Ocean multi-parameter precision measuring instrument wake-up signal intelligent identification method

Abstract

The application provides a marine multi-parameter precision measuring instrument wake-up signal intelligent identification method, and belongs to the technical field of marine multi-parameter precision measuring instrument wake-up signal intelligent identification. The original underwater acoustic sampling sequence collected by the underwater acoustic sensor is subjected to polarization noise adaptive cancellation based on the equivalent circuit parameters of a platinum black electrode to obtain a denoised underwater acoustic sequence. Meanwhile, the frame structure parameters are dynamically adjusted according to the channel quality comprehensive evaluation value. The shore-based server uses an asynchronous reasoning framework and a dynamic batch size adjustment strategy to summarize multi-node wake-up events, and periodically performs incremental fine-tuning on the artificial intelligence model with newly added labeled data, thereby solving the technical problem of low recognition accuracy of underwater acoustic wake-up signals in a low signal-to-noise ratio and strong multipath interference environment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recognition technology of wake-up signals for marine multi-parameter precision measuring instruments. Specifically, it relates to an intelligent recognition method for wake-up signals of marine multi-parameter precision measuring instruments. Background Technology

[0002] Marine multi-parameter precision measuring instruments rely on underwater acoustic wake-up signals to activate remote nodes, which is a core control method for underwater networked observation systems. Traditional wake-up signal identification methods typically employ single-domain processing frameworks based on matched filtering or energy detection, exhibiting low computational complexity under favorable channel conditions. Currently, in shallow-water acoustic communication scenarios, these methods have been widely deployed on marine observation buoys and seabed mooring nodes, using preset thresholds to determine the received signal.

[0003] However, traditional single-domain processing frameworks have significant drawbacks. The electrochemical polarization noise of the platinum black electrode highly overlaps with the wake-up signal frequency band within the sensor's operating frequency band, making it difficult for a single matched filter to effectively separate them; the multipath delay spread in shallow seas can reach hundreds of milliseconds, and fixed-frame structures cannot simultaneously achieve inter-symbol interference suppression and spectral efficiency; energy detection methods are extremely sensitive to noise power estimation errors under low signal-to-noise ratio conditions, leading to a significant increase in the false negative rate.

[0004] In existing technologies, due to the combined effects of polarization noise contamination, strong multipath spread, and time-varying Doppler effects on the underwater acoustic channel, single-domain feature extraction methods cannot simultaneously cover temporal causal local patterns and frequency-domain structured features. Existing methods struggle to adaptively allocate feature fusion weights under different channel conditions, resulting in low recognition accuracy of wake-up signals in environments with low signal-to-noise ratios and strong multipath interference. In other words, existing technologies suffer from the technical problem of low recognition accuracy of underwater acoustic wake-up signals in environments with low signal-to-noise ratios and strong multipath interference. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent recognition method for wake-up signals of marine multi-parameter precision measuring instruments, which can solve the technical problem of low recognition accuracy of underwater acoustic wake-up signals in environments with low signal-to-noise ratio and strong multipath interference in the prior art.

[0006] This invention is implemented as follows: This invention provides a method for intelligent identification of wake-up signals of a marine multi-parameter precision measuring instrument, comprising the following steps:

[0007] The equivalent circuit parameters of the platinum black electrode are identified online for the original underwater acoustic sampling sequence collected by the underwater acoustic sensor, and the polarization noise prediction waveform is generated. The negative noise prediction waveform is then extracted from the original underwater acoustic sampling sequence to obtain the denoised underwater acoustic sequence.

[0008] Sparse channel estimation is performed on the denoised underwater acoustic sequence to obtain multipath delay estimates and multipath complex amplitude estimates. An adaptive RAKE receiver is constructed to coherently superimpose each multipath component after delay compensation. The wake-up frame preprocessing is completed by combining the adaptive orthogonal frequency division multiplexing spread spectrum frame structure to obtain the preprocessed wake-up frame sequence.

[0009] The short-time Fourier transform time-frequency matrix and instantaneous frequency estimate are extracted from the preprocessed wake-up frame sequence. The real-time signal-to-noise ratio estimate is obtained simultaneously. The original underwater acoustic sampling sequence, short-time Fourier transform time-frequency matrix, instantaneous frequency estimate and real-time signal-to-noise ratio estimate are input into the physical prior constraint time-frequency dual-stream residual attention network, and the wake-up event recognition result and wake-up frequency estimate are output.

[0010] Based on the real-time signal-to-noise ratio estimate, multipath delay spread estimate, and Doppler frequency shift estimate, the channel quality comprehensive evaluation value is calculated through the adaptive frame structure adjustment function. Based on the interval to which the channel quality comprehensive evaluation value belongs, the subcarrier spacing and cyclic prefix length of the adaptive orthogonal frequency division multiplexing spread spectrum frame structure are dynamically adjusted.

[0011] The shore-based server uses an asynchronous inference framework to process wake-up event recognition results from multiple nodes in parallel. Data from nodes that have not completed uploading within a timeout period is filled with predicted values ​​of historical features by a physical prior constraint time-frequency dual-stream residual attention network. The batch inference scale of the graphics processor is adjusted through a dynamic batch size adjustment strategy, and the summary results of wake-up events from each node are output.

[0012] The wake-up event recognition results and historical wake-up event labels are stored in the event database. The newly labeled data in the event database are used periodically to perform incremental fine-tuning on the physical prior constraint time-frequency dual-stream residual attention network, and update the model parameters of the physical prior constraint time-frequency dual-stream residual attention network.

[0013] The step of identifying the equivalent circuit parameters of the platinum black electrode online further includes performing thermodynamic correction on the equivalent circuit parameters based on the Arrhenius temperature compensation equation.

[0014] The equivalent circuit parameters include double-layer capacitance, charge transfer resistance, and Warburg coefficient, which are obtained through online identification using a broadband excitation signal.

[0015] Specifically, the process of predicting the negative noise waveform from the original underwater acoustic sampling sequence employs a minimum mean square adaptive filter, using the polarization noise prediction waveform as a reference input and the original underwater acoustic sampling sequence as the desired response, to iteratively update the filter weight vector.

[0016] The sparse channel estimation employs an orthogonal matching pursuit algorithm to iteratively search for significant multipath components in the time delay-amplitude space, and outputs multipath time delay estimates and multipath complex amplitude estimates.

[0017] The physical prior-constrained time-frequency dual-stream residual attention network adopts a dual-stream parallel input structure. The time-domain stream extracts multi-scale temporal local features through one-dimensional dilated causal convolutional blocks, and the frequency-domain stream extracts joint time-frequency features through two-dimensional residual convolutional blocks.

