A DOA estimation method and system based on context learning and dynamic gated residual networks

By employing a DOA estimation method based on context learning and dynamic gated residual networks, we have solved the problems of weak target capture and robustness in DOA estimation in complex marine environments. This method achieves high continuity and high accuracy in DOA estimation, adapting to changes in complex marine environments.

CN123068930APending Publication Date: 2026-11-10ZHEJIANG UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611585359.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing DOA estimation techniques lack contextual correlation mechanisms in complex marine environments, resulting in poor weak target capture capabilities, rigid feature fusion strategies, insufficient robustness, and weak scene adaptability, making it difficult to adapt to nonlinear propagation distortions caused by changes in the marine environment.

Method used

A DOA estimation method based on context learning and dynamic gating residual network is adopted. Through signal conditioning module and edge processing unit, dynamic environmental prior constraints are constructed using support set. Combined with hierarchical gain hardware conditioning and dynamic gating feature fusion mechanism, adaptive adjustment of environmental background features and real-time signal quality optimization are achieved.

Benefits of technology

It improves the continuity and accuracy of DOA estimation, enhances the robustness of the system under extreme conditions, has good engineering application value and environmental scalability, and adapts to complex marine environmental changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068930A_ABST
    Figure CN123068930A_ABST
Patent Text Reader

Abstract

This invention discloses a DOA estimation method and system based on context learning and a dynamically gated residual network. The method includes: acquiring multi-channel synchronously output underwater acoustic digital snapshot data, extracting conjugate covariance features to obtain a multi-dimensional feature vector sequence, and dividing it into a support set and a query set in chronological order; performing context encoding on the support set to obtain an environment representation vector; dynamically gated weighted fusion of the original phase features and statistical smoothing features; inputting the fused features, environment representation vector, and query set into a depth residual regressor to obtain the DOA estimation result corresponding to the query set; and outputting the DOA monitoring trajectory after trajectory smoothing. This invention achieves high continuity and super-resolution accuracy in DOA estimation, suppresses broadband dynamic noise and sudden impulse interference, and enhances the system's robustness under extreme conditions such as array element channel gain or phase mismatch and Doppler frequency offset.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of underwater array signal processing and marine intelligent sensing, and particularly to a DOA estimation method and system based on context learning and dynamic gated residual networks. Background Technology

[0002] In complex marine environments, underwater target azimuth (DOA) estimation is the core of detection systems. Its working principle mainly relies on the phase information received by the sensor array, and establishes a nonlinear mapping from the array covariance matrix to spatial angles through a deep learning model.

[0003] Currently, some technologies have attempted to introduce multi-level networks or dynamic feature processing mechanisms to improve the accuracy and robustness of DOA estimation. For example, patent CN119337941A uses multi-level convolutional networks for layer-by-layer refinement, but due to the lack of long-term modeling of environmental background features, it is prone to trajectory jumps under low signal-to-noise ratio. Patent CN120724244A retains phase information through complex neural networks, but its fixed batch normalization layer is difficult to cope with the distribution shift in non-stationary sea trial data, resulting in limited generalization ability. Patent CN121457531A introduces global dynamic convolution to capture long-range dependencies, but lacks an adaptive adjustment mechanism for the original phase features and smoothing features, which easily leads to jitter in the calculated angle when facing impulse interference.

[0004] In summary, existing technologies still have several shortcomings: First, they lack a contextual association mechanism. Most existing models process sampled images independently and cannot use environmental background features (Support Set) to correct the current detection target, resulting in poor weak target capture capability. Second, the feature fusion strategy is rigid. The system cannot dynamically adjust the channel weights of "original phase features" and "smoothing features" based on real-time signal quality, resulting in insufficient robustness in the presence of Doppler frequency offset or hardware mismatch. Third, the scene adaptability is weak. The model operation relies on fixed normalized statistics, making it difficult to adapt to nonlinear propagation distortion caused by changes in ocean temperature, salinity, and depth, thus limiting the stability of the system when deployed across sea areas. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a DOA estimation method and system based on context learning and dynamic gated residual networks.

[0006] The specific technical solution is as follows:

[0007] A DOA estimation system based on context learning and dynamic gated residual networks includes: a hydrophone array for receiving underwater acoustic signals, a signal conditioning module, an edge processing unit, and an execution terminal; the signal conditioning module is used to perform physical layer gain conditioning, noise suppression, and multi-channel synchronous digital sampling on the acoustic signals; the edge processing unit is implemented using an FPGA hardware module, used to receive synchronous digital snapshot sequences and perform localized real-time inference, and output the target DOA estimation result; the execution terminal is electrically connected to the edge processing unit and is used to receive, display, store, or recall the target DOA estimation result and DOA monitoring trajectory.

