Model training method and device based on multi-channel detection system, equipment and medium

Abstract

The invention discloses a model training method and device based on a multi-channel detection system, equipment and a medium, and the method comprises the steps: obtaining environment parameters of the multi-channel detection system; simulating a radiation source signal and a receiving signal based on the environmental parameters, and determining a detection threshold as a first label based on the receiving signal and a preset false alarm probability; based on the detection threshold, detecting the generated receiving signal through a Monte Carlo simulation experiment, and counting the detection probability that the radiation source signal exists in the receiving signal as a second label; constructing a first training data set based on the environmental parameters and the first labels, and constructing a second training data set based on the environmental parameters and the second labels; and iteratively updating the threshold estimation model based on the first training data set, and iteratively updating the detection probability prediction model based on the second training data set. According to the embodiment of the invention, the effectiveness and robustness of the method in the aspects of predicting the detection performance and determining the threshold value are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of multi-channel signal detection, and in particular to a model training method and device based on a multi-channel detection system, equipment and medium. BACKGROUND

[0002] In the fields of wireless communication, radar signal processing, speech recognition and medical signal analysis, multi-channel detection technology plays a key role. Although traditional multi-channel detection methods such as energy detection, GLRT (Generalized Likelihood Ratio Test) detection, and eigenvalue-based detection are mature, their performance is significantly affected by environmental parameters such as signal-to-noise ratio, sampling rate, and number of nodes.

[0003] Existing theoretical analysis methods are based on ideal assumptions such as high signal-to-noise ratio, large sample size, and known noise power, making it difficult to accurately depict the relationship between detection probability and threshold in actual complex scenarios. To compensate for the shortcomings of theoretical methods, the academic and engineering communities often use Monte Carlo simulation to statistically analyze detection performance through a large number of random experiments. However, this method has the problems of large computational load and low efficiency, and it is difficult to obtain a mathematical mapping between parameters and performance. When parameters change, the simulation needs to be re-run, greatly limiting the flexibility and scalability of practical applications.

[0004] In recent years, with the development of artificial intelligence, machine learning methods have been introduced into the detection performance prediction task. However, traditional neural networks such as multi-layer perceptron (MLP) have the disadvantages of not being able to guarantee the physical monotonicity of the model output and input parameters, limited generalization ability, sensitivity to outliers, and poor interpretability in the multi-channel detection scenario. Currently, there is a lack of a detection performance prediction and threshold determination method that is efficient, robust, and physically consistent, and can adapt to complex and variable multi-channel signal detection systems, and related technical breakthroughs are urgently needed. SUMMARY

[0005] Therefore, it is necessary to provide a model training method based on a multi-channel detection system to solve at least one of the problems existing in the prior art.

[0006] In a first aspect, a model training method based on a multi-channel detection system includes:

[0007] Obtaining environmental parameters of a multi-channel detection system;

[0008] Simulating a radiation source signal and a received signal based on the environmental parameters, determining a detection threshold based on the received signal and a preset false alarm probability, and taking the detection threshold as a first label;

[0009] based on the detection threshold, performing multiple detections on the generated received signal through Monte Carlo simulation experiments, and counting a detection probability of the existence of the radiation source signal in the received signal as the second label;

[0010] based on the environment parameter and the first label, constructing a first training data set, and based on the environment parameter and the second label, constructing a second training data set;

[0011] iteratively updating a threshold estimation model based on the first training data set, and iteratively updating a detection probability prediction model based on the second training data set.

[0012] In a possible implementation, the simulation of the radiation source signal and the received signal in the real scene based on the environment parameter comprises:

[0013] constructing a multi-node received signal model;

[0014] based on the multi-node received signal model, modulating a low-frequency baseband signal to generate an analog signal with a frequency range greater than a preset frequency range, which is easy to transmit in an analog channel;

[0015] transmitting the analog signal through a channel to a receiving end for demodulation to obtain a received signal.

[0016] In a possible implementation, the multiple detections on the generated received signal through Monte Carlo simulation experiments based on the detection threshold, and the counting of the detection probability of the existence of the radiation source signal in the received signal comprise:

[0017] based on the detection threshold, determining a first probability density of the existence of the radiation source signal in the received signal and a second probability density of the non-existence of the radiation source signal;

[0018] based on a ratio between the first probability density and the second probability density, obtaining a likelihood ratio;

[0019] if the likelihood ratio is greater than a preset decision threshold, determining that the received signal contains the radiation source signal;

[0020] counting a detection probability of the determination of the existence of the radiation source signal in the received signal in a preset number of Monte Carlo simulation experiments.

[0021] In a possible implementation, the determination of the detection threshold based on the received signal and a preset false alarm probability comprises:

[0022] based on the received signal, determining noise;

[0023] based on the preset false alarm probability and the noise, determining the detection threshold.

[0024] In a possible implementation, the threshold estimation model and the detection probability prediction model each comprise an input layer, a hidden layer, and an output layer, wherein the hidden layer is composed of multiple sets of weight linear transformation units, each set containing several positive weight linear units and / or negative weight linear units, and the weight linear transformation units ensure monotonicity between input features and output features through weight constraints.

[0025] In a possible implementation, the iterative updating of the detection probability prediction model based on the second training data set comprises:

[0026] initial training data is randomly selected from the second training data set and input into the input layer;

[0027] the initial training data is input into the hidden layer through the input layer, processed in parallel through multiple sets of weight linear units, and the maximum response value is extracted from the output results of each set of weight linear units to obtain the output results of each set of weight linear units, wherein each set of weight linear units ensures that the weights are positive by squaring the weights, so that the input features and the output features are in a monotonically increasing relationship;

[0028] the minimum value of the output results of each set of weight linear units is taken through the output layer, and the detection probability is obtained through activation processing by a preset activation function;

[0029] a second loss value is calculated based on the detection probability, the second label, and a second loss function;

[0030] the detection probability prediction model is iteratively updated based on the second loss value until a preset convergence condition is met.

