Millimeter wave water surface flow velocity measurement radar echo attribute classification and identification method

Abstract

The invention discloses a millimeter wave river surface flow velocity monitoring radar echo attribute classification and identification method. The method specifically comprises the following steps: firstly, carrying out depth extraction on echo data features by adopting a time sequence convolutional network; then establishing a long-term dependency relationship of data by utilizing a double-layer gating circulation unit; finally, in order to improve the overall performance of the network and distribute computing resources emphatically, an attention-based channel sensing module is introduced, the network is focused on a key feature channel, and the convergence speed is increased. The method can show excellent recognition accuracy in the application of millimeter wave water surface flow velocity measurement radar echo attribute classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and in particular relates to a method for classifying and recognizing attributes of millimeter-wave water surface velocity measurement radar echoes. Background Art

[0002] In recent years, with the advancement of hydrological monitoring technology, various types of flow measurement equipment have been applied to river flow velocity measurement. Among them, radar velocimeters have garnered widespread attention due to their non-contact measurement capabilities. While traditional Doppler ultrasonic velocimeters can measure flow velocity, their accuracy is often compromised in waters with high sediment content or the presence of bubbles. In contrast, microwave and millimeter-wave radar velocimeters offer greater environmental adaptability and are capable of stable measurement in complex water conditions. Several typical radar flow velocity measurement systems have emerged in recent years. The Pisces HF radar system, installed along the coast of England and Wales in 2003, is a novel non-contact water velocity measurement device. Utilizing high-frequency microwave radar technology, it boasts high accuracy, strong anti-interference capabilities, and the ability to process data in real time and display it on a website. However, its period and directional parameters are easily contaminated by radar noise, resulting in poor performance in rough seas. In my country, the S3-SVRⅡ measurement system developed by Wang Wenhua's team meets the accuracy requirements for hydrological monitoring of small and medium-sized rivers. Operating on a dedicated cableway, it boasts fast speeds and short measurement times, making it adaptable to harsh environments such as flood surges and turbulent currents. However, to further improve velocity measurement accuracy, it is necessary to replace the radar with one with a smaller electromagnetic beam angle or lower the measurement altitude. Any strong reflection within the radar's projection range on the water surface may be mistaken for the flow velocity at the measuring point, resulting in greater uncertainty in the measurement location of the flow velocity. Existing non-contact measurement equipment can generally only measure the flow velocity at a single point, is easily interfered with by other reflection sources in the environment, and generally lacks distance resolution capabilities. In order to reduce repetitive work and improve velocity measurement accuracy, designing a device capable of distance resolution, velocity measurement, and target classification, and capable of simultaneously measuring the surface flow velocity at multiple points within the same section, has naturally become a popular research target. In addition to high accuracy and a wide measurement range, the device should also have the advantages of being easy to install, simple to operate, and adaptable to a variety of hydrological conditions.

[0003] In millimeter-wave flow velocity measurement, radar echo signals contain multiple components from the water surface, river channel structure, floating objects, and underwater obstacles. Accurately extracting water echoes becomes a key technical challenge. Radar Automatic Target Recognition (RATR) leverages the scattering characteristics of targets to extract and classify different types of echoes, thereby improving the accuracy of flow velocity measurements. With the development of machine learning and deep learning, intelligent classification methods have been widely applied to radar echo signal analysis. For example, a team from the UK developed a convolutional neural network model that uses micro-Doppler spectrum images to identify drones and birds. Field testing has shown that this approach achieves a recognition accuracy of 99%. A team in China has also designed a deep convolutional denoising encoder that effectively removes noise without suppressing micro-Doppler features. Combined with deep residual learning to train the network, this significantly reduces the training burden and improves learning efficiency. In millimeter-wave surface flow velocity measurement, the targets to be classified primarily include steady water flow, turbulent water flow, and non-flowing targets (such as ships and floating objects). Among them, the velocity information of steady water flow is the key data for flow velocity monitoring; the measurement results of turbulent water flow need to be filtered and data corrected due to its unstable flow state; and the echo information of non-water flow targets needs to be eliminated to avoid errors in flow velocity calculation. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for classifying and identifying echo attributes of a millimeter-wave river surface velocity monitoring radar.