[0008] The edge processing unit includes a signal digitization and preprocessing module and a localized DOA inference engine. The signal digitization and preprocessing module is used to extract multi-dimensional feature vector sequences from synchronous digitized snapshot data and divide them into support sets and query sets to achieve data splitting. The localized DOA inference engine includes: a dynamic gating unit for fusing original phase features and statistical smoothing features, a context encoding unit for generating environment representation vectors based on the support sets, and a deep residual regressor for performing DOA inference.

[0009] Furthermore, the signal conditioning module includes a pre-amplifier unit, a bandpass filter, a post-amplifier, and a multi-channel synchronous ADC connected in sequence;

[0010] The preamplifier unit is used to perform impedance matching and preliminary gain enhancement on the weak high-impedance signal output by the hydrophone.

[0011] The bandpass filter is used to suppress background noise outside the target frequency band and eliminate sampling aliasing interference;

[0012] The post-stage gain amplifier is used to dynamically adjust the output level according to the real-time amplitude of the signal to adapt to the full range of the multi-channel synchronous ADC.

[0013] The multi-channel synchronous ADC is used to control the aperture error of multi-channel synchronous sampling to the picosecond level.

[0014] Furthermore, the localized DOA inference engine also includes a method management and optimization module, and a mapping and output module;

[0015] The method management and optimization module is used to provide feedback and dynamically fine-tune the deep residual regressor.

[0016] The mapping and output module is used to realize angle mapping and trajectory smoothing, thereby outputting the DOA monitoring trajectory.

[0017] Furthermore, the signal digitization and preprocessing module includes, in sequence: an analog-to-digital conversion layer, an orthogonal down-conversion layer, a complex bandpass filter layer, and a feature vector extraction layer;

[0018] The digital signal output from the analog-to-digital conversion layer is fed into the quadrature downconversion layer to generate I and Q baseband signals; the I and Q complex baseband signals are input into the complex bandpass filter layer for complex bandpass filtering.

[0019] The feature vector extraction layer is used to perform complex conjugate multiplication and accumulation on the complex baseband sampling vectors of each channel within the observation window after filtering the complex baseband signal to obtain the sample complex covariance matrix; the elements of the complex covariance matrix and the phase difference information between array elements are vectorized to form the multidimensional feature vector sequence.

[0020] A DOA estimation method based on context learning and dynamic gated residual networks, implemented based on the aforementioned DOA estimation system based on context learning and dynamic gated residual networks, includes the following steps:

[0021] S1: Acquire underwater acoustic digital snapshot data from multiple channels simultaneously and perform preprocessing;

[0022] S2: Extract conjugate covariance features from the digital snapshot data, vectorize the conjugate covariance and phase difference information between array elements to obtain a multidimensional feature vector sequence, and divide the multidimensional feature vector sequence into a support set and a query set in chronological order.

[0023] S3: Context-encode the support set to obtain the environment representation vector;

[0024] S4: Dynamically gated weighted fusion of the original phase features and statistical smoothing features is performed to obtain the fused features;

[0025] S5: Input the fused features, environment representation vector and query set into the deep residual regressor to obtain the DOA estimation result corresponding to the query set;

[0026] S6: Perform trajectory smoothing processing on the DOA estimation results and output the DOA monitoring trajectory.

[0027] Furthermore, S2 is specifically implemented through the following sub-steps:

[0028] (2.1) For the multi-channel complex baseband snapshot data after orthogonal downconversion and bandpass filtering, the complex baseband sampling vectors of each channel are multiplied and summed within the preset observation window to obtain the sample complex covariance matrix;

[0029] (2.2) Vectorize the elements of the complex covariance matrix and the phase difference information between the array elements to form a multidimensional feature vector sequence;

[0030] (2.3) According to the time order, the multidimensional feature vector sequence is divided into a support set and a query set; the support set consists of historical snapshots and is used to characterize the long-term environmental background of the current sea area; the query set consists of current or subsequent short-term snapshots and is used to estimate the DOA of the current target.

[0031] Furthermore, the specific operation of S3 is as follows: linear transformation and nonlinear activation are performed on the support set, and an environment representation vector is obtained through global average pooling; the environment representation vector is used to characterize the statistical acoustic environment features of the support set within the corresponding time period, and serves as the environment prior constraint input of the deep residual regressor.

[0032] Further, the specific operation of S4 is as follows: generating a gated score based on the original phase features and statistical smoothing features; normalizing the gated score using the Softmax function to obtain the first fusion weight corresponding to the original phase features and the second fusion weight corresponding to the statistical smoothing features; and weighting and summing the original phase features and statistical smoothing features according to the first fusion weight and the second fusion weight to obtain the fusion features.

[0033] Furthermore, the deep residual regressor includes multiple cascaded residual blocks and an output layer; each residual block performs layer normalization, nonlinear transformation, and residual summation on the input features; the output layer generates the DOA estimation result based on the output of the last residual block and the environment representation vector.

[0034] Furthermore, S5 also includes feedback and update of the output layer parameters of the deep residual regressor: when obtaining a sample with a reliable reference azimuth, the Huber loss is calculated based on the residual between the DOA estimation result and the reliable reference azimuth.

