
[0024] Furthermore, the first Layers can learn to improve the dilation rate of one-dimensional depthwise separable convolutions in wavelet networks:
[0025]
[0026] in, Indicates the first Layers can be learned to improve the hole rate of one-dimensional depthwise separable convolutions in wavelet networks.
[0027] Preferably, the process of processing multi-scale subband tensors using a channel attention mechanism includes: batch normalization of the multi-scale subband tensors, learningable band weight and channel attention scaling to obtain channel scaling features.
[0028] Preferably, the process of performing time-varying gating on the channel scaling features includes:
[0029] The channel scaling features are processed by group depthwise convolution, batch normalization, and activation function to generate a signal. ;
[0030] For signal Perform residual gate injection to obtain gated output. The residual gate is represented as:
[0031]
[0032] in, Indicates channel scaling characteristics. Indicates the gating strength. This indicates a gating signal.
[0033] Preferably, the process of processing the gated output using a cross-subband multi-head attention mechanism includes: dividing the gated output into blocks along the time dimension, using time windows as units, to obtain multiple window blocks; converging the window blocks along the time dimension to obtain subband descriptions; calculating attention features based on the subband descriptions; restoring the attention features to subband weights through mapping and writing them back to the corresponding time positions along the time points within the window to generate initial cross-subband features; performing residual connection and normalization processing on the cross-subband features to obtain the final cross-subband features for that window; and concatenating the final cross-subband features of all windows to obtain complete cross-subband features.
[0034] Preferably, the temporal backbone network includes: a 1×1 convolutional layer, a convolutional positional coding layer, an L-layer Transformer encoder, and a convolutional refinement layer.
[0035] The beneficial effects of this invention are as follows: By incorporating learnable lifting wavelet transform as a front-end feature extractor into the model, this invention adaptively extracts the time-frequency features most relevant to the CCL well logging signal task and extracts the collar pulse features in the CCL signal at multiple scales, overcoming the inherent defects of rigid fixed wavelet transform basis functions and lack of task orientation in traditional methods. This invention combines learnable lifting wavelet transform with an attention mechanism. The multi-scale, physically meaningful (preserving perfect reconstruction characteristics) signal decomposition results of the learnable lifting wavelet transform provide sparse, structured, and high signal-to-noise ratio input for the subsequent attention mechanism, greatly alleviating the computational burden and noise interference problems when the attention mechanism directly processes the original complex signals. As a back-end processor, the attention mechanism can accurately focus on the feature channels (channel attention) and time segments (self-attention) most critical to the task decomposed by the lifting scheme and model long-range dependencies, thereby amplifying the effectiveness of learnable wavelet features. The two work together to form a powerful feature extraction and inference system with both local accuracy and global awareness. This invention combines wavelet multi-scale interpretability with end-to-end trainable adaptive capability, and can effectively reduce the false negative rate and improve the F1 score under complex well conditions and strong noise conditions, and has good application prospects. Attached Figure Description
[0036] Figure 1This is a schematic diagram of the CCL pulse recognition and positioning model structure in this invention;
[0037] Figure 2 This is a schematic diagram of the single-layer learnable lifting wavelet network structure in this invention. Detailed Implementation
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] This invention proposes a CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet, the method comprising the following:
[0040] This invention designs a CCL pulse recognition and positioning model, which acquires CCL voltage signals, feeds them into the model, and outputs pulse positioning results, such as... Figure 1 As shown, the model processes the CCL voltage signal as follows:
[0041] S1: Acquire the CCL voltage signal and preprocess it to obtain the preprocessed CCL voltage signal.
[0042] The CCL voltage signal is acquired and preprocessed. Specifically, the CCL voltage signal is subjected to amplitude mapping, optional window-by-window DC removal and window-by-window normalization (z-score) processing to obtain the preprocessed CCL voltage signal.
[0043] S2: Perform a learnable lifting wavelet transform on the preprocessed CCL voltage signal to obtain a multi-scale subband tensor.
[0044] The process of performing a learnable liftable wavelet transform (LLWT) on the preprocessed CCL voltage signal includes:
[0045] A multi-layered learnable boosting wavelet network is used to process the CCL voltage signal. The output of the previous learnable boosting wavelet network is used as the input of the next learnable boosting wavelet network. The detailed features of each previous learnable boosting wavelet network and the output signal of the last learnable boosting wavelet network are concatenated by channel to obtain a multi-scale subband tensor.
