An Automatic Method for Extracting Spectral Harmonic Features

Abstract

This invention discloses a method for automatically extracting harmonic features from a spectrum. The method includes frequency extraction, initial screening of harmonic group features, fine screening of harmonic group features, and extended Kalman tracking filtering. Finally, the feature information of the harmonic groups with the largest tracking counts in the tracking line spectrum set TrackS is output as the result. Based on the principle of harmonic frequency multiplication extraction, this method utilizes scaling search to reduce computational complexity and employs extended Kalman filtering to correlate and track multiple extraction results. This method is suitable for real-time extraction of combined line spectrum features from power and modulation spectra.

Description

Technical field:

[0001] The invention belongs to the field of underwater acoustics and underwater acoustic signal processing, and particularly relates to a method for automatically extracting spectrum harmonic features. Background technology:

[0002] Passive sonar target classification and identification is a primary function of sonar systems for submarines, surface ships, and shore-based surveillance systems. Its capabilities are a crucial factor in measuring naval combat effectiveness. Passive sonar target classification and identification involves analyzing the characteristics of a target's passively radiated noise, primarily extracting power spectrum and modulation spectrum features that characterize the target's physical properties. Line spectrum features are a crucial basis for target classification and identification, directly reflecting the target's physical characteristics and primarily arising from mechanical vibrations such as gears, pistons, drive shafts, propellers, and hulls. Considering the correlation between various excitation sources and the high-order harmonic components generated by line spectra, passive underwater acoustic target line spectra often exhibit a certain "clustering characteristic," whereby line spectra appear in groups and exhibit certain harmonic or other relationships. Harmonic combination characteristics of line spectra are an important identifiable attribute of targets, capable of characterizing target attributes and even individual information to a certain extent. Leveraging the harmonic information of target line spectra can, to a certain extent, facilitate target attribute determination. For example, by using the harmonic information of the modulation-based line spectrum, characteristic information about the target's propeller parameters can be estimated.

[0003] Furthermore, to enhance the operational capabilities of sonar systems, sonar equipment is currently developing towards autonomy and intelligence, and the ability to automatically extract passive sonar target features is urgently needed. When a target contains harmonic information in its power spectrum or modulation spectrum, automatically and accurately extracting this information is crucial for improving classification and recognition capabilities. The quality of harmonic group feature extraction is directly related to target recognition accuracy.

[0004] The collective form of a set of spectrum harmonic groups can be expressed as:

[0005]

[0006] where f off is the offset frequency, f div is the difference frequency, i is a natural number, representing the number of harmonics. In particular, for the DEMON harmonic group caused by the propeller, usually its offset frequency f off =0.

[0007] Existing harmonic group feature extraction methods are mainly developed for DEMON spectra, mainly including difference frequency extraction and octave frequency extraction. The difference frequency extraction method first extracts the line spectrum of the spectrum and extracts the harmonic relationship of the line spectrum group based on the difference distribution of the line spectrum. However, this method relies on the ability to extract the line spectrum. When the number of line spectra is large, it faces the problem of high algorithm complexity. The octave frequency extraction method is a search-scoring method that traverses the candidate harmonic group frequency points in the search space. The score value of the candidate harmonic is usually calculated based on the sum of the spectrum harmonic amplitude. Finally, the scores are sorted, and the harmonic that meets the threshold and has the highest score is obtained as the result of harmonic feature detection.

[0008] The frequency octave extraction method is less dependent on the performance of line spectrum extraction and is an effective and reliable method. Based on the idea of the frequency octave extraction method, this paper proposes a more universal spectrum harmonic group feature extraction method. Directly using the existing frequency octave extraction method to extract the harmonic features of a general spectrum mainly faces two difficulties:

[0009] 1) The algorithm has high computational complexity, because the harmonic group characteristics have two parameters to be estimated f off With f div , for a spectrum of length N, the complexity of the search space is O(N 2 ), the amount of computation required to traverse the search space is large, and the real-time performance of the method is difficult to meet.

