A method and apparatus for identifying water motion state based on spectral moment characteristics, and its readable storage medium.

Abstract

This invention proposes a method, device, and readable storage medium for identifying water body motion states based on spectral moment features, belonging to the field of hydrological measurement and intelligent analysis technology. The method utilizes an acoustic Doppler current meter to collect flow velocity sequences, converts them into frequency domain power spectral density via short-time Fourier transform, extracts multi-dimensional features including zero- to fourth-order spectral moments, frequency band energy, statistics, and spectral characteristics, constructs a feature time-time heatmap, and inputs it into a multi-scale residual convolutional neural network to identify laminar, turbulent, and recirculating flows. This invention accurately characterizes the essential differences between the three flow states through frequency domain features, and combines multi-scale convolutional kernels and residual connections to preserve weak signals, solving the problems of poor interpretability and insensitivity in traditional methods. It has the advantages of high accuracy, strong robustness, and applicability to various scenarios such as rivers and dams.

Description

Technical Field

[0001] This invention relates to the field of hydrological measurement and intelligent analysis technology, and in particular to a method, device and readable storage medium for identifying the motion state of water bodies based on spectral moment characteristics. Background Technology

[0002] Accurate identification of water movement states (such as laminar flow, turbulent flow, and backflow) is a key foundation for hydrological monitoring, safety assessment of water conservancy projects, and regulation of urban drainage systems. Traditional identification methods mainly rely on the temporal statistical characteristics (such as mean and variance) or spatial distribution characteristics of the original flow velocity, and classify it using shallow models such as support vector machines (SVM).

[0003] However, water flow is a typical non-stationary signal, and its core characteristics (such as energy distribution and frequency components) are more reflected in the frequency domain: laminar flow energy is concentrated in the low frequency and the spectrum shows a clear single peak, turbulent flow has a broad spectrum and the center of gravity frequency shifts upward, while backflow exhibits a multi-peak distribution. Existing technologies have two major limitations because they do not delve deeply into the frequency domain characteristics:

[0004] First, the data is poorly interpretable, making it difficult to reveal the spectral essence behind fluid motion;

[0005] Secondly, it is not sensitive to the frequency domain differences between laminar flow, turbulent flow, and recirculation, resulting in insufficient identification accuracy and robustness, especially in complex aquatic environments (such as high noise interference from urban drainage systems and natural disturbances in rivers).

[0006] Therefore, there is an urgent need for a water motion state identification method based on frequency domain characteristics, strong physical interpretability, and robustness to noise, in order to meet the high-precision requirements of actual hydrological monitoring. Summary of the Invention

[0007] This invention provides a method, device, and readable storage medium for identifying water motion states based on spectral moment characteristics. It addresses the problems of existing technologies that rely on time or spatial characteristics, making it difficult to reveal the spectral characteristics of water motion, and are not sensitive to the frequency domain differences between laminar, turbulent, and recirculating flows, resulting in insufficient identification accuracy and robustness.

[0008] The core technology of this invention mainly involves collecting flow velocity data using an acoustic Doppler current meter (ADCP), converting it into frequency domain power spectral density via short-time Fourier transform (STFT), extracting 12-dimensional features such as spectral moments (zero to fourth order), frequency band energy, and power gradient, and constructing a time heatmap. This heatmap is then input into a convolutional neural network containing a multi-scale residual structure (1×1, 3×3, and 5×5 convolutional kernels for parallel feature extraction + residual connections) to achieve accurate identification of laminar, turbulent, and recirculating flow.

[0009] In a first aspect, the present invention provides a method for identifying the motion state of water bodies based on spectral moment features, the method comprising the following steps:

[0010] Step 1: Collect flow velocity sequences using an acoustic Doppler current meter as input data;

[0011] Step 2: Convert the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation;

[0012] Step 3: Extract features from the frequency domain power spectral density. The features include spectral moment features, frequency band energy, statistical features, and spectral features. Among them, spectral moment features include zero-order spectral moment, first-order spectral moment, second-order spectral moment, third-order spectral moment, and fourth-order spectral moment; frequency band energy includes low-frequency energy, mid-frequency energy, and high-frequency energy; statistical features include maximum spectral power and spectral power standard deviation; and spectral features include frequency peak position and spectral power gradient.

[0013] Step 4: Construct the extracted features into a feature vector and input the feature vector into a preset classifier model. The classifier model identifies the current motion state of the water body, which includes laminar flow, turbulent flow, and backflow.

[0014] Furthermore, in step 2, the frequency domain transformation is a short-time Fourier transform. The process of converting the time domain signal into the frequency domain power spectral density through the short-time Fourier transform includes: dividing the original time domain signal into short time windows, performing a Fourier transform on each short time window to obtain the time-varying spectrum, averaging the squared amplitude of each frequency point to estimate the power spectral density, and then averaging the power spectral density along the time axis to obtain the overall power spectral density.

[0015] Furthermore, the window function used in the short-time Fourier transform is the Hanning window.

[0016] Furthermore, in step 3, the zeroth spectral moment is the total energy of the power spectral density, the first spectral moment is the spectral centroid, the second spectral moment is the dispersion, the third spectral moment is the skewness, and the fourth spectral moment is the kurtosis.

[0017] Furthermore, in step 3, the low-frequency energy is the integral value of the power spectral density in the low-frequency range, the mid-frequency energy is the integral value of the power spectral density in the mid-frequency range, and the high-frequency energy is the integral value of the power spectral density in the high-frequency range.

