Method and system for dynamically training and updating HMM model library based on unknown PRI information

Abstract

The invention discloses an unknown PRI information-based HMM model library dynamic training and updating method and system, and relates to the technical field of phased array radars, and the method comprises the steps: obtaining unknown PRI information for a plurality of wave position pulse sequences of radar signal echoes; performing binning processing on unknown PRI information, and determining an updated PRI typical value library by adopting a histogram and an EM algorithm; mapping a plurality of unknown paths based on the updated PRI typical value library, and determining each typical value mapping path; matching each typical value mapping path with each hidden Markov model in an HMM (Hidden Markov Model) library, and determining similar paths and dissimilar paths; and based on the similar paths and the dissimilar paths, carrying out dynamic training and updating on an HMM model library, and determining an updated HMM model library. According to the invention, the recognition processing efficiency and accuracy of radar signals of a complex system are improved.

Description

Technical Field

[0001] The present application relates to the technical field of phased array radar, and in particular to a method and system for dynamically training and updating an HMM model library based on unknown PRI information. Background Art

[0002] Radar technology plays a vital role in modern military, aerospace, weather forecasting, and other fields. Accurate radar signal processing is key to target detection, identification, tracking, and classification. Phased array radar, with its advantages such as beam agility, versatility, and high reliability, has become the mainstream architecture for modern radar systems. The pulse repetition interval (PRI) is a key characteristic parameter of radar signals, crucial for signal sorting, identification, and interference suppression. In particular, when extracting waveform features from phased array radar signals, the PRI variation pattern of the pulse train is one of the key parameters that distinguish waveforms. Statistical methods for analyzing PRI values ​​in radar signals typically rely on pre-established libraries of typical PRI values. However, in practical applications, a large number of PRI values ​​may not be covered by existing libraries for the following reasons: First, the diversity of radar systems: new radars are constantly emerging, and their PRI values ​​may exceed the range of existing libraries; second, the variability of radar operating modes: the same radar may use different PRI variation patterns (including the order of PRI values ​​and their relationships) in different operating modes; and third, the variability of radar targets: factors such as target type and motion state may cause dynamic changes in PRI parameters. These PRI parameter patterns, not covered by the typical value library, may contain important radar signal characteristics and patterns. Failure to effectively process this unknown PRI information will lead to a decrease in radar signal sorting and recognition efficiency, as well as problems such as an incomplete Hidden Markov Model (HMM) library and difficulty in dealing with interference from new radar signals. The Hidden Markov Model library is essentially the HMM model library. Furthermore, traditional PRI typical value libraries are mostly static and cannot be automatically updated, making them difficult to adapt to the dynamic changes in PRI values ​​in radar signals. Furthermore, most methods focus only on the statistical characteristics of PRI values ​​while ignoring their temporal variations, making it difficult to handle the complex and diverse PRI patterns. Consequently, accurate processing of complex radar signals and continuous improvement of the HMM model are impossible. Summary of the Invention

[0003] The purpose of this application is to provide a method and system for dynamic training and updating of an HMM model library based on unknown PRI information, which can solve the problem of low efficiency and accuracy in identifying and processing radar signals due to the inability to accurately process complex radar signals and continuously improve the HMM model.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides a method for dynamically training and updating an HMM model library based on unknown PRI information, comprising: for multiple wave position pulse sequences of radar signal echoes, obtaining the PRI values ​​of wave position pulse sequences that cannot be mapped in the PRI typical value library as unknown PRI information; the PRI typical value library is a plurality of PRI typical values ​​determined by historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each unknown path includes a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values; the unknown PRI information is binned, and an updated PRI typical value library is determined using a histogram and an EM algorithm; the plurality of unknown paths are mapped based on the updated PRI typical value library to determine each typical value mapping path; each typical value mapping path is matched with each hidden Markov model in the HMM model library to determine a similar path and a dissimilar path; the HMM model library includes a plurality of hidden Markov models determined based on the PRI information of the wave position pulse sequences that have been mapped in the PRI typical value library; based on the similar paths and the dissimilar paths, the HMM model library is dynamically trained and updated to determine an updated HMM model library to identify the working mode of complex radar signals. .

[0006] In the second aspect, the present application provides a dynamic training and updating system for an HMM model library based on unknown PRI information, including: an acquisition module for acquiring the PRI value of a wave position pulse sequence that cannot be mapped in the PRI typical value library as unknown PRI information for a plurality of wave position pulse sequences of radar signal echoes; the PRI typical value library is a plurality of PRI typical values ​​determined by historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each unknown path includes a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values; an update processing module is used to perform binning processing on the unknown PRI information, and adopt a histogram and an EM algorithm to determine An updated PRI typical value library; a determination module for mapping multiple unknown paths based on the updated PRI typical value library to determine each typical value mapping path; a matching module for matching each typical value mapping path with each hidden Markov model in the HMM model library to determine similar paths and dissimilar paths; the HMM model library includes multiple hidden Markov models determined based on the PRI information of the wave position pulse sequence mapped in the PRI typical value library; an updating module for dynamically training and updating the HMM model library based on the similar paths and the dissimilar paths to determine the updated HMM model library to identify the working mode of the complex radar signal.

[0007] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0008] This application first obtains the PRI values ​​of wave-position pulse sequences that cannot be mapped in the PRI typical value library as unknown PRI information by targeting multiple wave-position pulse sequences of radar signal echoes. Then, the unknown PRI information is binned and an updated PRI typical value library is determined using a histogram and the EM algorithm. That is, potential PRI typical values ​​are obtained from the unknown PRI information to update the typical value library. This increases the variety of PRI typical values ​​in the typical value library and achieves diversity in the PRI typical values ​​in the typical value library. In this way, by processing the PRI values ​​of wave-position pulse sequences that cannot be mapped, a wider variety of PRI typical values ​​are determined compared to processing only the PRI values ​​of wave-position pulse sequences that can be mapped. This also achieves more comprehensive and accurate processing of complex radar signals, laying the foundation for subsequent mapping of multiple unknown paths based on the updated typical value library to obtain more typical value mapping paths. Further, each typical value mapping path is matched with each hidden Markov model in the HMM model library respectively, similar path and dissimilar path are determined, then based on similar path and described dissimilar path, HMM model library is dynamically trained and updated, and the updated HMM model library is determined.Here, each original hidden Markov model library in the HMM model library is used to sort each typical value mapping path, and the existing HMM model library is trained as the similar path of each hidden Markov model to realize the updating of the HMM model library; If the dissimilar path of the hidden Markov model of the matching is not found, then it is used as the training set of the hidden Markov model corresponding to the unknown path (unknown PRI staggered pattern), the HMM model library is updated.So realize the continuous improvement of the HMM model.The updated HMM model library of the present application is conducive to the PRI value and different PRI staggered patterns in the novel radar signal that constantly emerge and is effectively and accurately identified, and improves the recognition processing efficiency and accuracy of the radar signal of complex constitution. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. 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 creative work.