[0010] 2) Lack of feature tracking capability. Existing frequency multiplication extraction methods are limited to extracting harmonic features within a single signal processing cycle. When the target harmonic frequency and intensity are relatively stable, existing methods can produce consistent results over a period of time. However, in reality, due to changes in the ship's operating conditions, motion state, underwater acoustic channel variations, and ambient noise fluctuations, the frequency and signal-to-noise ratio of the target harmonic line spectrum exhibit time-varying characteristics. This can lead to line spectrum broadening, frequency drift, and trajectory discontinuity. Existing methods will extract multiple different values over a period of time, hindering the subsequent target recognition process. Summary of the invention:

[0011] The technical problem to be solved by the present invention is how to quickly and accurately automatically extract spectrum harmonic features, especially solving the problem of real-time extraction of harmonic features with a large number of frequency points and frequency offset. The present invention provides a method for automatically extracting spectrum harmonic features. The method is based on the idea of frequency doubling extraction. First, spectrum scaling is used to perform coarse screening and detailed inspection of harmonic group features, which can accurately identify the harmonic group with the largest amplitude and score, effectively reducing the amount of computation. Secondly, an extended Kalman filter method is used to associate and track the harmonic group features extracted in each signal processing cycle, thereby enhancing the stability of harmonic feature extraction and realizing functions such as false alarm feature filtering and frequency-varying harmonic feature tracking.

[0012] The technical solution of the present invention is to provide a method for automatically extracting spectrum harmonic features, which can be found in Figure 1 , the method comprises,

[0013] Frequency point extraction, according to the sampling interval D = (2*N) 1 / 3 , select the largest point of the input spectrum x(i), i=1,2,...,N, and get the spectrum x d (k)=max{x((k-1)*D+1),...,x(k*D)};

[0014] Initial screening of harmonic group characteristics, first of all, x d (k) Perform preprocessing to remove the background spectrum, and then traverse the offset frequency and difference frequency combination search space <F divB ,F divE >, use the maximum harmonic amplitude and search scoring algorithm to find the frequency amplitude and the largest groups of harmonics Harmo c ;

[0015] Harmonic group characteristic fine screening, based on the results of the coarse screening of harmonic group characteristic Harmo c , taking x(i) as input, according to the range of fine screening of harmonic group characteristics, find the frequency amplitude and the largest groups of harmonics Harmo e ;

[0016] Extended Kalman tracking filter, based on the results of fine screening of harmonic group characteristics Harmo e , extract the offset frequency set and difference frequency set of a single extraction<BF,DF> ; For any element t∈TrackS, the prediction value of the tracking set is calculated according to the extended Kalman filter prediction model, and the prediction value is compared with<BF,DF> Matching, find out whether t has an observation value, and perform Kalman filtering on t based on the observation value; modify the tracking count of t according to the existence of the observation value. If the tracking count is too low, delete t; and for any element pt∈PTrackS, calculate the predicted value, find the observation value, perform Kalman filtering, modify the tracking count, and if the tracking count is greater than the set threshold, transfer pt to the TrackS set;

[0017] The characteristic information of the harmonic groups with the largest tracking counts in the tracking line spectrum set TrackS is output as the result.

[0018] The method proposed in the present invention is based on the principle of harmonic frequency multiplication extraction method, uses scaling search to reduce computational complexity, and uses extended Kalman filtering technology to achieve association and tracking of multiple extraction results. It is applicable to the problem of real-time extraction of combined line spectrum features of power spectrum and modulation spectrum. Starting from the spectrum of ship target radiation noise (including power spectrum and demodulation spectrum), the present invention automatically extracts and tracks spectrum harmonic features, can accurately extract the difference frequency and offset frequency of harmonic groups, has the ability to extract and track features of harmonic groups with weak harmonic line spectrum intensity, large frequency variation, and intermittent, and can automatically filter out false alarm noise. It provides a basis for multi-target line spectrum combination feature identification, target propeller parameter estimation, target classification and other processes, and has great engineering practical value.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] (1) To address the problem of high algorithm complexity when the number of spectrum frequency points is large, the method of the present invention uses the frequency point extraction and scaling method to effectively reduce the algorithm calculation complexity through coarse screening and fine screening ladder search of harmonic group characteristics;

