Machine learning based online detection of multi-function radar operating mode switching points

Abstract

The application discloses a kind of based on machine learning's multifunctional radar operating state switching point online detection method, comprising: step 1, the offline training of switching point prediction model: based on the radar intercepted signal sequence data set with operating state label, extract the feature vector with operating state switching point label, then this label sample data set is used to train the operating state switching classifier model based on machine learning, obtain model parameter;Step 2, the online prediction of operating state switching point: based on the radar signal intercepted in online flow mode, the operating state switching change feature vector of current time is extracted in real time, the operating state switching classifier model, model parameter and preset hyperparameter are combined, and the prediction result of the operating state switching point at this time is given.The application can automatically fit the distribution of each dimension of feature data and find the best judgment threshold, realize the accurate classification detection of multifunctional radar operating state switching point.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an online detection method for switching points of multifunctional radar operating states based on machine learning. Background Technology

[0002] To address the problem of online switching point detection for multi-functional radar operating states, Bao Jiadi et al. from Beijing Institute of Technology proposed a method for online switching point detection of multi-functional radar operating states under non-ideal observation. This method first utilizes the principle of autoregressive filtering to repair the pulse sequence through prediction and judgment, and then uses the model parameters estimated by maximum likelihood online to calculate the accumulated likelihood ratio to achieve online switching point detection of operating states. This method has the following three shortcomings: 1) It assumes that the multi-functional radar uses the same waveform pattern of inter-pulse modulation type throughout, while the difference in operating states mainly lies in the change of inter-pulse modulation parameters, which does not conform to the actual definition of radar operating states; 2) This method requires the use of expert domain knowledge to model the waveform pattern of each inter-pulse modulation type separately, which will face increasing challenges when dealing with the complex and diverse waveform patterns of digital programmable radar; 3) The online detection of operating state switching points is highly dependent on manually set accumulated likelihood ratio thresholds, and the rationality of the threshold setting directly affects the accuracy of switching point detection.

[0003] To overcome the need for separate modeling for different waveform patterns and modulation parameters, and to reduce reliance on human intervention, Li Yunjie et al. from Beijing Institute of Technology proposed a multifunctional radar pulse sequence operating state identification and switching point detection method based on a hierarchical sequence-to-sequence long short-term memory (LSTM) network. This method takes the pulse sequence as input, first using a bidirectional LSTM network at the pulse layer to capture sequence features within the same operating state. Then, based on this, a bidirectional LSTM network is used at the segment layer to further capture sequence features between operating states. By determining the operating state to which each pulse belongs, the method achieves accurate identification of the pulse sequence's operating state and detection of switching points. While this method, relying on a hierarchical LSTM deep network, can automatically extract sequence features within and between operating states, its unclear process mechanism can lead to poor robustness in representation performance. Furthermore, it requires a huge amount of training data and floating-point computing power, making deployment at the edge of electronic warfare in a military-game-like context difficult.

[0004] Existing likelihood estimation methods heavily rely on domain expert knowledge to construct representation models and struggle to set reasonable judgment thresholds, easily leading to false detections of operational state switching points in multi-functional radars. Existing deep learning methods, overly reliant on powerful computing resources, are difficult to deploy on resource-constrained electronic warfare edge environments, and insufficient data can result in poor robustness in detecting operational state switching points. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an online detection method for switching points of multi-functional radar operating states based on machine learning. First, it extracts discriminative features that characterize changes in operating states, reducing excessive reliance on data. Then, it utilizes a random forest learning method with lower computational requirements to automatically fit the distribution of feature data across various dimensions and find its optimal judgment threshold, thereby achieving accurate classification and detection of switching points of multi-functional radar operating states.

[0006] The technical solution adopted in this invention is as follows:

[0007] A machine learning-based online detection method for switching points of multi-functional radar operating states includes:

[0008] Step 1: Offline training of the switching point prediction model: Based on the radar intercept signal sequence dataset with working state labels, extract the representation feature vector with working state switching point labels, and then train the working state switching classifier model based on machine learning using this labeled sample dataset to obtain the model parameters.