[0031] In a possible implementation, the iterative updating of the threshold estimation model based on the first training data set comprises:

[0032] initial training data is randomly selected from the first training data set and input into the input layer;

[0033] the initial training data is input into the hidden layer through the input layer, processed in parallel through multiple sets of weight linear units, and the maximum response value is extracted from the output results of each set of weight linear units to obtain the output results of each set of weight linear units, wherein each set of weight linear units ensures that the input training data and the output results are in a monotonically increasing relationship and a monotonically decreasing relationship by squaring the weights and taking the negative square of the weights;

[0034] the minimum value of the output results of each set of weight linear units is taken through the output layer to obtain an optimal detection threshold;

[0035] calculate a first loss value based on the optimal detection threshold, the first label, and the first loss function;

[0036] perform a next round of iteration on the threshold estimation model based on the first loss value until a preset convergence condition is met.

[0037] In a second aspect, a model training apparatus based on a multi-channel detection system is provided, and includes:

[0038] an environment parameter acquisition unit configured to acquire an environment parameter of the multi-channel detection system;

[0039] a first label generation unit configured to simulate a radiation source signal and a received signal based on the environment parameter, determine a detection threshold based on the received signal and a preset false alarm probability, and take the detection threshold as a first label;

[0040] a second label generation unit configured to perform multiple detections on the generated received signal through a Monte Carlo simulation experiment based on the detection threshold, and take a detection probability of the received signal in which the radiation source signal exists as a second label;

[0041] a training data set construction unit configured to construct a first training data set based on the environment parameter and the first label, and construct a second training data set based on the environment parameter and the second label;

[0042] a training unit configured to iteratively update a threshold estimation model based on the first training data set, and iteratively update a detection probability prediction model based on the second training data set.

[0043] In a third aspect, a computer device is provided, which includes a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, and the processor implements the steps of the model training method based on the multi-channel detection system when executing the computer readable instructions.

[0044] In a fourth aspect, a readable storage medium is provided, which stores computer readable instructions, and the computer readable instructions implement the steps of the model training method based on the multi-channel detection system when executed by a processor.

[0045] The above-mentioned model training method, device, computer equipment and storage medium based on a multi-channel detection system are implemented by the following methods: obtaining environmental parameters of the multi-channel detection system; simulating a radiation source signal and a received signal based on the environmental parameters, determining a detection threshold based on the received signal and a preset false alarm probability, and using the detection threshold as a first label; performing multiple detections on the generated received signal through Monte Carlo simulation experiments based on the detection threshold, and statistically calculating the detection probability of the radiation source signal in the received signal, and using the detection probability as a second label; constructing a first training data set based on the environmental parameters and the first label, and constructing a second training data set based on the environmental parameters and the second label; iteratively updating a threshold estimation model based on the first training data set, and iteratively updating a detection probability prediction model based on the second training data set. In an embodiment of the present application, by modeling the monotonic relationship between the environmental parameters and the detection probability in the multi-channel detection system, it is ensured that the prediction results maintain physically reasonable monotonicity as key parameters change. Furthermore, a mapping relationship between the input parameters and the detection threshold is constructed to achieve adaptive threshold setting for different system configurations and target false alarm probability requirements. This achieves effectiveness and robustness in predicting detection performance and determining thresholds. It significantly improves the performance evaluation efficiency and actual deployment capabilities of multi-channel detection systems in complex environments, and provides key technical support for intelligent signal detection and system optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0047] Figure 1 This is a flow chart of a model training method based on a multi-channel detection system in one embodiment of the present application;

[0048] Figure 2 is a predicted detection probability and a true detection probability in an embodiment of the present application;

[0049] Figure 3 This is a structural diagram of a model training device based on a multi-channel detection system in one embodiment of the present application;

[0050] Figure 4 Schematic diagram of a computer device in one embodiment of the present application. DETAILED DESCRIPTION

[0051] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the scope of the present application.

[0052] In an embodiment, as shown in Figure 1 A model training method based on a multi-channel detection system is provided, including the following steps:

[0053] In step S110, an environment parameter of a multi-channel detection system is acquired.

[0054] The multi-channel detection system refers to a detection device / framework capable of simultaneously collecting and processing multiple channel signals in wireless communication, radar signal processing and other scenarios. The environment parameter can include signal bandwidth, sampling rate, false alarm probability, observation time, number of receiving nodes, and signal-to-noise ratio of receiving nodes. The receiving node usually refers to a terminal device or a base station, and the number of receiving nodes refers to the number of receiving terminals in wireless communication.

[0055] In step S120, a radiation source signal and a received signal are simulated based on the environment parameter, a detection threshold is determined based on the received signal and a preset false alarm probability, and the detection threshold is used as a first label.

[0056] Optionally, the radiation source signal generation process simulates a real communication signal processing flow, sequentially completes baseband signal generation, shaping filtering, QPSK modulation, and carrier shift. When modeling the received signal, Gaussian white noise conforming to complex Gaussian distribution is superimposed, and signal propagation is simulated through a far-field fast fading channel model to generate a corresponding received signal. After obtaining the received signal, noise characteristics are extracted in a signal-free period, a detection threshold is calculated in combination with a preset false alarm probability, and the detection threshold is used as a first label to participate in the training process of a threshold estimation model.

[0057] In step S130, the generated received signal is detected multiple times through a Monte Carlo simulation experiment based on the detection threshold, and a detection probability of the received signal containing a radiation source signal is counted, and the detection probability is used as a second label.

