Radar behavior recognition method and system based on mini rocket and channel attention mechanism

Abstract

The application discloses a radar behavior recognition method and system based on MiNiRocket and a channel attention mechanism, and specifically relates to the following steps: simulating a time-domain radar pulse description word (PDW) sequence of a radar under different behavior modes to construct a radar behavior dataset, and dividing the radar behavior dataset into a training sequence and a test sequence; constructing a deep learning network combined with MiNiRocket and a channel attention mechanism, and initializing the deep learning network; inputting the training sequence into the deep learning network, using MiNiRocket as a random feature extractor, extracting multi-dimensional quantile feature values, adaptively adjusting feature channel weights through the channel attention mechanism, and obtaining a trained radar behavior signal classification model; and inputting the test sequence into the trained signal classification model, testing the performance of the signal classification model, and obtaining a recognition result. The application has high calculation efficiency, small resource consumption, improved recognition accuracy and real-time performance of radar behavior, and improved informationization combat capability of the radar.

Description

Technical Field

[0001] This invention relates to the field of radar reconnaissance technology, and in particular to a radar behavior recognition method and system based on MiniRocket and channel attention mechanism. Background Technology

[0002] With the increasing demands of modern warfare on information-based combat capabilities, radar reconnaissance technology has become an important component of national defense. As a highly efficient detection and surveillance platform, radar can be widely used in various tasks such as airborne battlefield situational awareness, target identification, tracking, and attack. However, as radar system detection capabilities continue to improve, enemy target counter-detection capabilities are also constantly evolving, and target behavior patterns are becoming increasingly complex and diverse. Therefore, higher demands are placed on the accuracy and real-time performance of radar behavior pattern recognition.

[0003] In recent years, with the successful application of deep learning in various fields, radar behavior recognition methods based on deep neural networks have gradually attracted attention. In particular, convolutional neural networks (CNNs) have shown great potential in automatic feature extraction, capable of extracting useful features from raw data and adaptively learning the optimal combination of features through the model. However, deep learning methods are generally highly complex and computationally intensive, which greatly limits their practical applications.

[0004] To address the aforementioned issues, numerous lightweight time series classification methods have emerged in recent years. Among them, MiniRocket is an emerging time series classification algorithm that maintains high classification accuracy while reducing computational complexity through the design of sparse convolution and random convolution kernels. Furthermore, the efficient channel attention mechanism, as a lightweight attention mechanism, can adaptively adjust the weights of feature channels, enhancing attention to important features and improving the model's expressive power and classification performance. It has significant advantages in improving the accuracy, real-time performance, and computational efficiency of radar behavior pattern recognition, and can provide strong technical support for radar reconnaissance missions in complex battlefield environments. Summary of the Invention

[0005] The purpose of this invention is to provide a radar behavior recognition method and system based on MiniRocket and channel attention mechanism, which has high recognition accuracy, strong real-time performance, high computational efficiency, and low resource consumption.

[0006] The technical solution to achieve the purpose of this invention is: a radar behavior recognition method based on MiniRocket and channel attention mechanism, comprising the following steps:

[0007] Step 1: Simulate and generate time-domain radar pulse descriptors (PDW sequences) for radar under different behavior modes. Use the simulated PDW sequences to construct a radar behavior dataset. Add missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divide the radar behavior dataset into training sequences and test sequences.

[0008] Step 2: Construct a deep learning network that combines MiniRocket with a channel attention mechanism, and initialize the deep learning network.

[0009] Step 3: Input the training sequence into the deep learning network and use MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation;

[0010] Step 4: Input the obtained quantile feature values ​​into the channel attention mechanism to adaptively adjust the feature channel weights and obtain the trained radar behavior signal classification model.

[0011] Step 5: Input the test sequence into the trained signal classification model, test the performance of the signal classification model, and obtain the recognition result.

[0012] A radar behavior recognition system based on MiniRocket and channel attention mechanism is disclosed. This system implements the aforementioned radar behavior recognition method based on MiniRocket and channel attention mechanism. The system includes a dataset construction module, a deep learning network construction module, a feature extraction module, a channel attention mechanism module, and a recognition module, wherein:

[0013] The dataset construction module simulates and generates time-domain radar pulse descriptors (PDW sequences) under different radar behavior modes. It uses the simulated PDW sequences to construct a radar behavior dataset, adds missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divides the radar behavior dataset into training sequences and test sequences.