[0021] (2) To address the problem of unstable frequency and intensity of harmonic group line spectrum, the present invention uses the extended Kalman method to track and filter the characteristic parameters of the harmonic group, effectively reducing the false alarm rate and improving the accuracy of feature extraction;

[0022] (3) The results of simulation and actual sea trial data analysis show that the proposed method can automatically and accurately extract the harmonic group characteristics of underwater acoustic target spectrum, especially for the harmonic group characteristics of power spectrum with offset frequency. This method provides a way to automatically extract target combination line spectrum characteristics and automatically estimate propeller parameters, and provides support for autonomous target classification and identification. Description of the drawings:

[0023] Figure 1 It is a flowchart of the present invention;

[0024] Figure 2 LOFAR process diagram for simulation target 1;

[0025] Figure 3 This is the single batch extraction result of harmonic group features for simulation target 1;

[0026] Figure 4 Harmonic group feature extraction and tracking results for simulation target 1;

[0027] Figure 5 LOFAR process diagram for simulation target 2;

[0028] Figure 6 This is the single batch extraction result of harmonic group features for simulation target 2;

[0029] Figure 7 Harmonic group feature extraction and tracking results for simulation target 2;

[0030] Figure 8 LOFAR process map for a merchant ship target;

[0031] Figure 9 Extract and track the harmonic characteristics of a merchant ship target;

[0032] Figure 10 DEMON process diagram for a merchant ship target;

[0033] Figure 11 Harmonic group feature extraction and tracking results for a merchant ship target. Specific implementation method:

[0034] The present invention will be further described below with reference to the accompanying drawings:

[0035] A method for automatically extracting spectrum harmonic features. The specific implementation process of this method is as follows:

[0036] (1) Initialization: Define the tracking set TrackS as an empty set and the pre-tracking set PTrackS as an empty set. The elements in the tracking set contain harmonic group feature information and tracking count information.

[0037] (2) Frequency extraction: Receive the target spectrum data x(i), i = 1, 2, ..., N, and its lower frequency limit fl, resolution reso, and extract the maximum value from it. The extraction method is as follows:

[0038] 1) Determine the extraction multiple D = (2*N) 1 / 3 .

[0039] 2) Select a large draw point: x d (k) = max{x((k-1)*D+1),...,x(k*D)}, and at the same time determine the frequency lower limit fl of the new spectrum d , resolution reso d , the number of frequency points is N d .

[0040] (3) Initial screening of harmonic group characteristics: Based on the extracted spectrum x d (k) Find the line spectrum combination characteristics of harmonics.

[0041] 1) First use the line spectrum extraction method to extract the spectrum x d (k) Preprocessing: removing the background spectrum. The line spectrum extraction method is a multi-level decision method. The specific process is as follows:

[0042] Algorithm 1-Line spectrum extraction algorithm

[0043] Input: spectrum-x d (k).

[0044] Output: Only keep the spectrum of the line spectrum -x dl (k), whose length is the same as x d (k) Consistency.

[0045] ① Use the double-pass separation window method to estimate the continuous spectrum of the spectrum to obtain x s (k).

[0046] ②Signal-to-noise ratio judgment, screening the frequency point set that meets the signal-to-noise ratio conditions, the screening conditions are (x d (k)-x s (k)) / x s (k)>D er , the amplitude of the frequency points that do not meet the conditions is set to 0, and the frequency points x after screening are obtained r1 (k).

[0047] ③ Spectral peak judgment, screening x r1 The frequency points in (k) that meet the peak-like characteristics are screened by calculating the left and right slopes of the frequency points, and recording the left slope as Requirement D l (k)>Td,D r (k)>Td, the amplitude of the frequency point that does not meet the conditions is set to 0.

[0048] Compared with the simple line spectrum extraction function module in underwater acoustic target recognition, the signal-to-noise ratio judgment threshold and spectrum peak judgment threshold of line spectrum extraction here are more tolerant. The signal-to-noise ratio judgment threshold is generally D er =2~4, the spectrum peak judgment threshold is generally set to N=3. Finally, the spectrum x after background filtering is obtained. dl (k).

