Frequency spectrum segmentation method and system based on deep learning

Abstract

The invention relates to the technical field of communication signal separation, and discloses a frequency spectrum segmentation method and system based on deep learning, and the method comprises the following steps: obtaining a frequency spectrum image as input, and carrying out the feature extraction of the frequency spectrum image through a deep neural network; extracting a result based on the features; mapping the capacitor array control parameters into a frequency function form; transforming the frequency function through a preset nonlinear impedance response operator to obtain a corresponding impedance response; constructing a continuity loss item of the impedance response changing along with a network prediction path in a training process; according to the gradient change condition of the impedance response; and training the neural network through a joint loss function. According to the invention, by introducing the gradient compression strategy based on the derivative norm of the control parameter, the dynamic adjustment of the amplitude of the control parameter in the high-sensitivity frequency interval is realized, and the effects that the control process is smoother and the response is more stable are obtained.

Description

Technical Field

[0001] This invention relates to the field of communication signal separation technology, and in particular to a spectrum segmentation method and system based on deep learning. Background Technology

[0002] In the field of spectrum sensing and adjustable circuit control, common methods use deep neural networks or other machine learning models to predict the control parameters used to drive physical capacitor arrays. These control parameters are typically mapped to adjustable capacitor cells in the circuit and further influence the system's impedance response in the frequency dimension, thereby achieving dynamic adjustment and selective response to spectral signals.

[0003] In related research and applications, control parameters are usually directly output by neural networks, and the training objective focuses on a specific frequency response performance index, such as minimizing impedance or maximizing power throughput in the target frequency band. While this approach can achieve basic response matching, it commonly suffers from problems such as gradient imbalance and abrupt response changes during actual training. Especially in high-frequency sensitive regions, small changes in the control parameters' frequency function can trigger drastic fluctuations in the impedance function, leading to oscillating, unstable, or even non-convergent states in the entire control strategy. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a spectrum segmentation method and system based on deep learning, aiming to improve the existing technology, which lacks a mechanism to suppress the gradient sensitivity of control parameters and does not model or constrain the strength of the derivative of the response output with respect to the parameters. Therefore, in complex multi-band control tasks, problems such as sharp inflection points of the response function, gradient explosion, and slow training convergence often occur.

[0005] In a first aspect, the present invention provides the following technical solution: a spectrum segmentation method based on deep learning, comprising the following steps:

[0006] The spectrum image is obtained as input, and features are extracted from the spectrum image using a deep neural network;

[0007] Based on the feature extraction results, predict the spectrum segmentation mask and the corresponding capacitor array control parameters;

[0008] The control parameters of the capacitor array are mapped into a frequency function form;

[0009] The frequency function is transformed by a preset nonlinear impedance response operator to obtain the corresponding impedance response;

[0010] During training, a continuous loss term is constructed to describe how the impedance response changes as the network predicts the path.

[0011] Based on the gradient change of the impedance response, a compression adjustment operation is performed on the capacitance control parameters;

[0012] The neural network is trained using a joint loss function, and the trained capacitor control parameters are output as control commands to drive the physical capacitor array circuit for frequency division control.

[0013] Preferably, the spectral image is any one of the following transform spectra: short-time Fourier transform spectrum, Mel spectrum, or fast Fourier transform spectrum.

[0014] Preferably, the capacitor array control parameters are n-dimensional vectors, and the frequency function is generated by a linear combination of the vector and a set of preset basis functions, and the frequency function is defined in Hilbert space.

[0015] Preferably, the nonlinear operator constructs an integral relationship over the frequency range, which includes a combination of a frequency-coupled kernel function and a nonlinear function, used to transform the input function into an impedance response function.

[0016] Preferably, the frequency coupling kernel function is a kernel function defined in the frequency range that satisfies symmetry and positive definiteness, the nonlinear function is a monotonically bounded mapping of the input function value, and the control parameters are used to control the capacitance value of each capacitor unit in the capacitor array. The capacitor array contains multiple adjustable units and is connected to the frequency segmentation circuit.

[0017] Preferably, the impedance response constructs path continuity constraints at multiple consecutive time points, the constraint term being the sum of squares of the differences in impedance function values ​​at adjacent time points, and the neural network includes a multi-scale convolutional structure whose output is simultaneously connected to a spectral mask prediction branch and a control parameter prediction branch.