[0010] Figure 1 Schematic diagram of a flow chart of a method for dynamic training and updating of an HMM model library based on unknown PRI information provided in an embodiment of the present application;

[0011] Figure 2 A flowchart for dynamically updating and creating an HMM model library based on unknown PRI information provided in an embodiment of the present application;

[0012] Figure 3 This is the PRI outlier detection effect diagram; among them, Figure 3 (a) is the distribution diagram of the original data of unknown PRI values; Figure 3 (b) is a schematic diagram of PRI outlier detection and filtering;

[0013] Figure 4 Schematic diagram of data distribution of unknown PRI information provided for this application; wherein, Figure 4 (a) is a diagram showing the data distribution of unknown PRI information before outliers are removed; Figure 4 (b) is a diagram showing the data distribution of unknown PRI information after outliers are removed;

[0014] Figure 5 Updated the flow chart for the PRI typical value library;

[0015] Figure 6 is the original data distribution diagram and the candidate PRI distribution fitting diagram; among them, Figure 6 (a) is the original data distribution diagram; Figure 6 (b) in the figure is the distribution fitting diagram of the candidate PRI;

[0016] Figure 7 Schematic diagram of the visual graphical interface for manual selection of candidate PRIs;

[0017] Figure 8 This is a schematic block diagram of the dynamic training and updating of an HMM model library based on unknown PRI information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0020] like Figure 1 As shown, the present application provides a method for dynamic training and updating of an HMM model library based on unknown PRI information, including steps 101 to 105.

[0021] Step 101: For multiple wave position pulse sequences of radar signal echoes, PRI values ​​of wave position pulse sequences that cannot be mapped in a PRI typical value library are obtained as unknown PRI information; the PRI typical value library is a plurality of PRI typical values ​​determined based on historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each of which includes a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values;

[0022] Step 102: bin the unknown PRI information and use the histogram and EM algorithm to determine an updated PRI typical value library.

[0023] In some embodiments, step 102 specifically includes steps 201 and 202.

[0024] Step 201: Based on the quantile method, the unknown PRI information is binned to determine the final binning result.

[0025] Step 202: Based on the histogram generated by the final binning result, the EM algorithm is used to fit each unknown PRI value in each bin in the histogram to determine the fitting result.

[0026] Step 203: Determine an updated PRI typical value library based on the fitting result and a plurality of preset candidate PRI values.

[0027] In some embodiments, step 201 specifically includes: using the quantile method to eliminate outliers in the unknown PRI information to determine the main data; the main data includes multiple unknown PRI values; binning the main data to determine the binning results; based on the binning results, determining the variance and amount of the main data in each bin, and setting corresponding preset constraints for the variance and amount of the main data in each bin; determining the final binning results based on the preset constraints and the binning results.

[0028] In some embodiments, step 202 specifically includes: traversing the peak value corresponding to each bin of the histogram, and marking the peak value corresponding to the bin that meets the preset conditions as a candidate value; the preset conditions are that the peak value of the current bin is greater than the peak values ​​of the left and right adjacent bins and the peak value of the current bin reaches the preset peak value; using the EM algorithm, Gaussian distribution fitting is performed on each unknown PRI value in the bin corresponding to each candidate value to determine the fitting result; the fitting result is the mean and variance of each unknown PRI value in each bin.

[0029] In some embodiments, step 203 specifically includes step 301 -step 301.

[0030] Step 301: setting a plurality of preset candidate PRI values ​​according to each bin range; each bin range is a range to which all unknown PRI values ​​in the bin corresponding to each preset candidate value belong.

[0031] Step 302: Determine the mean and variance of the unknown PRI values ​​in the bins corresponding to each preset candidate PRI value, and use the mean and variance of the corresponding unknown PRI values ​​as the mean and variance of each preset candidate PRI value.

[0032] Step 303: Determine an updated PRI typical value library based on the mean and variance of each preset candidate PRI value and the mean and variance of the PRI typical values ​​in the PRI typical value library.

[0033] In some embodiments, step 303 specifically includes: determining whether the mean and variance of each preset candidate PRI value are consistent with the mean and variance of the PRI typical value in the PRI typical value library, and obtaining a first judgment result; if the first judgment result is yes, determining that the preset candidate PRI value is repeated with the PRI typical value in the PRI typical value library, and updating the mean and variance of the preset candidate PRI value based on the mean update trigger condition, and determining the updated PRI typical value library; the mean update trigger condition is determined based on the mean, variance, data volume of the PRI value and data volume of the PRI typical value of the preset candidate PRI value and the PRI typical value in the PRI typical value library; if the first judgment result is no, determining that the preset candidate PRI value is not repeated with the PRI typical value in the PRI typical value library, adding the preset candidate PRI value to the PRI typical value library, and obtaining an updated PRI typical value library.

[0034] Step 103: Map multiple unknown paths based on the updated PRI typical value library to determine each typical value mapping path.

[0035] Step 104: Match each typical value mapping path with each hidden Markov model in the HMM model library to determine similar paths and dissimilar paths; the HMM model library includes multiple hidden Markov models determined based on the PRI information of the wave position pulse sequence mapped in the PRI typical value library.

[0036] In some embodiments, step 104 specifically includes: grouping and matching the typical value mapping paths based on a preset path similarity verification method and each hidden Markov model in the HMM model library, and determining similar paths that match the HMM model library and dissimilar paths that do not match the HMM model library.

[0037] Step 105: Based on the similar paths and the dissimilar paths, dynamically train and update the HMM model library to determine an updated HMM model library.

[0038] In some embodiments, step 105 specifically includes steps 401 to 403 .