[0014] The deep learning network building module is used to build a deep learning network that combines MiNiRocket with a channel attention mechanism, and to initialize the deep learning network.

[0015] The feature extraction module inputs the training sequence into the deep learning network and uses MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation.

[0016] The channel attention mechanism module inputs the obtained quantile feature values ​​into the channel attention mechanism, adaptively adjusts the feature channel weights, and obtains a trained radar behavior signal classification model.

[0017] The recognition module inputs the test sequence into the trained signal classification model, verifies the performance of the signal classification model, and obtains the recognition result.

[0018] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the radar behavior recognition method based on MiniRocket and channel attention mechanism.

[0019] Compared with the prior art, the significant advantages of this invention are: (1) This invention adopts a deep learning network based on the combination of MiniRocket and channel attention mechanism, and introduces random convolution kernel mechanism and channel attention mechanism; (2) The random convolution kernel mechanism reduces the complexity of deep learning network and improves training speed, while the efficient channel attention mechanism improves attention to channel features, improves the ability to filter features, reduces computational cost, and improves recognition accuracy; (3) It improves the accuracy and real-time performance of radar behavior pattern recognition and improves information warfare capabilities. Attached Figure Description

[0020] Figure 1 This is a flowchart of the radar behavior recognition method based on MiniRocket and channel attention mechanism of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the radar behavior recognition model based on MiNiRocket and channel attention mechanism constructed in this embodiment of the invention.

[0022] Figure 3 This is a confusion matrix diagram of the recognition results under 10 radar behavior mode signals in an embodiment of the present invention. Detailed Implementation

[0023] It is readily understood that, based on the technical solution of this invention, those skilled in the art can conceive of various embodiments of this invention without altering its essential spirit. Therefore, the following specific embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of this invention or as limitations or restrictions on its technical solution.

[0024] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0025] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0026] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0027] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0028] Combination Figure 1 This invention discloses a radar behavior recognition method based on MiniRocket and channel attention mechanism, comprising the following steps:

[0029] Step 1: Simulate and generate time-domain radar pulse descriptors (PDW sequences) for radar under different behavior modes. Use the simulated PDW sequences to construct a radar behavior dataset. Add missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divide the radar behavior dataset into training sequences and test sequences.

[0030] Step 2: Construct a deep learning network that combines MiniRocket with a channel attention mechanism, and initialize the deep learning network.

[0031] Step 3: Input the training sequence into the deep learning network and use MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation;

[0032] Step 4: Input the obtained quantile feature values ​​into the channel attention mechanism to adaptively adjust the feature channel weights and obtain the trained radar behavior signal classification model.

[0033] Step 5: Input the test sequence into the trained signal classification model, test the performance of the signal classification model, and obtain the recognition result.

[0034] As a specific example, the radar behavior dataset in step 1 consists of pulse data from simulated radar emissions, and the radar behavior dataset is represented as follows:

[0035] x = [x0, x1, ..., x N-1 ]

[0036] In the formula, x is the pulse signal; N is the number of pulses; x0, x1, ..., x N-1 To simulate radar-transmitted pulse data; the radar behavior dataset contains the following features: pulse repetition period, pulse width, instantaneous bandwidth, number of pulses per CPI, and duty cycle.

[0037] As a specific example, step 3 involves inputting the training sequence into a deep learning network and using MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form an initial feature representation, as detailed below:

[0038] The MiNiRocket module randomly generates fixed-length convolution kernels and performs convolution operations in conjunction with dilation coefficients to extract multi-dimensional quantile features from radar pulse signal sequences. The convolution kernel length of the MiNiRocket module is 9, and the number of convolution kernels and dilation coefficients are dynamically adjusted according to the length of the input radar pulse signal sequence. The convolution kernels are used to capture the dimensional features of the radar pulse signal, thereby generating an initial feature representation of the radar pulse signal.

