Deep learning-based method and system for predicting coupling of liquid level fluctuation and air pressure in an air chamber of an owc device

Abstract

The application discloses a deep learning-based OWC device air chamber liquid level fluctuation and air pressure coupling prediction method and system, and belongs to the technical field of ocean renewable energy and intelligent prediction. The application aims at the continuous missing of OWC device air chamber liquid level and air pressure sensor data and the problem that existing interpolation methods destroy the liquid level-air pressure physical coupling relationship, and constructs a historical liquid level-air pressure sub-sequence mode library. Real-time detection of data missing, use of known data and local mean on both sides of the missing gap to construct a query sub-sequence, matching search and multi-level quality inspection based on joint similarity and time lag coupling consistency constraint, reconstruction of a complete sequence according to the matching result by using a scaling interpolation or a differentiated rollback strategy, input of the complete sequence into a deep learning prediction model containing a coupling regularization term, output of future multi-step liquid level and air pressure prediction values, and finally adaptive prediction through rolling update and coupling consistency verification. The application realizes high-fidelity repair and physically self-consistent prediction under data missing.

Description

Technical Field

[0001] This invention relates to the field of marine renewable energy technology and intelligent prediction, specifically to a method and system for predicting liquid level fluctuations and gas pressure coupling in the gas chamber of an OWC device based on deep learning. Background Technology

[0002] Oscillating water column (OWC) wave energy generation devices are one of the most researched technologies in the field of wave energy utilization. The fluctuation of liquid level and changes in gas pressure within the gas chamber constitute the core dynamic process of energy conversion in the device, and the two are strongly nonlinearly coupled.

[0003] Currently, the simulation and prediction of dynamics within OWC chambers mainly rely on the following methods: (1) Computational fluid dynamics (CFD) numerical simulation: high accuracy but computation time-consuming, making it difficult to use for real-time control.

[0004] (2) Simplified lumped parameter model: It has a small computational load but limited accuracy, and it is difficult to accurately reflect the nonlinear characteristics under complex sea conditions.

[0005] (3) Deep learning methods: Deep neural networks such as LSTM and CNN-LSTM have been introduced into the OWC air chamber state prediction task, but they have the following limitations: most of them only target the prediction of single variable air pressure and do not explicitly model the coupling relationship between liquid level and air pressure; they rely on synthetic data generated by CFD and it is difficult to make online predictions based on real-time sensor data; they assume that the input data is complete and do not take into account the reality of frequent loss of sensor data in the marine environment.

[0006] (4) Separation of missing data processing and prediction: Existing methods typically employ simple preprocessing techniques such as mean imputation and linear interpolation, neglecting the adaptability of the imputation method to the downstream prediction task. Full Subsequence Matching (FSM) performs well on strongly periodic data, but OWC cell dynamics are quasi-periodic stochastic processes, and the strong periodicity assumption of FSM does not always hold. Furthermore, existing deep learning multivariate imputation methods require a large amount of training data and computational resources, making them unsuitable for edge computing environments.

[0007] (5) Existing methods for multivariate time series similarity search: In the field of multivariate time series similarity search, there are mature methods such as multidimensional dynamic time warping and the MASS algorithm. However, multidimensional dynamic time warping suffers from waveform misalignment caused by time axis stretching and compression in handling continuous missing gaps, and the MASS algorithm has a large computational cost under the condition of a large-scale pattern library. In recent years, although deep learning-based multivariate interpolation methods can model the dynamic correlation between variables, they require a large amount of training data and computational resources, and are not suitable for real-time applications in edge computing environments.

[0008] In summary, there is currently a lack of an integrated method that can simultaneously handle missing OWC gas chamber data, fully model the liquid level-gas pressure coupling relationship, and achieve multi-step advance prediction. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based method and system for predicting the coupling of liquid level fluctuations and air pressure in the air chamber of an OWC device. This aims to solve the following technical problems: when sensor data is continuously missing, existing simple interpolation methods disrupt the physical coupling relationship between liquid level and air pressure, affecting prediction accuracy; existing deep learning prediction methods fail to explicitly model the time-delay coupling relationship between liquid level and air pressure; existing FSM methods rely on strong periodicity assumptions and are difficult to directly apply to quasi-periodic wave-driven processes. This invention embeds the physical coupling relationship between liquid level and air pressure into the entire process of missing value interpolation and deep learning prediction, achieving high-fidelity repair and physically self-consistent prediction under data loss, providing reliable status information for the control and monitoring of OWC devices.

[0010] In a first aspect, the present invention provides a method for predicting liquid level fluctuations and gas pressure coupling in an OWC device based on deep learning, comprising the following steps: S1: Obtain historical liquid level data and historical gas pressure data during normal operation of the OWC device, and construct a pattern library consisting of pairs of liquid level subsequences and gas pressure subsequences through sliding window slicing; S2: Receive sensor data streams in real time. When continuous gaps are detected in the level sensor and / or pressure sensor, determine the location and length of the gap. Take known data segments on both sides of the gap and fill the gap with the local mean to construct query subsequences for level and pressure respectively. S3: In the pattern library, based on the similarity between the query subsequence and each candidate subsequence in the two dimensions of liquid level and gas pressure, and combined with the time-delay coupling consistency constraint between liquid level and gas pressure, a matching search and multi-level quality inspection are performed, and the matching success or failure is determined according to the inspection results. S4: Based on the matching result, the reconstruction method corresponding to the matching result is used to generate the interpolation value of the missing part, and the interpolation value is spliced ​​with the known data segments on both sides of the missing gap to obtain the complete liquid level sequence and the complete gas pressure sequence; S5: Input the data from the most recent historical moment in the reconstructed complete liquid level sequence and complete gas pressure sequence into the deep learning prediction model. The loss function of the model includes a prediction accuracy term and a coupling regularization term used to constrain the physical consistency between the output liquid level prediction value and the gas pressure prediction value. The model outputs liquid level prediction values ​​and gas pressure prediction values ​​for multiple future moments.

[0011] Furthermore, in step S1, the length of the sliding window covers 3 to 6 typical wave cycles; an accelerated index structure is established for the pattern library, which adopts principal component analysis dimensionality reduction combined with KD tree, or adopts local sensitive hash index; the pattern library adopts an incremental update strategy based on a fixed time span.

[0012] Furthermore, in step S2, known data segments of equal length to the missing length are taken on both sides of the missing gap; the local mean is the arithmetic mean of all known data in the current query sequence; the constructed liquid level query subsequence and gas pressure query subsequence are standardized to eliminate dimensional differences.