[0039] Step 401: Using the similar path as a training set, retrain the hidden Markov model corresponding to the similar path in the HMM model library to determine a trained hidden Markov model; the trained hidden Markov model represents a PRI variation pattern; the PRI variation pattern includes multiple unknown PRI values ​​in the similar path and a sequential relationship between the multiple unknown PRI values.

[0040] Step 402: clustering the dissimilar paths to determine a plurality of new basic paths, and based on the plurality of new basic paths, determining a plurality of new hidden Markov models.

[0041] Step 403: Determine an updated HMM model library based on the multiple new hidden Markov models and the trained hidden Markov models.

[0042] In some embodiments, step 402 specifically includes: for dissimilar paths, mapping the dissimilar paths based on preset typical values ​​to determine multiple clusters; for the dissimilar paths in each cluster, selecting the path with the highest frequency of occurrence and the largest number of PRI value types as the new basic path of each cluster; using the new basic path of each cluster as the new PRI difference pattern corresponding to each cluster; using the new PRI difference pattern corresponding to each cluster as the PRI difference pattern corresponding to multiple new hidden Markov models.

[0043] Among them, before selecting the path with the highest occurrence frequency and the most PRI value types as the basic path for the dissimilar paths in each cluster, it also includes: taking multiple dissimilar paths corresponding to any cluster as a new training set, training the initial hidden Markov model, and determining the new hidden Markov model corresponding to the cluster; and taking the new hidden Markov model corresponding to each cluster as each new hidden Markov model.

[0044] Specifically, the training of similar paths is to update or expand the multiple paths included in the PRI variation pattern corresponding to the original hidden Markov model in the HMM model library, because the PRI variation pattern corresponding to the hidden Markov model is only a representative path, that is, the basic path, but there is no similar path corresponding to the unknown PRI value and the sequential relationship between the unknown PRI values ​​in the HMM model library, but the similar path also belongs to the basic path (or representative path). Therefore, here, training is carried out through similar paths to achieve the expansion and update of different paths included in the PRI variation pattern, so that the HMM model library can also recognize the PRI variation pattern that is the same as the similar path.

[0045] In practical applications, even after using a PRI representative value library to map typical values ​​for pulse sequence PRI parameters at each waveband, some waveband pulse sequences still contain PRI parameters that are not in the library and cannot be mapped. Consequently, subsequent waveform extraction of pulse sequences containing unknown PRI information is impossible. These unknown PRI pulse sequences may contain previously unrecognized hidden patterns that can enrich the PRI variation patterns in the HMM model library, facilitating subsequent phased array radar signal analysis. Therefore, it is necessary to discover the connection patterns, or sequential relationships, between the unknown PRI information. This can be modeled using an HMM model. Using the PRI representative value library, the detected PRI parameters can be quickly matched with known patterns, thereby identifying the corresponding waveform or type of the radar signal.

[0046] Reference Figure 2 The algorithm includes quantile outlier detection, dynamic histogram binning and peak detection, updating the PRI typical value library, and path similarity comparison and HMM model library update. The quantile outlier detection method extracts the main data based on quantile calculation, sets a dynamic threshold, and combines density secondary validation to filter out abnormally high PRI values. Dynamic histogram binning and peak detection use variance and data volume constraints to control the binning process and perform Gaussian fitting on high-frequency histogram bins to identify potential PRI typical values. The PRI typical value library update involves manually selecting candidate PRI values ​​for inclusion in the library after fitting the PRI histogram, and dynamically updating the mean and variance of the PRI typical value library using a visual graphical interface. Path similarity comparison and HMM model library update use multiple rules to detect the similarity between PRI mapping paths and existing PRI basic paths. Paths that meet the criteria are incorporated into the existing HMM model library training set. When a certain number of paths are accumulated, the model is retrained. If no matching basic path is found, it is used as the HMM model training set for unknown patterns. The specific process is as follows.

[0047] (1) Quantile method to detect outliers.

[0048] like Figure 3 and Figure 4 As shown in the figure, in the unknown PRI information, there may be abnormally high values ​​in the calculated PRI parameters due to human factors such as radar working mode switching and radar detection target changes. These abnormal values ​​have a small probability of repetition and interfere with the extraction of PRI jagged time series, so they need to be removed. Here, the quantile method is used to detect abnormal points. The steps are as follows, where: Figure 3 The raw data in refers to the unknown PRI information.

[0049] Main data truncation: truncate the first 95% of the main data X main , excluding the interference of extremely high values: X main={x∈X|x≤P 95 (X)}.

[0050] Among them, P 95 (X) represents the 95% quantile of the data set X, where the data set is the unknown PRI value and x is an unknown PRI value.

[0051] Quantile calculation: intercept X main Calculate the quartile (Q1): 25% quantile, the third quartile (Q3): 75% quantile, and the interquartile range (IQR): IQR = Q3 - Q1.

[0052] Dynamic threshold T h Setting: T h =Q3+k·IQR, where k is the expansion factor (k=3 by default).

[0053] Density verification: for X>T h Data is binned and counted (bin width = IQR / 2), and areas with single-bin counts < 3 are marked as outlier areas.

[0054] (2) Variance constrained histogram dynamic binning.

[0055] After removing outliers, the variance-data volume dual constraint (i.e., preset constraint) is used to control the binning process of the data set to ensure the consistency of data distribution in each bin. The steps are as follows.

[0056] Initialize the bin width: base width Δ = range(X) / 20, and generate bin boundaries: B0 = {min(X) + kΔ | k∈}. B0 is the set of initial bin boundaries; it is a set of positive integers = 1, 2, 3, ... N. N is the initial number of bins, which can be set arbitrarily and affects the calculation of base width Δ = range(X) / 20. Here, 20 is N.

[0057] Iterative optimization: Iteratively detect the variance of each bin and divide the bins that exceed the standard into two equal parts.

[0058] Variance test: calculate the variance of each box Adopting the variance-data volume dual constraint:

[0059]

[0060] Among them, SplitCondition(b) is the split flag, 1 represents the bin that needs to be split; 0 represents the bin that does not need to be split. max is the maximum variance of the data in the rated bin; ∧ is the quantity threshold, N bBins that exceed the threshold need to be split. Bins that exceed the constraint conditions are divided into two equal parts to obtain updated bins: Among them, Δ i+1 is the updated bin width and Δ i is the width of the current bin; i is the sequence number.