[0018] The time-domain stream and frequency-domain stream are fused through a cross-modal attention gating mechanism, and the attention weight vector is dynamically modulated by an affine transformation based on a real-time signal-to-noise ratio estimate.

[0019] The physical prior constraint time-frequency dual-stream residual attention network has an additional physical consistency constraint loss term at the top layer, which forces the estimated wake-up frequency to fall within the known frequency offset range of the crystal oscillator. The loss function is composed of a weighted sum of the cross-entropy classification loss and the physical consistency constraint loss term.

[0020] The adaptive frame structure adjustment function calculates the comprehensive channel quality assessment value and dynamically configures the subcarrier spacing and cyclic prefix length in three levels based on the relationship between the comprehensive channel quality assessment value and the high quality threshold and the low quality threshold.

[0021] The multipath delay spread estimate is calculated from the difference between the maximum and minimum values ​​of the multipath delay estimates, and the Doppler frequency shift estimate is obtained by frequency tracking processing of the denoised underwater acoustic sequence.

[0022] The dynamic batch size adjustment strategy calculates the optimal batch size in real time based on the inference queue length and the idle computing power of the graphics processor, increasing the batch size when the queue is crowded and decreasing the batch size when the queue is sparse.

[0023] The training dataset for the physical prior-constrained time-frequency dual-stream residual attention network is constructed by collecting raw underwater acoustic sampling sequences by deploying standard sound sources in a laboratory pool and shallow sea area, and then performing time delay offset, amplitude scaling, additive white Gaussian noise injection, and random multipath stacking data augmentation processing.

[0024] The training of the physical prior constraint time-frequency dual-stream residual attention network uses the Adam optimizer. After training, the network parameters are quantized with INT8 integers and then combined with the CMSIS-NN library to call single instruction multiple data instruction parallel computation and deployed to an embedded microcontroller.

[0025] In the asynchronous inference framework, feature uploading and batch inference scheduling of each node are decoupled. After a node completes feature uploading, it immediately enters the inference queue without waiting for other nodes.

[0026] In the adaptive frame structure adjustment function, the high-quality threshold for the comprehensive channel quality assessment value is 0.75, and the low-quality threshold is 0.40. When the comprehensive channel quality assessment value is greater than or equal to the high-quality threshold, a small subcarrier spacing and a short cyclic prefix are used; when the comprehensive channel quality assessment value is between the low-quality and high-quality thresholds, a medium configuration is used; and when the comprehensive channel quality assessment value is less than the low-quality threshold, a large subcarrier spacing and a long cyclic prefix are used. The weighting coefficient of the loss function for the physical prior-constrained time-frequency dual-stream residual attention network is in the range of 0.05 to 0.30, and the initial learning rate is... to Within the specified range, batch size is between 32 and 128, the total number of training epochs is no less than 100, and the training dataset size is no less than [missing information]. The sample size is 1, and the ratio of positive to negative samples is between 1:1 and 1:3.

[0027] This invention solves the technical problem of low recognition accuracy of underwater acoustic wake-up signals in environments with low signal-to-noise ratio and strong multipath interference by constructing a time-frequency dual-stream residual attention network based on physical prior constraints and combining adaptive polarization noise cancellation, sparse channel estimation and adaptive frame structure adjustment.

[0028] This invention utilizes a parameterized physical model of the Warburg impedance of platinum black electrodes to generate a polarization noise prediction waveform and adaptively cancel it, suppressing the polarization noise floor that overlaps with the wake-up frequency band at the signal source; sparse channel estimation only estimates the parameters of a few multipath components with concentrated energy, reducing the estimation variance, enabling the adaptive RAKE receiver to effectively aggregate multipath energy; the physical prior-constrained time-frequency dual-stream residual attention network uses a cross-modal attention gating mechanism to dynamically modulate the fusion weights of the time-domain and frequency-domain streams by the real-time signal-to-noise ratio estimate, adaptively selecting the most discriminative feature domain under different channel conditions.

[0029] In summary, the present invention solves the technical problem mentioned in the background art of low recognition accuracy of underwater acoustic wake-up signals in environments with low signal-to-noise ratio and strong multipath interference. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 The graph shows the weight variation curves of the time-domain and frequency-domain flow fusion for the cross-modal attention gating mechanism under different signal-to-noise ratio conditions. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1The diagram shown is a flowchart of an intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument provided by the present invention. This method includes the following steps:

[0034] S01. For the original underwater acoustic sampling sequence acquired by the underwater acoustic sensor, identify the equivalent circuit parameters of the platinum black electrode online, generate the polarization noise prediction waveform, and obtain the denoised underwater acoustic sequence from the negative noise prediction waveform in the original underwater acoustic sampling sequence.

[0035] S02. Perform sparse channel estimation on the denoised underwater acoustic sequence to obtain multipath delay estimates and multipath complex amplitude estimates. Construct an adaptive RAKE receiver to coherently superimpose each multipath component after delay compensation. Combine the adaptive orthogonal frequency division multiplexing spread spectrum frame structure to complete the wake-up frame preprocessing and obtain the preprocessed wake-up frame sequence.

[0036] S03. Extract the short-time Fourier transform time-frequency matrix and instantaneous frequency estimate from the preprocessed wake-up frame sequence, and simultaneously obtain the real-time signal-to-noise ratio estimate. Input the original underwater acoustic sampling sequence, short-time Fourier transform time-frequency matrix, instantaneous frequency estimate and real-time signal-to-noise ratio estimate into the physical prior constraint time-frequency dual-stream residual attention network, and output the wake-up event recognition result and wake-up frequency estimate.

[0037] S04. Based on the real-time signal-to-noise ratio estimate, multipath delay spread estimate, and Doppler frequency shift estimate, calculate the comprehensive channel quality assessment value through the adaptive frame structure adjustment function. Based on the interval to which the comprehensive channel quality assessment value belongs, dynamically adjust the subcarrier spacing and cyclic prefix length of the adaptive orthogonal frequency division multiplexing spread spectrum frame structure.

[0038] S05. The shore-based server uses an asynchronous inference framework to process the wake-up event recognition results from multiple nodes in parallel. The data of nodes that have not completed uploading within the timeout period is filled with the predicted values ​​of historical features by the physical prior constraint time-frequency dual-stream residual attention network. The batch inference scale of the graphics processor is adjusted through a dynamic batch size adjustment strategy, and the summary results of wake-up events of each node are output.

[0039] S06. Store the wake-up event recognition results and historical wake-up event labels into the event database. Periodically use the newly added labeled data in the event database to incrementally fine-tune the physical prior constraint time-frequency dual-stream residual attention network and update the model parameters of the physical prior constraint time-frequency dual-stream residual attention network.