[0018] Furthermore, in step 3, the maximum spectral power is the maximum power density value corresponding to a single frequency point in the spectrum, and the standard deviation of the spectral power is the degree of dispersion of the power spectral density value over the entire frequency range.

[0019] Furthermore, in step 3, the frequency peak position is the frequency point corresponding to the maximum value in the power spectral density, and the spectral power gradient is the average rate of change of the power spectral density along the frequency.

[0020] Furthermore, in step 4, the preset classifier model is a convolutional neural network.

[0021] Furthermore, the convolutional neural network includes a multi-scale residual extraction module, which is composed of multiple multi-scale residual structures connected in series. Each multi-scale residual structure has multiple parallel convolutional branches, and the multiple parallel convolutional branches use convolutional kernels with different receptive fields.

[0022] Furthermore, the convolution kernels for different receptive fields include 1×1, 3×3, and 5×5 convolution kernels.

[0023] Furthermore, in each multi-scale residual structure, the outputs of all convolutional branches are concatenated through channels, fused by a 1×1 convolution, and then batch normalized before being fused with the input signal via residual connection.

[0024] Furthermore, in step 4, the feature vectors are arranged in chronological order to form a two-dimensional feature time heatmap, and then the feature time heatmap is input into a convolutional neural network.

[0025] Furthermore, the identification results of water movement states include laminar flow, turbulent flow, and recirculation. Among them, the laminar flow state has a distinct single peak in its spectrum, concentrated energy, and a skewness close to 0; the turbulent flow state has a broadened spectrum, an upward shift in the center of gravity frequency, and a significant skewness; and the recirculation state has a multi-peaked spectrum with a high kurtosis.

[0026] Secondly, the present invention provides a water motion state identification device based on spectral moment features, comprising:

[0027] The data acquisition module uses an acoustic Doppler flow meter to collect flow velocity sequences as input data.

[0028] The frequency domain transformation module converts the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation;

[0029] The feature extraction module extracts features from the frequency domain power spectral density. These features include spectral moment features, frequency band energy, statistical features, and spectral features. The spectral moment features include zero-order, first-order, second-order, third-order, and fourth-order spectral moments. The frequency band energy features include low-frequency, mid-frequency, and high-frequency energy. The statistical features include maximum spectral power and spectral power standard deviation. The spectral features include frequency peak position and spectral power gradient.

[0030] The processing module constructs the extracted features into feature vectors and inputs the feature vectors into a preset classifier model. The classifier model identifies the current motion state of the water body, which includes laminar flow, turbulent flow, and backflow.

[0031] The output module outputs the recognition results of the classifier model.

[0032] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to execute the above-described method for identifying the water motion state based on spectral moment characteristics.

[0033] Fourthly, the present invention provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the above-described method for identifying the water motion state based on spectral moment characteristics.

[0034] The main contributions and innovations of this invention are as follows:

[0035] 1. Strong interpretability of features: Based on frequency domain spectral moment features (such as the zeroth-order spectral moment reflecting total energy and the third-order spectral moment reflecting skewness), it directly characterizes the essential differences between laminar flow (single peak, concentrated energy), turbulent flow (spectral broadening), and recirculation (multiple peaks), solving the problem of poor interpretability of traditional time domain features;

[0036] 2. High recognition accuracy: The multi-scale residual convolutional neural network captures frequency domain features at different scales in parallel using 1×1, 3×3, and 5×5 convolutional kernels. Combined with residual connections, it preserves weak signals (such as backflow multi-peaks). The F1-score of the three states on the validation set is over 95%, which is significantly better than the traditional time-domain statistics + SVM (average F1-score of only 62%).

[0037] 3. Excellent robustness: It maintains an accuracy of 86.5% even at low signal-to-noise ratio (SNR=0dB) and is applicable to multiple scenarios such as rivers, dams, and urban drainage (accuracy ≥93%). It has strong resistance to noise and environmental interference.

[0038] 4. No need for high-density sampling: Frequency domain features can be extracted based on single-point or small-scale sampling data of ADCP, reducing the dependence on spatial sampling density and making it suitable for complex hydrological environments;

[0039] 5. The model is efficient and flexible: end-to-end training does not rely on pre-trained models, the number of parameters is only 3.2M (far lower than AlexNet's 57M), the convergence speed is fast, and it is easy to deploy in engineering.

[0040] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0042] Figure 1 This is a flowchart of a water motion state identification method based on spectral moment features according to an embodiment of the present invention;

[0043] Figure 2 This is a comparison chart of validation set loss rates according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of a multi-scale residual structure according to an embodiment of the present invention;

[0045] Figure 4 This is a block diagram of the neural network structure used in this invention according to an embodiment of the present invention;

[0046] Figure 5 This is a characteristic time-temperature map of laminar flow according to an embodiment of the present invention;

[0047] Figure 6 This is a characteristic time-space heatmap of turbulence according to an embodiment of the present invention;

[0048] Figure 7 This is a characteristic time-temperature map of reflux according to an embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram of the loss curve during the training process according to an embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram of the accuracy curve of the training process according to an embodiment of the present invention;

[0051] Figure 10 This is a schematic diagram of a confusion matrix according to an embodiment of the present invention;

[0052] Figure 11 This is a schematic diagram illustrating the recognition accuracy under different signal-to-noise ratios according to an embodiment of the present invention;

[0053] Figure 12 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0054] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0055] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0056] Existing technologies rely on time-domain or spatial characteristics, making it difficult to reveal the spectral characteristics of water movement. Furthermore, they are not sensitive to the frequency domain differences between laminar, turbulent, and recirculating flows, resulting in insufficient accuracy and robustness in identification.