[0061] Set the iteration termination condition: Set the maximum number of iterations, generally set to 5, and stop the iteration when the number of iterations reaches the maximum or all bins meet the constraint conditions.

[0062] (3) Peak detection and multi-peak Gaussian fitting.

[0063] The binning results obtained above are subjected to peak detection, candidate peaks are selected, and Gaussian distribution parameters are fitted by maximum likelihood estimation in the detected histogram peak area. Among them, the binning results are data distribution histograms of various ranges, but not all data within the binning range need to be added to the PRI library. If the number is too small, the fitting effect will be poor. Therefore, only bins with a sufficient number of data need to be fitted with Gaussian distribution. The local maximum of the histogram is the binning histogram that meets the number requirements of the histogram peak in the binning result. Gaussian distribution fitting is performed on the above-mentioned histogram binning range that meets the requirements.

[0064] Detect local maxima in the histogram as candidate peaks.

[0065] The EM algorithm is executed on each peak area data to perform Gaussian distribution fitting, where the candidate peaks are those histogram ranges whose number of peaks in the histogram range reaches the required number, and the peak area here is these qualified histogram ranges.

[0066] Gaussian distribution probability density function:

[0067] Where x is an unknown PRI value in the data set; μ is the mean of the Gaussian distribution that the data in the region obeys; σ is the variance; and e is a natural constant.

[0068] Gaussian distribution parameter estimation:

[0069] in, and is the estimated result of the mean and variance parameters of the Gaussian distribution fitted to all unknown PRI values ​​in the candidate peak area, i.e., the fitting result; n is the number of x in the area; x i is the i-th unknown PRI value in the area.

[0070] Filter fitting results with excessive variance. To save computational costs and avoid excessive loops, a maximum number of iterations is set. If the number of iterations is exceeded and the variance-data size constraints are still not met, the results will not be split further and will be output. Therefore, it is necessary to filter out the results of the bin fitting that do not meet the constraints but have reached the maximum number of iterations and cannot be split again.

[0071] (4) PRI typical value library updated.

[0072] like Figure 5 As shown, based on each manually selected candidate PRI value, it is checked whether there are typical values ​​overlapping with the existing PRI typical value library, and the mean-variance update mechanism is performed on these overlapping typical values.

[0073] Given a significance level α (usually 0.05): if |t|>t α / 2,df Then reject the null hypothesis H0 and the mean needs to be updated.

[0074] If |t|≤t α / 2,df , then accept the null hypothesis H0 and do not need to update the mean. Where t is the absolute value of the hypothesis statistic calculated based on the sample data under the hypothesis test; |.| is the absolute value; t α / 2,df Find the critical t value for the set significance level α and the degrees of freedom df of the data.

[0075] Among them, α = 0.05 represents the probability of decision error, |t| > t α / 2,df It means that the probability of making an error in the decision to reject the null hypothesis does not exceed 0.05, that is, there is a 95% confidence that the H0 hypothesis is not true.

[0076] Mean update triggering mechanism: If the new mean is equal to the old mean, no mean update is performed; if the new mean is different from the old mean, a mean update is required. The candidate PRI value is first tested for the amount of data corresponding to it. If it exceeds a certain amount, a t-test is performed. If it is less than a certain amount, a Welch's t-test is required due to heterogeneity of variance. If the null hypothesis is rejected, meaning a mean update is required, Bayesian inference is used to update the mean, and dynamic hybrid updating is used to update the variance. If the null hypothesis is accepted, meaning no mean update is required, dynamic hybrid updating is used to update the variance alone, and the updated PRI representative value mean and variance are stored back in the representative value database.

[0077] For those candidate PRI values ​​that do not overlap with those in the database, their mean and variance are directly stored in the database to update and expand the PRI typical value database. The specific steps are as follows.

[0078] ① Overlap detection: Determine whether the candidate PRI value overlaps with the typical value in the library.

[0079] The essence of overlap detection is to measure the similarity of two statistical distributions and determine whether they represent the same type of PRI features. For PRI values ​​that obey the normal distribution, it can be determined by comparing the standardized distance, the standardized distance D std The calculation is as follows:

[0080]

[0081] Decision rules:

[0082] If D std <δ std , determined to be overlapping, D std ≥δ std , it is determined to be non-overlapping.

[0083] Threshold setting: δ std The typical value is 1.5-2.5, corresponding to a confidence level of approximately 87%-99%. It can be adjusted according to application requirements. A smaller threshold means a stricter overlap judgment standard.

[0084] ② Mean update trigger judgment: For the detected overlapping typical values, a hypothesis test is used to determine whether the mean needs to be updated. The original hypothesis H0: μ new =μ old (That is, the new mean is equal to the old mean and does not need to be updated), alternative hypothesis H1: μ new ≠μ old (That is, the new mean is different from the old mean and needs to be updated.) new is the mean of the Gaussian distribution in the fitting result, that is, the new mean; μ old is the mean of the Gaussian distribution in the PRI typical value library, that is, the old mean.

[0085] First determine the size of the sample: set a threshold n threshold (Default is 30), if n new ≥n threshold , then the standard t test is used; if n new <n threshold , then the variance homogeneity test is performed, and the Welch's t test (for unequal variances) or the standard t test (for equal variances) is used according to the situation. The sample is the data within the histogram range corresponding to each candidate PRI, where n new is the number of samples in the bin area.

[0086] Homogeneity of variance test: In the case of small samples, it is necessary to first test whether the variances of the two groups of data are equal, using the F test.

[0087] in, and are larger and smaller sample variances, respectively.

[0088] like Reject the hypothesis of equal variances if Then accept the assumption of equal variance. α is generally taken as 0.05 or 0.1. larger and n smaller are the data volume of a certain PRI value Gaussian distribution in the PRI typical value library and the larger and smaller values ​​of the sample data volume in the candidate bin area respectively; To show the error probability of α according to n larger and n smaller The calculated critical F value. Based on the set significance level α and the size of the two samples (n larger , n smaller ) The critical threshold value found from the F distribution is used to determine whether the calculated sample variance ratio has reached a statistically significant difference. That is, F is used to determine whether the two variances are consistent. This hypothesis test statistic is used.

[0089] Among them, F is a test method, which is a hypothesis test statistic constructed to determine whether the difference between two variances is large.