[0009] Step 2, Online prediction of working state switching points: Based on the radar signal intercepted in the online pipeline mode, the feature vector representing the working state switching change at the current moment is extracted in real time. Combined with the working state switching classifier model, model parameters and preset hyperparameters, the prediction result of the working state switching point at that moment is given.

[0010] Furthermore, the implementation methods of the working state switching classifier model include: random forest classification method, support vector machine classification method, shallow neural network classification method, and K nearest neighbor classification method.

[0011] Further, step 1 includes the following sub-steps:

[0012] Step 101: Extract the representational feature vector with labels for operating state switching points: for the intercepted signal sequence of multi-functional active phased array radar. and its working status label sequence The original dataset D = { <O (i) Q (i) >|i=1,2,…,I};For each intercepted signal sequence with a working status label <O (i) Q (i) > Extract the labels of adjacent coherent processing intervals, i.e., the working state switching points between adjacent CPI clusters. and the characteristic vector representing the change in working state Get data pairs The training and validation sample set S constitutes the working state switching points;

[0013] Step 102: Training the working state switching classifier model: For the previously constructed training and validation sample set S, first equalize the proportion of each label category and split it into training sample sets S′. train and test sample set S′ test Then, the working state switching classifier model based on the random forest ensemble learning method is iteratively trained and tested on these two sample sets. Where θ hp Let θ be the hyperparameter and the model parameter, respectively, and let y∈{0,1} be the label of the working state switching point. Finally, the model parameter θ of the working state switching classifier model is obtained by training.

[0014] Furthermore, in step 101, the O signal sequence intercepted by the multi-functional active phased array radar... (i) of Let represent the characteristic vector of the t-th CPI cluster in {1, 2, ..., T}, which is composed of the start arrival time ftoa, end arrival time ltoa, mean pulse amplitude pa, mean pulse repetition interval pti, and mean pulse width pw within the CPI. Working status label sequence Q (i) In This represents the working status label of the t-th CPI cluster in {1, 2, ..., T}.

[0015] Further, step 101 includes the following sub-steps:

[0016] Step 111: For each data pair in the original dataset D <O (i) Q (i) >, i∈{1,2,…,I}, for the corresponding intercepted signal sequence O (i) It can be described in the following formal way:

[0017]

[0018]

[0019]

[0020]

[0021] Extract the changes of two adjacent CPI clusters in each dimension, and use them to construct the working state switching point representation feature vector of the t∈{2,…,T}th CPI cluster.

[0022] Step 112: For each data pair in the original dataset D <O (i) Q(i) >, i∈{1,2,…,I}, for the corresponding working state label sequence Q (i) By calculating the difference in working status labels between adjacent CPI clusters Determine the working state switching point label of the t-th CPI cluster in {2, ..., T}. When Δq≠0, This indicates a switch in the working state.

[0023] Furthermore, in step 111, the changes of two adjacent CPI clusters in various dimensions include: signal energy, time-domain waveform pattern parameters, time interval dimension, frequency change, dwell time in the same state, pulse width and repetition interval value corresponding to the state.

[0024] Furthermore, in step 111, the working state switching point represents the feature vector. It is configured to capture various changes when the phased array radar switches between search and track modes, including the energy brought about by beam space agility pointing, the requirements of waveform pattern parameters for corresponding detection accuracy and stability, and the discrete arrangement of mission events.

[0025] Further, step 102 includes the following sub-steps:

[0026] Step 121: Regarding the phenomenon of a severely imbalanced ratio of working state switching to non-switching samples in the training sample set S, especially the case where non-switching samples constitute the vast majority, i.e., N... y＝0 >>N y＝1 The number N of samples needs to be switched according to the working status. y＝1 Without switching the sample set, an equal number of samples are randomly drawn, such that N′ y＝0 =N y＝1 Construct a balanced and uniformly mixed working state switching point hybrid sample set S′, and divide it into a training sample set S′ according to a preset ratio. train and test sample set S′ test ;

[0027] Step 122: Utilize the random forest ensemble learning method as a working state switching classifier model. Where θ hp Here, θ represents the hyperparameters, and θ represents the model parameters; using the training sample set S′ train The sample pairs <x j y j >,j∈{1,2,…1.6N y＝1} is the input, with hyperparameter θ hp Under the constraints, a divide-and-conquer approach is used for each decision tree in the random forest to train and generate model parameters θ = (θ n|n = 1, 2, ..., N tree ) and the membership classification labels y∈{0,1} of the leaf nodes, where θ i For the nth ∈ {1, 2, ..., N} tree Model parameters for decision trees;