[0018] Preferably, the compression strategy adjusts the value of the control parameters by constructing an exponential decay function of the control parameters with respect to the impedance response gradient norm, and the joint loss function includes spectral mask prediction error, impedance response error, path continuity term and gradient compression term.

[0019] Secondly, the present invention provides the following technical solution: a spectrum segmentation system based on deep learning, the system comprising:

[0020] The input module is used to receive spectrum images;

[0021] The feature extraction module is used to extract multi-scale convolutional features from the input spectral image;

[0022] A segmentation prediction module is used to generate a spectral mask based on the feature extraction results;

[0023] The control parameter generation module is used to generate control parameters for the capacitor array.

[0024] The function mapping module is used to map control parameters into frequency domain functions.

[0025] Impedance response module, used to generate the impedance response corresponding to the control output based on nonlinear impedance response operator;

[0026] The path control module is used to construct continuity constraints for the path of control parameter changes;

[0027] The gradient compression module is used to compress control parameters in high gradient regions.

[0028] The joint optimization module is used to jointly optimize the loss function of the entire network system.

[0029] An execution module is used to drive the physical capacitor array to perform spectrum segmentation control according to the optimized control parameters.

[0030] Thirdly, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned deep learning-based spectrum segmentation method.

[0031] Fourthly, the present invention provides the following technical solution: a readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned deep learning-based spectrum segmentation method.

[0032] The present invention has the following beneficial effects:

[0033] 1. In this invention, by introducing a gradient compression strategy based on the norm of the derivative of the control parameter, the amplitude of the control parameter in the high-sensitive frequency range is dynamically adjusted, resulting in a smoother control process and a more stable response.

[0034] 2. In this invention, by designing a nonlinear impedance response operator that includes a frequency coupling kernel function and a nonlinear transformation function, a continuous and differentiable mapping relationship from control parameters to impedance functions is achieved, resulting in a more controllable and expressive frequency response.

[0035] 3. In this invention, by constructing a path continuity constraint term based on the difference in impedance response at adjacent time points, the continuous evolution of the output function in the time dimension is realized, resulting in a more coherent and predictable dynamic change in the spectrum control process. Attached Figure Description

[0036] Figure 1 This is a flowchart of a deep learning-based spectrum segmentation method proposed in this invention.

[0037] Figure 2 This is a system architecture diagram of a deep learning-based spectrum segmentation system proposed in this invention. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1

[0040] Reference Figure 1 In the first embodiment of the present invention, a deep learning-based spectrum segmentation method is provided, comprising the following steps:

[0041] The spectrum image is obtained as input, and features are extracted from the spectrum image using a deep neural network;

[0042] Based on the feature extraction results, predict the spectrum segmentation mask and the corresponding capacitor array control parameters;

[0043] Map the capacitor array control parameters into a frequency function form;

[0044] The frequency function is transformed by a preset nonlinear impedance response operator to obtain the corresponding impedance response;

[0045] During training, a continuous loss term is constructed to reflect the change in impedance response as the network predicts the path.

[0046] Based on the gradient change of the impedance response, the capacitance control parameters are adjusted by compression.

[0047] The neural network is trained using a joint loss function, and the trained capacitor control parameters are output as control commands to drive the physical capacitor array circuit for frequency division control.

[0048] The spectral image is any one of the following transformations: short-time Fourier transform spectrum, Mel spectrum, or fast Fourier transform spectrum.