[0057] Based on this, the present invention addresses the problems existing in the prior art by using deep learning.

[0058] Example 1

[0059] This invention aims to propose a method for identifying the motion state of water bodies based on spectral moment features. Specifically, referring to... Figure 1 The method includes the following steps:

[0060] Step 1: Collect flow velocity sequences as input data using an acoustic Doppler current meter (ADCP);

[0061] In this embodiment, ADCP is based on the acoustic Doppler principle and can measure water flow velocity in a non-contact manner in various water environments such as rivers, waterways, dams, and urban drainage systems. The collected flow velocity sequence can reflect the dynamic changes of water movement and provide basic data support for subsequent conversion of time domain signals into frequency domain features and extraction of key features such as spectral moments. It is the primary data source for achieving accurate identification of water movement status.

[0062] Step 2: Convert the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation;

[0063] In this embodiment, the time-domain signal is converted into the frequency-domain power spectral density through short-time Fourier transform (STFT);

[0064] In the short-time Fourier transform, the original time-domain signal is divided into short time windows. A Fourier transform is performed on each window to obtain the time-varying spectrum. Then, the power spectral density is estimated by averaging the squared amplitudes at each frequency point. The specific formula is as follows:

[0065]

[0066] in This is the original time-domain signal; For window functions (in this method, the Hanning window is selected), and This represents a sliding window centered at t; Position of the sliding window; Indicates frequency; To estimate the power spectrum at that frequency at that moment, the power spectrum estimate is... The overall power spectral density (PSD) is obtained by averaging along the time axis.

[0067] Step 3: Extract features from the frequency domain power spectral density. The features include spectral moment features, frequency band energy, statistical features, and spectral features. Among them, the spectral moment features include the zeroth-order spectral moment (total energy), the first-order spectral moment (spectral centroid), the second-order spectral moment (dispersion), the third-order spectral moment (skewness), and the fourth-order spectral moment (kurtosis). The frequency band energy includes low-frequency energy, mid-frequency energy, and high-frequency energy. The statistical features include the maximum spectral power and the standard deviation of spectral power. The spectral features include the frequency peak position and the spectral power gradient.

[0068] In this embodiment, the formula for calculating the spectral moment is based on the power spectral density P(f) (i.e., the power distribution per unit frequency), by analyzing the frequency range, The power spectral density within the spectrum is integrated to quantize the energy distribution characteristics of the spectrum, as detailed below:

[0069] 1. The formula for the zeroth-order spectral moment (M0) is:

[0070]

[0071] This represents the integration of the power spectral density over the entire frequency range, resulting in the total energy (or total power) of the signal, reflecting the sum of the energy of the flow velocity signal across all frequencies.

[0072] 2. The first spectral moment (M1, spectral barycenter) is as follows:

[0073]

[0074] It is a weighted average of frequency f with power spectral density P(f) as the weight, i.e., the spectral centroid frequency, which reflects the central frequency position of the energy concentration in the spectrum (the energy is mainly distributed around this frequency).

[0075] 3. The second-order spectral moment (M2) is as follows:

[0076]

[0077] The dispersion of frequency components relative to the spectral centroid M1 is called spectral dispersion. The larger the value, the more dispersed the spectral energy distribution.

[0078] 4. The third spectral moment (M3, skewness) is as follows:

[0079]

[0080] Reflects the asymmetry (skewness) of the spectrum distribution: a positive value indicates that the spectrum is biased towards the high-frequency side, a negative value indicates that it is biased towards the low-frequency side, and a value close to 0 indicates that the distribution is symmetrical.

[0081] 5. The fourth spectral moment (M4, kurtosis) is as follows:

[0082]

[0083] Kurtosis is a measure of the sharpness of a spectrum. A higher value indicates the presence of obvious peaks in the spectrum (such as the multi-peak structure of recirculation) and a steeper distribution.

[0084] In summary, these spectral moments comprehensively characterize the frequency domain features of the power spectrum from dimensions such as total energy, concentration location, dispersion degree, symmetry, and sharpness, providing key quantitative basis for distinguishing the motion states of laminar flow, turbulent flow, and recirculation.

[0085] Step 4: Construct the extracted features into feature vectors and input the feature vectors into a preset classifier model (spatiotemporal feature heatmap multi-scale residual convolutional neural network model). The classifier model identifies the current motion state of the water body, which includes laminar flow, turbulent flow, and backflow.

[0086] In this embodiment, a feature vector is constructed and input into a classifier model (such as SVM, CNN, etc.) to identify the current motion state of the water body.

[0087] The final feature vector includes:

[0088] (1) Five spectral moment characteristics ( ), As given in step 3.

[0089] (2) Energy in three frequency bands ( Power spectral density (PSD) is a feature used to quantify the energy distribution of water flow velocity signals in different frequency ranges. It is used to characterize the energy proportion of water flow velocity signals in the low-frequency, mid-frequency, and high-frequency bands, as detailed below:

[0090] Low-frequency energy ( ): refers to the power spectral density in the low-frequency range The integral value within, i.e. This reflects the total energy in the low-frequency range. In laminar flow, the spectral energy is concentrated in the low frequencies, therefore... Usually larger.