[0035] When the preset update conditions are met, only the output layer weights of the deep residual regressor are updated; the update conditions are: the Huber loss is greater than the preset update threshold, and the reliable reference azimuth angle meets the preset confidence condition;

[0036] After performing fixed-point quantization and parameter range verification on the updated output layer weights, they are written to the weight storage area inside the FPGA through the bus interface, and switched to the effective output layer weights after the verification is passed and the current inference frame is finished.

[0037] If a reliable reference azimuth angle is not obtained and / or the update conditions are not met and / or parameter verification fails and / or the total loss function does not decrease after the update, then the update result is discarded and the original output layer weights are retained.

[0038] The beneficial effects of this invention are:

[0039] (1) This invention utilizes context learning technology to construct dynamic environmental prior constraints using support sets, enabling the system to "perceive environmental changes and autonomously correct deviations". Compared with traditional methods that can only handle isolated snapshots independently, this invention effectively overcomes the predicted trajectory jumps in complex multipath and fading sea areas, ensuring high continuity and super-resolution accuracy of DOA estimation.

[0040] (2) This invention combines a graded gain hardware conditioning link with a dynamic gating feature fusion mechanism, enabling the system to collaboratively optimize signal quality from the physical layer to the algorithm layer. By dynamically balancing the original phase features and spatial gradient smoothing features, it not only suppresses broadband dynamic noise and sudden pulse interference, but also significantly enhances the system's monitoring robustness under extreme conditions such as array element channel gain and phase mismatch, and Doppler frequency offset.

[0041] (3) By constructing an environmental representation vector, this invention can accurately compensate for the distortion of the sound velocity profile and the nonlinear propagation deviation caused by changes in temperature, salinity, and depth (CTD) during sea trials. This "environmentally aware" prediction mechanism eliminates the need for cumbersome manual parameter tuning when switching between heterogeneous sea areas, and has excellent engineering application value and environmental scalability. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall hardware architecture and signal conditioning link of the DOA estimation system based on context learning and dynamic gating residual network in this embodiment of the invention.

[0043] Figure 2 This is a schematic diagram of the logical architecture of the edge processing unit with built-in ContextDOA-Net neural network in an embodiment of the present invention.

[0044] Figure 3 This is a flowchart illustrating the DOA estimation method based on context learning and dynamic gated residual networks in an embodiment of the present invention.

[0045] Figure 4 This is a comparison chart of the mean absolute error (MAE) performance of the method of this invention and the traditional DOA estimation method under different operating conditions. Detailed Implementation

[0046] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The objectives and effects of the present invention will become clearer as a result. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0047] like Figure 1As shown, a DOA estimation system based on context learning and dynamic gating residual networks includes a hardware link consisting of a hydrophone array, a signal conditioning module, an edge processing unit (FPGA), and an execution terminal. The hydrophone array receives underwater acoustic signals. The signal conditioning module includes a preamplifier, a bandpass filter, a post-amplifier gain amplifier, and a multi-channel synchronous analog-to-digital converter (hereinafter referred to as the multi-channel synchronous ADC), connected in sequence, used for physical layer gain conditioning, noise suppression, and multi-channel synchronous digital sampling of the acoustic signals. The edge processing unit is implemented using an FPGA hardware module and is electrically connected to the multi-channel synchronous ADC. It receives synchronous digitized snapshot sequences, performs localized real-time inference, and outputs the target azimuth (DOA) estimation result. The edge processing unit incorporates a Context DOA neural network, which includes a dynamic gating unit, a context encoding unit, and a deep residual regressor. The execution terminal is connected to the edge processing unit and is used to receive, display, store, or recall the target azimuth estimation result and the DOA monitoring trajectory.

[0048] The hydrophone array uses a 12-channel equally spaced linear array (ULA) structure, with the element spacing d strictly following the... The criteria are set, where c≈1500m / s is the underwater speed of sound, and f max The highest operating frequency designed for the entire hydrophone system.

[0049] The signal conditioning module employs precise phase control and wide dynamic range gain design, using a hierarchical gain architecture. In this embodiment, the multi-channel synchronous ADC uses a 12-channel synchronous ADC, with phase consistency circuitry between each channel. The preamplifier unit is used for impedance matching and initial gain boosting of the weak high-impedance signal output from the hydrophone. This embodiment uses an instrumentation amplifier with ultra-low offset current and a common-mode rejection ratio (CMRR) greater than 110dB, designed to filter out induced power frequency interference in long cable transmission. The bandpass filter is used to suppress background noise outside the target frequency band and eliminate sampling aliasing interference. In this embodiment, a cascaded active filter is constructed using two operational amplifiers. By accurately calculating the RC parameters, the cutoff frequency is locked at the characteristic frequency band of the target sound source. Its transfer function H(s) satisfies:

[0050]

[0051] In the formula, s is the Laplace complex frequency domain variable, and G i Let Q be the passband gain of the i-th stage active filter unit. i Let be the quality factor of the i-th stage active filter unit, used to characterize the filter's selectivity and bandwidth. Let be the natural angular frequency of the i-th stage active filter unit.