[0005] The technical solution for achieving the purpose of the present invention is: a method for classifying and identifying echo attributes of millimeter-wave river surface velocity monitoring radar, comprising the following steps:

[0006] Step 1: transform and process the echo data collected from the millimeter-wave river surface velocity monitoring radar to obtain time domain series data, power spectrum data, and power spectrum power transform spectrum data as input data for the convolutional neural network;

[0007] Step 2: The three-channel input data converted in step 1 is fed into a convolutional neural network. The network consists of three temporal convolution modules to extract feature information from the input data.

[0008] Step 3: The feature information processed in step 2 is fed into the channel perception module, which assigns higher weights to the more critical data channels through calculation; the weights are then multiplied by the feature information to obtain the processed feature information;

[0009] Step 4: Feed the feature information from step 3 into a two-layer gated recurrent unit, which contains 50 positive and negative hidden units, for a total of 100-dimensional feature output, to extract global dependencies.

[0010] Step 5: After passing the feature data extracted in step 4 through the flattening layer and the fully connected layer, it is sent to the Softmax function to calculate the final classification result.

[0011] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.

[0012] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0013] A computer program product comprises a computer program, which implements the steps of the above method when executed by a processor.

[0014] Compared with the existing technology, the present invention has the following significant advantages: (1) The present invention adopts a temporal convolutional network based on the causal convolution method to efficiently extract the temporal relationship in the input data; (2) The present invention adopts DGRU to construct long-term dependency relationships, which greatly reduces the computational complexity and improves the operation efficiency while ensuring the efficient flow of gradient information; (3) The present invention adopts a simple attention mechanism and channel perception module to amplify the eigenvalues of important channels, allowing the network to focus more on key feature channels and accelerate the network convergence speed.

[0015] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a network structure diagram of the present invention.

[0017] Figure 2 It is the structure diagram of the temporal convolution block.

[0018] Figure 3 This is the structure diagram of the channel perception module.

[0019] Figure 4 is the confusion matrix obtained after the network is trained.

[0020] Figure 5 This is a comparison chart of the recognition accuracy curves of the network in the present invention and the multi-target classification method based on one-dimensional convolutional neural network (1D-CNN), the multi-target classification method based on long short-term memory network (LSTM), and the multi-target classification method based on the use of temporal convolutional network (TCN) alone, a total of four methods. DETAILED DESCRIPTION

[0021] This paper proposes a method for classifying and recognizing the echo attributes of millimeter-wave river surface velocity monitoring radar. First, a convolutional neural network (CNN) is used to extract data features. Causal convolution and residual connection methods are introduced to form a temporal convolutional network (TCN), which combines certain temporal feature capture capabilities with high computational efficiency. Subsequently, a double-layer gated recurrent unit (GRU) is used to reduce parameters and improve computational efficiency while retaining the ability to model long-term dependencies. A simple attention mechanism and channel perception module are added to dynamically calculate the weight of each channel to highlight the most important features and suppress redundant information, allowing the network to focus more on key feature channels and accelerate convergence. This method has achieved high recognition accuracy in the application of millimeter-wave river surface velocity monitoring radar echo attribute classification.

[0022] like Figure 1 As shown, this network uses a convolutional neural network for data feature extraction and introduces causal convolution and dilated convolution mechanisms to enhance its ability to construct temporal relationships. After three convolutional modules, it is fed into a bidirectional gated recurrent unit to establish more global, long-term dependencies. This allows gradient information to flow more freely, helps extract and compress higher-dimensional feature information and global dependencies, alleviates the problem of vanishing or exploding gradients, and effectively improves the convolutional network's ability to process sequential data. Furthermore, a simple attention mechanism and channel-aware modules are added to focus limited attention on information of interest, improving task efficiency and accelerating convergence.

[0023] like Figure 2 As shown in Figure 1, TCN, which combines causal convolution, dilated convolution, and residual connections, demonstrates strong performance in time series modeling. Causal convolution ensures the accuracy of time series predictions and prevents future information leakage; dilated convolution effectively expands the receptive field without increasing parameters; and residual connections improve gradient flow, enabling efficient training of deeper networks. These characteristics give TCN a significant advantage in time series tasks such as classification, regression, and forecasting.