[0091] Mid-frequency energy ( ): refers to the power spectral density in the mid-frequency range The integral value within, i.e. This reflects the total energy in the mid-frequency range. Due to the multi-peak distribution of the spectrum, the return current state often exhibits multiple energy peaks in the mid-frequency band. It will exhibit significant fluctuations.

[0092] High-frequency energy ( ): refers to the power spectral density in the high-frequency range The integral value within, i.e. This reflects the total energy in the high-frequency range. Turbulent states exhibit a broadened spectrum with abundant high-frequency components, therefore... Typically higher than laminar flow.

[0093] (3) Two statistics: maximum spectral power (max_power) and standard deviation of spectral power (std_power).

[0094] Specifically, the maximum spectral power (max_power) is:

[0095] set up For frequency The power spectral density value at that location (unit: power / Hz) is then:

[0096]

[0097] The maximum power spectral density is calculated by taking the maximum value from all frequency points. Physically, it represents the power density of the single frequency point with the strongest energy in the spectrum, corresponding to the energy intensity of the dominant frequency in the signal. Specifically, the maximum spectral power represents the maximum power density value corresponding to a single frequency point in the spectrum, reflecting the energy intensity of the main frequency component. It typically corresponds to the location of the dominant frequency of the signal. For laminar flow, the spectrum is concentrated at low frequencies, resulting in a larger maximum power value; for turbulent or recirculating flow, the spectrum broadens, and the maximum power value decreases.

[0098] Let there be a total of Frequency points The corresponding power spectral density is Average power for:

[0099]

[0100] Specifically, the standard deviation of the power spectrum (std_power) is:

[0101]

[0102] Here, std_power represents the dispersion (square root of variance) of the power spectral density value across the entire frequency range, reflecting the flatness or drastic variation of the spectral distribution. For laminar flow, the spectral energy is concentrated, and std_power is usually small; for turbulent or recirculating flow, the energy distribution is more dispersed, and std_power is larger.

[0103] In summary, the eigenvector is denoted as:

[0104]

[0105] Two spectral features (frequency peak position freq_at_max and spectral power gradient power_grad):

[0106] Here, freq_at_max represents the frequency point corresponding to the maximum value in the power spectral density, which is the location of the dominant frequency in the signal. It reflects the frequency component that contributes the most energy to the water's motion.

[0107] set up It is frequency The power spectral density at index [value], with the maximum value at index [value]. Location:

[0108]

[0109] but:

[0110]

[0111] Wherein power_grad represents the average rate of change of power spectral density along frequency, used to measure the slope trend of the spectral curve, and is an important indicator of whether the spectral change is gradual or rapidly decaying.

[0112] For laminar flow, freq_at_max is low, and the dominant frequency is in the low-frequency range. For turbulent flow, freq_at_max is in the mid-frequency range (spectral broadening, dominant frequency shifts upward). For recirculation, freq_at_max exhibits a multi-peak distribution, and the dominant frequency is not obvious or unstable.

[0113] Assume the frequency sequence is The corresponding power spectral density is Then the power gradient is:

[0114]

[0115] If the sampling is uniform, that is If constant, then it simplifies to:

[0116]

[0117] For laminar flow, energy is concentrated at low frequencies, with low energy at the tail end of the spectrum. The value is negative and decreases slowly. For turbulent flow, the high-frequency energy decreases rapidly, indicating spectral broadening. The value is negative, indicating a steep decline. For reflux, the spectrum is multi-peaked, with large local variations and an unclear overall trend. Fluctuations may approach zero.

[0118] In this embodiment, by classifying and identifying the spectral moment feature vectors, the spatiotemporal feature heatmap CNN model proposed in this method can effectively distinguish the water state into the following three categories:

[0119] (1) Laminar flow state: The spectrum has a clear single peak, concentrated energy, and skewness close to 0;

[0120] (2) Turbulent state: spectrum broadening, centroid frequency shifting upward, and significant skewness;

[0121] (3) Reflux state: the spectrum has a multi-peak distribution and a high kurtosis.

[0122] Preferably, the main steps in the spatiotemporal feature heatmap multi-scale residual convolutional neural network model proposed in this method are as follows:

[0123] Step 1: First, convert the above multidimensional feature vectors... Arranged chronologically to form a two-dimensional matrix, a feature time heatmap is constructed, where the horizontal axis is the time index and the vertical axis is the feature dimension index, and the matrix element values ​​are the values ​​under the corresponding time and feature dimension.

[0124] Step 2: To improve the classification accuracy of spatiotemporal feature heatmaps by classic convolutional neural networks, this method also proposes a multi-scale residual block (MSRB) structure for multi-scale feature extraction and enhancement of the input two-dimensional feature map. This is particularly suitable for frequency domain heatmap structures representing the motion state of water bodies, such as... Figure 3 As shown in the figure, BN represents Batch Normalization, and ReLU() represents the ReLU activation function. This structure processes the same input feature map in parallel by introducing multiple convolutional channels with different receptive fields (kernel sizes of 1×1, 3×3, and 5×5), thereby extracting feature information at different spatial scales. Each convolutional branch can capture local (small kernel) and global (large kernel) variation patterns, such as complex features like spectral energy concentration, energy diffusion, and frequency shift.