[0040] Step S01 further includes a step of thermodynamically correcting the equivalent circuit parameters according to the Arrhenius temperature compensation equation. The method used to predict the negativeed noise waveform from the original underwater acoustic sampling sequence is a minimum mean square adaptive filter.

[0041] The equivalent circuit parameters include double-layer capacitance, charge transfer resistance, and Warburg coefficient, which are obtained online through broadband excitation signal identification. Within each sampling period, based on the real-time electrode temperature and salinity measured by the underwater acoustic sensor, the double-layer capacitance, charge transfer resistance, and Warburg coefficient are thermodynamically corrected according to the Arrhenius temperature compensation equation to obtain the corrected equivalent circuit parameters, which are then used to construct the polarization noise prediction waveform. The least mean square adaptive filter uses the polarization noise prediction waveform as a reference input and the original underwater acoustic sampling sequence as the desired response, iteratively updating the filter weight vector to handle negative noise. After predicting the waveform, a denoised underwater acoustic sequence is output. The technical effect of the Warburg impedance physical model adaptive polarization noise cancellation algorithm is as follows: the Warburg impedance of the platinum black electrode satisfies a half-power law frequency dependence within the operating frequency band of the electrochemical sensor. The polarization noise prediction waveform is generated using a parameterized physical model and removed through adaptive cancellation, effectively suppressing the polarization noise floor noise within the wake-up signal frequency band. The method constructs a noise model based on the physical mechanism of electrode polarization. Compared with general filtering methods, it can accurately track the frequency structure changes of polarization noise caused by temperature and salinity changes, fundamentally reducing the probability of missed and false detections of weak wake-up frames.

[0042] The sparse channel estimation algorithm utilizes the sparsity of the impulse response of the underwater acoustic multipath channel in the time delay domain. It employs an orthogonal matched pursuit algorithm to iteratively search for significant multipath components in the time delay-amplitude space, outputting multipath delay estimates and multipath complex amplitude estimates. The adaptive RAKE receiver uses the multipath delay estimates and multipath complex amplitude estimates as parameters to coherently superimpose each multipath component after time delay compensation. The technical advantages of the sparse channel estimation algorithm are: the number of resolvable multipaths in the underwater acoustic multipath channel is much smaller than the channel impulse response length; sparse estimation only estimates parameters for a few time delay taps with concentrated energy, significantly reducing the estimation variance. This enables the adaptive RAKE receiver to effectively aggregate multipath energy in shallow sea scenarios where multipath delays extend to hundreds of milliseconds, improving the signal-to-noise ratio of the wake-up frame matched filter output, thereby keeping the wake-up frame false negative rate within an acceptable range.

[0043] The adaptive orthogonal frequency division multiplexing spread spectrum frame structure refers to the wake-up frame format that maps the spread spectrum sequence to multiple subcarriers in orthogonal frequency division multiplexing. The subcarrier spacing and cyclic prefix length are dynamically configured based on the comprehensive channel quality assessment value, so that the cyclic prefix length always covers the multipath delay spread estimate, thereby suppressing inter-symbol interference and inter-carrier interference.

[0044] The specific structure of the physical prior-constrained time-frequency dual-stream residual attention network is as follows: The network adopts a dual-stream parallel input structure. The time-domain stream takes the original underwater acoustic sampling sequence as input and extracts multi-scale temporal local features through 5 layers of one-dimensional dilated causal convolutional blocks. The dilation rates are 1, 2, 4, 8, and 16, respectively. Each layer contains a batch normalization layer and a linear rectified activation function. Causality ensures that the network does not introduce future time information. The frequency-domain stream takes the short-time Fourier transform time-frequency matrix as input and extracts joint time-frequency features through 4 sets of two-dimensional residual convolutional blocks. Each set of residual convolutional blocks contains two 3×3 convolutional layers and skip connections. The time-domain stream and the frequency-domain stream are fused in the intermediate layer through a cross-modal attention gating mechanism. The cross-modal attention gating mechanism concatenates the intermediate features of the time-domain stream and the intermediate features of the frequency-domain stream and then generates an attention weight vector through a linear transformation. The attention weight vector is dynamically modulated by an affine transformation based on a real-time signal-to-noise ratio estimate, enabling the network to automatically enhance its reliance on time-domain robust features under low signal-to-noise ratio conditions. The fused features are passed through two fully connected layers and a Softmax classification head to output wake-up event recognition results and wake-up frequency estimates. A physical consistency constraint loss term is added to the top layer of the network to force the wake-up frequency estimate to fall within the known frequency offset range of the crystal oscillator, and predictions that violate the physical constraint are suppressed by the gradient penalty term. The network internally sets up skip connections from the shallow layers of the time-domain flow directly to the Softmax classification head to avoid excessive abstraction of deep features and loss of fine frequency information. The loss function is composed of a weighted sum of cross-entropy classification loss and physical consistency constraint loss term, and the weighting coefficients are determined in the range of 0.05 to 0.30 through multiple rounds of ablation experiments on the validation set.

[0045] The specific steps for establishing the training dataset for the physical prior-constrained time-frequency dual-stream residual attention network include: deploying standard sound sources in a laboratory pool and typical shallow sea areas; transmitting reference wake-up frames using an adaptive orthogonal frequency division multiplexing spread spectrum frame structure; synchronously acquiring original underwater acoustic sampling sequences and recording corresponding historical wake-up event tags; extracting the short-time Fourier transform time-frequency matrix and instantaneous frequency estimate for each original underwater acoustic sampling sequence, and synchronously recording the real-time signal-to-noise ratio estimate; combining the four data items into multimodal samples; performing time delay offsetting, amplitude scaling, additive white Gaussian noise injection, and random multipath superposition data enhancement processing on the multimodal samples to ensure that the samples cover a dynamic range with a signal-to-noise ratio of 0–30 dB; maintaining the ratio of positive samples to negative samples between 1:1 and 1:3, where the ratio range is determined by testing on the validation set. Scores were obtained through multiple rounds of experiments; the training dataset size was no less than [number missing]. Sample.

[0046] The specific steps for training the physical prior-constrained time-frequency dual-stream residual attention network include: using the Adam optimizer, with an initial learning rate set to... to The initial learning rate range was determined through a learning rate warm-up scan experiment; the batch size was set to 32–128, and the batch size range was determined through a multi-round comparison experiment of memory capacity and convergence stability; the total number of training rounds was no less than 100, and an early stopping strategy was adopted, terminating training when the validation set loss did not decrease for 20 consecutive rounds; after training, the network parameters were quantized using INT8 integers, and the quantization perception training rounds were no less than 10. Combined with the CMSIS-NN library, single instruction multiple data instruction parallel computation was called to ensure that the inference latency met the real-time wake-up response requirements of the embedded microcontroller; finally, after verifying that the wake-up event recognition results met the standards on an independent test set, the quantized model parameters were deployed to the embedded microcontroller.