[0058] Optionally, whether there is a radiation source signal in the received signal can be detected by a signal detection algorithm, such as a generalized likelihood ratio test (GLR) method. A Monte Carlo experiment can be introduced, and the simulation experiment can be run a preset number of times, for example, 100 times, and the detection probability of the received signal existing a radiation source signal in the 100 simulation experiments can be counted. For example, the environmental parameters can be divided into a preset number of environmental parameter groups, each group containing different parameter combinations. In each simulation experiment, the real communication environment signal processing flow is simulated based on each group of environmental parameters, from baseband signal generation, shaping filtering, QPSK modulation, and carrier shift. When modeling the received signal, a Gaussian white noise with a complex Gaussian distribution is used, and a far field, fast fading, and other channel models are used to generate the corresponding received signal. Then, energy detection or GLRT algorithms are used to process the received signal to determine whether there is a radiation source signal therein. Finally, by a large number of independent simulations, the proportion of the number of correct detections under each group of parameters is counted to determine the detection probability.

[0059] In step S140, a first training data set is constructed based on the environmental parameters and the first label, and a second training data set is constructed based on the environmental parameters and the second label;

[0060] Optionally, after obtaining the environmental parameters, the environmental parameters that can be used to perform threshold estimation can be selected according to the input requirements of the threshold estimation model, such as the sampling rate of the receiver, the false alarm probability, the observation time, and the number of receivers. Then, the environmental parameters can be preprocessed, such as data cleaning, noise reduction, and normalization, and then the first training data set is constructed together with the first label. The first training data set is used for the training process of threshold estimation. It should be noted that the first training data set can include multiple training data combinations, and each training data combination can include the sampling rate of the receiver, the false alarm probability, the observation time, and the number of receivers. All data in the training data combination needs to be input during each model training.

[0061] In addition, after obtaining the environmental parameters, the environmental parameters that can be used to perform detection probability prediction can be selected according to the input requirements of the detection probability prediction model, such as the received signal bandwidth, the sampling rate, the observation time, the false alarm probability, and the signal-to-noise ratio feature. Then, the environmental parameters can be preprocessed, such as data cleaning, noise reduction, and normalization, and then the second training data set is constructed together with the second label. The second training data set is used for the training process of detection probability prediction. It should be noted that the second training data set can include multiple training data combinations, and each training data combination can include the received signal bandwidth, the sampling rate, the observation time, the false alarm probability, and the signal-to-noise ratio feature. All data in each training data combination needs to be input during each model training.

[0062] In step S150, the threshold estimation model is iteratively updated based on the first training dataset, and the detection probability prediction model is iteratively updated based on the second training dataset.

[0063] It should be noted that the threshold estimation model and the detection probability prediction model can both be APMNN neural network models (adaptive physical constraint neural network models), which can specifically include an input layer, a hidden layer, and an output layer, wherein the hidden layer is composed of multiple groups of weight linear transformation units, each group containing several positive weight linear units and / or negative weight linear units, and the weight linear transformation units ensure the monotonicity between the input features and the output features through weight constraints. Specifically, the hidden layer of the detection probability prediction model can be composed of multiple groups of weight linear transformation units, each group containing several positive weight linear units, and the weight linear transformation units ensure the increasing monotonicity between the input features and the output features through positive weight constraints. The hidden layer of the threshold estimation model can be composed of multiple groups of weight linear transformation units, each group containing several positive weight linear units and several negative weight linear units, and the weight linear transformation units ensure the increasing monotonicity and the decreasing monotonicity between the input features and the output features by taking the square or negative square operation of the weights, thereby forcing the weights to be positive or negative.

[0064] The first training dataset can include the sampling rate of the receiver, the false alarm probability, the observation time, and the number of receivers parameters, i.e., the input data input into the threshold estimation model each time is the sampling rate of the receiver, the false alarm probability, the observation time, and the number of receivers parameters. The second training dataset can include the received signal bandwidth, the sampling rate, the observation time, the false alarm probability, and the signal-to-noise ratio features, i.e., the input data input into the detection probability prediction model each time is the received signal bandwidth, the sampling rate, the observation time, the false alarm probability, and the signal-to-noise ratio features. Therefore, the input features of the threshold estimation model and the detection probability prediction model are different. For the detection probability prediction model, the output value is limited between 0 and 1 through the activation function in the output layer, such as the Sigmoid activation function, which is used to predict the detection probability. The threshold estimation model outputs the threshold, so the threshold estimation model is obtained by removing the Sigmoid activation function in the output layer of the detection probability prediction model, and at this time, the relu activation function is used for activation processing in the output layer.

[0065] The model training is achieved by optimizing the network parameters to minimize the prediction error. The model performance is tested by the validation set and the hyperparameters (such as the learning rate and the number of training rounds) are adjusted to ensure the generalization ability of the model on unseen data. After sufficient verification, the optimized model is saved for practical application.

[0066] In the training phase, the error between the predicted value of the model and the true value (label) can be calculated by the loss function (L1Loss). Then, through the backpropagation algorithm, the model updates the weight parameters according to the error (using the Adam optimizer). During the training process, the generalization ability of the model is evaluated according to its performance on the validation set, so as to select the optimal combination of hyperparameters.

[0067] The verification process is similar to the training process, but the verification data is not used to update the weight parameters of the model, only to evaluate the performance of the model. Through the evaluation on the validation set, the model can find the best parameter combination (such as the number of groups, the number of units, the learning rate, etc.), so as to obtain the lowest validation error.

[0068] The trained model uses fixed weight parameters in the prediction phase, and directly runs the forward propagation process when inputting features. Compared with the traditional Monte Carlo simulation-based method, the prediction process of the APMNN neural network model greatly reduces the computational cost. Due to the structural design of the model, the prediction phase does not need to perform complex simulation or iteration, and only one forward propagation is needed to complete the prediction of the detection probability or threshold. Therefore, the trained model can receive any parameter combination in real time and quickly calculate the output result.