[0028] Step 123: Switch the classifier model based on the trained working state. With test sample set S′ test Artificial feature vector x in j , j∈{1, 2, ..., 0.4N} y＝1 Given} as input, predict the probability of switching working states. Compare the labels y of the working state switching points in the test sample set. j , j∈{1, 2, ..., 0.4N} y＝1}, to determine whether the evaluation indicators predicted by the model meet the requirements; if they do not meet the requirements, repeat steps 121 to 123 until the requirements are met.

[0029] Furthermore, in step 122:

[0030] hyperparameter θ hp Includes: N decision trees tree The maximum depth N of the tree max_depth The number of optional features N for splitting max_featute Internal node splitting criterion f split ∈{gini, entropy}, the minimum number of samples N required to split internal nodes. min_smp_split Minimum number of samples N required for leaf nodes min_smp_leaf ;

[0031] Model parameters θ i Includes: the feature dimension d∈{1,2,…,N} for internal node selection splitting. max_featute} and splitting threshold The data pairs constitute

[0032] Furthermore, step 2 includes the following sub-steps:

[0033] Step 201: Extract the characteristic features of the working state switching: For the signal sequence O that intercepts a single radar radiation source in an online pipeline manner, use the adjacent previous CPI cluster o t-1 And the current CPI cluster t The feature vector x representing the change in working state is extracted according to step 122. t ;

[0034] Step 202: Online prediction of work state switching points: Based on the extracted feature vector x representing the change in work state switching... t This is used as the machine learning-based work state switching classifier model trained in step 123. The input is the probability prediction value of whether the working state will switch at the current time t, provided by the classifier model. like Then the current time t is the working state switching point, where α is the decision threshold.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention provides a multi-functional radar operating state switching point detection method based on machine learning. By utilizing the random forest ensemble learning algorithm, it automatically generates the optimal decision threshold for each feature dimension through fitting and learning the feature data with switching point labels in the training set. Under the same constraints of small sample data volume and computing resources, it achieves better multi-functional radar operating state switching point detection capability than segmentation methods such as unsupervised clustering and supervised deep learning. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of a reconnaissance simulation scenario for an airborne multi-functional AESAFCR radar.

[0038] Figure 2 This is a schematic diagram of the intercepted signal sequence with labels indicating the search and tracking status.

[0039] Figure 3 This is an offline training flowchart for the switching points of the multi-functional radar's operating state;

[0040] Figure 4 This is a flowchart of the online prediction process for the switching points of a multi-functional radar operating state;

[0041] Figure 5 This is a schematic diagram of the online detection results of the working state switching point of the intercepted signal sequence;

[0042] Figure 6 This is a magnified view of the online detection results of the working state switching point of the intercepted signal sequence. Detailed Implementation

[0043] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0044] Example 1

[0045] This embodiment provides an online detection method for switching points of multi-functional radar operating states based on machine learning, including:

[0046] Step 1: Offline training of the switching point prediction model: Based on the radar intercept signal sequence dataset with working state labels, extract the representation feature vector with working state switching point labels, and then train the working state switching classifier model based on machine learning using this labeled sample dataset to obtain the model parameters.

[0047] Step 2, Online prediction of working state switching points: Based on the radar signal intercepted in the online pipeline mode, the feature vector representing the working state switching change at the current moment is extracted in real time. Combined with the working state switching classifier model, model parameters and preset hyperparameters, the prediction result of the working state switching point at that moment is given.

[0048] Preferably, the implementation method of the working state switching classifier model can be: Random Forest (RFs), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), Shallow Neural Network (SNN), K-Nearest Neighbor (KNN) and other classification methods.

[0049] Specifically, taking the random forest classification method as an example, this embodiment will further illustrate the online detection method for switching points of multi-functional radar operating states.