[0052] The post-amplifier (PGA) is used to dynamically adjust the output level according to the real-time signal amplitude to adapt to the full input range of the multi-channel synchronous ADC. In this embodiment, an 8-bit parallel control word is used, and the FPGA of the edge processing unit performs dynamic closed-loop adjustment based on the effective value (RMS) of the previous frame signal. The multi-channel synchronous ADC adopts a 24-bit high-speed Sigma-Delta architecture and is driven by a unified sampling clock distributed by the FPGA chip. The clock distribution network ensures that the aperture error of the 12-channel synchronous sampling is controlled in the picosecond range, thereby ensuring that the digital snapshot sequence retains extremely weak angular phase deviation (i.e., true inter-element phase difference information) at the physical layer.

[0053] like Figure 2 As shown, the edge processing unit internally incorporates deeply customized acceleration logic, divided into a signal digitization and preprocessing module and a localized DOA inference engine. The signal digitization and preprocessing module includes: an analog-to-digital conversion layer, an orthogonal down-conversion layer, a complex bandpass filter layer, and a feature vector extraction layer; used to extract multi-dimensional feature vector sequences from synchronous digitized snapshot data. And divided into support sets and query set This enables data offloading. The localized DOA inference engine includes: a dynamic gating unit, a context encoding unit, a deep residual regressor, a mapping and output module, and a method management and optimization module. The dynamic gating unit is used to fuse the original phase features and statistical smoothing features; the context encoding unit is used to generate an environment representation vector based on the support set; the deep residual regressor is used to perform DOA inference based on the fused features, the environment representation vector, and the query set; the mapping and output module is used to implement angle mapping and trajectory smoothing, thereby outputting the DOA monitoring trajectory; the method management and optimization module is used to provide feedback and dynamically fine-tune the deep residual regressor.

[0054] In this embodiment, in the signal digitization and preprocessing module, the analog-to-digital conversion layer is configured with 16 parallel input channels, supporting 24-bit high-resolution sampling, and achieving real-time data stream throughput through the AXI4-Stream high-speed bus interface, ensuring the integrity of the original acoustic information. Subsequently, the quadrature downconversion layer (DDC) uses a digitally controlled oscillator (NCO) to perform complex baseband mapping, with its decimation factor set to 4, and the output frequency stably locked at 24kHz to compress the data bandwidth for subsequent processing. The complex bandpass filtering layer constructs a dedicated channel with a passband range of 3.5kHz to 7.5kHz through a 6th-order IIR or FIR filter bank, aiming to filter out out-of-band interference in the ocean background and accurately preserve the characteristic frequency band of the target sound source. Finally, in the feature vector extraction layer, the 16-channel complex baseband signal after quadrature downconversion and filtering is framed according to a preset observation window; let the 16-channel complex baseband sampling vector at the nth snapshot time be... ,in, Let be the complex baseband sample value of the i-th channel at the n-th snapshot. Within an observation window containing L snapshots, the complex baseband sample vectors of each channel are accumulated (in parallel) using a DSP48E hardware operator array to obtain the sample complex covariance matrix. In this embodiment, after one observation window ends, the accumulated result is normalized to obtain 256 elements of the complex covariance matrix. Further, the elements of the complex covariance matrix are expanded by real part, imaginary part, magnitude, or phase, and a 128-dimensional phase difference vector between array elements is extracted to form a multidimensional feature vector sequence for subsequent DOA inference. .

[0055] Within the localized DOA inference engine, a multi-dimensional feature vector sequence is constructed. Divide it into support sets (Long-time series) and query sets (Short time series); The dynamic gating network integrates a dual-branch dynamic gating mechanism, adaptively calculating the confidence weights of the original phase features and statistical smoothing features based on real-time signal quality; simultaneously, the context coding unit is configured with a support set of 128 historical snapshots. The current marine environment representation vector is aggregated with a 512-dimensional encoding dimension. Finally, the concatenated feature stream enters the deep residual regressor, which contains 16 cascaded residual blocks (ResBlock) and 2 regressor heads, and integrates the LayerNorm hardware operator to realize real-time standardization of the numerical values.

[0056] The execution terminal is connected to the edge processing unit to receive the DOA estimation results and DOA monitoring trajectory output by the edge processing unit, and to display, store, and retrieve the DOA estimation results and DOA monitoring trajectory in real time. The execution terminal can be a host computer, an embedded display controller, or a mission computer mounted on an underwater platform. In this embodiment, the execution terminal receives the target azimuth information after trajectory smoothing and outputs the target azimuth and its DOA monitoring trajectory changing over time according to a preset display period for use by operators or subsequent mission modules of the underwater platform.