[0024] like Figure 3 As shown in Figure 2, the core goal of channel perception is to dynamically calculate the weight of each channel to highlight the most important features and suppress redundant information. Because channel perception relies only on global pooling and two layers of full connectivity, it has a low computational load and can be easily embedded in various model structures. It can also be used with other temporal attention algorithms to further process data.

[0025] The specific steps of the millimeter wave river surface velocity monitoring radar echo attribute classification and identification method of the present invention are described in detail below. The method includes:

[0026] Step 1: The echo data collected from the millimeter-wave river surface velocity monitoring radar is transformed and processed to obtain time domain series data, power spectrum data and power spectrum power transform spectrum data as the input data of the network.

[0027] Step 2: Feed the three-channel input data converted in Step 1 into a convolutional neural network. The network consists of three temporal convolutional modules, which are used to extract feature information from the input data. The convolutional block structure specifically includes causal convolution, pooling layer, normalization layer, and activation layer, and gradients are propagated between the convolutional blocks using residual connections.

[0028] Step 2-1, the calculation method of causal convolution is:

[0029] Assume that CNN has L layers, and the first layer has K l filters with weight W (l,k) ∈R h×w×C , h and w represent the height and width of the filter, k represents the kth filter of the layer, and C is the number of channels; b (l,k) is the bias of the filter, and the convolution step size is s c , convolve the input data X to get:

[0030] Y (l,k) ∈f(X*W (l,k) +b (l,k) )

[0031] In causal convolution, the output y at time t is t Depends only on the current and past input x ≤t , will not use future information x ＞t The specific convolution algorithm is:

[0032]

[0033] Among them, k is the size of the convolution kernel, w i is the convolution kernel parameter. To increase the receptive field during convolution, the dilated convolution method is introduced to insert holes (intervals) between the convolution kernel elements. The calculation method is:

[0034]

[0035] Where d is the expansion rate, and the receptive field is expanded layer by layer through exponentially increasing expansion rates (such as d = 1, 2, 4, 8, ...).

[0036] Step 2-2, the calculation method of the pooling layer is:

[0037] Assume that the input data size before pooling is H×W and the size of the pooling kernel is k h ×k w , s p , then after pooling, the size of the data will be compressed to

[0038]

[0039] Step 2-3, the calculation method of the normalization layer is:

[0040] Assume that the input data set is B={x1,x2,...,x m}, at this time, the mean and variance of the samples are:

[0041]

[0042] Normalize the sample using its mean and standard deviation

[0043]

[0044] Where ε represents a small error to prevent the denominator from being zero.

[0045] Since the normalization process will destroy the learned features, the linear transformation method is introduced to add the learnable scaling coefficient γ and translation coefficient β. At this time, the normalization algorithm of the i-th input data is:

[0046]

[0047] Steps 2-4, the calculation method of the activation layer is:

[0048] The expression of LeakyReLU is:

[0049]

[0050] The expression of LeakyReLU derivative is:

[0051]

[0052] Among them, α is a very small value (such as 0.01), which represents the slope of the negative area to avoid potential neuron death problems; x is the function independent variable.

[0053] Steps 2-5: Residual connections are used between convolutional blocks to add the input directly to the output, enabling the network to transfer information across layers.

[0054] The calculation method of the residual connection is:

[0055] y t (res) =F(xt )+x t

[0056] Among them, F(x t ) represents the transformation of the convolutional layer. Residual connections provide an unobstructed gradient path, effectively alleviating the vanishing gradient problem and allowing the network to learn features close to the identity mapping, leading to more stable convergence.

[0057] Step 3: Send the feature information processed in step 2 to the channel perception module, which assigns higher weights to more critical data channels through calculation. The weights are then multiplied by the feature information to obtain the processed feature information. The calculation method of the channel perception module is:

[0058]

[0059] where s c is the global statistical feature of the cth channel, H, W are the height, width and number of channels of the input feature respectively, x i,j,c Represents the pixel value at position (i, j) in the cth channel; a two-layer fully connected network is used to perform nonlinear changes on the channel statistical feature s to obtain the channel weight:

[0060] w=σ(W2·δ(W1·s))

[0061] In the formula are all weight matrices of the fully connected layer, C is the number of channels, r is the compression factor to reduce computational complexity, δ is the ReLU activation function, and σ is the Sigmoid activation function.