[0125] This structure concatenates and fuses features at various scales along their channel dimensions at the output, then compresses the number of channels using a 1×1 convolution operation, ensuring the overall output dimension remains consistent with a single branch. To further optimize the information flow path and prevent feature degradation, the structure employs a residual connection mechanism, directly adding the input feature map element-wise to the multi-scale convolution output, thus achieving end-to-end information preservation and enhancement. This multi-scale residual structure combines the ability to extract diverse features with the stable training capabilities of deep networks, exhibiting good generalization ability and robustness when processing input data with multi-scale pattern variations, such as water velocity spectra and feature time heatmaps.

[0126] Step 3: The spatiotemporal feature heatmap multi-scale residual convolutional neural network mainly includes the following functional modules and structural layers:

[0127] (3-1) Input layer module

[0128] The system receives a two-dimensional feature heatmap tensor generated by the data preprocessing module as input. The horizontal axis of the heatmap represents the time step dimension, and the vertical axis represents the feature dimension. Each pixel in the heatmap reflects a quantitative value of a certain frequency domain feature at a specific time step. The input heatmap size is H×WH×WH×W. Typically, HHH is 12 (corresponding to 12 spectral statistical features), and WWW is 60 (representing 60 time steps). The input tensor shape is [B,C,H,W], where B is the batch size and C=1 indicates a single channel.

[0129] To further illustrate the specific optimization of the model proposed in this method, we used the classic AlexNet and trained and tested it on the same dataset. The test data is shown in Table 1 below:

[0130] Table 1

[0131]

[0132] As can be seen, under the same input data, the method of this invention improves the F1 score by about 4.3% compared to AlexNet; the number of model parameters is reduced by about 94.3%, making it significantly lighter; the number of convergence rounds is reduced by more than 60%, and the training efficiency is greatly improved.

[0133] (3-2) Multi-scale residual extraction module (backbone feature extractor)

[0134] This module is the core structure of this invention, responsible for extracting multi-level, multi-scale spectral variation features from the input heatmap. It consists of multiple multi-scale residual blocks (MSRBs) connected in series. Each MSRB module contains multiple parallel convolutional branches, each employing convolutional kernels with different receptive fields (e.g., 1×1, 3×3, 5×5) to extract features such as local edge changes, energy accumulation regions, and spectral abrupt changes within different spatial receptive ranges. The outputs of all branches are concatenated, fused through a 1×1 convolution, and then batch normalized. Finally, they are fused with the input signal via residual connections to preserve low-level features while introducing higher-order feature representation capabilities.

[0135] Multiple MSRBs are stacked sequentially, with pooling operations inserted in between to gradually compress the image spatial dimension while increasing the number of channels (i.e., feature dimension), thereby achieving a deeper semantic understanding.

[0136] Example hierarchical structure:

[0137] MSRB1 (input channel 1 → output channel 16) + max pooling;

[0138] MSRB2 (16→32) + max pooling;

[0139] MSRB3 (32→64) + max pooling.

[0140] Preferably, to illustrate the irreplaceable nature of multi-scale design, using the frequency domain feature time heatmap constructed by ADCP as input, convolutional neural networks (CNNs) with single convolutional kernel sizes (1×1, 3×3, 5×5) are constructed to classify and identify three types of flow regimes: laminar flow, turbulent flow, and recirculation. The model structures are shown in Table 2 below:

[0141] Table 2

[0142]

[0143] The final experimental results are compared in Table 3 below:

[0144] Table 3

[0145]

[0146] In practical applications, the spectral characteristics of different flow regimes exhibit structural differences, making it difficult for a single convolutional kernel to simultaneously adapt to these three different feature patterns. Experimental comparisons reveal that:

[0147] ① Advantages and limitations of 1×1 convolution kernels:

[0148] Because of its smallest receptive field, the 1×1 convolution kernel can only capture local intensity changes at a single point, making it particularly suitable for identifying single-peak structures with concentrated energy in the low-frequency region of laminar flow. In experiments, its accuracy in identifying laminar flow reached 93.2%, but its ability to identify turbulence and recirculation was weak, at only 68.4% and 72.3% respectively, demonstrating insufficient perception of spectral structure changes.

[0149] ②Sensitivity of 3×3 convolution kernels to reflow structures:

[0150] The 3×3 convolutional kernel possesses a moderate receptive field, capable of simultaneously sensing relative changes between two or three local frequency bands, and exhibits excellent recognition ability for multi-peak oscillation structures in the backflow spectrum. The recognition accuracy reaches 91.2%, outperforming other single-scale networks. Furthermore, it demonstrates a relatively balanced performance in laminar and turbulent flows, indicating that this scale possesses good local information fusion capabilities.

[0151] ③ The ability of a 5×5 convolution kernel to sense the diffusion of turbulent energy:

[0152] The 5×5 convolution kernel has a large receptive field, making it suitable for extracting global diffusion features such as spectral broadening and sideband enhancement in turbulent flow. It achieves a 92.4% accuracy rate in turbulent flow identification, but performs poorly in laminar flow scenarios, misinterpreting concentrated spectral energy as disturbance, with an accuracy of only 82.1%.

[0153] The above results indicate that:

[0154] Single-scale convolution kernels exhibit significant target bias; laminar flow requires a 1×1 kernel to extract concentrated main peak energy; turbulent flow requires a 5×5 kernel to integrate and broaden the structure; recirculation requires a 3×3 kernel to resolve mid-frequency perturbations and multi-peak information; any kernel size cannot simultaneously account for the characteristics of the three types of spectra, resulting in significant performance loss.