[0050] I. Offline Training of Switching Point Prediction Model

[0051] Preferably, offline training includes the following sub-steps:

[0052] Step 101: Extract the representational feature vector with labels for operating state switching points: for the intercepted signal sequence of multi-functional active phased array radar. and its working status label sequence The original dataset D = { <O (i) Q (i) >|i=1,2,…,I};For each intercepted signal sequence with working status label <O (i) Q (i) > Extract the labels of adjacent coherent processing intervals, i.e., the working state switching points between adjacent CPI clusters. and the characteristic vector representing the change in working state Get data pairs The training and validation sample set S constitutes the working state switching points;

[0053] Step 102: Training the working state switching classifier model: For the previously constructed training and validation sample set S, first equalize the proportion of each label category and split it into training sample sets S′. train and test sample set S′ test Then, the working state switching classifier model based on the random forest ensemble learning method is iteratively trained and tested on these two sample sets. Where θ hp Let θ be the hyperparameter and the model parameter, respectively, and let y∈{0,1} be the label of the working state switching point. Finally, the model parameter θ of the working state switching classifier model is obtained by training.

[0054] Preferably, in step 101, the O signal sequence intercepted by the multi-functional active phased array radar... (i) of Let represent the characteristic vector of the t-th CPI cluster in {1, 2, ..., T}, which is composed of the start arrival time ftoa, end arrival time ltoa, mean pulse amplitude pa, mean pulse repetition interval pti, and mean pulse width pw within the CPI. Working status label sequence Q (i) In This represents the working status label of the t-th CPI cluster in {1, 2, ..., T}.

[0055] Preferably, step 101 includes the following sub-steps:

[0056] Step 111: For each data pair in the original dataset D <O (i) Q (i) >, i∈{1,2,…,I}, for the corresponding intercepted signal sequence O (i) It can be described in the following formal way:

[0057]

[0058]

[0059]

[0060]

[0061] Extract the changes of two adjacent CPI clusters in various dimensions (e.g., signal energy, time-domain waveform pattern parameters, time interval, etc.), and use these to construct a feature vector representing the working state switching point of the t∈{2,…,T}th CPI cluster. in It is used to capture changes in energy caused by beam spatial agility pointing, the corresponding requirements for waveform pattern parameters for detection accuracy and stability, and the discrete arrangement of mission events when a phased array radar switches between search and track modes.

[0062] It should be noted that, apart from the aforementioned characteristics of signal energy, time-domain waveform pattern parameters, and time interval dimension, other characteristics such as frequency variation, dwell time in the same state, pulse width corresponding to the state, and repetition interval value are all within the protection range.

[0063] Step 112: For each data pair in the original dataset D <O (i) Q (i) >, i∈{1,2,…,I}, for the corresponding working state label sequence Q (i) By calculating the difference in working status labels between adjacent CPI clusters Determine the working state switching point label of the t-th CPI cluster in {2, ..., T}. When Δq≠0, This indicates a switch in the working state.

[0064] Preferably, in step 111, the changes of two adjacent CPI clusters in each dimension include: signal energy, time-domain waveform pattern parameters, time interval dimension, frequency change, dwell time of the same state, pulse width and repetition interval value corresponding to the state.

[0065] Preferably, in step 111, the working state switching point represents the feature vector. It is configured to capture various changes when the phased array radar switches between search and track modes, including the energy brought about by beam space agility pointing, the requirements of waveform pattern parameters for corresponding detection accuracy and stability, and the discrete arrangement of mission events.

[0066] Preferably, step 102 includes the following sub-steps:

[0067] Step 121: Regarding the phenomenon of a severely imbalanced ratio of working state switching to non-switching samples in the training sample set S, especially the case where non-switching samples constitute the vast majority, i.e., N... y＝0 >>N y＝1 The number N of samples needs to be switched according to the working status. y＝1 Without switching the sample set, an equal number of samples are randomly drawn, such that N′ y＝0 =N y＝1 Construct a balanced and uniformly mixed working state switching point hybrid sample set S′, and divide it into a training sample set S′ according to a preset ratio. train and test sample set S′ test ;