[0062] Step 4: Feed the feature information from step 3 into a double-layer gated recurrent unit, which contains 50 positive and negative hidden units, with a total of 100-dimensional feature output, to further extract the global dependency relationship. The update process of DGRU is:

[0063] Let the input sequence be x t , represents the input of the t-th time step. The hidden state sequence is h t . Reset the gate vector to:

[0064] r t =σ(W r x t +U r h t-1 +b r )

[0065] Where W r with U r is the input weight matrix and the weight matrix of the previous hidden state, b ris the bias term. σ(·) is the Sigmoid activation function. Select the hidden state Reset gate r t Impact:

[0066]

[0067] Where W h with U h is the input weight matrix and the weight matrix of the previous hidden state, b h is the bias term; the calculation method of the update gate is:

[0068] z t =σ(W z x t +U z h t-1 +b z )

[0069] Where W z with U z is the input weight matrix and the weight matrix of the previous hidden state, b z is the bias term; the hidden state is updated through the gate z t Candidate status and the previous hidden state h t-1 Weighted decision, calculated as follows:

[0070]

[0071] z t The closer it is to 1, the more historical information it tends to retain; t The closer it is to 0, the more attention is paid to the input of the current time step.

[0072] Step 5: After passing the feature data extracted in step 4 through the flattening layer and the fully connected layer, it is sent to the Softmax function to calculate the final classification result. The calculation method of the fully connected layer and Softmax is:

[0073]

[0074] Among them, FC w,b (x (l) ) is the input data x at the first moment (l) The fully connected layer expression, W FC is the weight matrix, b FC is the bias term, T is the transposed matrix;

[0075] The Softmax function is defined as:

[0076]

[0077] in, represents the probability that the input sample z is classified into the i-th category.

[0078] The network is used to classify water surface echoes into three categories: steady water flow, turbulent water flow and ships.

[0079] From the database, 8,000 samples from each of the three categories—steady current, choppy current, and boats—were randomly sampled, totaling 24,000 data samples to ensure the final model's high stability. 6,000 samples from each category were extracted to form the training set, and the remaining 2,000 samples were used as the test set. These two sets were independent of each other, ensuring that the test set accurately reflected the model's classification results. Each training session was repeated 500 times, allowing the model to fully learn the underlying features of the dataset and achieve excellent classification performance.

[0080] like Figure 4 As shown in Figure 2, the confusion matrix of the network shows that the final recognition accuracy of the hybrid model is about 91.6%.

[0081] like Figure 5 As shown, the accuracy change curves of the network in the present invention are compared with the multi-target classification method based on one-dimensional convolutional neural network (1D-CNN), the multi-target classification method based on long short-term memory network (LSTM), and the multi-target classification method based on the use of temporal convolutional network (TCN) alone. Observing the curve graph, it can be seen that the fusion model fully combines the advantages of convolutional network in parallel computing and the ability of recurrent neural network to capture global temporal features. Its performance is more superior than that of a single model structure, and the additional channel attention and simplified structure make its convergence speed significantly higher than other models, which can effectively improve the classification effect of the three types of targets.

Claims

1. A method for classifying and identifying echo attributes of millimeter-wave river surface velocity monitoring radar, characterized in that: The following steps are involved: Step 1: transform and process the echo data collected from the millimeter-wave river surface velocity monitoring radar to obtain time domain series data, power spectrum data, and power spectrum power transform spectrum data as input data for the convolutional neural network; Step 2: The three-channel input data converted in step 1 is fed into a convolutional neural network. The network consists of three temporal convolution modules to extract feature information from the input data. Step 3: The feature information processed in step 2 is fed into the channel perception module, which assigns higher weights to the more critical data channels through calculation; the weights are then multiplied by the feature information to obtain the processed feature information; Step 4: Feed the feature information from step 3 into a two-layer gated recurrent unit, which contains 50 positive and negative hidden units, for a total of 100-dimensional feature output, to extract global dependencies. Step 5: After passing the feature data extracted in step 4 through the flattening layer and the fully connected layer, it is sent to the Softmax function to calculate the final classification result.