[0155] Therefore, the multi-scale convolutional fusion structure (1×1+3×3+5×5) proposed in this invention can extract multiple scale feature responses in parallel under a single feature heatmap input. Through subsequent feature fusion and residual enhancement mechanisms, it can significantly improve the model's recognition accuracy and robustness for complex water body spectral structures.

[0156] (3-3) Classification Mapping Module (Fully Connected Layer)

[0157] The multi-scale residual feature maps extracted by the above convolutional modules are then flattened in space before being fed into the fully connected layer. This part typically includes:

[0158] A high-dimensional fully connected layer is responsible for integrating multi-channel convolutional features;

[0159] Dropout random deactivation mechanism to prevent overfitting;

[0160] The output layer is a fully connected layer activated by Softmax, and the output dimension is the number of predefined classification categories (3 types of flow states).

[0161] (3-4) Output module

[0162] The network output is the predicted probability distribution of each input heatmap sample in each category, which can be used to make the maximum probability judgment, thereby realizing the automatic classification of water body status.

[0163] The proposed neural network structure diagram is as follows: Figure 4 As shown.

[0164] Step 4: Neural Network Training Methods and Usage Procedures:

[0165] (a) Training data preparation

[0166] The original flow velocity signal is subjected to short-time Fourier transform (STFT); frequency domain statistical features (such as spectral moments, energy ratio, maximum spectral power, frequency shift, etc.) are extracted to form a multidimensional feature vector; after extracting features at each time step, they are combined in time sequence to form a feature time heatmap, which constitutes a two-dimensional tensor sample; the flow state label corresponding to each sample is labeled and used as a training supervision signal.

[0167] (b) Model initialization

[0168] Initialize neural network parameters, including convolutional kernel weights, Batch Normalization (BN) parameters, attention module weights, etc.; define loss function (such as cross-entropy loss); select optimizer (such as Adam) and learning rate strategy.

[0169] (c) Training iteration

[0170] In each training round, samples are fed into the network in batches; they pass through the multi-scale residual module, pooling module, and fully connected module in sequence to output the predicted probability; the loss value between the predicted result and the actual label is calculated; the network parameters are updated through the backpropagation algorithm; and the iteration is repeated until the loss converges or the accuracy is achieved.

[0171] (d) Model evaluation and reasoning

[0172] After training, the model is placed in evaluation mode and new heatmap sample data is input for identification. The output category prediction results can be used for further flow regime discrimination, hydraulic engineering monitoring, or control decision systems.

[0173] In this embodiment, to illustrate the effectiveness of the method proposed in this invention, an application example analysis is provided as follows:

[0174] Using raw flow velocity data acquired by an ADCP installed in a river channel, this method identifies the water movement state based on the collected flow velocity data and compares it with expert judgment results to demonstrate the effectiveness of the proposed method. Specific data are as follows:

[0175] The total dataset consists of data from 1 million sampling times, manually labeled and categorized into three flow states: laminar flow, turbulent flow, and recirculation. Each dataset contains 5000 data points. A portion of this data is presented as a feature time-space heatmap. Figure 5-7 As shown, the loss curve during the training process is as follows: Figure 8 As shown, the accuracy curve is as follows: Figure 9 As shown.

[0176] After training using the method of this invention, practical applications were conducted, and the accuracy of various indicators on the validation set is shown in Table 4 below:

[0177] Table 4

[0178]

[0179] The confusion matrix is ​​as follows Figure 10 As shown, by combining the above indicators and confusion matrix, it can be concluded that the method of the present invention can accurately distinguish the three states, and is effective and feasible in practical applications.

[0180] Preferably, to illustrate the superiority of this method over traditional methods, the effectiveness of the following two classification methods in identifying water movement states (laminar / turbulent / recirculation) is compared, as shown in Table 5 below:

[0181] Table 5

[0182]

[0183] The experimental results are compared in Table 6 below:

[0184] Table 6

[0185]

[0186] Comparative analysis of experimental results reveals the limitations of traditional time-domain methods:

[0187] 1. Weak ability to express features:

[0188] Time-domain statistics (such as mean and standard deviation) only capture the overall trend of the signal and cannot describe the spectral structure. For return current, which has a "multi-peak + dispersed" spectral state, the F1-score is only 50%.

[0189] 2. Severe information loss:

[0190] The differences in high-frequency / low-frequency / mid-frequency disturbances in water bodies were ignored, especially the key spectral line shapes such as energy abrupt change points.

[0191] 3. Limited model learning ability:

[0192] SVM is a shallow model that can only perform classification under linear or kernel function mappings and has poor perception of nonlinear spectral structures.

[0193] The significant advantages of the method of this invention are:

[0194] 1. Automatic multi-scale feature extraction:

[0195] Multi-scale convolution kernels (1×1, 3×3, 5×5) can learn spectral features such as local peaks, sideband broadening, and multi-peak oscillations; they are particularly suitable for distinguishing laminar flow (concentrated main peaks), turbulent flow (spectral broadening), and recirculation (multi-peak perturbation) states.

[0196] 2. Residual connectivity enhances weak feature preservation:

[0197] It effectively solves the problem of "weak peak energy being submerged by mainstream" such as backflow; shallow low-frequency signals can be directly transmitted to deep layers, improving discrimination ability.