[0049] Specifically, the Short Time Fourier Transform (SFT) spectrum analyzes the original signal by dividing it into frames and performing local Fourier analysis on each frame. Common processing methods include frame lengths of 20 to 30 milliseconds, frame shifts of up to 10 milliseconds, and the use of the Hamming window. Independent frequency domain analysis of each frame captures the spectral structure changing over time. The final output is a two-dimensional complex matrix, whose amplitude or power is used as image content, logarithmically transformed and normalized, and then used as model input. Mel-based spectrograms, building upon the SFT, introduce perceptual filtering. The core of this method is to use a set of equally spaced Mel-numbered filters on the frequency axis, distributed in a triangular pattern, covering the frequency range and weighted and integrated. Each filter outputs an energy value, thus compressing the frequency dimension of the spectrum. Typically, 40 to 128 Mel-numbered filters are used, resulting in a sparser two-dimensional matrix that emphasizes low-frequency variations. This is commonly used in audio and speech-related tasks. Fast Fourier Transform (FFT) spectrum processing is more direct. After the signal is divided into frames, a Fast Fourier Transform is directly performed on each frame. Unlike STFT, it typically does not incorporate a window function or introduce time shift. Its advantages lie in its computational simplicity and preservation of the full frequency composition, making it suitable for tasks that are sensitive to frequency distribution but not strongly dependent on timing characteristics.

[0050] All three types of spectra require parameter settings during generation, including sampling rate, frame length, frame shift, window function type, and FFT points. These parameters must remain consistent during model usage to avoid model structure errors caused by inconsistent dimensions. Typically, the amplitude information in the spectra is logarithmically transformed and standardized to a fixed range, making training more stable and feature extraction easier.

[0051] The control parameters of the capacitor array are n-dimensional vectors, and the frequency function is generated by a linear combination of this vector and a set of preset basis functions. The frequency function is defined in Hilbert space.

[0052] Specifically, firstly, the neural network model outputs a real-valued vector of length n based on the input spectral image. Each term in this vector corresponds to an initial control value for a specific channel or frequency band. To avoid inconsistencies or extreme values ​​interfering with subsequent modeling, this vector is typically normalized, limiting its value range to a relatively fixed numerical interval. Secondly, the normalized control parameters are fed into the function building unit. In this stage, the system calls a predefined set of basis functions to perform a weighted combination of each control parameter and its corresponding basis function. The number of basis functions is strictly consistent with the dimension of the control vector, and their arrangement is fixed, not allowing dynamic generation or adjustment. The combined result is a frequency function, continuously defined across the entire frequency domain and possessing integrability in the function space. Then, this frequency function is mathematically defined in a Hilbert space. That is, this function needs to satisfy the square integrability condition and form a complete space on its domain. This setting not only ensures the stability of the function in numerical analysis but also allows the system to perform further inner product, norm, or gradient operations on it. The Hilbert space itself does not depend on changes in the model structure; once it is determined during the system initialization phase, it remains unchanged.

[0053] Finally, the generated frequency function is passed to the next nonlinear processing module as input to construct the impedance response required for the capacitor array. It is important to note that this frequency function is not converted into a discrete vector, nor is it propagated back to the previous layer. The entire system relies on the continuity and structure of this function construction to complete subsequent impedance tuning modeling. Therefore, the construction step of this function cannot be omitted in the entire data flow, nor can it be replaced by other structures. Throughout the process, the system must meet the following preconditions: the control parameters are fixed-dimensional vectors; the set of basis functions used is unadjustable, untrainable, and must cover a sufficiently wide range in the frequency domain; the output frequency function must be continuous and integrable, and discontinuous, divergent, or high-frequency oscillating behavior is not allowed.

[0054] The nonlinear operator constructs an integral relation over the frequency range, which includes a combination of a frequency-coupled kernel function and a nonlinear function, used to transform the input function into an impedance response function. Specifically, the nonlinear operator module maps the input frequency function f(ω) to the impedance response function Z(ω). This transformation process occurs over the defined frequency range. The process is performed internally, employing an integral operator structure that combines a kernel function and a nonlinear function. The mapping relationship is defined as follows:

[0055] Z(ω)=∫ Ω k(ω,ω′)·σ(f(ω′))dω′;

[0056] in:

[0057] Z(ω) is the impedance response function of the output, which is continuously defined on the frequency variable ω;

[0058] f(ω′) is the frequency function derived from the mapping of control parameters;

[0059] K(ω,ω′) is the frequency coupling kernel function, used to describe the correlation between frequency points ω and ω′;

[0060] σ(·) is a nonlinear function, often chosen as the hyperbolic tangent function tanh, the sigmoid function, or ReLU activation type;

[0061] Ω is the defined range of frequency integration, which is usually the range from the lowest frequency to the highest frequency set by the system;

[0062] ω, ω′∈Ω are frequency variables.