[0047] The technical effects of the physical prior-constrained time-frequency dual-stream residual attention network are as follows: the dual-stream structure enables the network to simultaneously extract complementary information from the causal local patterns of the time-domain waveform and the structured features of the short-time Fourier transform time-frequency matrix; the cross-modal attention gating mechanism, through dynamic modulation of the real-time signal-to-noise ratio estimate, enables the network to adaptively allocate the fusion weights of the time-domain and frequency-domain features under different channel conditions, avoiding the performance degradation of single-domain features under extreme channel conditions; the physical consistency constraint loss term embeds the prior physical knowledge of the wake-up frequency into the training process, constraining the estimated wake-up frequency value to fall within a physically reasonable space, and suppressing false alarms caused by overfitting noise patterns; the shallow skip connections preserve fine frequency information, preventing the loss of narrowband frequency details that are crucial for wake-up frame recognition in deep abstract representations, and improving the overall robustness of the recognition method in low signal-to-noise ratio and strong multipath interference environments.

[0048] The adaptive frame structure adjustment function calculates the comprehensive channel quality assessment value based on the real-time signal-to-noise ratio estimate, the multipath delay spread estimate, and the Doppler frequency shift estimate. The formula is as follows: ;in This is a real-time estimate of the signal-to-noise ratio, in dB. This is an estimate of multipath delay spread, in milliseconds. This is an estimate of the Doppler frequency shift, in Hz. , , These are the standard reference values ​​for signal-to-noise ratio, multipath delay spread, and Doppler frequency shift, respectively. , , The weighting coefficients and The weighting coefficients are obtained through at least five rounds of orthogonal experimental design iterations under multiple typical channel conditions; when the comprehensive channel quality evaluation value... When using a smaller subcarrier spacing and a shorter cyclic prefix length, the overall channel quality assessment value is [value missing]. At that time, a medium subcarrier spacing and a medium cyclic prefix length configuration are adopted; when the overall channel quality assessment value is... At that time, a larger subcarrier spacing and a longer cyclic prefix length were configured; the interval boundary values ​​of 0.40 and 0.75 were obtained through multiple measurements in the experimental pool and typical shallow sea area, with the comprehensive index of wake-up event recognition result compliance rate and frame efficiency as the optimization target; the adjusted subcarrier spacing and cyclic prefix length were returned as configuration parameters to S02 to update the adaptive orthogonal frequency division multiplexing spread spectrum frame structure.

[0049] Among them, the multipath delay spread estimate is calculated by the difference between the maximum and minimum values ​​of the multipath delay estimates output by the sparse channel estimation algorithm; the Doppler frequency shift estimate is obtained by frequency tracking processing of the denoised underwater acoustic sequence; the above two parameters, together with the real-time signal-to-noise ratio estimate, serve as the input to the adaptive frame structure adjustment function.

[0050] The dynamic batch size adjustment strategy refers to a scheduling method that calculates the optimal batch size in real time based on the current inference queue length and the idle computing power of the graphics processor. When the queue is backlogged, the batch size is increased to improve throughput, and when the queue is sparse, the batch size is reduced to reduce latency. The inference queue length and the idle computing power of the graphics processor required for the optimal batch size calculation are monitored and obtained in real time by the task scheduling module of the shore-based server.

[0051] The Arrhenius temperature compensation equation is used to describe the exponential relationship between the electrochemical reaction rate and temperature. In this scheme, the real-time measured values ​​of electrode temperature and salinity are mapped to the correction values ​​of double-layer capacitance, charge transfer resistance and Warburg coefficient, thereby dynamically compensating for the parameter drift of the polarization noise prediction waveform.

[0052] One-dimensional dilated causal convolution refers to a convolution operation in which the kernel samples the input sequence at exponentially increasing intervals along the time axis and relies only on data from the current and historical moments. The dilation rate controls the size of the receptive field, and causality ensures that information from future moments is not introduced during real-time inference.

[0053] Among them, the cross-modal attention gating mechanism refers to the mechanism of concatenating intermediate features of the time-domain stream and intermediate features of the frequency-domain stream, generating an element-wise attention weight vector through linear transformation, and then using a real-time signal-to-noise ratio estimate to perform affine modulation on the attention weight vector, thereby dynamically controlling the contribution ratio of the time-domain stream and the frequency-domain stream during the feature fusion stage.

[0054] The asynchronous inference framework refers to an architecture in which feature uploading and batch inference scheduling of each node are decoupled, and a node immediately enters the inference queue after completing feature uploading without waiting for concurrent processing of other nodes.

[0055] Optionally, the present invention also provides a computer-based approach to form an intelligent recognition system for wake-up signals of a marine multi-parameter precision measuring instrument. The computer is equipped with a readable storage medium that stores program instructions, which can execute the above-described method when the computer is run.

[0056] The specific implementation of step S01 is as follows: The underwater acoustic sensor continuously acquires the original underwater acoustic sampling sequence, simultaneously measuring electrode temperature and salinity. Within each sampling period, a platinum black electrode is excited with a broadband excitation signal, and three equivalent circuit parameters—double-layer capacitance, charge transfer resistance, and Warburg coefficient—are extracted using a frequency domain identification algorithm. Since the Warburg impedance of the platinum black electrode satisfies a half-power-law frequency dependence within the operating frequency band of the electrochemical sensor, the above three parameters are sufficient to parameterize the frequency structure of the polarization noise. Subsequently, based on the Arrhenius temperature compensation equation, the real-time measured electrode temperature and salinity are mapped to correction values ​​for the three parameters, obtaining the corrected equivalent circuit parameters, and thus constructing the polarization noise prediction waveform. The least mean square adaptive filter uses the polarization noise prediction waveform as a reference input and the original underwater acoustic sampling sequence as the desired response, iteratively updating the filter weight vector according to the least mean square criterion, and outputting a denoised underwater acoustic sequence after negativeing ​​the noise prediction waveform.

[0057] The specific implementation of step S02 is as follows. Sparse channel estimation is performed on the denoised underwater acoustic sequence. The core basis is the sparsity of the underwater acoustic multipath channel impulse response in the time delay domain, meaning the number of resolvable multipaths is much smaller than the total length of the channel impulse response. The orthogonal matched pursuit algorithm iterates step-by-step in the time delay-amplitude space along the direction of maximum residual projection. In each iteration, the time delay tap most relevant to the current residual is selected and included in the support set until the stopping criterion is met. Finally, the multipath time delay estimate and the multipath complex amplitude estimate are output. The adaptive RAKE receiver uses the above two sets of estimates as parameters to apply corresponding time delay compensation to each multipath component, so that the components are aligned on the time axis and coherently superimposed, thereby aggregating multipath energy. The difference between the maximum and minimum values ​​in the multipath time delay estimates is the multipath time delay spread estimate, which is then passed to step S04 for frame structure adjustment. The aligned superimposed sequence, combined with the adaptive orthogonal frequency division multiplexing spread spectrum frame structure, completes wake-up frame preprocessing according to the frame boundaries, and outputs the preprocessed wake-up frame sequence.