[0039] As a specific example, the convolution operation of the MiNiRocket module is a sparse convolution operation, which expands the receptive field of the convolution kernel through a dilation coefficient. For a radar pulse signal sequence within a time period t, the output of the convolution operation is related to a fixed-length convolution kernel and a dilation coefficient d. The output formula of this convolution operation is:

[0040]

[0041] In the formula, X is the output of the convolution operation; x is the radar pulse signal; W M,d For convolution kernel; w m W is the m-th weight in the convolution kernel, used to calculate the convolution operation; M is the weight of the convolution kernel W. M,d The quantity; N is the number of radar pulse signals; m represents the number of W. M,d Each convolution kernel in the array is numbered; n represents the number of each radar pulse signal; d is the expansion coefficient.

[0042] As a specific example, the range of the expansion coefficient d is set as follows:

[0043]

[0044] In the formula, the value of the expansion coefficient d follows a 2... m Exponential distribution; m is the kernel number; l INPUT It is the length of the radar pulse signal input sequence; l KERNEL This is the length of the convolution kernel, with a value of 9.

[0045] As a specific example, the formula for calculating the initial characteristics of a radar pulse signal is:

[0046]

[0047] In the formula, PPV is the proportion of values ​​greater than 0 in the calculated output; X is the convolution output; k is the total number of convolution outputs; b is the bias value, which comes from the radar pulse signal convolution output and takes the output percentile values ​​of 0.25, 0.5 and 0.75.

[0048] As a specific example, step 4 involves inputting the obtained quantile feature values ​​into the channel attention mechanism, adaptively adjusting the feature channel weights, and obtaining a trained radar behavior signal classification model, as detailed below:

[0049] The channel attention mechanism reduces the dimensionality of radar pulse signal features through global average pooling and adaptively adjusts the weights of the radar pulse signal feature channels to enhance the contribution of important feature channels, increase attention to important features of the radar pulse signal, improve feature representation ability, and thus obtain a radar behavior signal classification model, the formula of which is:

[0050] α i =σ(Conv1D(z))

[0051] In the formula, z = [z1, z2, ..., z n ] is the vector obtained by channel average pooling of the input radar pulse signal eigenvalues; α i σ represents the attention weight of channel i; σ indicates the use of the Sigmoid activation function; Conv1D is a one-dimensional convolution operation used to calculate the attention weight of each channel.

[0052] As a specific example, the formula for choosing the kernel size in a one-dimensional convolution operation is:

[0053]

[0054] In the formula, C is the channel dimension; F is the kernel size; || odd This indicates taking the nearest odd number.

[0055] This invention also provides a radar behavior recognition system based on MiniRocket and channel attention mechanism. This system is used to implement the aforementioned radar behavior recognition method based on MiniRocket and channel attention mechanism. The system includes a dataset construction module, a deep learning network construction module, a feature extraction module, a channel attention mechanism module, and a recognition module, wherein:

[0056] The dataset construction module simulates and generates time-domain radar pulse descriptors (PDW sequences) under different radar behavior modes. It uses the simulated PDW sequences to construct a radar behavior dataset, adds missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divides the radar behavior dataset into training sequences and test sequences.

[0057] The deep learning network building module is used to build a deep learning network that combines MiNiRocket with a channel attention mechanism, and to initialize the deep learning network.

[0058] The feature extraction module inputs the training sequence into the deep learning network and uses MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation.

[0059] The channel attention mechanism module inputs the obtained quantile feature values ​​into the channel attention mechanism, adaptively adjusts the feature channel weights, and obtains a trained radar behavior signal classification model.

[0060] The recognition module inputs the test sequence into the trained signal classification model, verifies the performance of the signal classification model, and obtains the recognition result.

[0061] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the radar behavior recognition method based on MiniRocket and channel attention mechanism.

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Example

[0064] This embodiment simulates and generates temporal radar pulse descriptor (PDW) sequences under different radar behavior modes. The radar behavior recognition method based on MiniRocket and channel attention mechanisms of this invention is then used to identify the radar behavior. The specific steps are as follows:

[0065] Step 1: Simulate and generate time-domain radar pulse descriptor (PDW) sequences under different radar behavior modes, and construct a dataset based on them. Considering that the signals on the actual battlefield are complex and diverse, and are usually affected by enemy interference, signal loss and background noise, missing pulses, false pulses and measurement errors are added to the dataset, and the dataset is divided into training sequences and test sequences with a ratio of 7:3.

[0066] In this embodiment, the pulse repetition period (PRI), pulse width (PW), instantaneous bandwidth (BW), duty cycle (DR), and number of pulses per CPI (NUM) are set as characteristic parameters according to different radar behavior modes.