[0063] First, the input frequency function f(ω′) is constructed by a linear combination of control parameters and a set of basis functions, as described above. Second, the system inputs this function value into the nonlinear function σ(·), outputting the transformed frequency response amplitude. Then, for each target frequency point ω, the weighted coupling degree between it and all function values ​​at ω′ is calculated using the kernel function K(ω,ω′). Finally, the weighted outputs of the kernel functions at all frequency points are integrated and summed to obtain the final Z(ω) value. This structure as a whole constitutes a continuous kernel product operator, which essentially performs a "smooth transformation" of the input function through kernel weighting, while simultaneously incorporating a nonlinear response to characterize the impedance characteristics of the real physical system. It should be noted that the kernel function K(ω,ω′) used is generally a Gaussian kernel, a Laplace kernel, or a rectangular kernel, with a form as shown below (taking a Gaussian kernel as an example):

[0064]

[0065] Here, σ is the bandwidth parameter of the kernel function, which controls the frequency domain spread range of the response. The entire integration process needs to be implemented numerically using a discrete sampling integration method, such as Gaussian numerical integration or weighted summation with uniform sampling at frequency points. However, theoretically, it must be based on continuous frequency function modeling and cannot be directly replaced by finite-point interpolation.

[0066] The frequency coupling kernel function is a kernel function defined in the frequency range that satisfies symmetry and positive definiteness. The nonlinear function is a monotone bounded mapping of the input function value. The control parameters are used to control the capacitance value of each capacitor unit in the capacitor array. The capacitor array contains multiple adjustable units and is connected to the frequency segmentation circuit.

[0067] Specifically, first, the system initializes a two-dimensional kernel function K(ω,ω′), with its domain covering the frequency interval Ω. This function models the correlation between any two frequency points ω and ω′. To ensure the mathematical stability of the integration operation and the physical reasonableness of the result, the kernel function satisfies two fundamental properties: symmetry and positive definiteness. Symmetry means that for any ω, ω′ ∈ Ω, the following holds:

[0068] K(ω,ω′)=K(ω′,ω);

[0069] This ensures that the coupling between different frequency points is non-directional, and the mutual influence between any two points is the same. Positive definiteness requires that the Gram matrix formed by the kernel function on any finite set of sample points is positive definite, that is, all its eigenvalues ​​are positive. This is to ensure the numerical stability and invertibility of the integral mapping process, and also to ensure that the response of the output function on the frequency axis will not exhibit non-physical oscillations.

[0070] Secondly, the nonlinear function part is responsible for adjusting the amplitude of the input frequency function f(ω′), mapping it to a restricted interval. This mapping function must be monotonic, bounded, and continuous. Its function is to compress high-amplitude or low-amplitude regions, preventing excessive responses in certain frequency ranges during integration. Common choices include the hyperbolic tangent function tanh, the sigmoid function, or the softsign function, which suppress extrema and buffer sharp abrupt changes in the spectrum.

[0071] Then, the role of the control parameters in the entire mapping system is not only reflected in the frequency function construction stage, but also directly affects the final circuit response behavior. These parameters are output by the neural network, and after frequency mapping and nonlinear transformation, they ultimately participate in the actual configuration of the capacitor array;

[0072] Each control parameter value ci corresponds to a capacitor unit in the capacitor array. These capacitor units are adjustable, supporting dynamic adjustment based on the input signal or model output. In the specific implementation, the control parameters undergo certain numerical scaling and quantization processing, mapping them to specific capacitor value setting signals, such as voltage, charge control values, or switching thresholds.

[0073] Finally, the capacitor array is connected to a frequency segmentation circuit structure. This structure divides the entire frequency range into several sub-bands, each of which has its impedance or resonant frequency adjusted by a capacitor network. Adjustments to the control parameters cause changes in the capacitance of each sub-band, thereby altering the overall frequency response curve and achieving frequency-selective control of external signals.

[0074] The impedance response constructs path continuity constraints at multiple consecutive time points. The constraint term is the sum of the squares of the differences in impedance function values ​​at adjacent time points. The neural network includes a multi-scale convolutional structure, and its output is simultaneously connected to the spectral mask prediction branch and the control parameter prediction branch.