[0058] The specific implementation of step S03 is as follows: A short-time Fourier transform time-frequency matrix is ​​calculated for the preprocessed wake-up frame sequence. The window length and step size are configured based on the center frequency of the wake-up frame and the subcarrier spacing to achieve a balance between time resolution and frequency resolution. Simultaneously, instantaneous frequency estimates are extracted from the preprocessed wake-up frame sequence, achieved through a finite-difference approximation of the phase of the analytical signal. The real-time signal-to-noise ratio estimate is obtained from the power spectrum estimation of the preprocessed wake-up frame sequence. The physical prior-constrained time-frequency dual-stream residual attention network takes the original underwater acoustic sampling sequence as input in the temporal stream and extracts multi-scale temporal local features through five layers of one-dimensional dilated causal convolutional blocks with dilation rates of 1, 2, 4, 8, and 16. Each layer contains a batch normalization layer and a linear rectified activation function, and causality ensures that the network does not introduce future time information. The frequency stream takes the short-time Fourier transform time-frequency matrix as input and extracts joint time-frequency features through four sets of two-dimensional residual convolutional blocks. Each set contains two 3×3 convolutional layers and skip connections. The two features are fused in the middle layer through a cross-modal attention gating mechanism, and the attention weight vector is dynamically modulated by an affine transformation based on the real-time signal-to-noise ratio estimate. After fusion, the features are passed through two fully connected layers and a Softmax classification head to output the wake-up event recognition result and the wake-up frequency estimate. The physical consistency constraint loss term at the top layer of the network constrains the wake-up frequency estimate within the known frequency offset range of the crystal oscillator, and predictions that violate the constraint are suppressed by the gradient penalty term. The network internally sets skip connections from the shallow layers of the temporal stream directly to the Softmax classification head to preserve fine frequency information.

[0059] The specific implementation of step S04 is as follows. The adaptive frame structure adjustment function takes the real-time signal-to-noise ratio estimate, multipath delay spread estimate, and Doppler frequency shift estimate as inputs, and calculates them according to the weighted summation formula. Calculate the overall channel quality assessment value ,in The weighting coefficients were obtained through iterative orthogonal experimental design. When using a smaller subcarrier spacing and a shorter cyclic prefix length; when At that time, a medium configuration is used; when At this time, a larger subcarrier spacing and a longer cyclic prefix length are used to ensure that the cyclic prefix length always covers the multipath delay spread estimate, thereby suppressing inter-symbol interference and inter-carrier interference. The adjusted configuration parameters are fed back to step S02 to update the adaptive orthogonal frequency division multiplexing spread spectrum frame structure.

[0060] The specific implementation of step S05 is as follows. The asynchronous inference framework of the shore-based server decouples the feature uploading of each node from the batch inference scheduling. After a node completes feature uploading, it immediately enters the inference queue without waiting for other nodes, thereby achieving parallel processing. For nodes that have timed out and have not completed uploading, the predicted values ​​of the node's historical features are filled using a physical prior-constrained time-frequency dual-stream residual attention network to ensure the integrity of batch inference. The dynamic batch size adjustment strategy is implemented by the task scheduling module of the shore-based server, which monitors the inference queue length and the idle computing power of the graphics processor in real time. When the queue is backlogged, the batch size is increased to improve throughput; when the queue is sparse, the batch size is decreased to reduce latency. Finally, the wake-up event results of each node are aggregated and output.

[0061] The specific implementation of step S06 is as follows. After each recognition is completed, the wake-up event recognition result and the corresponding historical wake-up event label are written into the event database. Newly labeled data are periodically extracted from the event database, and the model parameters of the physical prior-constrained time-frequency dual-stream residual attention network are updated in an incremental fine-tuning manner. The fine-tuning process uses the same Adam optimizer and early stopping strategy as the initial training, so that the model can continuously adapt to the changes in channel distribution in the deployment environment.

[0062] It should be noted that the key technologies of this invention include: an adaptive polarization noise cancellation technology based on the Warburg impedance parameterization physical model, which utilizes the physical prior that the platinum black electrode satisfies the half-power law frequency dependence relationship in the operating frequency band to accurately model the frequency structure drift of polarization noise caused by temperature and salinity changes. Compared with general adaptive filtering methods, it can maintain cancellation accuracy when the noise frequency structure undergoes non-stationary drift, effectively reducing the risk of missed detection of weak wake-up frames; a sparse channel estimation technology based on orthogonal matched pursuit, which fully utilizes the sparsity of underwater acoustic multipath channels in the time delay domain, concentrates parameter estimation on a few energy-significant taps, significantly compresses the estimation variance, and makes the coherent superposition gain of the RAKE receiver approach the theoretical upper limit; and a cross-modal attention gating mechanism based on dynamic modulation of the real-time signal-to-noise ratio estimate, which enables the network to adaptively allocate the fusion weights of the time domain stream and the frequency domain stream under different channel conditions, avoiding the performance degradation of single-domain features under extreme channel conditions. The three technologies work together to form a complete processing chain from signal source noise reduction and channel equalization gain aggregation to feature adaptive fusion, so that the overall recognition accuracy remains stable even under the harsh conditions of polarization noise, multipath spread and low signal-to-noise ratio superposition.

[0063] It should be noted that this invention also solves the following technical problem: in multi-node concurrent reporting scenarios, the shore-based server experiences excessively long batch inference waiting times and reduced system throughput due to inconsistent node upload latency. This invention employs an asynchronous inference framework to decouple feature uploads from batch inference scheduling. After a node completes its upload, it is immediately enqueued, without blocking the inference process of other nodes. Timed-out nodes are filled with historical feature prediction values ​​to ensure batch inference integrity. A dynamic batch size adjustment strategy calculates the optimal batch size in real time based on the inference queue length and the idle computing power of the graphics processor. When the queue is congested, the batch size is increased to improve throughput; when the queue is sparse, the batch size is decreased to reduce latency. This mechanism eliminates inference waiting caused by asynchronous node reporting at the scheduling level, enabling the shore-based server to maintain stable batch inference throughput even as the number of nodes increases, effectively solving the technical problem of reduced system throughput in multi-node concurrent scenarios.