[0069] Compared with the model in the training phase, the prediction process of the trained model is more efficient, does not involve gradient calculation and weight update, and only needs to perform one forward propagation, which meets the requirements of real-time and high efficiency in actual application scenarios. This design significantly improves the prediction efficiency and real-time response ability of the model, avoids repeated simulation, and meets the practical application requirements of multi-channel detection systems in complex environments.

[0070] In the embodiments of the present application, a model training method based on a multi-channel detection system is provided, and the method implementation comprises: acquiring an environmental parameter of the multi-channel detection system; simulating a radiation source signal and a received signal based on the environmental parameter; determining a detection threshold based on the received signal and a preset false alarm probability, and taking the detection threshold as a first label; based on the detection threshold, performing multiple detections on the generated received signal through a Monte Carlo simulation experiment, and statistically determining a detection probability of the received signal in which the radiation source signal exists, and taking the detection probability as a second label; based on the environmental parameter and the first label, constructing a first training data set, and based on the environmental parameter and the second label, constructing a second training data set; iteratively updating a threshold estimation model based on the first training data set, and iteratively updating a detection probability prediction model based on the second training data set. And further construct the mapping relationship between the input parameters and the detection threshold, realize the adaptive setting of the threshold for different system configurations and target false alarm probability requirements. The effectiveness and robustness in predicting the detection performance and determining the threshold are realized. The performance evaluation efficiency and actual deployment ability of the multi-channel detection system in a complex environment are significantly improved, which provides key technical support for intelligent signal detection and system optimization.

[0071] In an embodiment of the present application, based on the environmental parameter, the radiation source signal and the received signal in the real scene are simulated and generated, which comprises:

[0072] Constructing a multi-node received signal model;

[0073] Based on the multi-node received signal model, modulating the low-frequency baseband signal to generate an analog signal with a frequency range greater than a preset frequency range, which is easy to transmit in the simulation channel;

[0074] The analog signal is transmitted to the receiving end through the channel and demodulated to obtain the received signal.

[0075] Optionally, the generation of the radiation source signal comprises baseband signal generation, shaping filtering and QPSK modulation implementation to simulate the common communication signal processing flow. The channel modeling adopts a far-field, fast-fading wireless channel model. In a simplified scenario, it is assumed that the transmitting end and the receiving end are respectively equipped with only one antenna, and the two ends are relatively static, at this time the channel model from the transmitting end to the i-th receiving end can be expressed as A i = b i e jφ The received signal of each perception node is:

[0076] y i (t) = A i g(t-D) + n i (t);

[0077] Wherein, Ai represents the complex channel gain from the signal source to the i-th receiving station, g(t) represents the emission signal of the radiation source at time t, D represents the signal time delay from the radiation source to the i-th receiving station, n i (t) represents the noise of the i-th receiving station, which is subject to a zero-mean complex Gaussian distribution.

[0078] In an embodiment of the present application, the detection threshold is determined based on the received signal and a preset false alarm probability, and the method comprises:

[0079] noise is determined based on the received signal;

[0080] The detection threshold is determined based on the preset false alarm probability and the noise.

[0081] Optionally, when the detector is constructed according to the energy detection method, the data collected by the sensing node for a certain length of time is first taken as noise samples when the radiation source is in the off state, and the energy accumulation calculation is performed on these noise data to obtain a test statistic representing the noise energy level. In order to ensure the accuracy of detection, a large number of noise energy samples are generated by using Monte Carlo experiments, and through statistical analysis of these samples, combined with the pre-set false alarm probability (i.e. the probability threshold of mistakenly judging that there is a signal when there is no signal), when the radiation source is in the off state, the data collected by the sensing node is taken as noise, and the test statistic is obtained from the noise data. The detection threshold is calculated through multiple Monte Carlo experiments and combined with the pre-set false alarm probability. The detection threshold can be used as a reference value to judge whether the radiation source is working, and when the received signal energy exceeds this threshold, it is determined that the radiation source is in the emission state; otherwise, it is considered that there is no signal, thereby constructing an energy detector that can adapt to different noise environments and has reliable detection performance, and effectively sensing the working state of the radiation source.

[0082] In an embodiment of the present application, the generated received signal is detected multiple times through Monte Carlo simulation experiments based on the detection threshold, and the detection probability of the received signal with the radiation source signal is counted, and the method comprises:

[0083] The first probability density that the received signal has the radiation source signal and the second probability density that the received signal does not have the radiation source signal are determined based on the detection threshold;

[0084] The likelihood ratio is obtained based on the ratio between the first probability density and the second probability density;

[0085] If the likelihood ratio is greater than a preset decision threshold, it is determined that the received signal has the radiation source signal;

[0086] The detection probability of the received signal with the radiation source signal in a preset number of Monte Carlo simulation experiments is counted.

[0087] Specifically, in distributed communication (such as multi-antenna, cooperative sensing), it is necessary to determine "whether there is a target signal" from the received signals of multiple nodes to control the false alarm probability, that is, the probability of misjudging as having a signal when there is no signal, and maximize the detection probability, that is, the probability of correct detection when there is a signal. Therefore, the generalized likelihood ratio test (GLR) method is used to detect the existence of a signal based on the received signal, and the Neyman-Pearson (NP) criterion is used for distributed communication signal detector design criterion. The NP criterion is the optimal criterion for hypothesis testing. Given the maximum acceptable false alarm probability, the detector is designed to maximize the detection probability. The NP criterion indicates that for a given false alarm probability, the optimal detection that maximizes the detection probability is defined as the LR detector:

[0088]

[0089] where p(g) represents the probability density function (PDF), y(t) represents the received signal vector, Y = [y(1),..., y(T)] ∈ N×T , y(t) = [y1(t),..., y N (t)] T ∈￡ N , t = 1,..., T, H0 represents a radiation source signal, and only noise, H1 represents a radiation source signal and noise, and p(Y | H1) p(Y|H1) represents the probability density of the observed data under H1, p(Y|H0) represents the probability density of the observed data under H0, and the statistical quantity ξ LR is greater than the threshold value to determine that there is a radiation source signal, and otherwise, it is determined that there is no radiation source signal. Finally, the behavior of the multi-channel detection system in a complex environment is simulated based on Monte Carlo experiments, and the proportion of the number of correct detections under each group of parameters is counted to determine the detection probability.