[0075] Specifically, at the raw data input end, the system processes a time-dimension sequence input. After generating impedance response functions for multiple time frames, the system performs differential processing on these functions along the time axis. The core idea is that if the response changes too much between adjacent time points, it indicates that the control strategy is unstable and the system behavior is discontinuous. Therefore, a continuity constraint term is introduced to force the network output to maintain a smooth transition in the time dimension. The mathematical form of this constraint term is as follows:

[0076] L path =∑ t ∫ Ω |Z t+1 (ω)-Z t (ω)| 2 dω;

[0077] in:

[0078] Z t (ω): represents the impedance response function at time t; Ω: frequency range;

[0079] L path : Path continuity loss term, used to measure the difference in response functions between adjacent time points.

[0080] Then, during the training phase, this term is added to the loss function and participates in backpropagation along with other losses (such as mask prediction loss and impedance error terms). This constraint avoids frequency response jitter over time, enhancing the system's robustness and inference controllability. Simultaneously, to support this temporal modeling, the neural network architecture employs a multi-scale convolutional module as the feature extraction backbone. This module integrates multiple sets of convolutional kernels with different receptive fields, such as 3×3, 5×5, and 7×7, enabling the model to capture local details while also possessing a certain degree of global structure awareness. The two branches share the previous feature extraction results, their parameters are updated completely independently, and the predicted targets do not interfere with each other, but their training process is uniformly scheduled at the loss function layer. This structure ensures minimal coupling between the masking task and the control task in feature utilization, while retaining the data utilization efficiency improvements brought by multi-task training.

[0081] The compression strategy adjusts the range of control parameters by constructing an exponential decay function of the control parameters with respect to the impedance response gradient norm. The joint loss function includes spectral mask prediction error, impedance response error, path continuity term, and gradient compression term.

[0082] Specifically, this compression strategy does not simply prune or weight the control parameters. Instead, it constructs a loss term with respect to the gradient norm of the control parameters and embeds it exponentially into the main loss function. This strategy is essentially a form of "indirect control": it does not directly constrain the parameter values, but rather influences the parameter amplitude by constraining their impact on the output. The system automatically calculates the partial derivative of the impedance function with respect to each control parameter after each response is generated.

[0083]

[0084] These partial derivatives are summed by uniform squares to form the gradient norm, which is then amplified by an exponential function to form the final compression loss term. The overall construction is as follows:

[0085]

[0086] in:

[0087] Z(ω): The impedance response generated by the current control parameters;

[0088] Ω: The frequency range within which integration is performed;

[0089] c i The i-th term in the control parameter vector output by the neural network;

[0090] α: Exponential weighting factor, used to control the intensity of the penalty.

[0091] n: Dimension of control parameters.

[0092] This loss term design has a hidden effect: when a certain frequency point is particularly sensitive to the control parameters (i.e., ... (If the value is large), it will make the entire L grad The number of terms increases rapidly. During training, the optimizer spontaneously compresses these highly sensitive parameters, tending to output more moderate value changes. In other words, this mechanism allows the system to automatically reduce the magnitude of a parameter when it "senses" that it will cause a strong reaction, thus achieving dynamic compression. This compression term, along with several other sub-loss terms, forms the complete joint loss function. To maintain the model's multi-task learning objective, the system uses the following loss structure:

[0093] L total =λ1L mask +λ2L resp +λ3L path +λ4L grad ;

[0094] Each of these serves a different sub-objective of the model:

[0095] L mask: Spectrum mask prediction error, used to monitor whether the model accurately identifies the signal frequency band;

[0096] L resp: Impedance response error directly measures the difference between the current response function and the target impedance;

[0097] L path : Path continuity term, ensuring the continuity of output impedance over time;

[0098] L grad The control parameter compression term, namely the gradient norm penalty term mentioned above, limits the oscillation amplitude of the parameter.

[0099] Each λ1 is a hyperparameter used to balance the training intensity among the terms. These weights can be specified statically or dynamically adjusted during training, such as using an uncertainty-driven weighting function to automatically allocate learning resources. This mechanism is particularly suitable for impedance response problems, as these problems are very sensitive to the smoothness of the frequency curve shape. Without such compression control, the frequency function will generate highly local extrema during optimization, causing control signal oscillations and making it difficult to stably configure the hardware capacitor array.