[0090] The variances of the two sets of data are the PRI variance in the library and the sample variance. In the case of a small sample, the Gaussian distribution parameters fitted are likely to be inaccurate, so the two variances are very different. This is called heterogeneity of variance. How to determine whether the difference between the two variances is large is to use the F test. and are the larger and smaller sample variances, respectively, When α = 0.05, it means that there is a 95% confidence that the variance gap is large, and vice versa.

[0091] Standard t-test (equal variances) implementation: When the variances of the two groups of data are approximately equal, the standard t-test is used.

[0092] Among them, n new is the number of PRI values ​​in the updated typical library; n old is the number of PRI values ​​in the typical library before updating; p is the pooled sample standard deviation; Degrees of freedom: df = n new +n old -2.

[0093] Welch's t test (unequal variance) implementation: When the variances of two groups of data are significantly different, Welch's t test is used.

[0094] Degrees of freedom (Welch-Satterthwaite approximation):

[0095] Decision rule: Given a significance level α (usually 0.05).

[0096] If |t|>t α / 2,df Then reject the null hypothesis H0 and the mean needs to be updated.

[0097] If |t|≤t α / 2,df , then accept the null hypothesis H0 and there is no need to update the mean.

[0098] ③Parameter update method.

[0099] Mean update: When the test results indicate that the mean needs to be updated, the Bayesian inference method is used to update the mean.

[0100] Bayesian reasoning uses existing information as priors and new data as likelihood, and calculates the posterior distribution through Bayes' theorem:

[0101] Among them, θ is the set of mean and variance of a Gaussian distribution of a PRI value in the existing PRI typical value library, D is the sample data in the candidate bin area, P(θ) is the prior distribution, P(D|θ) is the likelihood function, and P(θ|D) is the posterior distribution.

[0102] For the normal distribution, Bayesian update has an analytical solution. Assume that both the prior distribution and the likelihood function follow a normal distribution:

[0103] Posterior mean (precision-weighted average):

[0104] Posterior variance:

[0105]

[0106] Among them, μ′ is the posterior mean; σ′ 2 The mean and variance of the posterior distribution of the data in the candidate bin area are the result of updating the parameters of the Gaussian distribution obeyed by the PRI in the library. The updated mean and variance are μ′ and σ′ 2 .

[0107] Sample size update in PRI typical value database: n′=n old +n new .

[0108] Variance update: When the test results show that there is no need to update the mean but only the variance, or when the variance needs to be updated after the mean is updated, the dynamic hybrid method is used to update the variance.

[0109] Dynamic hybrid update is a weighted averaging method that takes into account the impact of mean differences on the total variance. Even if the mean is not updated, the mean difference should be considered in the variance calculation.

[0110] The mean remains unchanged: μ′ = μ old .

[0111] Corrected variance:

[0112] Update sample size: n′=n old +n new .

[0113] Among them, μ′, σ ′2 , n′ are the mean, variance and sample size in the updated PRI typical library respectively.

[0114] ④ For overlapping typical values, the updated mean, variance, and sample size are used to replace the corresponding mean, variance, and sample size in the library. For candidate PRI values ​​that do not overlap with those in the library, a candidate distribution fit is given, and then a visual graphical interface is used to assist in manually selecting the PRI value of interest and adding it to the typical value library: a new typical value ID is generated; a new record is created: μ new , n new ; Add to typical value library.

[0115] like Figure 6 and Figure 7 As shown in the figure, a visual graphical interface is introduced to assist manual selection of PRI typical values ​​and model update, thereby improving processing efficiency and accuracy.

[0116] (5) Use the updated PRI typical value library to map unknown PRI path nodes.

[0117] Specifically, for a phased array radar pulse group with the same waveform, its PRI variation patterns are complex and diverse, such as fixed, varying, etc. Since the PRI variation pattern is a key feature for waveform extraction, it is necessary to obtain and record the PRI variation value and its timing regularity, namely, the PRI variation pattern.

[0118] For the PRI staggered pattern, the next PRI value (i.e., the PRI staggered value) depends on the previous PRI value. If each PRI value is considered as a state, then the next state of the hidden Markov model depends only on the previous state. Therefore, it can be represented by a Gaussian hidden Markov model, and its parameter space is Θ = (A, B, π), where A = [a ij ] M×M is the transfer matrix of M states, π=(π1,π2,...,π M ) is the initial state probability distribution, represents the set of observation probability distributions, i.e. PRI sequence P k ={p1,p2,...,p k The probability distribution ξ(P k ) is as follows:

[0119] Among them, a ij is the state transition probability, which indicates the probability that the hidden state will transition from state i to state j at the next moment, i.e. the next state is j and the current state is i; i is the starting state index. ij represents the hidden state number at the current time t, from 1 to M; j is the target state index, in a ij represents the hidden state number at the next moment t+1, from 1 to M; is the observation (emission) probability distribution of state m. It describes the observed PRI value (p k ) follows a probability distribution. Here it is a Gaussian (normal) distribution, that is, μ m is the mean of the transfer matrix of the mth state; is the variance of the transfer matrix of the mth state; p k is the kth observation value. The kth PRI value in the sequence, p k ={p1, p2....p k}, that is, the observed PRI sequence, which contains k PRI values; α1(m) is the forward probability of belonging to a certain state m at the initial moment (t=1); α k (m) is the forward probability of belonging to a state m at the final moment (t = k), which refers to observing the entire sequence p k And the joint probability of being in state m at time k; α t+1 (m) is the forward probability of belonging to a certain state m at time t+1; a jm is the state transition probability. ij The meaning is the same, just the index letter is different. t+1 In the recursive formula of (m); Observe p in state m t+1 The probability density of . That is, when the hidden state is m, the observed PRI value p t+1 The probability density value of p t+1 is the observation value at the next moment t+1; is the probability density of observing p1 in state m. It indicates the probability density value of observing the first PRI value p1 when the hidden state is m; π m is the initial probability distribution of state m; α t(j) is the forward probability of state j at time t, that is, when calculating α t+1 (m), the forward probabilities of all possible previous states j at time t are required; t is the time index. It represents the time step or sequence number in the sequence (for example, the t-th PRI value); m is the state index, representing one of the M hidden states (from 1 to M), m = 1, 2, ..., M; M is the total number of hidden states, representing the number of different hidden states defined in the model; is the probability density function of the Gaussian distribution, that is, the PRI variation pattern is recorded as a state transition path. In other words, each PRI variation pattern can be considered as a path, and the states of each node in this path are the PRI values. The specific process of PRI variation pattern extraction and classification based on the hidden Markov model is as follows:

[0120] Step 1: Data preprocessing: path extraction and typical value mapping.