[0013] Further, step S3 includes: simultaneously calculating the similarity between the query subsequence and each candidate subsequence in the two dimensions of liquid level and air pressure in the pattern library, and introducing a time delay parameter characterizing the phase lag relationship between liquid level and air pressure for coupling consistency constraint screening to obtain a candidate pool; wherein, the similarity between the two dimensions of liquid level and air pressure is measured by calculating the joint distance, which is the weighted sum of the Euclidean distance of liquid level and the Euclidean distance of air pressure, wherein the coupling weight coefficient is adaptively determined according to the historical correlation coefficient of liquid level and air pressure, and the stronger the correlation between liquid level and air pressure, the greater the weight of the liquid level distance; the time delay coupling consistency constraint is: only retaining candidate subsequences whose time delay correlation coefficient exceeds a first preset threshold, and the time delay parameter is adaptively determined by the peak position of the cross-correlation function between the rate of change of liquid level and the rate of change of air pressure in historical data.

[0014] Furthermore, in step S3, the multi-level quality inspection includes: candidate pool quantity inspection: after obtaining the candidate pool through time-delay coupling consistency constraint screening, the number of candidate sub-sequences in the candidate pool is counted. If the number is lower than a preset lower limit, the first preset threshold is automatically reduced and time-delay coupling consistency constraint screening is performed again. If the number is still lower than the preset lower limit after re-screening, time-delay coupling consistency constraint screening is abandoned and all candidate sub-sequences are used as a new candidate pool.

[0015] Furthermore, in step S3, the matching search and multi-level quality inspection also include: selecting the subsequence with the smallest joint distance from the candidate pool after quality inspection as the most similar candidate match, and performing a similarity threshold test on the most similar candidate match; if the test passes, the match is determined to be successful, and if the test fails, the match is determined to be unsuccessful and a backoff strategy is triggered; the similarity threshold is adaptively determined based on the quantile of the distance distribution of self-matching of complete sequences in historical data.

[0016] Furthermore, in step S4, when a match is successful, the reconstruction method corresponding to the matching result is scaling interpolation, which specifically includes: calculating a scaling ratio based on the ratio of the data range of the known part in the query subsequence to the data range of the corresponding part in the most similar candidate match; if the data range of the corresponding part in the most similar candidate match is zero, then setting the scaling ratio to 1; adjusting the amplitude of the waveform corresponding to the missing position in the most similar candidate match according to the scaling ratio, and then shifting it based on the mean of the known part of the query subsequence to obtain the interpolated value of the missing part; When a matching failure triggers a fallback strategy, the reconstruction method corresponding to the matching result is determined differently based on the missing scenario: if only a single sensor data is missing, Kalman filter interpolation based on the OWC gas chamber lumped parameter model is used, and the state estimate is continuously corrected using the measurement value of the other sensor that is not missing, generating the interpolated value for the missing part; if both liquid level and gas pressure data are missing, open-loop propagation based on the physical model of the most recently known state is used, with the state estimate at the start of the missing time as the initial value, and the state is recursively calculated according to the dynamic equation of the lumped parameter model to generate the state trajectory of the missing period as the interpolated value.

[0017] Furthermore, in step S5, the deep learning prediction model adopts a hybrid architecture of convolutional neural network and long short-term memory network; the coupling regularization term is defined as: the square of the deviation between the time-delay correlation coefficient between the liquid level change rate sequence and the gas pressure change rate sequence output by the model and the pre-statistical historical time-delay correlation coefficient.

[0018] Furthermore, the method also includes the following steps: S6: When new sensor data arrives, slide the prediction window to update the input sequence with the latest data, and repeat step S5. S7: Perform a coupling consistency check on the liquid level prediction value and gas pressure prediction value output in step S5, and adaptively adjust the subsequent prediction step size based on the check result. Specifically, the coupling consistency verification involves: calculating the real-time correlation coefficient between the predicted liquid level change rate and the pressure change rate at multiple times corresponding to the time delay parameter; if the deviation between the real-time correlation coefficient and the historically statistical time delay correlation coefficient exceeds two standard deviations, a verification failure event is recorded; when the number of consecutive verification failures reaches 3, the prediction step size is automatically reduced to half of the original size; the count of verification failures is automatically cleared after any verification passes, and the original prediction step size is restored.

[0019] In a second aspect, the present invention also provides a deep learning-based prediction system for liquid level fluctuation and gas pressure coupling in an OWC device's gas chamber, for performing the method described in the first aspect, comprising: The data acquisition module is used to receive real-time data from the liquid level sensor and the air pressure sensor in the gas chamber of the OWC device. The historical pattern library module is used to store and manage liquid level-gas pressure sub-sequence pairs constructed from complete historical data and their corresponding acceleration indexes; The missing data handling and matching module is used to detect missing data, construct query sequences, perform joint similarity search and coupling constraint filtering, and trigger rollback strategies. The sequence reconstruction module is used to generate interpolation values ​​for missing parts based on matching results or backtracking strategies, and then concatenate the interpolation values ​​with known data to reconstruct complete liquid level and gas pressure sequences. The coupled prediction module has a built-in deep learning model with coupling regularization terms, which is used to output liquid level and gas pressure prediction values ​​for multiple future moments based on the complete historical sequence of the input. The rolling update and verification module is used to slide the prediction window, perform coupling consistency verification on the prediction results, and adaptively adjust the prediction step size according to the verification results, and finally output the liquid level prediction value and gas pressure prediction value at multiple future times.

[0020] In a third aspect, a control device is provided, including a processor and a storage device, the storage device being adapted to store a plurality of program codes, the program codes being adapted to be loaded and run by the processor to perform the method described in the second aspect.

[0021] In a fourth aspect, a computer-readable storage medium is provided having a plurality of program codes stored thereon, the program codes being adapted to be loaded and run by a processor to perform the method described in the second aspect.

[0022] The above-described technical solutions of the present invention have at least one or more of the following beneficial effects: 1. Within the sea state range covered by the historical database, this invention improves the joint similarity search and time-delay coupling consistency constraint of the FSM, ensuring that the interpolation results simultaneously maintain the respective fluctuation characteristics of liquid level and gas pressure, as well as their physical coupling relationship. In extreme or novel sea states not covered by the historical database, this invention automatically switches to Kalman filtering interpolation or open-loop propagation of the physical model based on the missing scenario. Although the accuracy decreases slightly, functional continuity is still guaranteed.

[0023] 2. The model of this invention can predict the state of the gas chamber in the next 12 to 24 steps (6 to 12 seconds with a 0.5-second sampling interval), providing reference state information for turbine control.

[0024] 3. The improved FSM interpolation method of this invention, through joint matching and adaptive scaling interpolation, can adapt to different wave conditions within the training data coverage area. Under extreme sea conditions not covered by the training data, performance exhibits an expected decrease.