[0057] Through the collaborative parallelism of the edge processing unit hardware pipeline, the system can output a high-precision G×1 dimensional orientation prediction trajectory, where G is the number of DOA estimation results within a preset time window, achieving low-latency and seamless DOA monitoring.

[0058] Based on the above system, this invention also proposes a DOA estimation method based on context learning and a dynamically gated residual network, such as... Figure 2 and Figure 3 As shown, the method includes the following steps:

[0059] S1: Hardware initialization and signal acquisition (acquiring underwater acoustic digital snapshot data from multiple channels simultaneously). After system power-on, the FPGA initializes the registers of the 16 ADC channels via the bus, enabling synchronous data access from multiple ADC channels (the output of the multi-channel synchronous ADC is underwater acoustic digital snapshot data). The ADC sampled data is pushed into the FPGA's internal FIFO with a 24-bit depth, and the frame synchronization circuit ensures that the phase relationship of the spatial manifold is not distorted due to transmission delay.

[0060] S2: Covariance Feature Extraction and Data Splitting. The edge processing unit reads the underwater acoustic digitized snapshot data acquired and synchronously processed by S1, performs parallel feature extraction, and obtains a multi-dimensional feature vector sequence. Specifically, this is achieved through the following sub-steps:

[0061] (2.1) For the multi-channel complex baseband snapshot data after orthogonal downconversion and bandpass filtering, the sample complex covariance matrix is ​​calculated within a preset observation window: Within an observation window containing L snapshot times, the complex baseband sampling vectors of each channel are multiplied and summed using a DSP48E hardware operator array to obtain the sample complex covariance matrix. for:

[0062]

[0063] In the formula, Let be the M-channel complex baseband sampling vector at the nth snapshot time, where n = 1, 2, ..., L, and L is the observation window length, i.e., the number of snapshots participating in one covariance estimation. ; This indicates the conjugate transpose. Let be the complex baseband sample value of the i-th channel at the n-th snapshot time; M is the number of channels in the hydrophone array, M=16 in this embodiment; R is a 16×16 dimensional sample complex covariance matrix with 16×16=256 complex covariance matrix elements, and its element in the i-th row and j-th column (i=1,2,…,M, j=1,2,…,M) is:

[0064]

[0065] In the formula, Indicates complex conjugation.

[0066] (2.2) Vectorize the elements of the complex covariance matrix (i.e., conjugate covariance) and the phase difference information between array elements to form a multidimensional feature vector sequence. .

[0067] (2.3) Arrange the multidimensional feature vector sequence in chronological order. Divided into support sets and query set Among them, support sets Composed of historical snapshots, used to characterize the long-term environmental background of the current sea area, query set It consists of current or subsequent short time-series snapshots, used for DOA estimation of the current target. In this embodiment, the length of both the support set and the query set is 128.

[0068] S3: Context awareness and prior environment constraint construction. Context coding units support a 128-bit support set. A nonlinear transformation is performed to obtain the environment representation vector. This embodiment utilizes Global Average Pooling (GAP) to extract the feature mean, and calculates a 512-dimensional environment representation vector. :

[0069]

[0070] In the formula, The support set feature matrix is ​​composed of support set feature vectors corresponding to 128 consecutively acquired historical snapshots, where d is the dimension of a single snapshot feature vector. For the trainable weight matrix of the context encoding unit, it is used to map d-dimensional support set features to a 512-dimensional environment feature space; This is the corresponding trainable bias vector matrix; It is a linear rectified activation function used to introduce nonlinear characterization capabilities; This is a global average pooling operator used to average the encoded features of 128 snapshots along the time dimension of the support set.

[0071] Environment Representation Vector Used to characterize the current sea area in the support set The statistical acoustic environment characteristics within the time window implicitly reflect the distortion of the sound velocity profile caused by changes in temperature, salinity, and depth in the current sea area; these acoustic environment characteristics include statistical features related to temperature, salinity, depth variations, multipath propagation, and background noise. The environment is represented as a vector. As a priori environmental constraint, and the query set The corresponding fused features are input into the deep residual regressor to complete the DOA inference for the current query target.

[0072] S4: Dynamically gated weighted fusion of the original phase features and statistical smoothing features is performed to obtain the fused features.

[0073] For query set The current snapshot features are used to extract the original phase features and statistical smoothing features by the dynamic gating unit, and the fusion weights are adaptively calculated based on the current signal quality of the two types of features. This is achieved through the following sub-steps.

[0074] (4.1) First, the two types of features are concatenated and input into the dynamic gating unit to obtain the gating scores corresponding to the two types of features:

[0075]

[0076] In the formula, This is the original phase eigenvector. For statistically smoothed feature vectors, Indicates feature concatenation operation; For the trainable weight matrix of the dynamic gating unit, This is the corresponding trainable bias vector; Z is a nonlinear mapping function; z1 represents the gate score corresponding to the original phase feature, and z2 represents the gate score corresponding to the statistical smoothing feature.