[0198] 3. Stronger training stability and generalization ability:

[0199] Multi-scale structures avoid overfitting at a single scale; mechanisms such as BatchNorm and Dropout ensure that the network adapts well to new samples.

[0200] In this embodiment, to verify the noise immunity performance of the present invention, such as... Figure 11 As shown, under different signal-to-noise ratio conditions, the method of this invention consistently maintains a high classification accuracy, especially achieving an accuracy of 86.5% even with SNR=0dB, which is far higher than the underwater acoustic signal model in document 2 (only 60.4%). This fully demonstrates the following key advantages:

[0201] Multi-scale convolution design forms a complementary structure that can effectively model key features under various perturbation frequency distributions, exhibiting good information redundancy and structural robustness. Detailed information representing water body characteristics (such as local peak-valley structures) in the original spectrum is easily "submerged" by noise signals. Residual connections allow shallow features (such as low-frequency energy concentration areas) to directly jump and propagate to higher layers, effectively preserving the original spectral structure information.

[0202] In this embodiment, to verify whether the method of the present invention is applicable to actual water scenarios with different physical conditions, flow velocity states, and boundary disturbance intensities, rather than being limited to a specific test tank or experimental environment, the following experiment was conducted:

[0203] (1) Test scenario and description:

[0204] River scene: Open waterway with natural laminar flow, backflow along the bank and turbulent flow after rain; water body is greatly affected by wind speed disturbance, and signal-to-noise ratio is unstable.

[0205] Dam spillway: high flow velocity, obvious turbulence, and frequent high-frequency scattering disturbances; the backflow zone and the discharge wave repeatedly intersect.

[0206] Urban drainage systems have complex pipe structures with obvious switching between intermittent laminar flow, backflow, and interference flow; the channel space is narrow and there is a lot of interference.

[0207] (2) Summary of recognition accuracy results

[0208] The recognition accuracy in various scenarios is shown in Table 7 below:

[0209] Table 7

[0210]

[0211] Analysis of the above recognition accuracy results shows that the accuracy rate is consistently above 93% in all scenarios; especially in urban drainage systems, this method still exhibits high robustness despite interference from factors such as high noise and multipath echoes.

[0212] Example 2

[0213] Based on the same concept, this invention also proposes a water motion state identification device based on spectral moment features, comprising:

[0214] The data acquisition module uses an acoustic Doppler flow meter to collect flow velocity sequences as input data.

[0215] The frequency domain transformation module converts the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation;

[0216] The feature extraction module extracts features from the frequency domain power spectral density. These features include spectral moment features, frequency band energy, statistical features, and spectral features. The spectral moment features include zero-order, first-order, second-order, third-order, and fourth-order spectral moments. The frequency band energy features include low-frequency, mid-frequency, and high-frequency energy. The statistical features include maximum spectral power and spectral power standard deviation. The spectral features include frequency peak position and spectral power gradient.

[0217] The processing module constructs the extracted features into feature vectors and inputs the feature vectors into a preset classifier model. The classifier model identifies the current motion state of the water body, which includes laminar flow, turbulent flow, and backflow.

[0218] The output module outputs the recognition results of the classifier model.

[0219] Preferably, to address the problem of "weak signal submersion" in water features (such as the dispersion of multi-peak signal energy in backflow, where residual connections can preserve shallow low-frequency characteristics), this invention solves the "weak signal submersion" problem through the following three methods:

[0220] (1) Retain low-frequency, low-amplitude characteristics:

[0221] Weak signals, such as the multi-peak spectrum of the reflow state, often have weak oscillations in the low-frequency region of the spectrum. These features can be detected well in the initial convolutional layer (smaller convolutional kernels or 1×1 convolutions are sensitive to local peaks), but after multiple downsampling (Pooling) and higher-order convolutions, this information often disappears.

[0222] Residual connections can directly sum and inject these shallow, low-frequency features into the deep output, preserving the in-situ information of these features in space, providing multi-view references in the channel dimension, and solving the problem of information attenuation in the propagation path.

[0223] (2) Avoid gradient vanishing and enhance gradient signal penetration through weak feature paths:

[0224] From a training perspective, residual connections form an "identity mapping path," allowing gradients in weak feature regions to be directly backpropagated to the previous layer without undergoing nonlinear transformations. This increases the probability of updating the weights in these regions, making networks with residual structures more robust to learning low-amplitude inputs (such as background changes or small peaks).

[0225] This is crucial for weak multi-peak patterns that are "masked" by high-power mainstream signals in the early stages of training, ensuring that the model does not get stuck in local optima where it "only learns strong signals".

[0226] (3) Enhanced detail expression through multi-scale perspectives:

[0227] In the multi-scale residual structure of this invention, convolutional branches of different scales are responsible for capturing:

[0228] 1×1 convolution: sensitive to pixel-level energy (differences in strength);

[0229] 3×3 convolution: local fluctuation features (such as two nearest neighbor frequency peaks);

[0230] 5×5 convolution: medium frequency structural variations (such as the spacing between the main and secondary peaks in the reflux spectrum);

[0231] By preserving the overall input information through residual connections and then fusing these scale features, the inter-spectral position information of multi-peak weak signals can be effectively preserved, avoiding the problems of "being averaged" and "being compressed".

[0232] Preferably, to verify the comparison of model loss curves with and without residual connections, such as... Figure 2 As shown, the accuracy of the validation set is improved by 6.35% when residual connections are present.