[0090] In an embodiment of the present application, the detection probability prediction model is iteratively updated based on the second training data set, comprising:

[0091] Selecting initial training data in the second training data set and inputting the initial training data into the input layer;

[0092] Inputting the initial training data into the hidden layer through the input layer, processing through multiple groups of weight linear units in parallel, and extracting the maximum response value in the output results of each group of weight linear units to obtain the output results of each group of weight linear units, wherein each group of weight linear units ensures that the weight is positive by squaring the weight, so that the input feature and the output feature have a monotonically increasing relationship;

[0093] The output results of the groups of weight linear units are minimized by an output layer, and are activated by a preset activation function to obtain a detection probability;

[0094] A second loss value is calculated based on the detection probability, the second label, and a second loss function;

[0095] The next round of iteration of the detection probability prediction model is performed based on the second loss value until a preset convergence condition is met.

[0096] Optionally, the detection probability prediction model can be an APMNN neural network model, which can specifically include an input layer, a hidden layer, and an output layer. The input layer receives signal bandwidth, sampling rate, observation time, false alarm probability, and signal-to-noise ratio features and inputs them to the hidden layer. The hidden layer is composed of multiple groups of weight linear transformation units, each group containing several positive weight linear units. The weight linear transformation unit ensures the monotonicity between the input features and the output features through weight constraints. The output layer uses a Sigmoid activation function to limit the output value to between 0 and 1, which is used to predict the detection probability.

[0097] The processing process is specifically as follows:

[0098] First, the training data in the second training data set is input into the input layer. The training data can include signal bandwidth, sampling rate, observation time, false alarm probability, and signal-to-noise ratio. The input layer inputs them to the hidden layer. The hidden layer is composed of multiple groups of PositiveLinear weight linear transformation units, each group containing several positive weight linear units. PositiveLinear is a special linear layer that ensures monotonicity between input features and output through weight constraints. For all features, the weight is constrained to be positive (achieved by taking the square), thereby achieving an increasing relationship. The output of each group of hidden layers is extracted by the torch.max operation to respond to the maximum value in the group, highlighting the key feature influence, and ensuring that the final output of each group can reflect the strongest feature response. After completing all weight linear transformation unit calculations, the output features of each weight linear unit can be input into the output layer. The output layer extracts the minimum value from all group outputs through the torch.min operation to ensure that the result is constrained by the weakest feature. This design ensures that the output value can consider the contribution of all features and maintain strict monotonicity in a physical sense. Finally, the output layer uses a Sigmoid activation function to limit the output value to the [0, 1] interval, representing the detection probability. Then, based on the detection probability, the second label, and the second loss function, the loss value is calculated, the model parameters are adjusted based on the loss value, and the next round of iteration is started based on the updated model. When the preset convergence condition is met, the trained detection probability prediction model is obtained.

[0099] wherein the positive weight linear unit can specifically be, determining an initial weight parameter, generating a non-negative effective weight through a square operation of the initial weight, performing a regular linear transformation on the input signal bandwidth, sampling rate, observation time, false alarm probability and signal-to-noise ratio using the effective weight (i.e. multiplying the effective weight with the input and adding a bias), so that the input feature and the output present a monotonically increasing relationship due to the non-negativity of the effective weight, and directly updating the initial weight by the gradient during training. After each group of positive weight linear units is processed, the value with the largest response in the group can be extracted by torch.max operation, which can be used as the output of the PositiveLinear (positive weight linear layer) weight linear transformation unit.

[0100] wherein the loss value between the predicted value and the true value of the model can be calculated by a loss function (L1Loss). Subsequently, the model updates the weight parameters according to the error by using the back propagation algorithm (using Adam optimizer).

[0101] In an embodiment of the present application, the iterative updating of the threshold estimation model based on the first training data set comprises:

[0102] Randomly selecting initial training data in the first training data set and inputting the initial training data into the input layer;

[0103] inputting the initial training data into the hidden layer through the input layer, processing in parallel through multiple groups of weight linear units, and extracting the maximum response value in the output results of each group of weight linear units to obtain the output results of each group of weight linear units, wherein each group of weight linear units processes the input training data and the output results in a monotonically increasing relationship and a monotonically decreasing relationship through square and negative square operations of the weight;

[0104] taking the minimum value of the output results of each group of weight linear units through the output layer to obtain the optimal detection threshold;

[0105] calculating a first loss value based on the optimal detection threshold, the first label and the first loss function;

[0106] performing the next round of iteration of the threshold estimation model based on the first loss value until the preset convergence condition is met.

[0107] ​Optionally, the threshold estimation model can be an APMNN neural network model, which can specifically include an input layer, a hidden layer, and an output layer. The input layer receives features such as sampling rate, false alarm probability, observation time, and receiver number parameters, and inputs them to the hidden layer. The hidden layer is composed of multiple groups of weight linear transformation units, each group containing several positive weight linear units and several negative weight linear units. The weight linear transformation unit ensures the monotonicity between the input features and the output features through weight constraints. The output layer discards the Sigmoid activation function and directly linearly transforms the hidden layer output features for threshold estimation.