[0046] LLWT can be viewed as making prediction / update a learnable convolution within a lifting framework, implemented in the style of SWT, thus allowing the wavelet front-end to adapt to the task. To address the problems of fixed wavelets being unlearnable and difficult to adapt to non-stationary conditions, while avoiding the shift sensitivity caused by downsampling, this invention proposes a lifting-based learnable wavelet with a stationary implementation, thereby improving the learnability of the "prediction-update" process in the decomposition; a single-layer learnable lifting wavelet network is shown below. Figure 2 As shown, the processing steps of a single-layer learnable boosting wavelet network include:
[0047] The input signal is split into odd and even sequences; specifically:
[0048] The input signal is split into two "odd / even phase zero-filled sequences" (with the same length as the original sequence and complementary): .
[0049] Phase mask and Defined as:
[0050]
[0051] The signal is the same as the original signal at even indices and 0 at odd indices. Conversely, the lengths of the two signals are... Same, and .
[0052] The equivalent discrete "zero-filling" representation of the two split signals is:
[0053]
[0054]
[0055] The odd-even sequence is used for prediction processing to obtain detailed features; specifically, the prediction processing uses learnable prediction convolutional kernels. Processing the even sequence yields the predicted odd sequence, and subtracting it from the odd sequence obtained after splitting gives the details, represented as:
[0056]
[0057] in, Indicates the first Layers can learn to enhance the detailed features of wavelet networks. Represents the first energy normalization factor. No. Layers can learn to improve the output of the wavelet network after splitting into odd sequences. Indicates the first Layers can learn to improve the even sequences output by the wavelet network after splitting. Indicates the first Layers can learn to improve the prediction steps of wavelet networks using one-dimensional deep separable convolutions.
[0058] The detailed features and even sequences are updated to obtain the output signal of the learnable boosting wavelet network at this layer. Specifically: the update process applies the detailed features to the learnable update kernel. By re-feeding back to the even phase, we obtain the approximation for this layer, and the update process is expressed as follows:
[0059]
[0060] in, Indicates the first Layers can learn to improve the output signal of a wavelet network. This represents the second energy normalization factor. Indicates the first Layers can learn to improve the update step of wavelet networks and one-dimensional depthwise separable convolutions.
[0061] To avoid downsampling and maintain the same length alignment between each subband and the input, the first... Layers can learn to improve the dilation rate of one-dimensional depthwise separable convolutions in wavelet networks:
[0062]
[0063] In some embodiments of the present invention, reflection filling is used to avoid the effect of discontinuity in the signal at the boundary, for example, for a one-dimensional signal. When performing convolution, for These boundary locations use reflected boundary values:
[0064]
[0065] Reflection filling avoids anomalous signal changes at the boundaries by ensuring the boundary symmetry of the convolution calculation.
[0066] Output equal length layer by layer With the final And spliced together according to the channel ,form Multiscale subband tensor.
[0067] The model of the present invention can use a dynamic weighted loss including BCE and Dice during training, and perform learning rate scheduling and early stopping with F1 as the metric; in some preferred embodiments, zero DC (DC) constraints and vanishing moment constraints are added to the loss function during model training.
[0068] Implementation of zero DC constraint:
[0069] Implementation: During the transformation process, a loss function is designed to ensure that the average value of the transformed signal is close to zero. This can be achieved by constraining the average value of the output signal.
[0070]
[0071] in It is the first The layer approximates the signal. This constraint forces the elimination of the DC component in the output of each layer, making the signal transformation more focused on the changing parts.
[0072] Add this constraint as a regularization term to the total loss function:
[0073]
[0074] By minimizing this value during the optimization process, the DC component of the signal is ensured to be removed.
[0075] Implementation of vanishing moment constraints:
[0076] Vanishing moment constraints are typically added as a soft constraint to the loss function. For example, for a first-order moment constraint, the following can be used:
[0077]
[0078] By minimizing this value, we can ensure that the wavelet transform can effectively remove the low-frequency components and focus on the changing parts of the signal.
[0079] This invention incorporates a learnable wavelet transform into the model as a front-end feature extractor, enabling the learnable enhancement scheme to adaptively extract the time-frequency features most relevant to the CCL well logging signal task and extract the coupling pulse features in the CCL signal at multiple scales. This overcomes the inherent defects of traditional fixed wavelet transform basis functions being rigid and lacking task orientation.
[0080] S3: The channel attention mechanism is used to process the multi-scale subband tensor to obtain the channel scaling feature.
[0081] The multi-scale subband tensor is processed using a channel attention mechanism, specifically:
[0082] Batch normalization, learnable band weights, and channel attention scaling are applied to the multi-scale subband tensor to obtain channel-scaled features. Specifically, batch normalization of the multi-scale subband tensor makes each dimension of the input data more stable during training, which helps accelerate convergence and prevent gradient explosion / vanishing. The input multi-scale subband tensor is represented as follows: ,in: It refers to the batch size. It is a sub-band number. This refers to the time step. Batch normalization normalizes the C-dimensional (subband dimension):
[0083]
[0084] in, and These are the mean and variance of each channel (sub-band). Used to prevent division by zero errors.
