Millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment

Abstract

The invention relates to the technical field of signal reconstruction, in particular to a complex electromagnetic environment-oriented millimeter wave signal blind source separation and reconstruction system, which comprises a frequency spectrum trend division module, a path fading construction module, an initial cluster label generation module, a multi-solution path screening module and a fusion reconstruction execution module. According to the method, power spectral density sequence processing is carried out on millimeter wave frequency domain data, a multi-dimensional feature group is formed according to a path loss factor, an angle of arrival and the like, and an initial feature cluster is screened through an Euclidean distance, so that the accuracy of signal source classification is effectively improved; after the frequency domain response and the phase contour are continuously subjected to point comparison, path screening is completed by integrating a mean square error residual value, the path misjudgment probability is reduced, time domain resampling and phase frequency offset standardization are executed in fusion reconstruction, the consistency and fidelity of signal reconstruction are enhanced, and the reconstruction precision is improved. The whole process improves the separation accuracy and reconstruction precision of mixed signals in a complex electromagnetic environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal reconstruction, in particular to a millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment. BACKGROUND

[0002] The technical field of signal reconstruction mainly studies how to accurately recover the original signal based on limited measurement information under incomplete or damaged signal observation conditions. This field covers methods such as sparse representation, compressed sensing, blind source separation, dictionary learning, and low-rank matrix recovery, and combines statistical inference, optimization theory, and deep learning to realize the restoration of multi-dimensional signals such as time domain, frequency domain, and spatial domain. This technology has key application value in communication, radar, medical imaging, audio processing, image restoration, cognitive radio, and electromagnetic countermeasures, etc. The goal is to improve the accuracy and robustness of signal restoration under conditions of low signal-to-noise ratio, incomplete data, or complex signal source mixing.

[0003] Among them, the millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment is a system for automatically separating and accurately restoring millimeter wave signals in environments with strong electromagnetic interference, mixed multiple signal sources, and serious channel fading. Its purpose is to separate multiple independent signal sources from unlabeled and unknown statistical characteristic received mixed signals, and to reconstruct the original signal based on the estimated source signal and the mixing model. It is widely used in millimeter wave radar communication, electronic reconnaissance, unmanned system perception, interference identification and countermeasures, etc. high-frequency high-speed scenarios, to enhance the system's ability to counter complex interference and enhance signal interpretation.

[0004] Traditional reconstruction systems directly process mixed signals with power spectral density, lack of segment classification of spectral change trends, resulting in overly rough spectral feature extraction, and fail to fully consider the quantitative expression of path loss factor and reflection path complexity in the construction of link propagation model, which can easily cause problems such as ambiguous path attribute description. In the initial clustering stage, only the statistical characteristics are used to separate the signal sources, and the structural feature differences between signals cannot be accurately captured, causing clustering risk. In the path screening process, the joint judgment of frequency response deviation and residual error is insufficient, resulting in invalid paths mixed into the reconstruction step. For example, in signal reconstruction, if the signal segment contains path overlap or severe fading, it will cause distortion of the reconstructed output waveform, affecting the system's analysis ability and result stability in the interference scenario. SUMMARY

[0005] In order to solve the technical problems existing in the prior art, the embodiments of the present application provide a millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment. The technical solution is as follows: On the one hand, a millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment is provided, which comprises: The spectrum trend segmentation module obtains the power spectral density sequence in the millimeter-wave frequency domain data frame, calls the frequency segmentation values ​​and the sampling point sequence within the bandwidth, performs trend identification of the frequency band, and generates millimeter-wave spectrum slope interval label information; Based on the millimeter-wave spectral slope interval label information, the path fading construction module calculates the product of the link propagation distance and the density of obstructions to obtain the segment path loss factor, marks the segments with a reflection number greater than the path complexity benchmark value, and establishes millimeter-wave channel path damping segment information. The initial cluster label generation module calls the millimeter-wave channel path damping segmentation information, and constructs a two-dimensional feature vector group by combining the angle of arrival vector and the loss factor. Feature groups with a distance less than the vector similarity threshold are aggregated into the same initial cluster, forming the initial clustering label set of the signal source. The multi-solution path filtering module calculates the frequency response deviation between matching points based on the initial clustering label set of the signal source, compares the mean square error residual value with the set mean square error residual tolerance threshold, filters out paths whose deviation value and residual value both exceed the threshold, and obtains a set of blind source candidate reconstruction path numbers. The fusion reconstruction execution module calls the blind source candidate reconstruction path number set, resamples and adjusts the time domain amplitude estimate, maps the minimum mean square phase estimate to the frame-level time axis and superimposes it with the amplitude value, standardizes the frequency offset vector value in each path and then weights and superimposes it to form a signal reconstruction set, thus constructing and generating millimeter-wave blind source separation and reconstruction waveform information.