[0100] Example 2:

[0101] Reference Figure 2 In a second embodiment of the present invention, the present invention provides a spectrum segmentation system based on deep learning, the system comprising:

[0102] The input module is used to receive spectrum images;

[0103] The feature extraction module is used to extract multi-scale convolutional features from the input spectral image;

[0104] The segmentation prediction module is used to generate a spectral mask based on the feature extraction results;

[0105] The control parameter generation module is used to generate control parameters for the capacitor array.

[0106] The function mapping module is used to map control parameters into frequency domain functions.

[0107] Impedance response module, used to generate the impedance response corresponding to the control output based on nonlinear impedance response operator;

[0108] The path control module is used to construct continuity constraints for the path of control parameter changes;

[0109] The gradient compression module is used to compress control parameters in high gradient regions.

[0110] The joint optimization module is used to jointly optimize the loss function of the entire network system.

[0111] The execution module is used to drive the physical capacitor array to perform spectrum segmentation control based on optimized control parameters.

[0112] Specifically, the input module is mainly responsible for receiving spectral image data. The spectral image can be a two-dimensional frequency-time graph, with the horizontal axis representing the time series and the vertical axis representing the frequency distribution. The pixel intensity of the image represents the signal energy or power spectral density. The input format supports single-frame images or continuous frame sequences. The system can process the data frame by frame during the inference phase, and during the training phase, a sliding window input method can be used to construct a time-dependent structure. Before entering the network, the spectral image can be preprocessed, such as by normalization and logarithmic compression, to improve training stability.

[0113] The feature extraction module is responsible for extracting deep-level structures from the input spectral image, employing a multi-scale convolution approach. The convolution path is formed by stacking multiple convolutional kernels with different receptive fields in parallel, commonly using sizes such as 3x3, 5x5, and 7x7. This parallel perception of local and global frequency structures improves the completeness and stability of the feature representation. The output features, after channel fusion, form a unified-dimensional embedded feature tensor, which serves as the shared input for subsequent prediction modules.

[0114] The segmentation prediction module maps features to a binary mask of the same size as the input image through a set of pixel-wise prediction convolutional layers. The output of each frequency point represents the probability that it belongs to the target frequency band. During the training phase, binary cross-entropy is used as the loss function for optimization. This module directly outputs the mask without additional post-processing.

[0115] The control parameter generation module is used to convert shared feature maps into control parameter vectors. Each dimension of this parameter vector corresponds to an independent control unit in the capacitor array. The system generates this vector through a fully connected layer or an attention compression structure. To ensure controllability, the output parameter range can be limited to a certain preset interval, such as using the Tanh activation function to compress the output to the [-1,1] interval before performing linear mapping.

[0116] After the function mapping module generates the control parameter vector, it converts it into a continuous function over the frequency domain. This is achieved using a linear combination method:

[0117]

[0118] Where φ i (ω) is a pre-defined set of basis functions, typically Gaussian, sine, cosine, or B-spline functions, where ω represents the frequency variable. This function characterizes the influence of the current control parameters in the frequency space and is a crucial intermediate step in subsequent impedance generation.

[0119] The impedance response module is responsible for mapping the frequency function to the impedance response function. Its core implementation structure is a nonlinear kernel integral operator. The specific formula is: Z(ω)=∫ Ω K(ω,ω′)·σ(f(ω′))dω′ is used to connect the input control and physical response in the model;

[0120] The path regulation module introduces path continuity constraints along the time-series dimension of the impedance response. This module constructs a difference loss term for the response function across all consecutive time frames. Its purpose is to limit drastic changes in the model output over time, enhance the controllability of the system response, and avoid control instability caused by sudden disturbances. This term is embedded as one of the loss terms in the joint optimization module during the training phase.

[0121] To further improve the stability of control parameters and prevent runaway control in high-frequency sensitive regions, the gradient compression module calculates the partial derivatives of the impedance response function with respect to the control parameters and introduces an exponential penalty function. This module dynamically suppresses parameters that trigger drastic responses in highly sensitive regions, thereby constraining the oscillation range of the overall control strategy. This module requires no additional explicit gradient calculation; the automatic differentiation system can support its inference process.