[0049] 2) Maximum harmonic energy and search, at this time for the spectrum x dl (k), the frequency range corresponding to its frequency point is [fl d ,fh d ]=[fl d ,fl d +reso d ·N d ], determine the candidate offset frequency set F off ={reso d ,2·reso d ,...,N d ·reso d}, and the difference frequency search condition range mapping relationship F div ={f divB ,f divB +reso d ,...,f divE |foff ∈F off}, where for any fo ff ∈Fo ff , take f divB =0,f divE =(fh d -f divB ) / 2.

[0050] At this time, the candidate offset frequency f off ∈F off And the difference frequency search condition range mapping relationship F div As input, the maximum harmonic amplitude and search algorithm are used to obtain <f off ,F div ,x dl (k)> group energy and the characteristic information of the three largest harmonic groups under the condition of condition , harmonic energy and search specific algorithm are as follows:

[0051] Algorithm 2 - Maximum Harmonic Amplitude and Search Algorithm

[0052] Input: bias frequency -f off , difference frequency search range - Spectrum x dl (k).

[0053] Output: Maximum energy sum -sp, difference frequency -f div , harmonic number - nDiv.

[0054] ①For each candidate value within the difference frequency search range Calculate the harmonic group line spectrum set Find the frequency point position corresponding to the harmonic line spectrum m(F)={m1,m2,...,m k}, k is an adjustable parameter, indicating the number of harmonics, and is generally taken as k=7.

[0055] ②Estimation of harmonic group F * The magnitude and score value of , the scoring function s(·) is as follows:

[0056]

[0057] in

[0058] ③Traverse F div Calculate the amplitude and score of all elements in , and you can get the score set scoSet and its harmonic frequency point position set mSet. The two set elements have a one-to-one correspondence. Find the largest element sco* in scoSet and its harmonic frequency point position m*, let x dl (m*)=0.

[0059] ④ Repeat the above process three times to get the three largest amplitudes and scores The corresponding difference frequency And its harmonic frequency position Output as the final result.

[0060] 3) Traverse F off The maximum harmonic amplitude and search of the elements in can be obtained by 3·N d The search results of harmonic groups are sorted and the three harmonic group features with the largest global amplitude and maximum harmonic group features are obtained. The harmonic group features include offset frequency and difference frequency. The result expression is:

[0061]

[0062] (4) Fine screening of harmonic group characteristics:

[0063] 1) Results of coarse screening based on harmonic group characteristics <Fc off ,Fc div > Perform fine screening of harmonic group characteristics, Determine the offset frequency search range where reso d is the spectrum resolution after the frequency points are extracted in the previous text. Within this range, the offset frequency search interval is the original spectrum resolution reso, and the calculation can be obtained The CCP has N rs elements. At this time, any candidate frequency offset in the range Select the same difference frequency search range

[0064] 2) Similar to the coarse screening process of harmonic group characteristics, the triplet As input, the harmonic amplitude and search algorithm are used to obtain the group energy and the maximum three harmonic group characteristic information, where x(k) is the original input spectrum. The three largest amplitudes and scores can be obtained. The corresponding difference frequency And its harmonic frequency position

[0065] Traversal The elements in can be obtained as 3·N rs harmonic group search results, including the three harmonic group information sets with the largest amplitude and score <Fr off ,Fr div , Mr> is the harmonic group search result of the current spectrum x(k).

[0066] (5) Harmonic group feature tracking filtering:

[0067] 1) Description of the Kalman filter algorithm of the present invention

[0068] Feature tracking state variables: X = [Bf, Bf′, Df, Df′, z] T , respectively represent the offset frequency, offset frequency change rate, difference frequency, and difference frequency change rate, where z∈{-1,0,1} represents the offset frequency correction factor to determine the frequency position of the first line spectrum of the harmonic group. The prediction model is Among them, P k To predict the model error, the initial state is P = 0.01·eye(5), where eye(·) represents the diagonal identity matrix.