[0077] (4.2) Softmax normalization is applied to the gating scores to obtain the adaptive fusion weights for the two types of features:

[0078]

[0079] In the formula, w1 is the adaptive fusion weight of the original phase feature, w2 is the adaptive fusion weight of the statistical smoothing feature, and w1+w2=1.

[0080] (4.3) Perform weighted fusion of the two types of features to obtain the fused features. for:

[0081]

[0082] The fusion feature is used to enhance the contribution of the statistical smoothing feature when there is impulse noise or signal quality degradation, while preserving the detailed information of the original phase feature when the signal quality is good, and is combined with the environment representation vector obtained by S3. Common input depth residual regressor.

[0083] The context encoding process of S3 and the dynamic gating process of S4 can be executed in parallel.

[0084] S5: Input the fused features, environment representation vector and query set into the deep residual regressor, perform deep residual inference and bias correction, and obtain the DOA estimation result corresponding to the query set.

[0085] The deep residual regressor consists of multiple cascaded residual blocks and an output layer; in this embodiment, there are 16 residual blocks. The fused features corresponding to the query set are then... The environment representation vector obtained with S3 The data is then concatenated, and the input consists of a depth residual regressor composed of 16 residual blocks connected in series. For the k-th residual block, The residual mapping output is added to the input feature vector to obtain the intermediate feature vector to be normalized. :

[0086]

[0087] in, Let be the input feature vector of the k-th residual block. Let be the residual mapping function for the k-th residual block. The set of trainable parameters for implementing residual mapping includes the weights and bias parameters of fully connected or convolutional layers within the residual block.

[0088] For intermediate feature vectors By performing layer normalization and nonlinear activation, the output feature vector y of the k-th residual block is obtained. k+1 The calculation expression is:

[0089]

[0090] In the formula, For layer normalization operators.

[0091] Let the d-dimensional input feature vector be... Then the normalized expression for the execution layer is:

[0092]

[0093]

[0094]

[0095] Where μ is the input feature vector Mean across the feature dimension; Input feature vector Variance in the feature dimension; ε is a preset positive number to prevent the denominator from being zero; γ is the trainable scaling parameter of the layer normalization operator, and β is the trainable bias parameter of the layer normalization operator; This indicates element-wise multiplication.

[0096] By using a series of nonlinear transformations and layer normalization processes on multiple residual blocks, the impact of changes in the distribution of input data on the regression results is reduced. The output layer generates the DOA estimation result based on the output of the last residual block and the environment representation vector.

[0097] On the other hand, bias correction is achieved through a method management and optimization module, which provides feedback and dynamic fine-tuning updates to the output layer parameters of the depth residual regressor based on a reliable reference azimuth angle. The depth residual regressor outputs the target azimuth angle prediction value through the regressor head. When the azimuth of the calibrated sound source, the true value of the navigation track, or other reliable reference azimuth angles are obtained... At time (i.e., as the measured reference value), calculate the prediction residual for the q-th sample:

[0098]

[0099] The Huber loss function serves as the target for edge-side fine-tuning. The expression is:

[0100]

[0101] In the formula, δ is the preset residual threshold.

[0102] For multiple samples within the current update window, the average of the Huber loss functions for each sample is taken as the total loss function. When the total loss function is greater than a preset update threshold, and the reliable reference DOA meets the preset reliability condition, only the output layer weights of the deep residual regressor are updated with lightweight quantization. Let the output layer weights at update t be... The gradient of the total loss function with respect to the output layer weights is:

[0103]

[0104] In the formula, The total loss function relative to the output layer weights at the t-th update. gradient, Represents the weights relative to the output layer. Find the gradient. Let t be the total loss function corresponding to the t-th update.

[0105] The AdamW optimization algorithm is used to calculate the first-order moment estimate. and second-order moment estimation :

[0106]

[0107]

[0108] in, The momentum coefficient of the first moment. Let be the momentum coefficient of the second moment. After bias correction, we get:

[0109]

[0110]

[0111] Update the output layer weights according to the following formula:

[0112]

[0113] Where η is the learning rate, λ is the weight decay coefficient, and ε2 is a preset positive number to prevent the denominator from being zero.

[0114] After fixed-point quantization, parameter range verification, and deep residual regressor version verification, the updated output layer weights are written to the double-buffered weight storage area inside the FPGA via the AXI interface. The current inference frame continues to use the output layer weights before the update. After the current inference frame is processed, it switches to the updated output layer weights to avoid the output layer weight switching affecting the DOA inference of the current frame. If a reliable reference DOA is not obtained, the update conditions are not met, parameter verification fails, or the total loss function is not reduced after the update, the update result is discarded and the original output layer weights are retained.

[0115] The method management and optimization module is also used to record the deep residual regressor version, update count, Huber loss function value, and DOA estimation results, so as to realize the rollback, retrospection, and stable operation of the deep residual regressor output layer parameters.