[0006] As a further embodiment of the present invention, the millimeter-wave spectral slope interval label information includes a frequency domain partition number, a fitting slope level identifier, and a power spectrum trend type; the millimeter-wave channel path damping segmentation information includes a propagation segment path number, a path damping factor grouping label, and a multipath propagation segment marker; the signal source initial clustering label set includes a signal source cluster number, a feature vector aggregation number, and an initial cluster center vector identifier; the blind source candidate reconstruction path number set includes a reconstruction path sequence number, a path amplitude and phase error label, and a path screening marker identifier; and the millimeter-wave blind source separation and reconstruction waveform information includes a waveform amplitude sequence set, a phase recovery function set, and a frequency offset fusion vector set.

[0007] As a further aspect of the present invention, the spectrum trend division module includes: The data frame receiving submodule acquires the power spectral density sequence in the millimeter-wave frequency domain data frame, extracts the frequency distribution information and corresponding power spectral density sequence within the frequency domain bandwidth, serializes each data frame according to the center frequency arrangement, extracts continuous sampling point data groups within a fixed interval length based on the frequency point index, and generates a frequency band power sampling sequence set. The power sequence processing submodule calls the frequency band power sampling sequence set, extracts the power value sequence and frequency point sequence in each segment, and performs linear regression fitting with the power value sequence as the dependent variable and the frequency value sequence as the independent variable according to the order of the frequency point sequence. It extracts the absolute value of the regression slope of each frequency segment and assigns a corresponding index to generate a segmented fitting slope value set. The trend interval labeling submodule calls the power spectrum slope segment threshold based on the segmented fitted slope value set, determines the interval of the slope value of each frequency segment according to the range of the value and the threshold, sets the interval number and binds the frequency segment sequence number, and generates millimeter wave spectrum slope interval label information.

[0008] As a further aspect of the present invention, the path fading construction module includes: The environmental element extraction submodule extracts the propagation path number corresponding to the millimeter wave spectrum slope interval label information, and collects the corresponding link propagation distance, reflection count, and obstruction density three types of environmental parameter values. It performs unified conversion and unit standardization on the data types of link propagation distance and obstruction density, establishes the number mapping relationship between propagation path and environmental parameters, and generates a link environmental parameter index table. The path damping calculation submodule calls the link environment parameter index table to obtain the link propagation distance and obstruction density corresponding to each path number, performs product calculation on the two parameters, and binds the product value with the path number to form a path factor structure. Based on the path factor structure, the path number is mapped and updated to obtain the path loss factor dataset. The multipath segment labeling submodule reads the reflection count value corresponding to each path based on the path loss factor dataset, and compares the reflection count value with the set path complexity benchmark value. Path numbers with a reflection count greater than the path complexity benchmark value are marked as multipath propagation segments. Multipath attribute labeling and damping value archiving are performed on the path numbers to generate millimeter-wave channel path damping segment information.

[0009] As a further aspect of the present invention, the process of determining the relationship between the reflection count value and the set path complexity benchmark value specifically involves obtaining the reflection count value corresponding to each path, using a greater than or equal to determination logic, and marking the path number with a reflection count value greater than or equal to the path complexity benchmark value as a path number with multipath propagation characteristics. The number of reflections is an integer-order single-path propagation relay point count, the path complexity benchmark is a fixed integer threshold, and the path complexity benchmark is three.

[0010] As a further aspect of the present invention, the initial cluster tag generation module includes: The feature vector construction submodule calls the millimeter-wave channel path damping segmentation information to obtain the loss factor value, angle of arrival vector value and spectrum trend category label number corresponding to each path segment. It combines the loss factor value as the first dimension component and the angle of arrival vector value as the second dimension component to construct a two-dimensional feature vector group for the path and generate a path feature vector set. The label condition filtering submodule extracts a subset of feature vectors with the same spectral trend category label number based on the path feature vector set, groups the feature set according to the label number as the condition filtering parameter, and retains the number index information for the feature vectors in each group to obtain a label mapping vector group set. The Euclidean clustering labeling submodule obtains the feature vector coordinate values ​​in each group according to the label mapping vector group set, calculates the Euclidean distance between each pair of vectors, and judges the relationship between them with the set vector similarity threshold. It then aggregates the vector number groups with distance values ​​less than the threshold to construct a set of the same cluster category, forming the initial clustering label set of the signal source.

[0011] As a further aspect of the present invention, the multi-solution path filtering module includes: The spectrum feature extraction submodule extracts the frequency domain amplitude response sequence, phase profile point series and mean square error residual value data corresponding to the signal segments in the clusters based on the initial clustering label set of the signal source. It performs time axis alignment on the signal segment frame sequence and synchronously numbers and indexes it to establish a set of spectrum response structure of the signal segments under each cluster and generate a clustered signal spectrum feature set. The error index calculation submodule calls the clustered signal spectrum feature set, extracts the frequency domain amplitude response value and phase profile point pair in each signal segment, calculates the frequency deviation value and phase difference value between each pair of points synchronously, and performs normalization processing. The residual value and frequency response deviation value in each signal segment are numerically superimposed to obtain the multipath error evaluation value group. The path number filtering submodule performs interval judgment on the evaluation value group of each path segment and the set mean square error residual tolerance threshold and frequency response error tolerance threshold, respectively, based on the multi-path error evaluation value group. It filters out path segment numbers that exceed both thresholds at the same time, retains path numbers that do not exceed the thresholds, and performs number classification and sorting to obtain a set of blind source candidate reconstruction path numbers.