[0025] 4. On typical edge computing hardware (2.3 GHz CPU, 16 GB RAM), the average inference latency of the interpolation stage is about 30~80 ms, and that of the prediction stage is about 50~120 ms, which fully meets the time requirements for data processing with a sampling interval of 0.5~2 seconds.

[0026] 5. The residual sequence between the predicted and actual measured values ​​of this invention can serve as a reference indicator for device performance degradation. Coupling consistency verification can further identify sensor faults or abnormal physical conditions.

[0027] 6. The effectiveness of this invention depends on the extent to which the historical model library covers the current sea state characteristics. Under extreme sea conditions, the system automatically switches to a backoff strategy, which reduces the accuracy of interpolation and prediction. This is a common limitation of data-driven methods. Attached Figure Description

[0028] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a flowchart illustrating the deep learning-based prediction method for liquid level fluctuation and gas pressure coupling in the gas chamber of an OWC device according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-chamber OWC wave energy device in an embodiment of the present invention; Figure 3 This is a schematic diagram of the improved FSM joint similar subsequence search in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the scaling interpolation principle of an embodiment of the present invention; Figure 5 This is a schematic diagram of a deep learning-based prediction system for liquid level fluctuation and gas pressure coupling in an OWC device according to an embodiment of the present invention. Detailed Implementation

[0029] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0030] In the description of this invention, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, memory, and may also include software components, such as program code, or a combination of software and hardware. A processor can be a central processing unit, microprocessor, image processor, digital signal processor, or any other suitable processor. The processor has data and / or signal processing capabilities. The processor can be implemented in software, in hardware, or a combination of both. Non-transitory computer-readable storage media includes any suitable medium capable of storing program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random access memory, etc. The term "A and / or B" means all possible combinations of A and B, such as only A, only B, or A and B. The terms "at least one A or B" or "at least one of A and B" have a similar meaning to "A and / or B" and can include only A, only B, or A and B. The singular terms "a" or "this" can also include plural forms.

[0031] This invention provides a deep learning-based method and system for predicting the coupling of liquid level fluctuations and air pressure in the gas chamber of an OWC (Oscillating Water Column) wave energy generator. Specifically, it addresses the missing data repair and future state coupling prediction of liquid level fluctuations and air pressure changes within the gas chamber of an OWC. By combining similarity calculation and time-delay coupling consistency constraints, the physical coupling relationship between liquid level and air pressure is embedded into the missing value imputation process, achieving high-fidelity repair of continuous missing segments. A CNN-LSTM coupling prediction model is constructed, and a coupling consistency regularization term is introduced into the loss function to ensure that the prediction results meet practical requirements in terms of both numerical accuracy and physical consistency. Multi-level quality checks and differentiated backoff strategies ensure the robustness and functional continuity of the method under conditions of insufficient historical database coverage or extreme sea conditions. Ultimately, it provides reliable and real-time information on the coupling of liquid level and air pressure in the gas chamber for turbine control, condition monitoring, and health management of OWC devices. In particular, given that the dynamics of the OWC air chamber are driven by random ocean waves and belong to a quasi-periodic random process, the implicit assumption in the full subsequence matching FSM method that "highly similar subsequences of waveforms can always be found in the historical database" is not always valid, this invention makes the following targeted improvements when porting FSM: the physical coupling relationship between liquid level and air pressure is used as a proxy criterion when waveform similarity is insufficient, and the physical rationality of the matching results is ensured by time-delay correlation constraints; a similarity threshold control is introduced, which automatically switches to physical model interpolation or pure prediction extrapolation when there are not enough similar patterns in the historical database; a time-delay coupling loss function is constructed, and a liquid level-air pressure coupling regularization term is embedded in the prediction stage to ensure that the prediction results are physically self-consistent.

[0032] Figure 1This is a flowchart illustrating a deep learning-based prediction method for the coupling of liquid level fluctuations and gas pressure within the gas chamber of an OWC device, according to an embodiment of the present invention. Figure 1 As shown, a deep learning-based method for predicting liquid level fluctuations and gas pressure coupling in an OWC device includes the following steps: S1: Obtain historical liquid level data and historical gas pressure data during normal operation of the OWC device, and construct a pattern library consisting of pairs of liquid level subsequences and gas pressure subsequences through sliding window slicing; Step S1 is used to build the historical pattern library and index acceleration structure.

[0033] like Figure 2 The diagram shows a schematic of a multi-chamber OWC wave energy device. Liquid level and pressure data collected during normal operation of the OWC device are compiled into a historical database. Based on the sampling requirements for wave dynamics, the sampling interval is set to 0.5–2 seconds to meet Nyquist sampling requirements and retain high-frequency disturbance information within the chambers. The historical data is sliced ​​into sliding windows, with each window containing a liquid level subsequence of length L and a corresponding pressure subsequence, constructing a paired "liquid level-pressure" pattern library M, defined as: (1) in, Represents the liquid level subsequence. This represents the corresponding pressure subsequence. This represents the number of samples in the pattern library. Let i be the liquid level subsequence of the i-th sample in the pattern library. The barometric subsequence of the i-th sample in the model library is retained only in complete, unmissing segments.

[0034] To meet real-time retrieval requirements, an accelerated index structure is constructed for the pattern library M. This accelerated index structure employs principal component analysis (PCA) dimensionality reduction combined with a KD-tree, or a locality-sensitive hash index. The pattern library uses an incremental update strategy based on a fixed time span. The preferred sliding window length covers 3 to 6 typical wave cycles, with a window length L of 60–120, which falls into the mid-to-high dimensionality range. The following combined strategy is prioritized: PCA dimensionality reduction is performed on the window subsequences, retaining the principal components corresponding to 90% of the variance contribution rate; after dimensionality reduction, a KD-Tree is constructed in the low-dimensional space, achieving a query time complexity that is approximately [missing information - likely a time complexity threshold]. When the pattern library exceeds 10,000 windows or the sequence dimension exceeds 10, it automatically switches to a locality-sensitive hash index, with a query time complexity of O(log n). Furthermore, an incremental update strategy based on sliding windows is adopted—when new data arrives, only the newly generated window subsequence is inserted into the index, while the oldest windows that are outside the valid historical range are discarded. The valid historical range is defined as a fixed time span of 30 days to ensure that the pattern library covers a sufficiently long sea state change cycle; if the number of windows within the time span is less than the lower limit of 5,000, the time span is automatically expanded until the quantity requirement is met.

[0035] It should be noted that the liquid level data used in this invention is obtained by inversion calculation based on gas chamber pressure data through a hydrodynamic model, or by simulation data generated based on CFD numerical simulation. Because the gas chamber hydrodynamic model incorporates fluid inertia terms, gas compressibility, and gas chamber geometric boundary conditions, the inverted liquid level sequence is not a simple functional mapping of gas pressure, but rather an independent state variable containing the dynamic response characteristics of the system, which, together with the gas pressure, constitutes a complete state-space description of the gas chamber state.