2. The millimeter wave river surface velocity monitoring radar echo attribute classification and identification method according to claim 1 is characterized in that: The network structure of step 2 specifically includes causal convolution, pooling layer, normalization layer and activation layer, and the gradient is propagated between each convolution block through the residual connection method; (1) The calculation method of causal convolution is: Assume that CNN has L layers, and the first layer has K l filters with weight W (l,k) ∈R h×w×C , h and w represent the height and width of the filter, k represents the kth filter of the layer, and C is the number of channels; b (l,k) is the bias of the filter, and the convolution step size is s c , convolve the input data X to get: Y (l,k) ∈f(X*W (l,k) +b (l,k) ) In causal convolution, the output y at time t is t Depends only on the current and past input x ≤t , will not use future information x ＞t ; The specific convolution algorithm is: Among them, k is the size of the convolution kernel, w i is the convolution kernel parameter; in order to increase the receptive field during convolution, the dilated convolution method is introduced to insert holes between the convolution kernel elements. The calculation method is: Where d is the expansion rate, and the receptive field is expanded layer by layer through the exponentially growing expansion rate; (2) The calculation method of the pooling layer is: The input data size before pooling is H×W, and the size of the pooling kernel is k h ×k w , the pooling step size is s p , then after pooling, the size of the data will be compressed to (3) The calculation method of the normalization layer is: Assume that the input data set is B={x1,x2,...,x m }, at this time, the mean μ of the sample B and variance They are: Normalize the sample using its mean and standard deviation In the formula, ε represents a small error to prevent the denominator from being zero; Since the normalization process will destroy the learned features, the linear transformation method is introduced to add the learnable scaling coefficient γ and translation coefficient β. At this time, the normalization algorithm of the i-th input data is: (4) The calculation method of the activation layer is: The expression of LeakyReLU is: The expression of LeakyReLU derivative is: Among them, α is a very small value, which represents the slope of the negative area, and x is the independent variable of the function; (5) The calculation method of residual connection is: y t (res) =F(x t )+x t Among them, F(x t ) represents the transformation of the convolutional layer.

3. The millimeter wave river surface velocity monitoring radar echo attribute classification and identification method according to claim 1 is characterized in that: The calculation method of the channel perception module in the network structure of step 3 is: where s c is the global statistical feature of the cth channel, H, W are the height, width and number of channels of the input feature respectively, x i,j,c Represents the pixel value at position (i, j) in the cth channel; a two-layer fully connected network is used to perform nonlinear changes on the channel statistical feature s to obtain the channel weight: w=σ(W2·δ(W1·s)) In the formula are all weight matrices of the fully connected layer, C is the number of channels, r is the compression factor, δ is the ReLU activation function, and σ is the Sigmoid activation function.

4. The millimeter wave river surface velocity monitoring radar echo attribute classification and identification method according to claim 1 is characterized in that: The update process of the double-layer gated recurrent unit in step 4 is: Let the input sequence be x t , represents the input of the t-th time step; the hidden state sequence is h t ; Reset gate vector to: r t =σ(W r x t +U r h t-1 +b r ) Where W r with U r is the input weight matrix and the weight matrix of the previous hidden state, b r is the bias term; σ(·) is the Sigmoid activation function; select the hidden state Reset gate r t Impact: Where W h with U h is the input weight matrix and the weight matrix of the previous hidden state, b h is the bias term; the calculation method of the update gate is: z t =σ(W z x t +U z h t-1 +b z ) Where W z with U z is the input weight matrix and the weight matrix of the previous hidden state, b z is the bias term; the hidden state is updated through the gate z t Candidate status and the previous hidden state h t-1 Weighted decision, calculated as follows: z t The closer it is to 1, the more historical information it tends to retain; t The closer it is to 0, the more attention is paid to the input of the current time step.

5. The millimeter wave water surface velocity measurement radar echo attribute classification and identification method according to claim 1 is characterized in that: The calculation method of the fully connected layer and Softmax in step 5 is: Among them, FC w,b (x (l) ) is the input data x at the first moment (l) The fully connected layer expression, W FC is the weight matrix, b FC is the bias term, T is the transposed matrix; The Softmax function is defined as: in, represents the probability that the input sample z is classified into the i-th category.

6. The millimeter wave river surface velocity monitoring radar echo attribute classification and identification method according to claim 1 is characterized in that: Convolutional neural networks are used to classify water surface echoes into three categories: steady water flow, choppy water flow, and ships.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images