[0067] The radar behavior dataset consists of pulse data simulating radar emissions, represented as:

[0068] x = [x0, x1, ..., x N-1 ]

[0069] In the formula, x is the pulse signal; N is the number of pulses; x0, x1, ..., x N-1 To simulate the pulse data emitted by radar;

[0070] The characteristic data of the nth pulse is described as follows:

[0071] x = {PRI, PW, BW, NUM, DR}

[0072] In the formula, PRI is the pulse repetition period, PW is the pulse width, BW is the instantaneous bandwidth, NUM is the number of pulses within CPI, and DR is the duty cycle.

[0073] The radar behavior patterns in the dataset are shown in Table 1: each behavior pattern simulates 4000 samples, and each sample contains 8000 pulse sequences; considering that the signals on the actual battlefield are complex and diverse, and are usually affected by enemy interference, signal loss and background noise, missing pulses, false pulses and measurement errors are added to the dataset.

[0074] Measurement errors include frequency measurement error, pulse width error, and pulse repetition interval error. The frequency measurement error is 0.7MHz to 2MHz, the pulse width error is 0.1μs to 0.3μs, and the pulse repetition interval error is 10% to 30% of the pulse rise edge width.

[0075] To enable the deep learning network to adapt to and learn the impact of noise on pulses, 20% of the samples in the training set are randomly selected, and 10% to 30% of the missing pulses and fake pulses are added in stages as data augmentation measures.

[0076] The formula for the lost pulse interference term is:

[0077] x = [x0, x1, Δ, x3, ..., x N-1 ]

[0078] Where x represents the sequence of lost radar pulse signals after a certain number of pulses are lost, Δ represents the lost pulses, and N is the number of pulses;

[0079] The formula for the spurious impulse interference term is:

[0080] x = [x0, x1, ... x m ...,x N-1 ]

[0081] Where x represents the lost radar pulse signal sequence after a certain number of pulses have been lost, x m This represents a spurious pulse, where N is the number of pulses.

[0082] Table 1 Radar Behavior Pulse Signal Parameter Settings

[0083] behavioral patterns PRI / μs PW / μs Duty cycle / % Number of pulses in CPI Instantaneous bandwidth / MHz Frequency repetition modulation VS 3.3～10 1～3 10～30 500～2000 0.3～10 fixed RWS 3.3～10 1～3 10～30 500～2000 0.3～10 Group change VRS 50～165 1～20 1～25 30～256 1～10 Group change MTT 3.3～125 0.1～20 0.1～25 1～64 1～50 slip BR 3.3～125 0.1～20 0.1～25 1～64 1～50 Sine period GMTI 120～500 2～60 0.1～25 20～256 0.5～15 Uneven GMTT 62～160 2～40 0.1～25 20～256 0.5～15 Uneven SSS 1000～2000 1～200 0.1～10 1～8 0.2～500 Uneven SST 500～1000 1～200 0.1～20 20～256 0.2～10 Uneven SAR 100～1000 3～60 1～25 70～20000 10～500 fixed

[0084] Step 2: Construct a deep learning network that combines MiniRocket with a channel attention mechanism, such as... Figure 2 As shown, the deep learning network is initialized.

[0085] Step 3: Input the training sequence into the deep learning network and use MiniRocket as a random feature extractor. The MiniRocket module randomly generates different fixed-length convolution kernels and combines them with the dilation coefficient to perform sparse convolution, extracting multi-dimensional quantile feature values ​​to form the initial feature representation.

[0086] For a radar pulse signal sequence within a time period t, the output of the convolution operation is related to a fixed-length convolution kernel and an expansion coefficient d. For the input signal x = [x0, x1, ..., x...], the output is related to the input signal x = [x0, x1, ..., x...]. N-1 ] and kernel W m,d =[w0,w1,...,w M-1 To perform a convolution operation, the expression is:

[0087]

[0088] In the formula, X is the output of the convolution operation; x is the radar pulse signal; W M,d For convolution kernel; w m W is the m-th weight in the convolution kernel, used to calculate the convolution operation; M is the weight of the convolution kernel W. M,d The quantity; N is the number of radar pulse signals; m represents the number of W. M,d Each convolution kernel in the array is numbered; n represents the number of each radar pulse signal; d is the expansion coefficient.