[0069] The linearized state transfer matrix φ and the state transfer covariance matrix Q are shown as follows:

[0070]

[0071] The observation model is Z k =HX k +R, where the observed variable Z = [Bf, Df, z] T , the observation transfer matrix H and the observation noise variance are divided into levels:

[0072]

[0073] The Kalman update formula is:

[0074] P t - =φP t-1 φ T +Q

[0075]

[0076]

[0077]

[0078] If the observation is missing, then:

[0079]

[0080] 2) Update the tracking set TrackS: For all s*∈TrackS, calculate its Kalman filter prediction value X - , check if <fr off ,fr div >∈ <Fr off ,Fr div >, making |fr off -Δf - |<delta1±k·Df - |k∈{0,1}, and |fr div -Df -|<delta2, if it exists, then the observed value of s* is recorded as <fr off ,fr div >, while from the collection <Fr off ,Fr div >Remove <fr off ,fr div >, otherwise the observation is missing.

[0081] Where count* is the counter of s*. If the above observation value exists and count*<mvalve, then count* is increased by 1. Otherwise, if the observation value is missing, count* is decreased by 1. When count*<tvalve1, s* is removed from TrackS.

[0082] 3) Update the pre-tracking set PTrackS: For all s*∈PTrackS, perform the steps similar to updating the tracking set TrackS. The difference is that when the count is updated, if count*>avalve, s* is moved into TrackS.

[0083] (6) Output the harmonic group tracking information in the tracking set TrackS, and take the tracking result with the most cumulative tracking cycles as the optimal detection result.

[0084] The above implementation process was simulated and tested with sea trial data, and the results are as follows:

[0085] On the basis of the above implementation process, the actual sea trial data of the simulation and admission are analyzed and processed to give the processing results of the present invention. Figures 2 to 5 As shown, the actual sea trial processing results of the merchant ship target are as follows Figures 6 to 9 shown.

[0086] Computer simulation 1: The simulation contains 1 target, which contains 6 line spectra. The spectra are 135Hz, 170Hz, 240Hz, 310Hz, 380Hz, and 450Hz respectively. 170Hz, 240Hz, 310Hz, 380Hz, and 450Hz are a group of harmonics. The offset frequency is 100Hz and the difference frequency is 70Hz. Figure 2 The results show that the target line spectrum has weak phase stability and wide spectrum, and it is difficult to extract the harmonic group features by line spectrum extraction. Using the harmonic group feature extraction method proposed in this invention, the harmonic group feature extraction results can be obtained as follows: Figure 3 and Figure 4 shown. Figure 3 This is the result of feature extraction of harmonic groups without tracking processing. For simplicity of display, only the first and last line spectra of the harmonic group are shown in the figure, represented by circular symbols and star symbols respectively. Figure 3The results show that the harmonic group characteristics extracted from a single cycle are unstable, where the extracted value of the offset frequency is 30 and the difference frequency is 35, which does not meet the set conditions. The method based on the extended Kalman filter proposed in this invention is used to track and filter the harmonic group characteristics, and the tracker result with the most cumulative tracking cycles is output. The results are as follows: Figure 4 As shown, it can be seen that after tracking filtering, the harmonic group characteristics can be stably extracted, and the extracted line spectrum frequency is consistent with the set conditions.

[0087] Computer simulation 2: The simulation includes a target with five line spectra, with center frequencies of 170Hz, 240Hz, 310Hz, 380Hz, and 450Hz. This clearly forms a harmonic group with an offset frequency of 100Hz and a difference frequency of 70Hz. The line spectra vary in a consistent cosine pattern. Figure 5 The LOFAR process diagram shows that the target line spectrum frequency changes greatly, and its feature extraction is more difficult than the line spectrum harmonic group with stable frequency. Using the harmonic group feature extraction method proposed in this invention, the harmonic group feature extraction results can be obtained as follows: Figure 6 and Figure 7 shown. Figure 6 This is the result of harmonic group feature extraction without tracking processing. Similar to the result of computer simulation 1, the stability of the harmonic group features extracted from a single batch of data is low. Figure 7 After the extended Kalman filter tracking process, the stable harmonic group feature extraction result is obtained. It can be seen that the Kalman filter method proposed in the present invention can effectively solve the problem caused by the unstable frequency of the harmonic group line spectrum and filter out the erroneous extraction results.