[0068] Step 122: Utilize the random forest ensemble learning method as a working state switching classifier model. Where θ hp Here, θ represents the hyperparameters, and θ represents the model parameters; using the training sample set S′ train The sample pairs <x j y j >,j∈{1,2,…1.6N y＝1} is the input, with hyperparameter θ hp Under the constraints, a divide-and-conquer approach is used for each decision tree in the random forest to train and generate model parameters θ = (θ n |n = 1, 2, ..., N tree ) and the membership classification labels y∈{0,1} of the leaf nodes, where θ i For the nth ∈ {1, 2, ..., N} tree Model parameters for decision trees;

[0069] Step 123: Switch the classifier model based on the trained working state. With test sample set S′ test Artificial feature vector x in j , j∈{1, 2, ..., 0.4N} y＝1 Given} as input, predict the probability of switching working states. Compare the labels y of the working state switching points in the test sample set. j , j∈{1, 2, ..., 0.4N} y＝1}, to determine whether the evaluation indicators predicted by the model meet the requirements; if they do not meet the requirements, repeat steps 121 to 123 until the requirements are met.

[0070] Preferably, in step 122:

[0071] hyperparameter θ hp Includes: N decision trees tree The maximum depth N of the treemax_depth The number of optional features N for splitting max_feature Internal node splitting criterion f split ∈{gini, entropy}, the minimum number of samples N required to split internal nodes. min_smp_split Minimum number of samples N required for leaf nodes min_smmp_leaf ;

[0072] Model parameters θ i Includes: the feature dimension d∈{1,2,…,N} for internal node selection splitting. max_feature} and splitting threshold The data pairs constitute

[0073] II. Online Prediction of Working State Switching Points

[0074] Preferably, online prediction includes the following sub-steps:

[0075] Step 201: Extract the characteristic features of the working state switching: For the signal sequence O that intercepts a single radar radiation source in an online pipeline manner, use the adjacent previous CPI cluster o t-1 And the current CPI cluster t The feature vector x representing the change in working state is extracted according to step 122. t ;

[0076] Step 202: Online prediction of work state switching points: Based on the extracted feature vector x representing the change in work state switching... t This is used as the machine learning-based work state switching classifier model trained in step 123. The input is the probability prediction value of whether the working state will switch at the current time t, provided by the classifier model. like Then the current time t is the working state switching point, where α is the decision threshold.

[0077] Example 2

[0078] This embodiment is based on embodiment 1:

[0079] This embodiment provides an online detection method for switching points of multi-functional radar operating states based on machine learning. It constructs reconnaissance scenarios for a simulated model of an airborne multi-functional active phased array fire control radar (AESAFCR) according to different head-on or tail-chase relative situations between enemy and friendly forces. Figure 1As shown in Tables 1 and 2, the airborne AESAFCR simulation model searches for targets in head-on or tail-chase situations using five sets of variable-stiffness high-repetition-frequency (HPRF) or medium-repetition-frequency (MPRF) waveform pattern terms, and uses one of these waveform pattern terms to achieve tracking under prior conditions. The airborne AESAFCR model can randomly select a frequency point from the frequency set {9470,9490,9530,9550,9590,9610,9630,9660,9680,9510,9540,9570,9640,9700,9720,9740,9760} (in MHz) and combine it with a waveform pattern term to complete target search and tracking detection.

[0080] Table 1 - HPRF waveforms of the variable type under head-on situation

[0081]

[0082] Table 2 - MPRF waveforms of variable-staggered components under tail-chase situation

[0083]

[0084]

[0085] For different simulation scenarios, the intercepted signal sequences of the radar model and the corresponding actual search and tracking state sequences provided by the radar model were collected and recorded. <O (i) Q (i) >,i∈{1,2,…,I}, such as Figure 2 As shown. A total of I=50 intercepted signal sequence data with working status tags were collected, and the duration of each sequence data was approximately 110 seconds, that is... The amount It involves two working states: search and tracking, that is... To construct the original dataset D.

[0086] according to Figure 3 The offline training process for the working state switching point shown in this embodiment can be roughly described as follows:

[0087] Step 1: Extract labeled features representing changes in operating state: Based on the previously collected raw dataset D = {...} of the airborne AESA FCR radar model... <o (i) Q (i) For each intercepted signal sequence with a working status label, |i = 1, 2, ..., 50}. <O (i) Q i On the one hand, extract the labels of the working state switching points between adjacent CPI clusters. When Δq≠0, On the other hand, to capture the switching of search and follow-up working status events, feature vectors are extracted from dimensions such as time of arrival (TOA), repetition interval (PRI), pulse width (PW), and pulse amplitude (PA). Get from The data pairs constitute the training and validation sample set S of the working state switching points.