[0085] Next, learnable band weights are applied, meaning each sub-band (frequency band) is trained to obtain an importance weight. The specific process is as follows:
[0086] These weights are updated via backpropagation, and the optimization objective is typically to minimize the task-related loss. These weights are defined as a trainable vector. (size is) ), and on the input tensor Weighting:
[0087]
[0088] In this way, the representation of each subband will be based on its learned weights. Scaling is applied to highlight important frequency band information.
[0089] Scaling further weights each sub-band using Channel Attention (ECA). ECA adjusts the weights of each channel (sub-band) through local context. ECA employs a local convolution-based approach, weighting each channel as follows:
[0090]
[0091] here It is the convolution kernel corresponding to each sub-band. This represents the convolution operation. It is an activation function. In this way, the weights of each subband are dynamically adjusted based on its context information.
[0092] Finally, after batch normalization, learnable band weighting, and channel attention, the adjusted subband representation is obtained. The final output subband tensor is the channel scaling feature. This will be used for subsequent tasks:
[0093]
[0094] here, These are the attention features obtained for each channel (sub-band) after ECA.
[0095] S4: Perform time-varying gating on the channel scaling features to obtain the gated output.
[0096] A time-varying gating network is used to generate gate functions for each time point. Specifically:
[0097] The channel scaling features are processed by group depthwise convolution, batch normalization, and the sigmoid function to generate a signal. ;
[0098] For signal Perform residual gate injection to obtain gated output. The residual gate is represented as:
[0099]
[0100] in, This indicates the input signal from the previous level (channel scaling feature). Indicates the gating strength. This indicates a gating signal.
[0101] This invention utilizes time-varying gating to preserve short-term transients and suppress smoothing damage, creating a supermodel that combines local multi-scale analysis capabilities with global context modeling capabilities. This significantly reduces the missed detection of short pulses and high-noise pulses.
[0102] S5: The gating output is processed using a cross-subband multi-head attention mechanism to obtain cross-subband features.
[0103] Gating output in units of time windows (in This refers to the number of batches processed simultaneously. For sub-numbers, The duration is divided into blocks along the time dimension to obtain multiple window blocks;
[0104]
[0105] window block By converging along the time dimension, we obtain the subband description. Among these, convergence employs learnable projection without altering the number of subbands. ;
[0106] Attention features are calculated based on subband descriptions. :
[0107]
[0108]
[0109]
[0110]
[0111] In the formula, The number of heads in the bullish attention. The total embedding dimension described by the subband. The embedding dimension is for a single attention head. Attention features are... The sub-band weights are restored by mapping. It then broadcasts / writes back to the corresponding time position along the time points within the window to generate initial cross-subband features. .
[0112] By performing residual connection and normalization on the cross-subband features, the final cross-subband features for this window are obtained:
[0113]
[0114] After performing the above processing on each window block, the final cross-subband feature of each window is obtained; the final cross-subband features of all windows are then concatenated according to the stride to obtain the complete cross-subband feature. .
[0115] S6: Use a temporal backbone network to process cross-subband features to obtain time-point features.
[0116] Cross-subband features obtained after processing by a cross-subband multi-head attention mechanism in a temporal backbone network The input is modeled along the time axis, passing through a 1×1 convolutional layer, a convolutional positional encoding layer, an L-layer Transformer encoder, and a convolutional refinement layer in sequence, outputting point-by-point logits, i.e., time-by-time features; the convolutional refinement layer is a convolutional layer added after the Transformer encoder to optimize the features.
[0117] This invention maintains higher recall and threshold stability under high-speed and noisy conditions by explicitly modeling frequency band coupling across subband attention.
[0118] S7: Input the time-point features into the classifier for processing to obtain the CCL pulse localization results.
[0119] The time-point features are processed by the Sigmoid function to obtain the probability curve; threshold scanning or adaptive thresholding is performed and temperature scaling is combined for calibration to generate a point-by-point mask and event list (start position, duration, confidence level) to obtain the CCL pulse localization result.