[0121] ①Extract path data.

[0122] Cluster multiple historical paths to determine multiple clusters, and use the multiple historical paths corresponding to any cluster as the training set. Extract the PRI difference values ​​of all paths in the training set. Assume that the training set contains N paths, and the path set is recorded as Each path i Represented as a sequence of nodes: The node is a PRI difference value or a PRI value.

[0123] Among them, x j Indicates the jth PRI value in the path; n i is the number of PRI values, j∈n i ; j is the serial number.

[0124] ② The visual graphical interface assists in manually selecting possible typical values ​​of PRI parameters and typical value mapping.

[0125] A visual graphical interface software is designed to intuitively display all PRI values ​​in all wave positions. Then, by manually selecting the possible ranges of PRI typical value clusters, the mean of the node values ​​of each cluster is calculated as its typical value.

[0126] After obtaining the possible PRI typical values, they are stored in a unified data file. For each PRI value in the PRI path corresponding to each wave position, the likelihood distance is calculated to determine whether it can be mapped to a typical value. After the mapping is completed, the typical value mapping path is obtained.

[0127] (6) After mapping, typical value mapping paths are obtained. These paths are sorted using a multi-rule-based path similarity test method to determine the hidden Markov model to which they belong, thereby serving as the retraining and new training sets for each hidden Markov model.

[0128] The multi-rule-based path similarity verification method achieves path matching in complex scenarios through a three-level linkage rule system, so as to find the corresponding hidden Markov model after mapping the PRI and thus construct a training set for retraining the model. Specifically, it includes: first verifying the consistency of node order and allowing subsequence matching; secondly, introducing a merge replacement mechanism to intelligently merge adjacent node pairs that meet the sum value neighbor conditions; and finally allowing a limited number of end extensions. By dynamically generating virtual paths and verifying subsequent connections, the inclusive identification of reasonable variants is achieved while maintaining the core order constraints. Among them, the nodes are the PRI staggered values, and the consistency of the node order means that the order relationship between the PRI values ​​is consistent. The specific similarity judgment rules are as follows.

[0129] ①Subsequence matching rules.

[0130] The node order of the test path must strictly follow the node order of the base path. It is allowed to skip some nodes in the base path, but the original order cannot be changed. The rule matching steps are as follows:

[0131] 1. Initialize pointer: set the basic path pointer p base =1.

[0132] 2. Traverse the test path: for each node t in the test path i , move backward p base Until a matching node b is found j =t i If no match is found after traversing the basic path, it is judged as dissimilar.

[0133] 3. Termination condition: If all nodes in the test path match in order, it is considered similar. Let the basic path be Test path T test for There exists a set of strictly increasing position indices satisfy: Among them, t k is the kth node in the test path; t k The corresponding basic path node; is arbitrary; that is, for any test node in the test path, there is a basic path node corresponding to the node and the preceding node in the basic path, and the order of the corresponding nodes is strictly increasing.

[0134] It can be understood that existence: each node of the test path must appear in the basic path; order: the order of nodes in the test path must be consistent with the order of corresponding nodes in the basic path.

[0135] ②Merge and replace rules.

[0136] It is allowed to use the sum of two adjacent nodes to approximately replace the original node pair, but the connection relationship between subsequent nodes must be maintained. The rule matching steps are as follows.

[0137] 1. Merge condition detection: for each node t in the test path i , check whether there is an adjacent node pair of the basic path (b j ,b j+1 ) satisfies: |t i -(b j +b j+1 )|≤∈·(b j +b j+1 ).

[0138] Where ∈ is the merging error threshold (default ∈ = 0.1).

[0139] 2. Generate a virtual path: replace b in the base path j and b j+1 Replace with t i , and obtain the virtual path V.

[0140] V=[v1,...,v j-1 ,t i ,v j+1 ,...,v n ].

[0141] in,

[0142] 3. Connection verification: Verify whether the nodes of the test path after the merge position are consistent with the virtual path:

[0143] t i+1 =v k+1 ,t i+2 =v k+2 ,...; where t i+1 is the i+1th test path node; t i+2 is i+2 test path nodes; v k+2 is the k+2th virtual path node; v k+1 is the k+1th virtual path node.

[0144] Among them, when a node t on the test path i Detects two adjacent nodes b on a certain basic path j and b j+1Satisfaction|t i -(b j +b j+1 )|≤∈·(b j +b j+1 ), it is considered that t i It is b j and b j+1 The combined replacement value of these two nodes, but in order to judge the test path and the representative path as similar, it is also necessary to add t i and b j 、b j+1 In addition to the corresponding, other nodes must be equal to each other, that is, t1 = b1, t2 = b2... In order to verify this point, it is necessary to create a virtual path V based on the basic path to represent b in the basic path B. j and b j+1 In the case of merging these two nodes, the virtual path includes b in the basic path. j and b j+1 Replace with t i Except for the virtual path, all other nodes are completely equal, so the node value in V can be represented by b. After the virtual path is obtained, the test path is matched with the virtual path according to the subsequence matching rule. If the test path matches the virtual path, it means that the test path can match the basic path.

[0145] ③End expansion rules.

[0146] A limited number of new nodes can be appended to the end of the base path, but the order of the original nodes cannot be modified.

[0147] The rule matching steps are as follows:

[0148] 1. Prefix matching verification: test the first n of the path before The nodes must be exactly the same as the base path: t i =b i (n before =len(B base )); where len(.) represents the length, i.e. the number of nodes; stands for "any"; b i represents the i-th basic path; n before Represents the length of the base path.

[0149] 2. New node restrictions: The test path length must not exceed the basic path length plus the maximum allowed number of new nodes K (usually K = 2).

[0150] len(T test )-len(B base )≤K

[0151] ④Definition of fault tolerance and illegal operations.

[0152] 1. Illegal intermediate insertion: inserting a new node at a non-merge position, destroying the original order:

[0153] 2. Sequence violation: The order in which the base path nodes appear in the test path is inconsistent with the original order:

[0154] t p+1 ≠b j+1 And t p+1 Does not meet the merge replacement rule; among them, If it is "existent", that is, there is a test path node p whose subsequent node p+1 does not match the subsequent node j+1 of the basic path node j and does not satisfy the 2. merge replacement rule, then it is an illegal operation that destroys the order; t p For a test path node p; b j is a basic path node j; t p+1 is the next test path node p+1; b j+1 is the next basic path j+1.