[0088] Step 2: Train the work state switching point classifier: For the previously constructed training and validation sample set S, first refer to the number of work state switching samples N. y＝1 The proportions of each label category are balanced, and the dataset is divided into training sets S′ in an 8:2 ratio. train and the test sample set S′ test Subsequently, Random Forests (RFs) were used as the classification and detection method for work state switching points. And according to N tree =50, N max_depth =8, N max_feature ='auto', f split ='entropy', N min_smp_split =2, N min_smp_leaf =1 etc. set its hyperparameter θ hp Then at S′ train Train the model parameters θ on S′ and in S′ test Verification Evaluation metrics such as prediction accuracy, recall, and F1 score are used to repeat the training and validation process until the requirements are met.

[0089] according to Figure 4 The online prediction and processing flow of the working state switching point shown in this embodiment can be roughly described as follows:

[0090] Step 1: Extracting the characteristics of the working state switching: For the signal sequence O radiated by the airborne multi-function AESAFCR radar model intercepted in an online pipeline manner, the adjacent previous CPI cluster o is used. t-1 And the current CPI cluster t Following the manual feature extraction method mentioned in the aforementioned offline training process, the switching point representation feature vector x is still extracted from dimensions such as TOA, PRI, PW, and PA. t =(Δpa) t ,Δpri t ,Δpw t ,Δtoa t ) T .

[0091] Step 2: Online prediction of working state switching points: Based on the extracted feature vector x of the working state switching points... tThis is used as the working state switching point classifier trained in the aforementioned offline training process based on random forest. The input is the predicted probability value of the working state switching at the current time t, which is given by the classifier model. like Therefore, the current time t is considered to be the working state switching point.

[0092] Using the method of this invention, the feature vector representing the switching point of the airborne AESAFCR radar model at the current time t is extracted sequentially, and the switching point is predicted or determined, thus obtaining the following... Figure 5 The switching point detection situation of the entire intercepted signal sequence O is shown, where Figure 6 for Figure 5 A magnified view of a portion of the image. (From...) Figure 6 As can be seen, the method of the present invention can effectively capture the switching points between the search and tracking working states, especially the tracking state segments embedded in the search state waveform sequence.

[0093] Example 3

[0094] This embodiment is based on embodiment 1:

[0095] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the online detection method for switching points of a multi-functional radar based on machine learning in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0096] Example 4

[0097] This embodiment is based on embodiment 1:

[0098] This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the steps of the online detection method for switching points of a multi-functional radar operating state based on machine learning as described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content contained in the storage medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.