[0036] S2: Receive sensor data streams in real time. When continuous gaps are detected in the level sensor and / or pressure sensor, determine the location and length of the gap. Take known data segments on both sides of the gap and fill the gap with the local mean to construct query subsequences for level and pressure respectively. Step S2 is used to implement real-time missing detection and query sequence construction.

[0037] Receive sensor data streams in real time, and perform the following operations when a continuous loss of liquid level sensor and / or air pressure sensor is detected.

[0038] First, determine the start and end points of the missing gap and the length of the missing gap. (i.e., the number of consecutively missing data points). Known data segments of equal length to the missing length are taken on both sides of the missing gap; the local mean is the arithmetic mean of all known data in the current query sequence; the constructed liquid level query subsequence and gas pressure query subsequence are standardized to eliminate dimensional differences. Specifically: For the liquid level, take the m known liquid level values ​​to the left of the missing gap and the n known liquid level values ​​to the right of the missing gap, and take the local average value for the T missing positions in the middle. Filling; for air pressure, take m known air pressure values ​​to the left of the missing gap and n known air pressure values ​​to the right of the missing gap, and take the local average of the T missing positions in the middle. Filling. Preferably, m=n=T, that is, taking known segments of equal length on both the left and right sides. Local mean. Defined as The arithmetic mean of all known liquid level values. Defined as The arithmetic mean of all known air pressure values. The query sequence can be represented as: (2) (3) in, This is a sequence of known liquid level values ​​to the left of the missing gap. This is a sequence of known liquid level values ​​to the right of the missing gap. The sequence of known air pressure values ​​is to the left of the missing gap. This is a sequence of known air pressure values ​​to the right of the missing gap. : Liquid level query subsequence; : Barometric pressure query subsequence.

[0039] More preferably, the query sequence is Z-score standardized before matching to eliminate the influence of the difference in dimensions between liquid level and gas pressure on distance calculation. The standardization method is as follows: and Subtract each local mean and then divide by its local standard deviation.

[0040] S3: In the pattern library, based on the similarity between the query subsequence and each candidate subsequence in the two dimensions of liquid level and gas pressure, and combined with the time-delay coupling consistency constraint between liquid level and gas pressure, a matching search and multi-level quality inspection are performed, and the matching success or failure is determined according to the inspection results. Step S3 is used to perform joint search and matching verification. This step S3 includes the following three sub-steps.

[0041] 3.1: Joint Similarity Calculation In step S3.1, the similarity between the query subsequence and each candidate subsequence in the pattern library M is calculated simultaneously along two dimensions: liquid level and air pressure. The similarity in these two dimensions is measured by calculating the joint distance, which is a weighted sum of the Euclidean distance between liquid level and air pressure. The coupling weight coefficient is adaptively determined based on the historical correlation coefficients between liquid level and air pressure; the stronger the correlation between liquid level and air pressure, the greater the weight of the liquid level distance. Specifically: Define joint distance for: (4) in, Represents Euclidean distance. and These are the overall standard deviations (used for normalization) of historical data for liquid level and gas pressure, respectively. These are the coupling weight coefficients. For general query sequence representation, including and . For a general candidate subsequence representation, including and . The k-th candidate subsequence (containing the liquid level portion) in the pattern library M and air pressure section ). α is determined by the correlation coefficient between liquid level and gas pressure history. Adaptive determination: (5) in This represents the Pearson correlation coefficient between the liquid level series and the gas pressure series in the complete historical data.

[0042] 3.2: Coupling Consistency Constraint Screening After calculating the similarity between the query subsequence and each candidate subsequence in terms of liquid level and gas pressure in step S3.1, a time delay parameter characterizing the phase lag relationship between liquid level and gas pressure is introduced for coupling consistency constraint screening to obtain a candidate pool, which is a set of candidate subsequences filtered by a time delay correlation coefficient threshold. The time delay coupling consistency constraint is: only candidate subsequences whose time delay correlation coefficient exceeds a first preset threshold are retained. Specifically: Considering the phase lag between the liquid level and gas pressure in the OWC gas chamber, a time lag parameter is introduced. Calculate the rate of change of liquid level and hysteresis. Time-delay correlation coefficient between the rate of change of air pressure at time t_ ... : (6) in, Indicates the rate of change of liquid level. This represents the rate of change of air pressure. This represents the Pearson correlation coefficient. The time lag parameter τ is adaptively determined by the peak position of the cross-correlation function between the rate of change of liquid level and the rate of change of gas pressure in historical data.

[0043] Only retain the time-delay correlation coefficient Historical segments exceeding a preset threshold θ are entered into the matching candidate pool. The threshold θ is adaptively determined using the 70th percentile of the historical correlation coefficient distribution.

[0044] Finally, through step S3.2, historical segments with time delay correlation coefficients ρ(τ) lower than the preset threshold θ are removed from the pattern library M, and only candidate subsequences that meet the coupling consistency constraint are retained to enter the subsequent matching candidate pool.

[0045] 3.3: Matching Search and Multi-level Quality Inspection After obtaining the candidate pool in step S3.2, step S3.3 performs matching search and multi-level quality checks, specifically including: selecting the subsequence with the smallest joint distance from the candidate pool after being filtered by time-delay coupling consistency constraints as the most similar candidate match, and performing a similarity threshold check on the most similar candidate match; if the check passes, the match is determined to be successful, and if the check fails, the match is determined to be unsuccessful and a backoff strategy is triggered; the similarity threshold is adaptively determined based on the quantile of the distance distribution of self-matching of complete sequences in historical data.

[0046] 1. Candidate Pool Quantity Verification: Count the number of candidate subsequences in the candidate pool that satisfy the time-delay coupling consistency constraint. If this number is lower than a preset lower limit, the threshold for the time-delay coupling consistency constraint is automatically relaxed (i.e., lowered), and the time-delay coupling consistency constraint screening is performed again to expand the candidate pool. If the number is still lower than the preset lower limit after rescreening, the time-delay coupling consistency constraint screening is abandoned, and all candidate subsequences are used as a new candidate pool. For example: if the number of candidate sequences satisfying the coupling constraint is lower than the preset lower limit... (If the value is 5), the threshold θ will be automatically relaxed to the 50th percentile for a re-search; if it is still lower than... If so, it reverts to matching based solely on joint distance without imposing coupling constraints.