[0155] If the above illegal operations exist in the test path, it is judged as dissimilar.

[0156] According to the above rules, a linkage rule system is implemented to achieve path matching in complex scenarios. Similar paths are classified into this category as a training set for retraining. When a certain number of paths are accumulated, the corresponding hidden Markov model is retrained. For dissimilar paths, the path clustering, representative path selection, hidden Markov model training and other steps are re-performed.

[0157] This application uses the quantile method to detect outliers, dynamic binning and peak detection of histograms, updates of the PRI typical value library, and path similarity comparison and HMM library model updates. Among them, the quantile method for detecting outliers is to extract the main data based on the quantile calculation method, set a dynamic threshold, and combine the density secondary verification to screen out abnormal high values ​​in the unknown PRI values; the histogram dynamic binning and peak detection uses the variance-data volume dual constraint conditions to control the binning process, and Gaussian fitting is performed on the high-frequency histogram bins to identify potential PRI typical values; the PRI typical value library update is to manually select the PRI candidate values ​​to be stored after fitting the PRI histogram, and complete the dynamic update of the PRI typical value library in combination with a visual graphical interface. The path similarity comparison and HMM model library update detects the similarity between the PRI mapping path and the existing PRI basic path through multiple rules, incorporates the paths that meet the conditions into the existing HMM model library, and retrains the model when a certain number is accumulated; if no matching basic path is found, it is used as the training set of the hidden Markov model corresponding to the unknown PRI jagged pattern.

[0158] The beneficial effects of this application are as follows.

[0159] 1. Improved the accuracy and robustness of PRI typical value extraction: Compared with the traditional histogram fixed binning method, this application detects outliers through the quantile method and the variance-constrained histogram dynamic binning method, which can effectively eliminate outliers, analyze the PRI data distribution more reasonably and accurately, and improve the accuracy of typical value extraction.

[0160] 2. Enhanced the adaptability and scalability of the PRI typical value library: Through mechanisms such as overlap detection, mean update trigger judgment, Bayesian reasoning and dynamic variance hybrid update, the parameters of existing typical values ​​in the library are updated and new typical values ​​are added. Compared with the traditional static PRI typical value library, this application uses an automated update mechanism to enable the PRI typical value library to dynamically adjust as the radar signal changes, enabling it to adapt to the dynamic changes of radar signals and the emergence of new radar radiation sources.

[0161] 3. Improved the adaptability and generalization ability of the PRI parameter HMM model library: By using a multi-rule-based path similarity test method to sort unknown paths, the hidden Markov model to which they belong is determined and used as a retraining and new training set for the HMM model library, so that the HMM model library can continuously learn and adapt to new PRI variation patterns, thereby improving the model's adaptability and generalization performance to unknown PRI variation patterns.

[0162] The following alternatives can also be used in this application. The quantile method can also be used to detect outliers using clustering-based methods such as DBSCAN to identify outliers as noise, but the parameter selection of the clustering algorithm requires manual real-time adjustment, otherwise it will affect the detection effect. The dynamic binning of the histogram constrained by variance and data volume can also use the traditional fixed-width binning method, but it is difficult to adapt to changes in the distribution of PRI data, or the equal-frequency binning method adjusts the binning boundaries so that each bin contains approximately the same number of data points, but may divide the area originally belonging to the same PRI typical value into different bins. When overlapping PRI typical values ​​are detected in the PRI typical value library update, the new typical value parameters can be directly used to replace the overlapping PRI typical value parameters in the library, so that its mapping ability to this data is enhanced, but historical information will be lost, resulting in instability of the PRI typical library and decreased generalization ability. The path similarity test based on multiple rules can use a distance model (such as Euclidean distance, cosine similarity, etc.) as a single rule to calculate the similarity between the test path and the basic path, but it is difficult to handle changes in complex PRI staggered patterns, such as special cases such as PRI value merging.

[0163] Reference Figure 8, provides a dynamic training and updating system for an HMM model library based on unknown PRI information, including: an acquisition module, for obtaining, for multiple wave position pulse sequences of radar signal echoes, PRI values ​​of wave position pulse sequences that cannot be mapped in the PRI typical value library as unknown PRI information; the PRI typical value library is a plurality of PRI typical values ​​determined by historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each unknown path includes a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values; an update processing module, for binning the unknown PRI information, and determining an updated PRI typical value library using a histogram and an EM algorithm; a determination module, for mapping the plurality of unknown paths based on the updated PRI typical value library, and determining each typical value mapping path; a matching module, for matching each typical value mapping path with each hidden Markov model in the HMM model library respectively, and determining similar paths and dissimilar paths; the HMM model library includes a plurality of hidden Markov models determined based on the PRI information of the wave position pulse sequences that have been mapped in the PRI typical value library. An updating module is used to dynamically train and update the HMM model library based on the similar paths and the dissimilar paths, and determine an updated HMM model library.

[0164] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0165] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for dynamic training and updating of an HMM model library based on unknown PRI information, characterized in that: include: For multiple wave position pulse sequences of radar signal echoes, PRI values ​​of wave position pulse sequences that cannot be mapped in a PRI typical value library are obtained as unknown PRI information; the PRI typical value library is a plurality of PRI typical values ​​determined based on historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each unknown path includes a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values; The unknown PRI information is binned and the updated PRI typical value library is determined using histogram and EM algorithm; Map multiple unknown paths based on the updated PRI typical value library to determine the mapping path of each typical value; Matching each typical value mapping path with each hidden Markov model in the HMM model library to determine similar paths and dissimilar paths; the HMM model library includes multiple hidden Markov models determined based on the PRI information of the wave position pulse sequence mapped in the PRI typical value library; Based on the similar paths and the dissimilar paths, the HMM model library is dynamically trained and updated, and an updated HMM model library is determined to identify the working mode of the complex radar signal.

2. The method for dynamic training and updating of an HMM model library based on unknown PRI information according to claim 1, characterized in that: The unknown PRI information is binned and the histogram and EM algorithm are used to determine the updated PRI typical value library, including: Based on the quantile method, the unknown PRI information is binned to determine the final binning result; Based on the histogram generated by the final binning result, the EM algorithm is used to fit the unknown PRI values ​​in each bin in the histogram to determine the fitting result; An updated PRI typical value library is determined based on the fitting result and a plurality of preset candidate PRI values.