[0108] The processing process is specifically as follows:

[0109] First, the training data in the first training data set is input into the input layer,

[0110] The training data can include the sampling rate, false alarm probability, observation time, and receiver number parameters. The sampling rate, false alarm probability, and observation time can be training data in an increasing relationship, and the receiver number can be training data in a decreasing relationship. Through the input layer, they are input into the hidden layer. The hidden layer is composed of multiple groups of PositiveLinear weight linear transformation units, each group containing several positive weight linear units and negative weight linear units. The positive / negative weight linear units are generated through weight square / negative square operations, respectively corresponding to increasing / decreasing mapping. PositiveLinear is a special linear layer that ensures the monotonicity between the input features and the output through weight constraints. For all features, the weights are constrained to be positive (achieved by taking square) and negative (achieved by taking negative square), such as the sampling rate, false alarm probability, observation time, and threshold in an increasing relationship, processed through positive weight units; the receiver number and the threshold in a decreasing relationship, processed through negative weight units, thereby realizing increasing relationship and decreasing relationship. The output of each group of hidden layers is extracted through the torch.max operation to respond to the maximum value in the group, ensuring that the final output of each group can reflect the strongest feature response. The output layer extracts the minimum value in all group outputs through the torch.min operation to obtain the optimal detection threshold, ensuring that the result is constrained by the weakest feature. This design ensures that the output value can consider the contribution of all features and maintain strict monotonicity in a physical sense.

[0111] Then, based on the optimal detection threshold, the first label, and the first loss function, the loss value is calculated. If the loss value is greater than the preset loss threshold, the model parameters are adjusted, and the next iteration is started. Until the preset convergence condition is met, such as the number of iterations reaching the preset number, or the loss value is less than or equal to the preset loss threshold, the trained threshold estimation model can be obtained.

[0112] The positive weight linear unit can specifically be determining an initial weight parameter, generating a non-negative effective weight by squaring the initial weight parameter, and generating a negative effective weight by negative squaring the initial weight parameter, performing a regular linear transformation on the input received sampling rate, false alarm probability, and observation time using the non-negative effective weight (i.e., multiplying the input by the effective weight and adding a bias), and performing a regular linear transformation on the receiver quantity parameter through the negative effective weight (i.e., multiplying the input by the effective weight and adding a bias). The input feature and the output are in a monotonically increasing relationship due to the non-negative effective weight, and the input feature and the output are in a monotonically decreasing relationship due to the negative effective weight, and the gradient directly updates the initial weight during training. After each group of positive weight linear units is processed, the value with the largest response in the group can be extracted through a torch.max operation, which can be used as the output of the PositiveLinear weight linear unit.

[0113] The loss value between the predicted value of the model and the true value can be calculated by a loss function (L1Loss). Subsequently, the model updates the weight parameters according to the error through a back propagation algorithm (using an Adam optimizer).

[0114] To further verify the technical effects of the present application, the following experiment is designed: It should be noted that in order to capture the key characteristics of the signal environment and the receiver settings, the feature vectors of the data set are carefully designed to cover signal bandwidth, sampling rate, observation time, false alarm probability, and signal-to-noise ratio of each node. These features ensure that the model can fully reflect the environmental characteristics of the multi-channel detection system and its impact. The data label represents the detection probability obtained through 100 Monte Carlo experiments, which is true and reliable. The radiation source signal is generated in QPSK modulation mode to simulate the signal in the actual communication scenario. The generation of the signal is completed using the MATLAB communication toolbox to ensure data quality and consistency. The training data set generated in this way can provide a wide range of feature distributions for the model to support detection performance modeling for various parameter combinations.

[0115] The network architecture of the detection probability prediction model can include an input layer, a hidden layer, and an output layer, aiming to achieve nonlinear modeling and prediction of detection performance. The input layer is responsible for receiving the feature vector, including signal bandwidth, sampling rate, observation time, false alarm probability, and signal-to-noise ratio of each node, ensuring that the model can learn the complex relationship between environmental parameters and detection performance. The hidden layer adopts a full positive weight constraint to ensure that the output maintains monotonicity with respect to the input parameters, consistent with physical laws. The output layer limits the output to between 0 and 1 through a Sigmoid activation function, which is used to represent the detection probability.

[0116] During the model training process, the mean square error (MSE) is used as the main performance evaluation index. In the optimization stage, the stochastic gradient descent (SGD) optimizer is used, and the learning rate is adjusted in an adaptive manner to ensure that the model parameters converge quickly and efficiently. To optimize the model performance, grid search is used to optimize the hyperparameters, including the learning rate, the number of training rounds, and other parameters. Finally, the optimal hyperparameter combination (such as setting the learning rate to 0.01, the number of training rounds to 2000, and dividing the neurons in the hidden layer into 3 groups to associate specific features or feature combinations with specific network units to improve the model's interpretability and enhance monotonicity control) is selected for model training. By systematically traversing the parameter combinations and comparing their performance on the validation set, the model with the smallest validation error is selected for final training to ensure that the model performance is optimal.

[0117] As shown in Figure 2 , a comparison graph of predicted detection probability and true detection probability is provided, from which it can be seen that the performance of the monotonic network in the detection performance prediction task is very close to the result of Monte Carlo simulation. When other parameters are fixed, the detection probability gradually increases with the increase of signal-to-noise ratio, which conforms to the physical law. This fully verifies the monotonicity constraint and prediction accuracy of the network. The experiment also shows that the network can accurately capture the influence of environmental parameters on detection performance as the signal-to-noise ratio changes, further proving the effectiveness of the proposed architecture and optimization process.

[0118] In the embodiments of the present application, the monotonic relationship between environmental parameters and detection probability in a multi-channel detection system is modeled to ensure that the prediction result remains physically reasonable and monotonic as the key parameters change. Furthermore, a mapping relationship between input parameters and detection thresholds is constructed to achieve adaptive threshold setting for different system configurations and target false alarm probability requirements. The effectiveness and robustness of the proposed architecture and optimization process in predicting detection performance and determining thresholds are achieved. The performance evaluation efficiency and actual deployment capability of the multi-channel detection system in complex environments are significantly improved, providing key technical support for intelligent signal detection and system optimization.