[0047] 2. Similarity Threshold Check: From the candidate pool that has undergone the quality check above, the subsequence with the smallest joint distance is selected as the most similar candidate match, and its normalized joint distance is calculated. If the joint distance exceeds a preset similarity threshold, the match is considered a failure, i.e., "no sufficiently similar historical patterns," triggering a backoff strategy; otherwise, the match is considered successful. For example: Select the historical subsequence with the smallest joint distance from the candidate pool. As a candidate matching, its normalized joint distance is denoted as The similarity threshold η is adaptively determined by the 90th percentile of the distance distribution of self-matching complete sequences in historical data, with a recommended range of 0.3 to 0.5. If... If the value exceeds η, the match is considered a failure; if If the number of matches does not exceed η, the match is considered successful.

[0048] Figure 3 This illustrates the principle of the improved FSM joint similar subsequence search of the present invention. Figure 3 The upper and lower channels correspond to the liquid level sequence and the air pressure sequence, respectively. The horizontal axis is the time axis, and the vertical axes are the liquid level value and the air pressure value, respectively. The query sequence Q is composed of known data segments (solid lines) on both sides of the missing gap and the missing segment (flat segment) filled with local means. Historical candidate subsequences in the pattern library. The final selected most similar match is shown by a dashed line. Indicated by a solid green line. Unlike the traditional univariate FSM method, which calculates Euclidean distance in only a single channel, this invention simultaneously calculates the normalized Euclidean distance in both the liquid level and gas pressure channels, and obtains the joint distance through a weighted summation using an adaptive weight α. This is used to measure the overall similarity between the candidate subsequence and the query sequence.

[0049] S4: Based on the matching result, the reconstruction method corresponding to the matching result is used to generate the interpolation value of the missing part, and the interpolation value is spliced ​​with the known data segments on both sides of the missing gap to obtain the complete liquid level sequence and the complete gas pressure sequence; Step S4 is used to generate missing values ​​based on improved scaling interpolation.

[0050] S4.1: Scaling interpolation like If the value does not exceed η, and the match is considered successful, the reconstruction method corresponding to this match result is scaling interpolation. Specifically, the scaling interpolation method is used to generate the missing value. This includes: calculating a scaling ratio based on the ratio of the data range of the known portion in the query subsequence to the data range of the corresponding portion in the most similar candidate match; adjusting the amplitude of the waveform corresponding to the missing position in the most similar candidate match according to the scaling ratio; and then shifting it based on the mean of the known portion of the query subsequence to obtain the interpolated value of the missing portion. The details are as follows: Taking liquid level as an example, let's assume a liquid level query subsequence. The known part is (i.e., the actual observations on both sides of the missing gap), the corresponding part of the matching sequence is The missing part is Calculate the scaling ratio. : (7) in, The liquid level portion of the most similar historical subsequence selected in step S3; Matching liquid level subsequences Zhongyu The part corresponding to the location; Matching liquid level subsequences The portion corresponding to the location of the missing gap; : Liquid level scaling ratio; : Retrieves the maximum value in the sequence; : Take the minimum value in the sequence.

[0051] If the denominator is zero (i.e., the corresponding part of the matching sequence is a constant), then let Scaled interpolation values ​​for missing parts for: (8) in, : The sequence of interpolated values ​​for the missing liquid level portion; for The arithmetic mean (mean). for The arithmetic mean (mean) of the data. This scaling interpolation method takes into account both amplitude scaling and baseline shifting, so that the interpolated value preserves the fluctuation pattern of the matched sequence while adapting to the absolute value level of the current query sequence.

[0052] Interpolation sequence of missing air pressure portions Calculate the gas compression and release ratio using the same method. (by and) (Calculate in the same way) and generate interpolated values.

[0053] Figure 4 This is a schematic diagram of the scaling interpolation principle. Figure 4 The present invention illustrates the improved scaling interpolation principle based on amplitude scaling and baseline translation, using liquid level variables as an example. Figure 4 The solid black line in the middle represents the known liquid level sequence on both sides of the missing gap. The black horizontal dashed line represents its mean. . Figure 4 The gray dashed line in the middle represents the most similar matching subsequence retrieved from the pattern database. The original waveform at the position corresponding to the missing gap As can be seen, the mean and amplitude of the original waveform differ from the known sequence. Directly using it for interpolation would result in significant jumps at the boundaries of the missing gaps. The red solid line represents the interpolation result after adjustments for both amplitude scaling and baseline shifting. First, calculate the scaling factor by comparing the known amplitude range of the sequence with the known amplitude range of the matched sequence. ,Will The fluctuation pattern according to Perform amplitude compression or amplification; then use The scaled waveform is baseline-shifted to ensure the mean level of the interpolated values ​​matches the known sequence. Figure 4 As can be seen, the adjusted interpolation result achieves a smooth connection at the left and right boundaries, while preserving the fluctuation characteristics of the matched sequence, demonstrating the advantages of this invention in waveform fidelity compared to simple mean filling and linear interpolation.

[0054] S4.2: Determine the interpolation method based on the differences in missing scenarios. like If the value exceeds η, the match is considered a failure. In this case, FSM scaling interpolation is not performed. Instead, a differentiated backoff strategy is triggered based on the missing scenario. The reconstruction method corresponding to this matching result is determined based on the differences in the missing scenario, including: if only a single sensor data is missing, Kalman filter interpolation based on the OWC gas chamber lumped parameter model is used. The state estimate is continuously corrected using the measurement value of the other sensor that is not missing, generating the interpolated value for the missing part; if both liquid level and gas pressure data are missing, open-loop propagation based on the physical model of the most recently known state is used. The state estimate at the start of the missing data is used as the initial value, and the state is recursively calculated according to the dynamic equation of the lumped parameter model to generate the state trajectory of the missing period as the interpolated value. The details are as follows: Scenario A (Single Sensor Missing Only): If only the liquid level or only the gas pressure is missing, the process reverts to Kalman filter interpolation. The state equation is constructed based on the OWC gas chamber lumped parameter model. The state variables include the liquid level h and its rate of change v, while the observed variables are the measurements from the other sensor that is not missing. The state transition relationships are described by the gas chamber hydrodynamic and thermodynamic equations, and the state estimate is continuously corrected using the measured data from the sensor that is not missing.

[0055] Scenario B (Simultaneous Loss of Both Sensors): If both liquid level and gas pressure are simultaneously and continuously lost, the system reverts to open-loop propagation based on the most recently known physical model. Using the state estimate at the start of the loss as the initial value, pure state recursion is performed according to the dynamic equations of the lumped-parameter model of the gas chamber to generate the state trajectory for the lost period as the interpolation value. In this scenario, the interpolation accuracy gradually decreases as the loss length increases. When the continuous loss length T exceeds half the window length L, a warning signal is generated to indicate significant interpolation uncertainty.