[0088] Sea trial data processing 1: Actual recorded merchant ship target signal. Figure 8 The LOFAR spectrum of the merchant ship target shows that the target has obvious LOFAR line spectrum harmonic group line spectrums of 112Hz, 172Hz, 232Hz, 292Hz, 352Hz, and 412Hz. The harmonic group characteristics of the target are extracted using the method proposed in this invention. The results are as follows: Figure 9 As shown, it can be seen that for practical purposes, the present invention can stably and accurately extract LOFAR harmonic group features.

[0089] Sea trial data processing 2: In order to verify the ability of the present invention to extract the harmonic characteristics of the DEMON spectrum, a target signal of a merchant ship was actually recorded. The target DEMON process diagram is as follows: Figure 10 As shown in the figure, there are obvious harmonic group features, the shaft frequency is 2Hz, and the blade frequency is 8Hz. The method proposed by the present invention is used to extract the harmonic group features of DEMON, and the line spectrum with the largest amplitude in the harmonic is taken as the blade frequency. The results are as follows Figure 11It can be seen that when the target DEMON modulation features are clear, the method of the present invention can effectively extract the shaft-blade frequency features.

[0090] The above-mentioned simulation analysis and sea trial data processing results demonstrate that the proposed automatic harmonic group feature extraction method can extract line spectrum harmonic group features from LOFAR and DEMON spectra, effectively overcoming the difficulty in line spectrum extraction caused by large line spectrum broadening. It is capable of stably extracting and tracking harmonic group features with large line spectrum frequency variations and weak intensity, while also filtering out the influence of false alarm noise, effectively improving feature extraction accuracy. This method provides technical support for line spectrum combination feature identification, propeller parameter estimation, and autonomous underwater acoustic target recognition, and has promising application prospects.

[0091] In other words, this method can accurately extract the difference and offset frequencies of harmonic groups. It is capable of extracting and tracking the characteristics of harmonic groups with weak line spectrum intensity, large spectrum width, large frequency variation, and intermittent transmission. It can also automatically filter out false alarm noise to achieve accurate feature extraction. This method provides a foundation and support for multi-target line spectrum combination feature identification, target propeller parameter estimation, and target classification, and has great engineering practical value.

[0092] The above description is only for the preferred embodiment of the present invention, which should not be understood as limiting the claims. Any equivalent process changes made using the present invention description are included in the patent protection scope of the present invention.

Claims

1. A method for automatically extracting spectrum harmonic features, characterized by: The method includes, Frequency point extraction, according to the sampling interval D = (2*N) 1 / 3 , select the largest point of the input spectrum x(i), i=1,2,...,N, and get the spectrum x d (k)=max{x((k-1)*D+1),...,x(k*D)}; Initial screening of harmonic group characteristics, first of all, x d (k) Perform preprocessing to remove the background spectrum, and then traverse the offset frequency and difference frequency combination search space <F divB ,F divE >, use the maximum harmonic amplitude and search scoring algorithm to find the frequency amplitude and the largest groups of harmonics Harmo c ; Harmonic group characteristic fine screening, based on the results of the coarse screening of harmonic group characteristic Harmo c , taking x(i) as input, according to the range of fine screening of harmonic group characteristics, find the frequency amplitude and the largest groups of harmonics Harmo e ; Extended Kalman tracking filter, based on the results of fine screening of harmonic group characteristics Harmo e , extract the offset frequency set and difference frequency set of a single extraction<BF,DF> ; For any element t∈TrackS, the prediction value of the tracking set is calculated according to the extended Kalman filter prediction model, and the prediction value is compared with<BF,DF> Matching, find out whether t has an observation value, and perform Kalman filtering on t based on the observation value; modify the tracking count of t according to the existence of the observation value. If the tracking count is too low, delete t; and for any element pt∈PTrackS, calculate the predicted value, find the observation value, perform Kalman filtering, modify the tracking count, and if the tracking count is greater than the set threshold, transfer pt to the TrackS set; The characteristic information of the harmonic groups with the largest tracking counts in the tracking line spectrum set TrackS is output as the result.