[0119] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0120] In an embodiment, a model training device based on a multi-channel detection system is provided, which corresponds to the model training method based on the multi-channel detection system in the above-mentioned embodiments. As shown in the following figure, the model training device based on the multi-channel detection system includes an environment parameter acquisition unit 10, a first label generation unit 20, a second label generation unit 30, a training data set construction unit 40, and a training unit 50. The functions of each module are described in detail as follows: Figure 3

[0121] The environment parameter acquisition unit 10 is configured to acquire environment parameters of the multi-channel detection system.

[0122] The first label generation unit 20 is configured to simulate a radiation source signal and a received signal based on the environment parameters, determine a detection threshold based on the received signal and a preset false alarm probability, and use the detection threshold as a first label.

[0123] The second label generation unit 30 is configured to perform multiple detections on the generated received signal through Monte Carlo simulation experiments based on the detection threshold, and count a detection probability of the received signal containing the radiation source signal, and use the detection probability as a second label.

[0124] The training data set construction unit 40 is configured to construct a first training data set based on the environment parameters and the first label, and construct a second training data set based on the environment parameters and the second label.

[0125] The training unit 50 is configured to iteratively update a threshold estimation model based on the first training data set, and iteratively update a detection probability prediction model based on the second training data set.

[0126] In an embodiment of the present application, the first label generation unit 20 is further configured to:

[0127] construct a multi-node received signal model;

[0128] modulate a low-frequency baseband signal based on the multi-node received signal model to generate an analog signal with a frequency range greater than a preset frequency range, which is easy to transmit in an analog channel;

[0129] demodulate the analog signal transmitted through the channel to the receiving end to obtain a received signal.

[0130] In an embodiment of the present application, the second label generation unit 30 is further configured to:

[0131] determine a first probability density of the received signal containing the radiation source signal and a second probability density of the received signal not containing the radiation source signal based on the detection threshold;

[0132] ​obtaining a likelihood ratio based on a ratio between the first probability density and the second probability density;

[0133] if the likelihood ratio is greater than a preset decision threshold, determining that the received signal contains a radiation source signal;

[0134] counting a detection probability of determining that the received signal contains a radiation source signal in a preset number of Monte Carlo simulation experiments.

[0135] In an embodiment of the present application, the first label generating unit 20 is further configured to:

[0136] determining noise based on the received signal;

[0137] determining the detection threshold based on the preset false alarm probability and the noise.

[0138] In an embodiment of the present application, the threshold estimation model and the detection probability prediction model each include an input layer, a hidden layer, and an output layer, wherein the hidden layer is composed of multiple sets of weight linear transformation units, each set containing several positive weight linear units and / or negative weight linear units, and the weight linear transformation units ensure the monotonicity between input features and output features through weight constraints.

[0139] In an embodiment of the present application, the training unit 40 is further configured to:

[0140] randomly selecting initial training data in the first training data set and inputting the initial training data into the input layer;

[0141] inputting the initial training data into the hidden layer through the input layer, processing the initial training data in parallel through multiple sets of weight linear units, and extracting a maximum response value from output results of each set of weight linear units to obtain the output results of each set of weight linear units, wherein each set of weight linear units makes the input training data and the output results in a monotonically increasing relationship and a monotonically decreasing relationship through a weight squaring operation and a weight negative squaring operation;

[0142] obtaining an optimal detection threshold by taking a minimum value of the output results of each set of weight linear units through the output layer;

[0143] calculating a first loss value based on the optimal detection threshold, the first label, and a first loss function;

[0144] performing a next round of iteration on the threshold estimation model based on the first loss value until a preset convergence condition is met.

[0145] In an embodiment of the present application, the training unit 40 is further configured to:

[0146] selecting initial training data in the second training data set and inputting the initial training data into the input layer;

[0147] inputting the initial training data to a hidden layer through the input layer, processing in parallel through multiple sets of weight linear units, and extracting a maximum response value in output results of the multiple sets of weight linear units to obtain the output results of the multiple sets of weight linear units, wherein each set of weight linear units ensures weights to be positive by squaring the weights, so that the input features and the output features have a monotonically increasing relationship;

[0148] taking a minimum value of the output results of the multiple sets of weight linear units through an output layer, and performing activation processing through a preset activation function to obtain a detection probability;

[0149] calculating a second loss value based on the detection probability, a second label, and a second loss function;

[0150] performing a next round of iteration on the detection probability prediction model based on the second loss value until a preset convergence condition is met.

[0151] In the embodiments of the present application, the monotonic relationship between the environmental parameters and the detection probability in the multi-channel detection system is modeled, so as to ensure that the prediction result remains physically reasonable and monotonous with the change of the key parameters. Furthermore, a mapping relationship between the input parameters and the detection threshold is constructed to realize adaptive setting of the threshold for different system configurations and target false alarm probability requirements. The effectiveness and robustness of the prediction detection performance and the determination of the threshold are realized. The performance evaluation efficiency and the actual deployment capability of the multi-channel detection system in a complex environment are significantly improved, which provides key technical support for intelligent signal detection and system optimization.

[0152] The specific limitations of the model training device based on the multi-channel detection system can be referred to the limitations of the model training method based on the multi-channel detection system in the foregoing, which will not be described herein. Each module in the model training device based on the multi-channel detection system can be realized by software, hardware, and combinations thereof, in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0153] In one embodiment, a computer device is provided, which can be a terminal device, and an internal structure diagram of the computer device can be as shown in Figure 4As shown. The computer device includes a processor, a memory, a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium. The readable storage medium stores computer readable instructions. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer readable instructions are executed by the processor to implement a model training method based on a multi-channel detection system. The readable storage medium provided in the embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0154] In the embodiments of the present application, a computer device is provided, including a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, and the processor executes the computer readable instructions to implement the steps of the model training method based on the multi-channel detection system as described above.