[0056] S5: Input the data from the most recent historical moment in the reconstructed complete liquid level sequence and complete gas pressure sequence into the deep learning prediction model. The loss function of the model includes a prediction accuracy term and a coupling regularization term used to constrain the physical consistency between the output liquid level prediction value and the gas pressure prediction value. The model outputs liquid level prediction values ​​and gas pressure prediction values ​​for multiple future moments.

[0057] Step S5 is used to build and train the deep learning prediction model.

[0058] A multi-input, multi-output deep learning prediction model is constructed, employing a hybrid architecture of convolutional neural networks and long short-term memory networks (CNN-LSTM hybrid architecture). The inputs are the liquid level and gas pressure sequences from the past p time steps, and the output is the predicted liquid level values ​​for the next q time steps. and air pressure forecast .

[0059] The model introduces a coupling consistency regularization term into the loss function, and the formula for the loss function is: (9) in, This is the sum of the mean square errors of the predicted liquid level and gas pressure values. For coupling regularization weights. Defined as the time-lag correlation between the rate of change of liquid level and the rate of change of gas pressure in the predicted sequence and historical statistics. The square of the deviation. To ensure the differentiability and numerical stability of the loss term, a sliding window intra-batch correlation coefficient is used for calculation, and a numerical stability term is added. (Values ​​range from 1e-8): (10) in, ; Prediction step size, i.e., the number of future time steps output by the model; : Time delay parameter, representing the number of time delay steps between liquid level change and gas pressure change; Time step index; : No. The first difference of the predicted liquid level at a given time; : No. First difference of the predicted air pressure at time point, ; : No. Predicted liquid level at any given time; : The predicted air pressure at time t.

[0060] Historical statistics The calculation method is as follows: historical data is segmented into statistical windows of 30 minutes each (with 50% overlap between windows), and the time-lag correlation coefficient is calculated for each window. The average of the correlation coefficients of all windows is then taken as the mean. When running online, the system continuously maintains rolling statistics for the most recent 30 minutes. ,like and If the deviation exceeds twice the standard deviation, it will trigger... Adaptive update - Incorporation by exponentially weighted average (attenuation coefficient γ=0.9) .in, : The correlation coefficient of flow statistics with time lag during online operation over the most recent 30 minutes; The time-lag correlation coefficient between the historical statistical rates of change in liquid level and the rates of change in gas pressure; : The decay coefficient of the exponentially weighted average, with a value of 0.9; : The square norm of the vector difference (i.e., the squared error).

[0061] Preferably, in some embodiments, the method further includes the following steps: S6: When new sensor data arrives, slide the prediction window to update the input sequence with the latest data, and repeat step S5. Step S6 is used to implement multi-step rolling prediction and sliding window incremental update. Specifically: After obtaining the complete liquid level and gas pressure sequence including interpolated values, perform multi-step rolling prediction: [The last part is incomplete and likely refers to a specific timeframe or time period]. p The liquid level and gas pressure data at each moment are input into the trained coupled prediction model to obtain future... q The prediction window is set at a given time step; when new sensor data arrives, the prediction window is moved forward one time step to update the input sequence with the latest observation or interpolated value, and prediction is performed again.

[0062] S7: Perform a coupling consistency check on the liquid level prediction value and gas pressure prediction value output in step S5, and adaptively adjust the subsequent prediction step size based on the check result. Step S7 is used to perform coupling consistency verification.

[0063] The coupling consistency verification specifically involves calculating the predicted liquid level change rate and the hysteresis time delay parameter. τ The real-time correlation coefficient between the rate of change of air pressure at any given time, and if this real-time correlation coefficient is correlated with the time-lag correlation coefficient of historical statistics. If the deviation exceeds twice the standard deviation, a verification failure event is recorded; when the number of consecutive verification failures reaches 3, the prediction step size is automatically adjusted. q Reduced to half of its original size q / 2; The count of failed verifications is automatically cleared to zero after any successful verification and restored to the original prediction step size.

[0064] Figure 5 This is a schematic diagram of a deep learning-based prediction system for liquid level fluctuation and gas pressure coupling in an OWC device gas chamber, according to an embodiment of the present invention. A deep learning-based prediction system for liquid level fluctuation and gas pressure coupling in an OWC device gas chamber, used to execute the aforementioned method, includes: Data acquisition module 210 is used to receive liquid level sensor data and air pressure sensor data in the gas chamber of the OWC device in real time; Historical pattern library module 220 is used to store and manage liquid level-gas pressure sub-sequence pairs constructed from complete historical data and their corresponding acceleration indexes; The missing data handling and matching module 230 is used to detect missing data, construct query sequences, perform joint similarity search and coupling constraint filtering, and trigger a rollback strategy. The sequence reconstruction module 240 is used to generate interpolation values ​​for missing parts based on matching results or backoff strategies, and to concatenate the interpolation values ​​with known data to reconstruct complete liquid level and gas pressure sequences. The coupling prediction module 250 has a built-in deep learning model with coupling regularization terms, which is used to output liquid level prediction values ​​and gas pressure prediction values ​​for multiple future moments based on the complete historical sequence of the input. The rolling update and verification module 260 is used to slide the prediction window, perform coupling consistency verification on the prediction results, and adaptively adjust the prediction step size according to the verification results, and finally output the liquid level prediction value and gas pressure prediction value at multiple future times.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0066] 1. Data Sources and Processing: The following data sources were used to verify the method in this embodiment.

[0067] (1) Air pressure data: Based on the air chamber parameters of the Mutriku wave power station (total installed capacity 296 kW, air chamber size 4.5 m wide × 3.1 m long, opening diameter 750 mm, see BiMEP official technical data and PRIMRE database for details), the JONSWAP irregular wave spectrum (effective wave height) was adopted. =1.5~3.5 m, energy cycle =6~10 s, peak enhancement factor γ=3.3) driving the OWC gas chamber lumped parameter hydrodynamic model, generation time 30 days (approximately 2.59×10 6 The simulated data sequence of air chamber pressure (seconds) with a sampling interval of 0.5 seconds.

[0068] (2) Liquid level data: Based on the simulated gas chamber pressure, the position of the free liquid surface inside the gas chamber is synchronously output through the same lumped parameter model to obtain the corresponding liquid level simulation data sequence. The state equation of the lumped parameter model is in the form of: ; in, For liquid level, The rate of change of liquid level. For the effective mass of the water column, External wave-induced pressure, The pressure in the air chamber. The density of seawater, It is the acceleration due to gravity. This is a hydrodynamic loss term. This is a function of gas thermodynamic processes. Under this model, the liquid level... With air pressure The instantaneous Pearson correlation coefficient is approximately This indicates that there is a strong correlation between the two, but it is not a deterministic functional mapping relationship. The liquid level contains dynamic response characteristics independent of the gas pressure.