[0116] S6: Perform angle mapping and trajectory smoothing on the DOA estimation results, and output a G×1 dimensional DOA monitoring trajectory to achieve high-smoothness DOA output.

[0117] To compare DOA estimation performance, the mean absolute error (MAE) under different signal-to-noise ratio (SNR) conditions was used for evaluation. The formula for calculating MAE is:

[0118]

[0119] In the formula, N represents the number of Monte Carlo simulations. This is the actual DOA value. This is the estimated DOA value for the k-th Monte Carlo trial.

[0120] like Figure 4As shown, the MAE simulation curves of the proposed method and the traditional MUSIC algorithm under different signal-to-noise ratio (SNR) conditions demonstrate that the proposed method (Meta-AI) outperforms the traditional MUSIC algorithm across the entire SNR range. In particular, under high SNR (SNR>5dB) conditions, the prediction error of the proposed method consistently remains below 0.7°, exhibiting extremely high positioning accuracy. Under extreme conditions of low SNR (SNR=-20dB), the MAE of the traditional algorithm approaches 10°, while the proposed method still manages to control the error to around 6°. Furthermore, the convergence speed with increasing SNR is significantly faster than the traditional method, proving the strong robustness of the Context DOA neural network to environmental noise. In addition, in terms of computational efficiency, the average inference time of the proposed method is only 0.003ms, compared to 0.163ms for the traditional MUSIC algorithm, representing an improvement of approximately 54 times. This fully verifies that the proposed method, while ensuring high accuracy, possesses excellent real-time processing performance, greatly adapting to the low-latency requirements of edge processors.

[0121] This invention employs a lightweight residual network with a parallel pipeline design, coupled with a normalized architecture, to eliminate the dependence of traditional deep learning on the statistical characteristics of batch samples. This design is highly compatible with edge hardware with limited computing resources, such as FPGAs and embedded GPUs, enabling extremely low-latency local inference and providing a reliable decentralized autonomous orientation solution for unmanned underwater vehicles, underwater towed platforms, and other applications.

[0122] This invention utilizes an edge-side dynamic fine-tuning mechanism based on Huber loss to perform lightweight updates on the output layer weights of the depth residual regressor when the total loss function is greater than a preset update threshold and the credible reference DOA meets the preset confidence conditions. This enables adaptive updates of the model under different sea areas and different detection depths, thereby improving the long-term working stability and environmental adaptability of the system.

[0123] Through the above implementation methods, this invention achieves high-precision, low-latency orientation for acoustic targets in complex sea areas. This solution deeply integrates the Context DOA neural network with an FPGA parallel architecture, significantly reducing the system's computational load and power consumption while ensuring detection accuracy. It is suitable for the autonomous orientation needs of resource-constrained platforms such as unmanned underwater vehicles (UAVs).

[0124] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A DOA estimation system based on context learning and dynamically gated residual networks, characterized in that, include: The system comprises a hydrophone array for receiving underwater acoustic signals, a signal conditioning module, an edge processing unit, and an execution terminal. The signal conditioning module performs physical layer gain conditioning, noise suppression, and multi-channel synchronous digital sampling on the acoustic signals. The edge processing unit, implemented using an FPGA hardware module, receives synchronous digital snapshot sequences, performs localized real-time inference, and outputs target DOA estimation results. The execution terminal is electrically connected to the edge processing unit and is used to receive, display, store, or recall the target DOA estimation results and DOA monitoring trajectory. The edge processing unit includes a signal digitization and preprocessing module and a localized DOA inference engine. The signal digitization and preprocessing module is used to extract multi-dimensional feature vector sequences from synchronous digitized snapshot data and divide them into support sets and query sets to achieve data splitting. The localized DOA inference engine includes: a dynamic gating unit for fusing original phase features and statistical smoothing features, a context encoding unit for generating environment representation vectors based on the support sets, and a deep residual regressor for performing DOA inference.

2. The DOA estimation system based on context learning and dynamic gated residual network according to claim 1, characterized in that, The signal conditioning module includes a pre-amplifier unit, a bandpass filter, a post-gain amplifier, and a multi-channel synchronous ADC connected in sequence. The preamplifier unit is used to perform impedance matching and preliminary gain enhancement on the weak high-impedance signal output by the hydrophone. The bandpass filter is used to suppress background noise outside the target frequency band and eliminate sampling aliasing interference; The post-stage gain amplifier is used to dynamically adjust the output level according to the real-time amplitude of the signal to adapt to the full range of the multi-channel synchronous ADC. The multi-channel synchronous ADC is used to control the aperture error of multi-channel synchronous sampling to the picosecond level.

3. The DOA estimation system based on context learning and dynamic gated residual networks according to claim 1, characterized in that, The localized DOA inference engine also includes a method management and optimization module, and a mapping and output module; The method management and optimization module is used to provide feedback and dynamically fine-tune the deep residual regressor. The mapping and output module is used to realize angle mapping and trajectory smoothing, thereby outputting the DOA monitoring trajectory.