[0089] The MiNiRocket convolutional kernel has a length of 9, and the weights are constrained to α = -1 and β = 2. The kernel is represented using a combination of α and β, as follows:

[0090] [α,α,α,α,α,β,α,β,β]

[0091] [α,α,α,α,β,α,α,β,β]

[0092] [α,α,α,α,α,α,β,β,β]

[0093] Each convolutional kernel is assigned the same set of fixed dilation rates d, which is adjusted according to the length of the input time series; the range of dilation rate d is set as follows:

[0094]

[0095] In the formula, the value of the expansion coefficient d follows a 2... mExponential distribution; m is the kernel number; l INPUT It is the length of the radar pulse signal input sequence; l KERNEL It is the length of the convolution kernel, which takes the value 9.

[0096] The formula for the initial characteristic PPV of a radar pulse signal is:

[0097]

[0098] In the formula, PPV is the proportion of values ​​greater than 0 in the calculated output; X is the convolution output; k is the total number of convolution outputs; b is the bias value, which comes from the radar pulse signal convolution output and takes the output percentile values ​​of 0.25, 0.5 and 0.75.

[0099] Step 4: Input the obtained quantile feature values ​​into the channel attention mechanism. The efficient channel attention mechanism adaptively adjusts the feature channel weights to enhance the contribution of important feature channels, increase attention to important features, improve feature representation ability, and thus obtain a well-trained radar behavior signal classification model. The formula is as follows:

[0100] α i =σ(Conv1D(z))

[0101] In the formula, z = [z1, z2, ..., z n ] is the vector obtained by channel average pooling of the input radar pulse signal eigenvalues; α i σ represents the attention weight of channel i; σ indicates the use of the Sigmoid activation function; Conv1D is a one-dimensional convolution operation used to calculate the attention weight of each channel.

[0102] The formula for selecting the kernel size in the one-dimensional convolution operation is as follows:

[0103]

[0104] In the formula, C is the channel dimension; F is the kernel size; and odd represents the nearest odd number.

[0105] Step 5: Input the test sequence into the trained signal classification model to test the performance of the signal classification model and obtain the recognition results. The recognition results for 10 radar behavior pattern signals are as follows: Figure 3 As shown in Table 2, the recognition performance of airborne radar behavior models under different network architectures is as follows. It can be seen that the method of this invention (MiNiRocket-ECA) significantly improves the recognition accuracy and enhances the information warfare capability.

[0106] Table 2. Recognition performance of airborne radar behavior models under different network architectures

[0107] Network Model AlexNet ConvNet-18 ResNet-18 LSTM MiNiRocket-ECA accuracy 88.59％ 66.21％ 88.55％ 84.43％ 98.02％

[0108] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A radar behavior recognition method based on MiniRocket and channel attention mechanism, characterized in that, Includes the following steps: Step 1: Simulate and generate time-domain radar pulse descriptors (PDW sequences) for radar under different behavior modes. Use the simulated PDW sequences to construct a radar behavior dataset. Add missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divide the radar behavior dataset into training sequences and test sequences. Step 2: Construct a deep learning network that combines MiniRocket with a channel attention mechanism, and initialize the deep learning network. Step 3: Input the training sequence into the deep learning network and use MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation; Step 4: Input the obtained quantile feature values ​​into the channel attention mechanism to adaptively adjust the feature channel weights and obtain the trained radar behavior signal classification model. Step 5: Input the test sequence into the trained signal classification model, test the performance of the signal classification model, and obtain the recognition result.

2. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 1, characterized in that, The radar behavior dataset in step 1 consists of pulse data from simulated radar emissions. The radar behavior dataset is represented as follows: x=[x0,x1,…,x N-1 ] In the formula, x is the pulse signal; N is the number of pulses; x0, x1, ..., x N-1 To simulate radar-transmitted pulse data; the radar behavior dataset contains the following features: pulse repetition period, pulse width, instantaneous bandwidth, number of pulses per CPI, and duty cycle.

3. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 2, characterized in that, Step 3 involves inputting the training sequence into the deep learning network and using MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form an initial feature representation, as detailed below: The MiNiRocket module randomly generates fixed-length convolution kernels and performs convolution operations in conjunction with dilation coefficients to extract multi-dimensional quantile features from radar pulse signal sequences. The convolution kernel length of the MiNiRocket module is 9, and the number of convolution kernels and dilation coefficients are dynamically adjusted according to the length of the input radar pulse signal sequence. The convolution kernels are used to capture the dimensional features of the radar pulse signal, thereby generating an initial feature representation of the radar pulse signal.

4. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 3, characterized in that, The convolution operation of the MiNiRocket module is a sparse convolution operation, which expands the receptive field of the convolution kernel through a dilation coefficient. For a radar pulse signal sequence within a time period t, the output of the convolution operation is related to a fixed-length convolution kernel and a dilation coefficient d. The output formula of this convolution operation is as follows: In the formula, X is the output of the convolution operation; x is the radar pulse signal; W M,d For convolution kernel; w m W is the m-th weight in the convolution kernel, used to calculate the convolution operation; M is the weight of the convolution kernel W. M,d The quantity; N is the number of radar pulse signals; m represents the number of W. M,d Each convolution kernel in the array is numbered; n represents the number of each radar pulse signal; d is the expansion coefficient.

5. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 4, characterized in that, The range of the expansion coefficient d is set as follows: In the formula, the value of the expansion coefficient d follows a 2... m Exponential distribution; m is the kernel number; l INPUT It is the length of the radar pulse signal input sequence; l KERNEL This is the length of the convolution kernel, with a value of 9.

6. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 5, characterized in that, The formula for calculating the initial characteristics of a radar pulse signal is: In the formula, PPV is the proportion of values ​​greater than 0 in the calculated output; X is the convolution output; k is the total number of convolution outputs; b is the bias value, which comes from the radar pulse signal convolution output and takes the output percentile values ​​of 0.25, 0.5 and 0.

75.

7. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 6, characterized in that, Step 4 involves inputting the obtained quantile feature values ​​into the channel attention mechanism to adaptively adjust the feature channel weights, resulting in a trained radar behavior signal classification model, as detailed below: The channel attention mechanism reduces the dimensionality of radar pulse signal features through global average pooling and adaptively adjusts the weights of the radar pulse signal feature channels to enhance the contribution of important feature channels, increase attention to important features of the radar pulse signal, improve feature representation ability, and thus obtain a radar behavior signal classification model, the formula of which is: a i =σ(Conv1D(z)) In the formula, z = [z1, z2, ..., z n ] is the vector obtained by channel average pooling of the input radar pulse signal eigenvalues; α i σ represents the attention weight of channel i; σ indicates the use of the Sigmoid activation function; Conv1D is a one-dimensional convolution operation used to calculate the attention weight of each channel.

8. The radar behavior recognition method based on MiniRocket and channel attention mechanism according to claim 7, characterized in that, The formula for choosing the kernel size in a one-dimensional convolution operation is: In the formula, C is the channel dimension; F is the kernel size; || odd This indicates taking the nearest odd number.

9. A radar behavior recognition system based on MiniRocket and channel attention mechanism, characterized in that, This system is used to implement the radar behavior recognition method based on MiniRocket and channel attention mechanism as described in any one of claims 1 to 8. The system includes a dataset construction module, a deep learning network construction module, a feature extraction module, a channel attention mechanism module, and a recognition module, wherein: The dataset construction module simulates and generates time-domain radar pulse descriptors (PDW sequences) under different radar behavior modes. It uses the simulated PDW sequences to construct a radar behavior dataset, adds missing pulses, false pulses, and measurement errors to the radar behavior dataset, and divides the radar behavior dataset into training sequences and test sequences. The deep learning network building module is used to build a deep learning network that combines MiNiRocket with a channel attention mechanism, and to initialize the deep learning network. The feature extraction module inputs the training sequence into the deep learning network and uses MiniRocket as a random feature extractor to extract multi-dimensional quantile feature values ​​to form the initial feature representation. The channel attention mechanism module inputs the obtained quantile feature values ​​into the channel attention mechanism, adaptively adjusts the feature channel weights, and obtains a trained radar behavior signal classification model. The recognition module inputs the test sequence into the trained signal classification model, verifies the performance of the signal classification model, and obtains the recognition result.

10. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the radar behavior recognition method based on MiNiRocket and channel attention mechanism as described in any one of claims 1 to 8.

Images