[0069] Data preprocessing: Remove unstable data during the simulation startup phase (first 2 hours); mark transient outliers (exceeding ±3σ range) and treat them as missing values; select a complete data segment of 20 consecutive days for pattern library construction, and then select a data segment of 5 consecutive days for testing; standardize the data to make its mean 0 and variance 1.

[0070] Missing data simulation: In the test data segment, two methods were used to simulate missing data. (1) Random missing: Data points were randomly deleted at proportions of 10%, 20%, 30%, 40%, and 50%, respectively. (2) Continuous missing: Sensor interruption failure was simulated, and continuous missing blocks were deleted at time steps of 12, 24, 48, 96, and 144 (corresponding to 6, 12, 24, 48, and 72 seconds, respectively). For the scenario of simultaneous missing data in both channels, an additional test group was set up in the continuous missing simulation to delete liquid level and gas pressure at the same position.

[0071] 2. Parameter settings: Table 1 Parameter Settings

[0072] 3. Quantitative Experimental Results: The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used to measure the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The liquid level-gas pressure prediction (R) is used as an evaluation index. 2 Defined as liquid level prediction R 2 With air pressure prediction R 2 The arithmetic mean of each variable R 2 =1– / , To predict the sum of squared residuals, This is the sum of squares of the total deviations from the true values.

[0073] Set the following comparison method: Benchmark 1: Linear interpolation + univariate LSTM Benchmark 2: Original FSM interpolation + Univariate LSTM Benchmark 3: MissForest interpolation + univariate LSTM Benchmark 4: Multivariate LSTM (MV-LSTM) Benchmark 5: Multidimensional Dynamic Time Warping Interpolation + MV-LSTM Benchmark 6: Kalman filter interpolation + MV-LSTM This invention (complete): Improved FSM joint interpolation + CNN-LSTM coupled prediction Ablation 1: This invention removes the coupling loss term (λ=0). Ablation 2: This invention removes the similarity threshold and backoff strategy (η=∞). The results of routine sea state experiments (30% random missing rate) are shown in Table 2: Table 2 Results of experiments in normal sea states

[0074] The results of the extreme sea state experiment (significant wave height exceeding the training set by more than 20%) are shown in Table 3: Table 3 Results of extreme sea state experiments

[0075] The results of the dual-path simultaneous continuous deletion experiment (deletion length T=48 steps) are shown in Table 4: Table 4 Results of simultaneous continuous deletion experiment in dual-pathway network

[0076] Experimental results show that under normal sea conditions, the method of this invention significantly outperforms all benchmarks in all evaluation metrics (Wilcoxon rank-sum test, p<0.05). Ablation version 2 performs slightly better than the full version of this invention under normal sea conditions because the historical pattern library is sufficiently covered, the similarity threshold η is rarely triggered, and the backoff strategy is almost not activated. Under extreme sea conditions, the performance of ablation version 2 without a backoff strategy deteriorates significantly, while this invention controls the RMSE deterioration to about 97% through the backoff strategy, and the level-pressure prediction R... 2 The value remains at 0.61. In scenarios where both paths are simultaneously missing, this invention provides relatively stable interpolation results through open-loop propagation of the physical model, R... 2 It is approximately 27% higher than ablation version 2. The introduction of the coupling loss term increases the level-pressure prediction R... 2 An increase of approximately 7-10%.

[0077] 4. Applicable Boundaries and Limitations: The effectiveness of this method depends on the coverage of the historical model library with the current sea state characteristics. When the effective wave height exceeds the maximum wave height in the training data by more than 20%, or the wave period offset exceeds 30%, the joint distance usually exceeds the similarity threshold η, and the system automatically triggers a backoff strategy, resulting in a corresponding decrease in interpolation accuracy. When the consecutive missing length T exceeds 50% of the window length L, the effective information on both sides of the query sequence is severely insufficient, and it is recommended to directly backoff to physical model interpolation or open-loop propagation. In actual deployment, if the gas chamber is equipped with an independent liquid level measurement device, the measured liquid level data can be used directly, and the methodological framework remains consistent.

[0078] Furthermore, embodiments of the present invention also include a control device. In one embodiment of the control device according to the present invention, the control device includes a processor and a storage device. The storage device may be configured to store a program for executing the deep learning-based prediction method for liquid level fluctuation and pressure coupling in the air chamber of an OWC device according to the above-described method embodiments. The processor may be configured to execute the program in the storage device, which includes, but is not limited to, the program for executing the deep learning-based prediction method for liquid level fluctuation and pressure coupling in the air chamber of an OWC device according to the above-described method embodiments. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The control device may be a control device device comprising various electronic devices.

[0079] Furthermore, embodiments of the present invention also include a computer-readable storage medium. In one embodiment of the present invention, the computer-readable storage medium may be configured to store a program that executes the deep learning-based prediction method for liquid level fluctuation and pressure coupling in the gas chamber of an OWC device according to the above-described method embodiments. This program may be loaded and run by a processor to implement the deep learning-based prediction method for liquid level fluctuation and pressure coupling in the gas chamber of an OWC device. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The computer-readable storage medium may be a storage device comprising various electronic devices. Optionally, in the embodiments of the present invention, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0080] Furthermore, it should be understood that since the various modules are only provided to illustrate the functional units of the device of the present invention, the physical devices corresponding to these modules may be the processor itself, or a part of the processor's software, hardware, or a combination of software and hardware. Therefore, the number of modules shown in the figures is merely illustrative.

[0081] Those skilled in the art will understand that the various modules in the device can be adaptively split or combined. Such splitting or combining of specific modules will not cause the technical solution to deviate from the principles of the present invention; therefore, the technical solutions after splitting or combining will fall within the protection scope of the present invention.

[0082] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A deep learning-based method for predicting liquid level fluctuations and gas pressure coupling in an OWC device's gas chamber, characterized in that... Includes the following steps: S1: Obtain historical liquid level data and historical gas pressure data during normal operation of the OWC device, and construct a pattern library consisting of pairs of liquid level subsequences and gas pressure subsequences through sliding window slicing; S2: Receive sensor data streams in real time. When continuous gaps are detected in the level sensor and / or pressure sensor, determine the location and length of the gap. Take known data segments on both sides of the gap and fill the gap with the local mean to construct query subsequences for level and pressure respectively. S3: In the pattern library, based on the similarity between the query subsequence and each candidate subsequence in the two dimensions of liquid level and gas pressure, and combined with the time-delay coupling consistency constraint between liquid level and gas pressure, a matching search and multi-level quality inspection are performed, and the matching success or failure is determined according to the inspection results. S4: Based on the matching results, the corresponding reconstruction method is used to generate the interpolation value of the missing part, and the interpolation value is spliced ​​with the known data segments on both sides of the missing gap to obtain the complete liquid level sequence and the complete gas pressure sequence. S5: Input the data from the most recent historical moment in the reconstructed complete liquid level sequence and complete gas pressure sequence into the deep learning prediction model. The loss function of the model includes a prediction accuracy term and a coupling regularization term used to constrain the physical consistency between the output liquid level prediction value and the gas pressure prediction value. The model outputs liquid level prediction values ​​and gas pressure prediction values ​​for multiple future moments.