4. The DOA estimation system based on context learning and dynamic gated residual network according to claim 1, characterized in that, The signal digitization and preprocessing module includes, in sequence: an analog-to-digital conversion layer, an orthogonal down-conversion layer, a complex bandpass filter layer, and a feature vector extraction layer; The digital signal output from the analog-to-digital conversion layer is fed into the quadrature downconversion layer to generate I and Q baseband signals; the I and Q complex baseband signals are input into the complex bandpass filter layer for complex bandpass filtering. The feature vector extraction layer is used to perform complex conjugate multiplication and accumulation on the complex baseband sampling vectors of each channel within the observation window after filtering the complex baseband signal to obtain the sample complex covariance matrix; the elements of the complex covariance matrix and the phase difference information between array elements are vectorized to form the multidimensional feature vector sequence.

5. A DOA estimation method based on context learning and dynamic gated residual networks, implemented based on the DOA estimation system based on context learning and dynamic gated residual networks as described in any one of claims 1-4, characterized in that, Includes the following steps: S1: Acquire underwater acoustic digital snapshot data from multiple channels simultaneously and perform preprocessing; S2: Extract conjugate covariance features from the digital snapshot data, vectorize the conjugate covariance and phase difference information between array elements to obtain a multidimensional feature vector sequence, and divide the multidimensional feature vector sequence into a support set and a query set in chronological order. S3: Context-encode the support set to obtain the environment representation vector; S4: Dynamically gated weighted fusion of the original phase features and statistical smoothing features is performed to obtain the fused features; S5: Input the fused features, environment representation vector and query set into the deep residual regressor to obtain the DOA estimation result corresponding to the query set; S6: Perform trajectory smoothing processing on the DOA estimation results and output the DOA monitoring trajectory.

6. The DOA estimation method based on context learning and dynamically gated residual networks according to claim 5, characterized in that, S2 is specifically implemented through the following sub-steps: (2.1) For the multi-channel complex baseband snapshot data after orthogonal downconversion and bandpass filtering, the complex baseband sampling vectors of each channel are multiplied and summed within the preset observation window to obtain the sample complex covariance matrix; (2.2) Vectorize the elements of the complex covariance matrix and the phase difference information between the array elements to form a multidimensional feature vector sequence; (2.3) According to the time order, the multidimensional feature vector sequence is divided into a support set and a query set; the support set consists of historical snapshots and is used to characterize the long-term environmental background of the current sea area; the query set consists of current or subsequent short-term snapshots and is used to estimate the DOA of the current target.

7. The DOA estimation method based on context learning and dynamic gated residual networks according to claim 5, characterized in that, The specific operation of S3 is as follows: linear transformation and nonlinear activation are performed on the support set, and the environment representation vector is obtained by global average pooling; the environment representation vector is used to characterize the statistical acoustic environment features of the support set within the corresponding time period, and serves as the environment prior constraint input of the deep residual regressor.

8. The DOA estimation method based on context learning and dynamic gated residual networks according to claim 5, characterized in that, The specific operation of S4 is as follows: a gated score is generated based on the original phase features and statistical smoothing features; the gated score is normalized by the Softmax function to obtain the first fusion weight corresponding to the original phase features and the second fusion weight corresponding to the statistical smoothing features. The original phase features and statistical smoothing features are weighted and summed according to the first fusion weight and the second fusion weight to obtain the fused features.

9. The DOA estimation method based on context learning and dynamic gated residual networks according to claim 5, characterized in that, The deep residual regressor comprises multiple cascaded residual blocks and an output layer; each residual block performs layer normalization, nonlinear transformation, and residual summation on the input features; the output layer generates the DOA estimation result based on the output of the last residual block and the environment representation vector.

10. The DOA estimation method based on context learning and dynamic gated residual networks according to claim 5, characterized in that, The S5 also includes feedback and update of the output layer parameters of the deep residual regressor: when a sample with a reliable reference azimuth is obtained, the Huber loss is calculated based on the residual between the DOA estimation result and the reliable reference azimuth. When the preset update conditions are met, only the output layer weights of the deep residual regressor are updated; the update conditions are: the Huber loss is greater than the preset update threshold, and the reliable reference azimuth angle meets the preset confidence condition; After performing fixed-point quantization and parameter range verification on the updated output layer weights, they are written to the weight storage area inside the FPGA through the bus interface, and switched to the effective output layer weights after the verification is passed and the current inference frame is finished. If a reliable reference azimuth angle is not obtained and / or the update conditions are not met and / or parameter verification fails and / or the total loss function is not reduced after the update, then the update result is discarded and the original output layer weights are retained.

Citation Information

Patent Citations

  • Multi-level convolutional neural network DOA estimation method based on spatial feature feedback

    CN119337941A

  • Multi-signal DOA estimation method and device based on complex neural network

    CN120724244A

  • DOA estimation method and system based on global dynamic convolutional neural network

    CN121457531A