[0099] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for online detection of switching points in the operating state of a multi-functional radar based on machine learning, characterized in that, include: Step 1: Offline training of the switching point prediction model: Based on the radar intercept signal sequence dataset with working state labels, extract the representation feature vector with working state switching point labels, and then train the working state switching classifier model based on machine learning using this labeled sample dataset to obtain the model parameters. Step 2, Online prediction of working state switching point: Based on the radar signal intercepted in the online pipeline mode, extract the feature vector representing the working state switching change at the current moment in real time, and combine the working state switching classifier model, model parameters and preset hyperparameters to give the prediction result of the working state switching point at that moment. Step 1 includes the following sub-steps: Step 101: Extract the representational feature vector with labels for operating state switching points: for the intercepted signal sequence of multi-functional active phased array radar. and its working status label sequence The original dataset For each intercepted signal sequence with a working status tag Extract the labels of the working state switching points between adjacent coherent processing intervals, i.e., between adjacent CPI clusters. and the characteristic vector representing the change in working state , to obtain data pairs The training and validation sample set for the working state switching points ; Step 102: Train the working state switching classifier model: using the previously constructed training and validation sample set. First, equalize the proportion of each label category and then divide it into training sample sets. and test sample set Then, iteratively train and test the working state switching classifier model based on the random forest ensemble learning method on these two sample sets. ,in and These are hyperparameters and model parameters, respectively. Labeling the work state switching points, the model parameters of the work state switching classifier model are finally obtained through training. ; In step 101, the multi-functional active phased array radar intercepts the signal sequence. of Indicates the first The characteristic vector representing each CPI cluster is determined by the start and arrival times within the CPI. End arrival time Average pulse amplitude Mean pulse repetition interval and average pulse width Composition, that is Working status label sequence In Indicates the first The working status label of each CPI cluster; Step 101 includes the following sub-steps: Step 111: For the original dataset Each data pair For the corresponding intercepted signal sequence It can be described in the following formal way: Extract the changes in each dimension of two adjacent CPI clusters, and construct the first... The working state switching points of each CPI cluster represent feature vectors. ; Step 112: For the original dataset Each data pair For the corresponding working status label sequence By calculating the difference in working status labels between adjacent CPI clusters , thus arriving at the first Labels indicating the working status switching points of individual CPI clusters ,when hour, Indicates a switch in working status; Step 102 includes the following sub-steps: Step 121: For the training sample set There is a severe imbalance in the proportion of samples that switch working states and those that do not, with samples that do not switch states accounting for the vast majority. The number of samples needs to be switched according to the working status. Without switching the sample set, an equal number of samples are randomly drawn, such that Construct a balanced and uniformly mixed sample set of working state switching points. And divide it into training sample sets according to a preset ratio. and test sample set ; Step 122: Utilize the random forest ensemble learning method as a working state switching classifier model. , of which For hyperparameters, For model parameters; using the training sample set The sample pairs As input, in hyperparameters Under the constraints, a divide-and-conquer approach is used for each decision tree in the random forest to train and generate model parameters. and the category labels of the leaf nodes ,in For the first Model parameters for decision trees; Step 123: Switch the classifier model based on the trained working state. to test sample set Artificial feature vectors in Given the input, predict the probability of switching working states. Compare the labels of working state switching points in the test sample set. To determine whether the evaluation indicators predicted by the model meet the requirements; if they do not meet the requirements, repeat steps 121 to 123 until the requirements are met.

2. The online detection method for switching points of multi-functional radar operating states based on machine learning according to claim 1, characterized in that, The implementation methods of the working state switching classifier model include: random forest classification method, support vector machine classification method, shallow neural network classification method and K nearest neighbor classification method.

3. The online detection method for switching points of multi-functional radar operating states based on machine learning according to claim 1, characterized in that, In step 111, the changes of two adjacent CPI clusters in various dimensions include: signal energy, time-domain waveform pattern parameters, time interval dimension, frequency change, dwell time of the same state, pulse width and repetition interval value corresponding to the state.

4. The online detection method for switching points of multi-functional radar operating states based on machine learning according to claim 1, characterized in that, In step 111, the working state switching point represents the feature vector. It is configured to capture various changes when the phased array radar switches between search and track modes, including the energy brought about by beam space agility pointing, the requirements of waveform pattern parameters for corresponding detection accuracy and stability, and the discrete arrangement of mission events.

5. The online detection method for switching points of multi-functional radar operating states based on machine learning according to claim 1, characterized in that, In step 122: Hyperparameters Includes: the number of decision trees Maximum depth of the tree Optional feature number of splits Internal node splitting criteria Minimum number of samples required to split internal nodes Minimum number of samples required for leaf nodes ; Model parameters Includes: Feature dimensions for internal node selection and splitting and splitting threshold The data pairs constitute .

6. The online detection method for switching points of multi-functional radar operating states based on machine learning according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 201: Extracting the characteristics of working state switching: For signal sequences intercepted by a single radar radiation source in an online pipeline manner. Utilizing the adjacent previous CPI cluster and current CPI clusters The feature vector representing the change in working state is extracted according to step 122. ; Step 202: Online prediction of work state switching points: Based on the extracted feature vector representing the changes in work state switching... This is used as the machine learning-based work state switching classifier model trained in step 123. The input is given by the classifier model at the current time step. Predicted probability of switching working status ,like Then at the current moment This is the point where the working state is switched, where This is the threshold for the judgment.

Images