2. The method according to claim 1, characterized in that, In step S1, the length of the sliding window covers 3 to 6 typical wave cycles; an accelerated index structure is established for the pattern library, which adopts principal component analysis dimensionality reduction combined with KD tree, or uses local sensitive hash index; the pattern library adopts an incremental update strategy based on a fixed time span.

3. The method according to claim 1, characterized in that, In step S2, known data segments of equal length to the missing length are taken on both sides of the missing gap; the local mean is the arithmetic mean of all known data in the current query sequence; the constructed liquid level query subsequence and gas pressure query subsequence are standardized to eliminate dimensional differences.

4. The method according to claim 1, characterized in that, Step S3 includes: In the pattern library, the similarity between the query subsequence and each candidate subsequence in two dimensions, liquid level and gas pressure, is calculated simultaneously. A time delay parameter, which characterizes the phase lag relationship between liquid level and gas pressure, is introduced to perform coupling consistency constraint screening to obtain the candidate pool. The similarity between liquid level and air pressure is measured by calculating the joint distance, which is a weighted sum of the Euclidean distance between liquid level and air pressure. The coupling weight coefficient is adaptively determined based on the historical correlation coefficient between liquid level and air pressure. The stronger the correlation between liquid level and air pressure, the greater the weight of the liquid level distance. The time-delay coupling consistency constraint is to retain only candidate subsequences whose time-delay correlation coefficient exceeds a first preset threshold. The time-delay parameter is adaptively determined by the peak position of the cross-correlation function between the rate of change of liquid level and the rate of change of air pressure in historical data.

5. The method according to claim 4, characterized in that, In step S3, the multi-level quality inspection includes: candidate pool quantity inspection: after obtaining the candidate pool through time-delay coupling consistency constraint screening, the number of candidate sub-sequences in the candidate pool is counted. If the number is lower than the preset lower limit, the first preset threshold is automatically reduced and time-delay coupling consistency constraint screening is performed again. If the number is still lower than the preset lower limit after re-screening, the time-delay coupling consistency constraint screening is abandoned and all candidate sub-sequences are used as a new candidate pool.

6. The method according to claim 5, characterized in that, In step S3, the matching search and multi-level quality inspection further include: selecting the subsequence with the smallest joint distance from the candidate pool after quality inspection as the most similar candidate match, and performing a similarity threshold test on the most similar candidate match; if the test passes, the match is determined to be successful, and if the test fails, the match is determined to be unsuccessful and a backoff strategy is triggered; the similarity threshold is adaptively determined based on the quantile of the distance distribution of self-matching of complete sequences in historical data.

7. The method according to claim 6, characterized in that, In step S4, when a match is successful, the reconstruction method corresponding to the matching result is scaling interpolation, which specifically includes: calculating a scaling ratio based on the ratio of the data range of the known part in the query subsequence to the data range of the corresponding part in the most similar candidate match; if the data range of the corresponding part in the most similar candidate match is zero, then setting the scaling ratio to 1; adjusting the amplitude of the waveform corresponding to the missing position in the most similar candidate match according to the scaling ratio, and then shifting it based on the mean of the known part of the query subsequence to obtain the interpolated value of the missing part; When a matching failure triggers a fallback strategy, the reconstruction method corresponding to the matching result is determined differently based on the missing scenario: if only a single sensor data is missing, Kalman filter interpolation based on the OWC gas chamber lumped parameter model is used, and the state estimate is continuously corrected using the measurement value of the other sensor that is not missing, generating the interpolated value for the missing part; if both liquid level and gas pressure data are missing, open-loop propagation based on the physical model of the most recently known state is used, with the state estimate at the start of the missing time as the initial value, and the state is recursively calculated according to the dynamic equation of the lumped parameter model to generate the state trajectory of the missing period as the interpolated value.

8. The method according to claim 1, characterized in that, In step S5, the deep learning prediction model adopts a hybrid architecture of convolutional neural network and long short-term memory network; the coupling regularization term is defined as the square of the deviation between the time-delay correlation coefficient between the liquid level change rate sequence and the gas pressure change rate sequence output by the model and the pre-statistical historical time-delay correlation coefficient.

9. The method according to claim 4, characterized in that, The method further includes the following steps: S6: When new sensor data arrives, slide the prediction window to update the input sequence with the latest data, and repeat step S5. S7: Perform a coupling consistency check on the liquid level prediction value and gas pressure prediction value output in step S5, and adaptively adjust the subsequent prediction step size based on the check result. Specifically, the coupling consistency verification involves: calculating the real-time correlation coefficient between the predicted liquid level change rate and the pressure change rate at multiple times corresponding to the time delay parameter; if the deviation between the real-time correlation coefficient and the historically statistical time delay correlation coefficient exceeds two standard deviations, a verification failure event is recorded; when the number of consecutive verification failures reaches 3, the prediction step size is automatically reduced to half of the original size; the count of verification failures is automatically cleared after any verification passes, and the original prediction step size is restored.

10. A deep learning-based prediction system for liquid level fluctuation and gas pressure coupling in an OWC device's gas chamber, characterized in that, For performing the method according to any one of claims 1 to 9, comprising: The data acquisition module is used to receive real-time data from the liquid level sensor and the air pressure sensor in the gas chamber of the OWC device. The historical pattern library module is used to store and manage liquid level-gas pressure sub-sequence pairs constructed from complete historical data and their corresponding acceleration indexes; The missing data handling and matching module is used to detect missing data, construct query sequences, perform joint similarity search and coupling constraint filtering, and trigger rollback strategies. The sequence reconstruction module is used to generate interpolation values ​​for missing parts based on matching results or backtracking strategies, and then concatenate the interpolation values ​​with known data to reconstruct complete liquid level and gas pressure sequences. The coupled prediction module has a built-in deep learning model with coupling regularization terms, which is used to output liquid level and gas pressure prediction values ​​for multiple future moments based on the complete historical sequence of the input. The rolling update and verification module is used to slide the prediction window, perform coupling consistency verification on the prediction results, and adaptively adjust the prediction step size according to the verification results, and finally output the liquid level prediction value and gas pressure prediction value at multiple future times.

Images