[0120] In summary, this invention provides an effective solution to the technical challenges of low signal-to-noise ratio, difficulty in manual interpretation, and low accuracy of automated analysis in the field of oilfield geology. This invention performs amplitude mapping and optional window-by-window DC removal and Z-score processing on the original voltage sequence; based on a stationary multi-scale lifting wavelet front-end, it employs one-dimensional depth-separable convolution to implement the prediction / update operator, and sets the void ratio layer by layer to maintain consistency between each sub-band and the input length, outputting details of each layer and the final approximation; the sub-bands are sequentially input into channel attention, time-varying gating, and cross-sub-band multi-head attention, and then temporal dependencies are extracted through convolutional position encoding and a Transformer encoder, outputting point-by-point pulse probabilities through a convolutional classification head. This invention can more accurately identify coupling positions and determine the status of downhole tools, providing reliable technical support for the efficient development and safe production of oil and gas fields, and has significant practical application value and economic benefits.
[0121] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet, characterized in that, include: S1: Acquire the CCL voltage signal and preprocess it to obtain the preprocessed CCL voltage signal; S2: Perform a learnable lifting wavelet transform on the preprocessed CCL voltage signal to obtain a multi-scale subband tensor; S3: The channel attention mechanism is used to process the multi-scale subband tensor to obtain the channel scaling feature; S4: Perform time-varying gating processing on the channel scaling features to obtain the gated output; S5: A cross-subband multi-head attention mechanism is used to process the gated output to obtain cross-subband features; S6: A temporal backbone network is used to process the cross-subband features to obtain time-point features; S7: Input the time-point features into the classifier for processing to obtain the CCL pulse localization result.
2. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet as described in claim 1, characterized in that, The process of performing a learnable lifting wavelet transform on the preprocessed CCL voltage signal includes: A multi-layer learnable boosting wavelet network is used to process the CCL voltage signal, with the output of the previous layer serving as the input to the next layer. The detailed features of each preceding layer are concatenated with the output signal of the last layer to obtain a multi-scale subband tensor. The processing steps of a single-layer learnable boosting wavelet network include: The input signal is split into odd and even sequences; Detailed features are obtained by predicting even sequences based on odd sequences. The detailed features and even sequences are updated to obtain the output signal of the learnable boosting wavelet network of this layer.
3. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 2, characterized in that, Predicting even sequences based on odd sequences is represented as follows: ； in, Indicates the first Layers can learn to enhance the detailed features of wavelet networks. Represents the first energy normalization factor. Indicates the first Layers can learn to improve the output of the wavelet network after splitting into odd sequences. Indicates the first Layers can learn to improve the even sequences output by the wavelet network after splitting. Indicates the first Layers can learn to improve the prediction steps of wavelet networks using one-dimensional deep separable convolutions.
4. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 2, characterized in that, The update processing of detail features and even sequences is represented as follows: ； in, Indicates the first Layers can learn to improve the output signal of a wavelet network. This represents the second energy normalization factor. Indicates the first Layers can learn to improve the even sequences output by the wavelet network after splitting. Indicates the first Layers can learn to enhance the detailed features of wavelet networks. Indicates the first Layers can learn to improve the update step of wavelet networks and one-dimensional depthwise separable convolutions.
5. A CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 3 or 4, characterized in that, No. Layers can learn to improve the dilation rate of one-dimensional depthwise separable convolutions in wavelet networks: ； in, Indicates the first Layers can be learned to improve the hole rate of one-dimensional depthwise separable convolutions in wavelet networks.
6. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 1, characterized in that, The process of using the channel attention mechanism to process multi-scale subband tensors includes: batch normalization of the multi-scale subband tensors, learningable band weight and channel attention scaling to obtain channel scaling features.
7. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 1, characterized in that, The process of performing time-varying gating on channel scaling features includes: The channel scaling features are processed by group depthwise convolution, batch normalization, and activation function to generate a signal. ; For signals Perform residual gate injection to obtain gated output. The residual gate is represented as: ； in, Indicates channel scaling characteristics. Indicates the gating strength. This indicates a gating signal.
8. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 1, characterized in that, The process of processing the gated output using a cross-subband multi-head attention mechanism includes: dividing the gated output into blocks along the time dimension, using time windows as units, to obtain multiple window blocks; converging the window blocks along the time dimension to obtain subband descriptions; calculating attention features based on the subband descriptions; restoring the attention features to subband weights through mapping and writing them back to the corresponding time positions along the time points within the window to generate initial cross-subband features; performing residual connection and normalization processing on the cross-subband features to obtain the final cross-subband features for that window; and concatenating the final cross-subband features of all windows to obtain the complete cross-subband features.
9. The CCL pulse recognition and localization method based on attention mechanism and adaptive wavelet according to claim 1, characterized in that, The temporal backbone network consists of: 1×1 convolutional layers, convolutional positional coding layers, L-layer Transformer encoders, and convolutional refinement layers.