[0155] In the embodiments of the present application, a readable storage medium is provided, and the readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the steps of the model training method based on the multi-channel detection system as described above.

[0156] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by computer readable instructions instructing related hardware, and the computer readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium, and when executed, can include the processes of the above-mentioned embodiments of each method. Among them, any reference to memory, storage, database or other medium used in each embodiment of the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM) and the like.

[0157] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0158] The above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A model training method based on a multi-channel detection system, characterized in that: The method comprises: Obtain environmental parameters of the multi-channel detection system; Simulating a radiation source signal and a received signal based on the environmental parameters, determining a detection threshold based on the received signal and a preset false alarm probability, and using the detection threshold as a first label; Based on the detection threshold, the generated received signal is detected multiple times through a Monte Carlo simulation experiment, and the detection probability of the radiation source signal in the received signal is calculated, and the detection probability is used as the second label; Constructing a first training data set based on the environmental parameters and the first label, and constructing a second training data set based on the environmental parameters and the second label; The threshold estimation model is iteratively updated based on the first training data set, and the detection probability prediction model is iteratively updated based on the second training data set.

2. The model training method based on the multi-channel detection system according to claim 1, characterized in that: The simulating and generating the radiation source signal and the receiving signal in the real scene based on the environmental parameters includes: Construct a multi-node receiving signal model; Based on the multi-node received signal model, modulating the low-frequency baseband signal to generate an analog signal that is easy to transmit in an analog channel and has a frequency range greater than a preset frequency range; The analog signal is transmitted to the receiving end through a channel for demodulation to obtain a received signal.

3. The model training method based on the multi-channel detection system according to claim 1, characterized in that: The method of performing multiple detections on the generated received signal based on the detection threshold through a Monte Carlo simulation experiment and calculating the detection probability of the radiation source signal in the received signal includes: Based on the detection threshold, determining a first probability density of the presence of the radiation source signal and a second probability density of the absence of the radiation source signal in the received signal; Obtaining a likelihood ratio based on a ratio between the first probability density and the second probability density; If the likelihood ratio is greater than a preset decision threshold, it is determined that a radiation source signal exists in the received signal; The detection probability of the radiation source signal existing in the received signal is determined by statistically performing a preset number of Monte Carlo simulation experiments.

4. The model training method based on the multi-channel detection system according to claim 1, characterized in that: The determining a detection threshold based on the received signal and a preset false alarm probability includes: determining noise based on the received signal; The detection threshold is determined based on the preset false alarm probability and the noise.

5. The model training method based on the multi-channel detection system according to claim 1, characterized in that: The threshold estimation model and the detection probability prediction model both include an input layer, a hidden layer, and an output layer, wherein the hidden layer is composed of multiple groups of weighted linear transformation units, each group containing several positive weighted linear units and / or negative weighted linear units, and the weighted linear transformation units ensure the monotonicity between input features and output features through weight constraints.

6. The model training method based on a multi-channel detection system according to claim 1 or 5, characterized in that: The iterative updating of the threshold estimation model based on the first training data set includes: Randomly selecting initial training data from the first training data set and inputting it into the input layer; Inputting the initial training data into the hidden layer through the input layer, processing them in parallel through multiple groups of weighted linear units, and extracting the maximum response value from the output results of each group of weighted linear units to obtain the output results of each group of weighted linear units, wherein each group of weighted linear units performs weight squaring and weight negative squaring operations so that the input training data and the output results are in a monotonically increasing relationship and a monotonically decreasing relationship; The output layer takes the minimum value of the output results of each group of weighted linear units to obtain the optimal detection threshold; Calculating a first loss value based on the optimal detection threshold, the first label, and the first loss function; The threshold estimation model is iterated for the next round based on the first loss value until a preset convergence condition is met.

7. The model training method based on a multi-channel detection system according to claim 1 or 5, characterized in that: The iteratively updating the detection probability prediction model based on the second training data set includes: Selecting initial training data from the second training data set and inputting it into the input layer; Inputting the initial training data into the hidden layer through the input layer, processing the data in parallel through multiple groups of weighted linear units, and extracting the maximum response value from the output results of each group of weighted linear units to obtain the output results of each group of weighted linear units, wherein each group of weighted linear units squares the weights to ensure that the weights are positive, so that the input features and the output features are in a monotonically increasing relationship; The output layer takes the minimum value of the output results of each group of weighted linear units, and performs activation processing through a preset activation function to obtain the detection probability; Calculating a second loss value based on the detection probability, the second label, and the second loss function; The detection probability prediction model is iterated for the next round based on the second loss value until a preset convergence condition is met.

8. A model training device based on a multi-channel detection system, characterized in that: The device comprises: An environmental parameter acquisition unit, used to acquire environmental parameters of the multi-channel detection system; A first label generating unit is configured to simulate a radiation source signal and a received signal based on the environmental parameters, determine a detection threshold based on the received signal and a preset false alarm probability, and use the detection threshold as a first label; a second label generating unit, configured to perform multiple detections on the generated received signal through a Monte Carlo simulation experiment based on a detection threshold, and calculate a detection probability of a radiation source signal being present in the received signal, and use the detection probability as a second label; a training data set construction unit, configured to construct a first training data set based on the environmental parameter and the first label, and to construct a second training data set based on the environmental parameter and the second label; A training unit is used to iteratively update the threshold estimation model based on the first training data set, and iteratively update the detection probability prediction model based on the second training data set.

9. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein: When the processor executes the computer-readable instructions, the steps of the model training method based on the multi-channel detection system as described in any one of claims 1 to 7 are implemented.

10. A readable storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps of the model training method based on a multi-channel detection system as described in any one of claims 1 to 7 are implemented.

Images