[0064] Specifically, the principle of this invention is as follows: The core reason why this invention can solve the above-mentioned technical problems is that the physical modeling of the three aspects—polarization noise, multipath spread, and feature fusion—is highly coupled with the underwater acoustic propagation mechanism. The Warburg impedance satisfies a half-power-law frequency dependence, and the parameterized model can accurately track the frequency structure drift of polarization noise caused by temperature and salinity changes. Based on this, the least mean square adaptive filter achieves noise cancellation, avoiding sideband damage to the wake-up signal caused by the lack of physical constraints in general filtering methods. The impulse response of the underwater acoustic multipath channel is sparsity in the time delay domain. The orthogonal matched pursuit algorithm utilizes this prior to concentrate parameter estimation on a few significant taps, significantly reducing the estimation variance, thereby bringing the coherent superposition gain of the RAKE receiver close to the theoretical upper limit. The time-frequency dual-stream network integrates the temporal local patterns extracted by one-dimensional dilated causal convolution with the short-time Fourier transform time-frequency structural features extracted by two-dimensional residual convolution. The cross-modal attention gating mechanism uses affine modulation of the attention weights by the real-time signal-to-noise ratio estimate, which enables the network to automatically enhance its dependence on the time-domain robust features when the signal-to-noise ratio is low. The physical consistency constraint loss term further constrains the wake-up frequency estimate within the known frequency offset range of the crystal oscillator, suppressing false alarms caused by noise overfitting. The overall logic is self-consistent and conforms to the laws of underwater acoustics.

[0065] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0066] The specific implementation of step S01 is as follows: the underwater acoustic sensor acquires the original underwater acoustic sampling sequence. ,in This serves as an index for sampling times. Within each sampling period, the equivalent circuit parameters of the platinum black electrode, including the double-layer capacitance, are identified online using the broadband excitation signal. (Unit: F) Charge transfer resistance (Unit: Ω) and Warburg coefficient (unit Based on the Arrhenius temperature compensation equation, thermodynamic corrections are made to the above three parameters to obtain the corrected double-layer capacitance. (Unit: F) Corrected charge transfer resistance (Unit: Ω) and the corrected Warburg coefficient (unit The three correction formulas are expressed as follows:

[0067] A method for intelligent recognition of wake-up signals of a marine multi-parameter precision measuring instrument

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, The activation energy (in J) corresponds to the charge transfer resistance. This represents the activation energy (in J) corresponding to a double-layer capacitor. The activation energy (in J) corresponds to the Warburg coefficient. Boltzmann constant (values) J / K), The real-time temperature of the electrode (in K). This is a reference temperature (unit: K, usually taken as 298K). The measured salinity value is in per mille (‰). This is the salinity equivalent temperature correction factor (unit: K / ‰, empirical value: 0.04 K / ‰), exponential term. Since it is a dimensionless quantity, the dimensions of the right and left sides of the three equations are the same. Based on the corrected equivalent circuit parameters, the Warburg impedance polarization noise prediction waveform is constructed. Its frequency domain expression is:

[0072] ;

[0073] In the formula, Frequency domain representation (in V) of the polarization noise prediction waveform. Frequency (unit: Hz) The imaginary unit, This is the frequency domain representation (in meters) of the broadband excitation signal. The unit is , The unit is , The unit is Ω, and The unit of the product of (unit A) is V, and Dimensionally consistent. The least mean square adaptive filter uses... For reference input, with To obtain the desired response, iteratively update the filter weight vector. ,in Given the filter order (empirical values ​​range from 32 to 256), the weight update formula is:

[0074] ;

[0075] In the formula, The reference input vector (in units V) This is the error signal (unit: V). Step size factor (dimensionless, empirical value is) to ), The weight vector is a dimensionless vector, and both sides of the equation are dimensionless vectors, representing the error signal. That is, the denoised underwater acoustic sequence (Unit: V).

[0076] The specific implementation of step S02 is as follows: For the denoised underwater acoustic sequence... Sparse channel estimation is performed by leveraging the sparsity of the underwater acoustic multipath channel impulse response in the time-delay domain. An orthogonal matched pursuit algorithm is employed to iteratively search for significant multipath components in the time-delay-amplitude space. Let the channel impulse response vector be... ( (Total number of delay taps), received signal vector is ,in The sensing matrix (composed of transmitted pilot signals). For pilot symbols, This is a noise vector. The orthogonal matching pursuit algorithm selects the column index with the largest inner product with the current residual in each iteration, accumulating the support set. The non-zero magnitude vector is updated using least squares estimation, as shown in the formula:

[0077] ;

[0078] In the formula, For the perception matrix Central support set The corresponding column submatrix, For the multipath complex magnitude estimation vector corresponding to the support set, For transpose operation, For matrix inversion, the residual is updated as follows: Iterate to the residual Less than the preset threshold or the number of iterations reaches the upper limit (With an empirical value of 20 to 50), the final output is a set of multipath delay estimates. and multipath complex amplitude estimates set ,in This represents the estimated number of multipath components. The adaptive RAKE receiver coherently superimposes each multipath component after time-delay compensation, and the output signal is described below:

[0079] ;

[0080] In the formula, This is the output of the RAKE receiver (in V). For the first The conjugate (dimensionless) of the multipath complex amplitude estimate. For the first The time delay estimates of the multipath propagation (in units of sampling points) are combined with the adaptive orthogonal frequency division multiplexing spread spectrum frame structure to complete the wake-up frame preprocessing, resulting in a preprocessed wake-up frame sequence. (Unit: V).

[0081] The specific implementation of step S03 is as follows: preprocess the wake-up frame sequence. Extracting the short-time Fourier transform time-frequency matrix ,in Represents the real number field. For frequency bin number, For the time frame number, the matrix number is... The elements are described as follows:

[0082] ;

[0083] In the formula, For the first Index of frequency bins ( ), For time frame indexing, It is a window function (usually the Hanning window). For the index of sampling points within the window, Window length (unit: number of sampling points) Frame shift (unit: number of sampling points) The unit is Synchronously extract instantaneous frequency estimates. (Unit: Hz) and real-time signal-to-noise ratio estimate (Unit: dB). The original underwater acoustic sampling sequence Time-frequency matrix Instantaneous frequency estimate and real-time signal-to-noise ratio estimator Input a time-frequency dual-stream residual attention network with physical prior constraints, and output the wake-up event recognition result. Compared with the estimated wake-up frequency (Unit: Hz). The network loss function consists of cross-entropy classification loss. Loss term related to physical consistency constraints The weighted summation is expressed by the following formula:

[0084] ;

[0085] ;

[0086] In the formula, Total loss (dimensionless). Cross-entropy is the classification loss (dimensionless). The physical consistency constraint loss weight coefficient (dimensionless, ranging from 0.05 to 0.30). This refers to the nominal frequency of the crystal oscillator (in Hz). This represents the allowable frequency deviation range of the crystal oscillator (in Hz). middle and All units are in Hz. After taking the positive value For dimensionless normalized quantities, predictions that violate physical constraints are penalized by gradient penalty terms. suppress.