3. The method for dynamic training and updating of HMM model library based on unknown PRI information according to claim 2, characterized in that: Based on the quantile method, the unknown PRI information is binned to determine the final binning results, including: Using the quantile method to eliminate outliers in the unknown PRI information to determine the main data; the main data includes multiple unknown PRI values; Perform binning processing on the subject data and determine the binning results; Based on the binning results, determining the variance and amount of the main data in each bin, and setting corresponding preset constraints for the variance and amount of the main data in each bin; The final binning result is determined based on the preset constraints and the binning result.

4. The method for dynamic training and updating of an HMM model library based on unknown PRI information according to claim 2, characterized in that: Based on the histogram generated by the final binning result, the EM algorithm is used to fit the unknown PRI values ​​in each bin in the histogram to determine the fitting results, which specifically include: Traverse the peak values ​​corresponding to each bin of the histogram and mark the peak value corresponding to the bin that meets the preset conditions as a candidate value; the preset conditions are that the peak value of the current bin is greater than the peak values ​​of the left and right adjacent bins and the peak value of the current bin reaches the preset peak value; The EM algorithm is used to perform Gaussian distribution fitting on each unknown PRI value in the bin corresponding to each candidate value to determine a fitting result; the fitting result is the mean and variance of each unknown PRI value in each bin.

5. The method for dynamic training and updating of HMM model library based on unknown PRI information according to claim 2, characterized in that: Determining an updated PRI typical value library based on the fitting result and multiple preset candidate PRI values ​​specifically includes: According to each bin range, a plurality of preset candidate PRI values ​​are set; each bin range includes a range to which all unknown PRI values ​​in the bin corresponding to each preset candidate value belong; Determine the mean and variance of the unknown PRI values ​​in the bins corresponding to each preset candidate PRI value, and use the mean and variance of the corresponding unknown PRI values ​​as the mean and variance of each preset candidate PRI value; An updated PRI typical value library is determined based on the mean and variance of each preset candidate PRI value and the mean and variance of the PRI typical values ​​in the PRI typical value library.

6. The method for dynamic training and updating of HMM model library based on unknown PRI information according to claim 5, characterized in that: An updated PRI typical value library is determined based on the mean and variance of each preset candidate PRI value and the mean and variance of the PRI typical values ​​in the PRI typical value library, specifically including: Determine whether the mean and variance of each preset candidate PRI value are consistent with the mean and variance of the PRI typical values ​​in the PRI typical value library, and obtain a first determination result; If the first judgment result is yes, determining that the preset candidate PRI value is repeated with the PRI typical value in the PRI typical value library, and updating the mean variance of the preset candidate PRI value based on the mean update trigger condition to determine an updated PRI typical value library; the mean update trigger condition is determined based on the mean, variance, data volume of the PRI value, and data volume of the PRI typical value of the preset candidate PRI value and the PRI typical value in the PRI typical value library; If the first judgment result is no, it is determined that the preset candidate PRI value is not repeated with the PRI typical value in the PRI typical value library, and the preset candidate PRI value is added to the PRI typical value library to obtain an updated PRI typical value library.

7. The method for dynamic training and updating of HMM model library based on unknown PRI information according to claim 1, characterized in that: Each typical value mapping path is matched with each hidden Markov model in the HMM model library to determine similar paths and dissimilar paths, specifically including: Based on a preset path similarity test method and each hidden Markov model in the HMM model library, the typical value mapping paths are grouped and matched to determine similar paths that match the HMM model library and dissimilar paths that do not match the HMM model library.

8. The method for dynamic training and updating of HMM model library based on unknown PRI information according to claim 1, characterized in that: Based on the similar paths and the dissimilar paths, the HMM model library is dynamically trained and updated to determine an updated HMM model library, specifically including: Using the similar path as a training set, retraining the hidden Markov model corresponding to the similar path in the HMM model library to determine a trained hidden Markov model; the trained hidden Markov model represents a PRI variation pattern; the PRI variation pattern includes multiple unknown PRI values ​​in the similar path and a sequential relationship between the multiple unknown PRI values; Clustering the dissimilar paths to determine a plurality of new basis paths, and determining a plurality of new hidden Markov models based on the plurality of new basis paths; An updated HMM model library is determined according to the multiple new hidden Markov models and the trained hidden Markov model.

9. The method for dynamic training and updating of an HMM model library based on unknown PRI information according to claim 8, characterized in that: Clustering the dissimilar paths to determine multiple new base paths, and determining multiple new hidden Markov models based on the multiple new base paths, specifically including: For dissimilar paths, mapping the dissimilar paths based on preset typical values ​​to determine a plurality of clusters; For the dissimilar paths in each cluster, the path with the highest frequency and the largest number of PRI value types is selected as the new basic path of each cluster; The new basic path of each cluster is used as the new PRI staggered pattern corresponding to each cluster; The new PRI staggered patterns corresponding to the clusters are used as the PRI staggered patterns corresponding to multiple new hidden Markov models.

10. A dynamic training and updating system for an HMM model library based on unknown PRI information, characterized in that: include: An acquisition module is configured to obtain, for multiple wave position pulse sequences of radar signal echoes, PRI values ​​of wave position pulse sequences that cannot be mapped in a PRI typical value library as unknown PRI information; the PRI typical value library is a plurality of PRI typical values ​​determined based on historical PRI information and prior knowledge; the unknown PRI information includes a plurality of unknown paths, each unknown path including a plurality of unknown PRI values ​​and a sequential relationship between the unknown PRI values; The update processing module is used to bin the unknown PRI information and determine the updated PRI typical value library using histogram and EM algorithm; A determination module, configured to map multiple unknown paths based on the updated PRI typical value library and determine each typical value mapping path; A matching module is used to match each typical value mapping path with each hidden Markov model in the HMM model library to determine similar paths and dissimilar paths; the HMM model library includes multiple hidden Markov models determined based on the PRI information of the wave position pulse sequence mapped in the PRI typical value library; The updating module is used to dynamically train and update the HMM model library based on the similar path and the dissimilar path, and determine the updated HMM model library to identify the working mode of the complex radar signal.

Images