2. The method for automatically extracting spectrum harmonic features according to claim 1, wherein: Before frequency extraction, an initialization operation is performed. Specifically, the tracking set TrackS is defined as an empty set, the pre-tracking set PTrackS is defined as an empty set, and the elements in the tracking set include harmonic group feature information and tracking count information.

3. The method for automatically extracting spectrum harmonic features according to claim 1, wherein: When extracting frequency points, assume that the target spectrum data x(i), i = 1, 2, ..., N, and its lower frequency limit fl, resolution reso, are received and extracted. The extraction method is as follows: Determine the extraction multiple D = (2*N) 1 / 3 ; Choose a big draw point: x d (k) = max{x((k-1)*D+1),...,x(k*D)}, and at the same time determine the frequency lower limit fl of the new spectrum d , resolution reso d , the number of frequency points is N d .

4. The method for automatically extracting spectrum harmonic features according to claim 1, wherein: In the initial screening operation of harmonic group characteristics, the spectrum x is first extracted using the line spectrum extraction method. d (k) Preprocessing: removing the background spectrum. The line spectrum extraction method is a multi-level decision method. The specific process is as follows: First, the double-pass separation window method is used to estimate the continuous spectrum of the spectrum to obtain x s (k); Next, the signal-to-noise ratio is judged, and the frequency point set that meets the signal-to-noise ratio condition is screened. The screening condition is (x d (k)-x s (k)) / x s (k)>D er , the amplitude of the frequency points that do not meet the conditions is set to 0, and the frequency points x after screening are obtained r1 (k); Then, peak judgment, screening x r1 The frequency points in (k) that meet the peak-like characteristics are screened by calculating the left and right slopes of the frequency points, and recording the left slope as Requirement D l (k)>Td,D r (k)>Td, the amplitude of the frequency point that does not meet the conditions is set to 0; among them, the signal-to-noise ratio judgment threshold is D er =2~4, the spectrum peak decision threshold is N=3, and finally the spectrum x after background filtering is obtained. dl (k).

5. The method for automatically extracting spectrum harmonic features according to claim 1, wherein: In the initial screening operation of harmonic group characteristics, the maximum harmonic amplitude and search scoring algorithm are as follows: For each candidate value within the difference frequency search range Calculate the harmonic group line spectrum set Find the frequency point position corresponding to the harmonic line spectrum m(F)={m1,m2,...,m k }, k is an adjustable parameter, indicating the number of harmonics, and k=7; Estimated harmonic group F * The magnitude and score value of , the scoring function s(·) is as follows: in Traverse F div Calculate the amplitude and score of all elements in , and get the score set scoSet and its harmonic frequency point position set mSet. Find the largest element sco* in scoSet and its harmonic frequency point position m*. Let x dl (m*)=0; Repeat the above process three times to get the three largest amplitudes and scores. The corresponding difference frequency And its harmonic frequency position Output as the final result.

6. The method for automatically extracting spectrum harmonic features according to claim 5, wherein: The specific operation of harmonic group characteristic fine screening is as follows: Results of coarse screening based on harmonic group characteristics <Fc off ,Fc div > Perform fine screening of harmonic group characteristics, Determine the offset frequency search range where reso d is the spectrum resolution after frequency extraction. Within this range, the offset frequency search interval is the original spectrum resolution reso. The calculation can be obtained The CCP has N rs elements, then any candidate offset frequency in the range Select the same difference frequency search range; Triple As input, the harmonic amplitude and search algorithm are used to obtain the group energy and the three largest harmonic group feature information, where x(k) is the original input spectrum, and the three largest amplitudes and scores can be obtained. The corresponding difference frequency And its harmonic frequency position as well as, Traversal The elements in can be obtained as 3·N rs harmonic group search results, including the three harmonic group information sets with the largest amplitude and score <Fr off ,Fr div , Mr> is the harmonic group search result of the current spectrum x(k).

7. The method for automatically extracting spectrum harmonic features according to claim 1, wherein: The Kalman filter prediction model is Among them, P k To predict the model error, the initial state is P = 0.01·eye(5); where eye(·) represents the diagonal unit matrix.

Images