[0087] The specific implementation of step S04 is: multipath delay spread estimation value It is calculated from the difference between the maximum and minimum values ​​in the set of multipath delay estimates output by the orthogonal matching pursuit algorithm, i.e. (Unit: ms). Doppler frequency shift estimate. Obtained by frequency tracking processing of the denoised underwater acoustic sequence (unit: Hz). Overall channel quality assessment value. The formula is expressed as follows:

[0088] ;

[0089] In the formula, This is a dimensionless comprehensive evaluation value for channel quality. This is a real-time estimate of the signal-to-noise ratio (in dB). This is the standard reference value for signal-to-noise ratio (in dB). This is the estimated multipath delay spread (in milliseconds). This is a standard reference value for multipath delay spread (unit: ms). This is an estimate of the Doppler frequency shift (in Hz). This is the standard reference value for Doppler frequency shift (unit: Hz). , , The coefficients are dimensionless weighted coefficients and satisfy the following conditions: All terms are dimensionless ratios, and the weighting coefficients mentioned above were obtained through at least five rounds of orthogonal experimental design iterations under multiple typical channel conditions. When using a smaller subcarrier spacing and a shorter cyclic prefix length; when When a medium subcarrier spacing and a medium cyclic prefix length are used; when When the subcarrier spacing and cyclic prefix length are used, the adjusted parameters are returned to step S02 to update the adaptive orthogonal frequency division multiplexing spread spectrum frame structure.

[0090] The specific implementation of step S05 is as follows: The shore-based server uses an asynchronous inference framework to process the wake-up event recognition results from multiple nodes in parallel. Each node immediately enters the inference queue after completing feature upload. Data from nodes that fail to upload within the timeout period is filled with the predicted values ​​of historical features from a physical prior-constrained time-frequency dual-stream residual attention network. The dynamic batch size adjustment strategy is based on the inference queue length. With graphics processor idle computing power (Unit: TFLOPS) Calculates optimal batch size in real time Expand when queue backlog occurs To improve throughput, reduce the size when the queue is sparse. To reduce latency, the final output summarizes the wake-up events of each node.

[0091] The specific implementation of step S06 is as follows: The wake-up event recognition result... Tags related to historical awakening events The data is stored in an event database, and the physical prior-constrained time-frequency dual-stream residual attention network is incrementally fine-tuned periodically with newly added labeled data. The Adam optimizer is used, and the initial learning rate ranges from [value missing]. to The batch size ranges from 32 to 128, the total number of training rounds is no less than 100, an early stopping strategy is adopted, and training is terminated when the validation set loss does not decrease for 20 consecutive rounds. After training is completed, the network model parameters are updated and redeployed.

[0092] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians deployed 5 marine multi-parameter precision measurement instrument nodes in a shallow sea test area. The underwater acoustic wake-up signal was intelligently identified using the method of this invention. The test lasted for 72 hours, and the channel conditions covered three typical scenarios: low multipath delay spread, high multipath delay spread, and strong Doppler frequency shift.

[0093] Node sensors at 50 The sampling rate continuously acquires raw underwater acoustic sampling sequences. Electrode temperature varies within the range of 18–24°C, and salinity ranges from 31–34%. The frequency range varies within a certain range. In step S01, the frequency range of the broadband excitation signal is 1 to 20 Hz. The typical value of the identified double-layer capacitor is 85. The typical charge transfer resistance is 120Ω. The typical value of the Warburg coefficient is The Arrhenius temperature compensation equation corrects the three parameters mentioned above based on the real-time electrode temperature, with a correction margin of approximately 12% when the temperature changes by 6°C. The step size of the minimum mean square adaptive filter is set to... After cancellation, the polarization noise floor noise of the denoised underwater acoustic sequence is effectively suppressed in the wake-up signal frequency band.

[0094] In step S02, the orthogonal matching pursuit algorithm iteratively searches in the time delay-amplitude space, and finally extracts 3 to 7 multipath components. The distribution of multipath time delay spread estimates under different test scenarios is shown in Table 1.

[0095] Table 1. Distribution of Multipath Delay Spread Estimates in Three Typical Scenarios

[0096]

[0097] The adaptive RAKE receiver coherently superimposes the delay compensation of each component based on the multipath delay estimate. The signal-to-noise ratio of the preprocessed wake-up frame sequence output signal-to-noise ratio is significantly improved compared with single-path receiver.

[0098] In step S03, the window length of the short-time Fourier transform time-frequency matrix is ​​256 points, the step size is 64 points, and the frequency resolution is approximately 195. Physically prior-constrained time-frequency dual-stream residual attention network on the training set Training with multiple modal samples, the sample signal-to-noise ratio covers 0–30. The dynamic range is set with a positive to negative sample ratio of 1:2. The training configuration is shown in Table 2.

[0099] Table 2 Training Configuration Table for Time-Frequency Dual-Stream Residual Attention Network with Physical Prior Constraints

[0100]

[0101] like Figure 2 As shown, cross-modal attention gating mechanisms are effective when the signal-to-noise ratio is below 5. When the weights of time-domain flow feature fusion are significantly higher than those of frequency-domain flow, the network automatically switches to a working mode that relies on time-domain robust features. However, when the signal-to-noise ratio is higher than 15... At this time, the frequency domain flow weight increases, and the network makes full use of the structured information of the short-time Fourier transform time-frequency matrix.

[0102] In step S04, the standard reference value for the adaptive frame structure adjustment function is set as: signal-to-noise ratio reference value. 15 Multipath delay spread reference value For 200 Doppler frequency shift reference value 50 Weighting coefficients , , The comprehensive channel quality evaluation values ​​and frame structure configuration results for the three test scenarios are shown in Table 3.

[0103] Table 3. Comprehensive Channel Quality Evaluation Values ​​and Frame Structure Configurations for Three Typical Scenarios

[0104]

[0105] The cyclic prefix length covers the measured maximum multipath delay spread in high multipath delay spread scenarios, effectively suppressing inter-symbol interference.

[0106] In step S05, the timeout threshold for the shore-based server's asynchronous inference framework is set to 200. The average feature upload latency of the 5 nodes ranges from 80 to 160. During the 72-hour test, 47 node timeout events occurred, all of which were filled with historical feature prediction values, ensuring the integrity of batch inference. The dynamic batch size adjustment strategy increased the batch size to 128 when the queue was backlogged and decreased it to 16 when the queue was sparse, maintaining a high level of GPU utilization.