[0233] Example 3

[0234] This embodiment also provides an electronic device, see reference. Figure 12 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0235] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0236] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0237] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0238] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the water motion state identification methods based on spectral moment features in the above embodiments.

[0239] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0240] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0241] Input / output device 408 is used to input or output information.

[0242] Example 4

[0243] This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the water motion state identification method based on spectral moment features according to Embodiment 1.

[0244] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0245] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0246] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0247] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0248] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for identifying the motion state of water bodies based on spectral moment features, characterized in that, Includes the following steps: Step 1: Collect flow velocity sequences using an acoustic Doppler current meter as input data; Step 2: Convert the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation; Step 3: Extract features from the frequency domain power spectral density. The features include spectral moment features, frequency band energy, statistical features, and spectral features. The spectral moment features include zero-order spectral moment, first-order spectral moment, second-order spectral moment, third-order spectral moment, and fourth-order spectral moment. The frequency band energy includes low-frequency energy, mid-frequency energy, and high-frequency energy. The statistical features include maximum spectral power and spectral power standard deviation. The spectral features include frequency peak position and spectral power gradient. Step 4: Construct the extracted features into feature vectors, arrange the feature vectors in chronological order to form a two-dimensional feature time heatmap, and then input the feature time heatmap into a preset convolutional neural network classifier model. The classifier model identifies the current motion state of the water body, which includes laminar flow, turbulent flow, and backflow. The convolutional neural network includes a multi-scale residual extraction module, which is composed of multiple multi-scale residual structures connected in series. Each multi-scale residual structure has multiple parallel convolutional branches, and the multiple parallel convolutional branches use convolutional kernels with different receptive fields.

2. The water motion state identification method based on spectral moment features as described in claim 1, characterized in that, In step 2, the frequency domain transformation is a short-time Fourier transform. The process of converting the time-domain signal into the frequency-domain power spectral density through the short-time Fourier transform includes: The original time-domain signal is divided into short time windows. A Fourier transform is performed on each short time window to obtain the time-varying spectrum. The squared amplitude of each frequency point is averaged to estimate the power spectral density. The power spectral density is then averaged along the time axis to obtain the overall power spectral density.

3. The water motion state identification method based on spectral moment features as described in claim 1, characterized in that, In step 3, the zeroth spectral moment is the total energy of the power spectral density, the first spectral moment is the spectral centroid, the second spectral moment is the dispersion, the third spectral moment is the skewness, and the fourth spectral moment is the kurtosis.

4. The water motion state identification method based on spectral moment features as described in claim 1, characterized in that, In step 3, the low-frequency energy is the integral value of the power spectral density in the low-frequency range, the mid-frequency energy is the integral value of the power spectral density in the mid-frequency range, and the high-frequency energy is the integral value of the power spectral density in the high-frequency range. The maximum spectral power is the maximum power density value corresponding to a single frequency point in the spectrum, and the standard deviation of the spectral power is the degree of dispersion of the power spectral density value over the entire frequency range; The frequency peak position is the frequency point corresponding to the maximum value in the power spectral density, and the spectral power gradient is the average rate of change of the power spectral density along the frequency.

5. The water motion state identification method based on spectral moment features as described in claim 1, characterized in that, The convolution kernels with different receptive fields include 1×1, 3×3 and 5×5 convolution kernels; In each of the multi-scale residual structures, the outputs of all convolutional branches are concatenated through channels, fused by a 1×1 convolution, and then batch normalized before being fused with the input signal via residual connection.

6. The water motion state identification method based on spectral moment features as described in any one of claims 1 to 5, characterized in that, The identification results of the water movement state include laminar flow, turbulent flow, and backflow. Among them, the laminar flow state has a distinct single peak in the spectrum, concentrated energy, and a skewness close to 0; the turbulent flow state has a broadened spectrum, an upward shift in the center of gravity frequency, and a significant skewness; and the backflow state has a multi-peaked spectrum with a high kurtosis.

7. A water motion state identification device based on spectral moment characteristics, characterized in that, include: The data acquisition module uses an acoustic Doppler flow meter to collect flow velocity sequences as input data. The frequency domain conversion module converts the time-domain signal of the flow velocity sequence into the frequency-domain power spectral density through frequency domain transformation; The feature extraction module extracts features from the frequency domain power spectral density. The features include spectral moment features, frequency band energy, statistical features, and spectral features. The spectral moment features include zero-order spectral moment, first-order spectral moment, second-order spectral moment, third-order spectral moment, and fourth-order spectral moment. The frequency band energy includes low-frequency energy, mid-frequency energy, and high-frequency energy. The statistical features include maximum spectral power and spectral power standard deviation. The spectral features include frequency peak position and spectral power gradient. The processing module constructs the extracted features into feature vectors, arranges the feature vectors in chronological order to form a two-dimensional feature time heatmap, and then inputs the feature time heatmap into a preset convolutional neural network classifier model to identify the current motion state of the water body, wherein the motion state of the water body includes laminar flow, turbulent flow and backflow; The output module outputs the recognition results of the classifier model; The convolutional neural network includes a multi-scale residual extraction module, which is composed of multiple multi-scale residual structures connected in series. Each multi-scale residual structure has multiple parallel convolutional branches, and the multiple parallel convolutional branches use convolutional kernels with different receptive fields.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the water motion state identification method based on spectral moment features as described in any one of claims 1 to 6.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program code for controlling the process to execute the process, the process including the water motion state identification method based on spectral moment features according to any one of claims 1 to 6.

Images