[0122] The joint optimization module integrates the loss functions of each task to form an overall optimization objective, which is ultimately used to control the overall training direction and ensure that the objectives of each sub-module converge collaboratively.

[0123] Finally, the control parameters are sent to the execution module to drive the capacitor array hardware system. Each parameter ci controls the bias state of the corresponding adjustable capacitor. After the capacitor array structure is connected to the frequency segmentation circuit, impedance changes are achieved within different frequency ranges. This module is directly coupled to the physical layer signal chain, providing physical response feedback for the model output.

[0124] Example 3

[0125] In a third embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a deep learning-based spectrum segmentation method and system described in the above embodiments.

[0126] Example 4

[0127] In the fourth embodiment of the present invention, based on the same inventive concept, a computer device is proposed, comprising: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the spectrum segmentation method and system based on deep learning of the above embodiment.

[0128] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0129] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A spectrum segmentation method based on deep learning, characterized in that, Includes the following steps: The spectrum image is obtained as input, and features are extracted from the spectrum image using a deep neural network; Based on the feature extraction results, predict the spectrum segmentation mask and the corresponding capacitor array control parameters; The control parameters of the capacitor array are mapped into a frequency function form; The frequency function is transformed by a preset nonlinear impedance response operator to obtain the corresponding impedance response; During training, a continuous loss term is constructed to describe how the impedance response changes as the network predicts the path. Based on the gradient change of the impedance response, a compression adjustment operation is performed on the capacitance control parameters; The neural network is trained using a joint loss function, and the trained capacitor control parameters are output as control commands to drive the physical capacitor array circuit for frequency division control.

2. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The spectral image is any one of the following transformation spectra: short-time Fourier transform spectrum, Mel spectrum, or fast Fourier transform spectrum.

3. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The capacitor array control parameters are n-dimensional vectors, and the frequency function is generated by a linear combination of this vector and a set of preset basis functions. The frequency function is defined in Hilbert space.

4. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The nonlinear operator constructs an integral relationship over the frequency range, which includes a combination of a frequency coupling kernel function and a nonlinear function, used to transform the input function into an impedance response function.

5. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The frequency coupling kernel function is a kernel function defined on the frequency range that satisfies symmetry and positive definiteness. The nonlinear function is a monotonically bounded mapping of the input function value. The control parameters are used to control the capacitance value of each capacitor unit in the capacitor array. The capacitor array contains multiple adjustable units and is connected to the frequency segmentation circuit.

6. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The impedance response constructs path continuity constraints at multiple consecutive time points, and the constraint term is the sum of squares of the differences in impedance function values ​​at adjacent time points. The neural network includes a multi-scale convolutional structure, and its output is simultaneously connected to a spectral mask prediction branch and a control parameter prediction branch.

7. The spectrum segmentation method based on deep learning according to claim 1, characterized in that, The compression strategy adjusts the range of control parameters by constructing an exponential decay function of the control parameters with respect to the impedance response gradient norm. The joint loss function includes spectral mask prediction error, impedance response error, path continuity term, and gradient compression term.

8. A spectrum segmentation system based on deep learning, characterized in that, The system for a deep learning-based spectrum segmentation method according to any one of claims 1-7, the system comprising: The input module is used to receive spectrum images; The feature extraction module is used to extract multi-scale convolutional features from the input spectral image; A segmentation prediction module is used to generate a spectral mask based on the feature extraction results; The control parameter generation module is used to generate control parameters for the capacitor array. The function mapping module is used to map control parameters into frequency domain functions. Impedance response module, used to generate the impedance response corresponding to the control output based on nonlinear impedance response operator; The path control module is used to construct continuity constraints for the path of control parameter changes; The gradient compression module is used to compress control parameters in high gradient regions. The joint optimization module is used to jointly optimize the loss function of the entire network system. An execution module is used to drive the physical capacitor array to perform spectrum segmentation control according to the optimized control parameters.

9. A computer device 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 computer program, it implements a deep learning-based spectrum segmentation method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements a deep learning-based spectrum segmentation method as described in any one of claims 1 to 7.

Images