[0107] In step S06, a total of 1840 newly labeled samples were accumulated during the 72-hour test, triggering two incremental fine-tuning rounds. Each fine-tuning terminated when the validation set loss did not decrease for 20 consecutive rounds. The model parameters were continuously updated according to the changes in the channel distribution of the test sea area. The wake-up event identification statistics of the 5 nodes within 72 hours are shown in Table 4.

[0108] Table 4. Statistical table of wake-up events for each node during the 72-hour test.

[0109]

[0110] Compared to traditional single-domain recognition methods based on matched filtering and fixed-frame structures, the technological advancements of this invention are reflected in the following aspects: Traditional methods lack physical modeling of platinum black electrode polarization noise, leading to a decrease in polarization noise cancellation capability with changes in temperature and salinity. This invention, however, uses the Arrhenius temperature compensation equation to correct equivalent circuit parameters in real time, ensuring stable polarization noise cancellation accuracy regardless of environmental changes. Traditional fixed-frame structures cannot simultaneously accommodate multiple scenarios with significant differences in multipath delay spread. This invention uses adaptive frame structure adjustment driven by a comprehensive channel quality assessment value to ensure that the cyclic prefix always covers the measured multipath delay spread, thus eliminating inter-symbol interference at its mechanism. The physically prior-constrained time-frequency dual-stream residual attention network, through dynamic modulation via a cross-modal attention gating mechanism, enables a single model to adaptively select the most discriminative feature domain under wide dynamic range signal-to-noise ratio conditions, avoiding the feature failure problem of traditional single-domain methods under extreme channel conditions.

[0111] It should be noted that the variables involved in this invention are explained in detail in Tables 5 and 6.

[0112] Table 5. Variable Explanation Table (Part 1)

[0113]

[0114] Table 6. Variable Explanation Table (Part Two)

[0115]

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent recognition of wake-up signals of a marine multi-parameter precision measuring instrument, characterized in that, Includes the following steps: The equivalent circuit parameters of the platinum black electrode are identified online for the original underwater acoustic sampling sequence collected by the underwater acoustic sensor, and the polarization noise prediction waveform is generated. The negative noise prediction waveform is then extracted from the original underwater acoustic sampling sequence to obtain the denoised underwater acoustic sequence. Sparse channel estimation is performed on the denoised underwater acoustic sequence to obtain multipath delay estimates and multipath complex amplitude estimates. An adaptive RAKE receiver is constructed to coherently superimpose each multipath component after delay compensation. The wake-up frame preprocessing is completed by combining the adaptive orthogonal frequency division multiplexing spread spectrum frame structure to obtain the preprocessed wake-up frame sequence. The short-time Fourier transform time-frequency matrix and instantaneous frequency estimate are extracted from the preprocessed wake-up frame sequence. The real-time signal-to-noise ratio estimate is obtained simultaneously. The original underwater acoustic sampling sequence, short-time Fourier transform time-frequency matrix, instantaneous frequency estimate and real-time signal-to-noise ratio estimate are input into the physical prior constraint time-frequency dual-stream residual attention network, and the wake-up event recognition result and wake-up frequency estimate are output. Based on the real-time signal-to-noise ratio estimate, multipath delay spread estimate, and Doppler frequency shift estimate, the channel quality comprehensive evaluation value is calculated through the adaptive frame structure adjustment function. Based on the interval to which the channel quality comprehensive evaluation value belongs, the subcarrier spacing and cyclic prefix length of the adaptive orthogonal frequency division multiplexing spread spectrum frame structure are dynamically adjusted. The shore-based server uses an asynchronous inference framework to process wake-up event recognition results from multiple nodes in parallel. Data from nodes that have not completed uploading within a timeout period is filled with predicted values ​​of historical features by a physical prior constraint time-frequency dual-stream residual attention network. The batch inference scale of the graphics processor is adjusted through a dynamic batch size adjustment strategy, and the summary results of wake-up events from each node are output. The wake-up event recognition results and historical wake-up event labels are stored in the event database. The newly labeled data in the event database are used periodically to perform incremental fine-tuning on the physical prior constraint time-frequency dual-stream residual attention network, and update the model parameters of the physical prior constraint time-frequency dual-stream residual attention network.

2. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 1, characterized in that, The step of identifying the equivalent circuit parameters of the platinum black electrode online also includes performing thermodynamic correction on the equivalent circuit parameters based on the Arrhenius temperature compensation equation.

3. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 2, characterized in that, The equivalent circuit parameters include double-layer capacitance, charge transfer resistance, and Warburg coefficient. The equivalent circuit parameters are obtained through online identification using a broadband excitation signal.

4. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 3, characterized in that, The method for predicting the negative noise waveform from the original underwater acoustic sampling sequence employs a minimum mean square adaptive filter, using the polarization noise prediction waveform as the reference input and the original underwater acoustic sampling sequence as the desired response, to iteratively update the filter weight vector.

5. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 4, characterized in that, The sparse channel estimation employs an orthogonal matching pursuit algorithm to iteratively search for significant multipath components in the time delay-amplitude space, and outputs multipath time delay estimates and multipath complex amplitude estimates.

6. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 5, characterized in that, The physical prior-constrained time-frequency dual-stream residual attention network adopts a dual-stream parallel input structure. The time-domain stream extracts multi-scale temporal local features through one-dimensional dilated causal convolutional blocks, and the frequency-domain stream extracts joint time-frequency features through two-dimensional residual convolutional blocks.

7. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 6, characterized in that, The time-domain stream and frequency-domain stream are fused through a cross-modal attention gating mechanism, and the attention weight vector is dynamically modulated by an affine transformation based on a real-time signal-to-noise ratio estimate.

8. The intelligent identification method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 7, characterized in that, The physical prior constraint time-frequency dual-stream residual attention network has an additional physical consistency constraint loss term at the top layer, which forces the estimated wake-up frequency to fall within the known frequency offset range of the crystal oscillator. The loss function is composed of a weighted sum of the cross-entropy classification loss and the physical consistency constraint loss term.

9. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 8, characterized in that, The adaptive frame structure adjustment function calculates the comprehensive channel quality assessment value, and dynamically configures the subcarrier spacing and cyclic prefix length in three levels based on the relationship between the comprehensive channel quality assessment value and the high quality threshold and the low quality threshold.

10. The intelligent recognition method for wake-up signals of a marine multi-parameter precision measuring instrument according to claim 9, characterized in that, The multipath delay spread estimate is calculated from the difference between the maximum and minimum values ​​of the multipath delay estimates, and the Doppler frequency shift estimate is obtained by frequency tracking processing of the denoised underwater acoustic sequence.

Images