[0012] As a further aspect of the present invention, the mean square error residual tolerance threshold is set by identifying the upper limit of the deviation range after statistical normalization of the residual distribution corresponding to each sampling point in the frequency domain amplitude response sequence of the signal segment. The frequency response error tolerance threshold is set based on the maximum allowable offset range of the frequency drift amplitude change rate of the phase profile point array within adjacent frames.

[0013] As a further aspect of the present invention, the fusion reconstruction execution module includes: The amplitude data correction submodule calls the blind source candidate reconstruction path number set, reads the time domain amplitude estimation value sequence under the corresponding path, performs time axis resampling on each amplitude value according to the frame index order, performs equal-interval interpolation on the sampling point spacing and generates a resampled amplitude data structure of uniform length to obtain path amplitude resampling information. The phase data mapping submodule obtains the minimum mean square phase estimate sequence corresponding to the path number based on the path amplitude resampling information, performs frame-level index mapping processing on the sequence with the resampling time axis, maps the phase estimate of each frame to the amplitude value at the same frame position, performs point-to-point superposition, and generates a phase amplitude fusion frame sequence group. The frequency offset superposition and reconstruction submodule extracts the frequency offset vector value corresponding to each path segment based on the phase amplitude fusion frame sequence group and performs zero-mean normalization processing. It then performs weighted average superposition of the frequency offset vectors according to the path number order, and performs one-dimensional structure splicing of the fusion frame sequence and the frequency offset superposition vector to form a continuous reconstruction signal, thereby constructing millimeter-wave blind source separation and reconstruction waveform information.

[0014] As a further aspect of the present invention, the weighted average superposition process specifically involves assigning weights to the standardized vectors according to the path bandwidth ratio in the order of path numbers, and performing component-level summation on the weighted vectors to obtain the fused frequency offset superposition vector. The process of performing one-dimensional structure splicing of the fused frame sequence and the frequency offset superposition vector is as follows: according to the frame index, the frequency offset superposition vector component under the corresponding frame index is added to the end of the phase amplitude fused frame sequence, and the one-dimensional continuous output format of the data structure is maintained.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This system processes millimeter-wave frequency domain data using power spectral density sequences, identifies spectral interval trend characteristics by dividing the data into frequency trends, and constructs path damping segmentation information by combining environmental factors such as link propagation distance, obstruction density, and reflection count. This enables a quantitative expression of path complexity. Multi-dimensional feature groups are constructed based on path loss factors and angle of arrival, and initial feature clusters are selected using Euclidean distance, effectively improving the accuracy of signal source classification. Subsequently, point comparisons are performed on the frequency domain response and phase profile, and the mean square error residual value is used to complete path selection, reducing the probability of path misjudgment. In the fusion reconstruction, time-domain resampling and phase frequency offset standardization are performed to enhance the consistency and fidelity of signal reconstruction. The overall process improves the accuracy of separating and reconstructing mixed signals in complex electromagnetic environments and ensures stable and controllable output of the reconstructed waveform under multi-source co-frequency interference. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a millimeter-wave signal blind source separation and reconstruction system for complex electromagnetic environments provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the spectrum trend division module in this invention; Figure 4 This is a flowchart of the path fading construction module in this invention; Figure 5 This is a flowchart of the initial cluster tag generation module in this invention; Figure 6 This is a flowchart of the multi-solution path filtering module in this invention; Figure 7 This is a flowchart of the fusion and reconstruction execution module in this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] like Figures 1-2 As shown, this embodiment of the invention provides a millimeter-wave signal blind source separation and reconstruction system for complex electromagnetic environments. The system includes a spectrum trend division module, a path fading construction module, an initial cluster label generation module, a multi-solution path screening module, and a fusion reconstruction execution module. The spectrum trend segmentation module obtains the power spectral density sequence in the millimeter-wave frequency domain data frame, calls the frequency segmentation value and the sampling point sequence within the bandwidth, calculates the linear regression coefficient for the power sequence of each segment sampling point in the order of the frequency axis, performs interval positioning based on the absolute value of the regression coefficient and the set power spectrum slope segmentation threshold, uses the interval label as the trend identifier of the current frequency band, and numbers and organizes the trend identifiers corresponding to the frequency segments to generate millimeter-wave spectrum slope interval label information; The power spectrum slope segmentation threshold is the interval boundary for judging the spectral trend based on the absolute value of the linear fitting slope. Its value usually depends on the amount of power change per unit frequency under the system bandwidth. The path fading construction module extracts three environmental factors corresponding to each type of interval: link propagation distance, number of reflections, and density of obstructions, based on the millimeter-wave spectrum slope interval label information. It calculates the link propagation distance and obstruction density by multiplying them to obtain the segment path loss factor. For segments with a number of reflections greater than the path complexity benchmark value, they are marked as multipath propagation segments to establish millimeter-wave channel path damping segment information. Link propagation distance is the total propagation length along the path from the transmitter to the receiver; reflection count is the number of times the electromagnetic wave is reflected after encountering an object, which is an environmental factor modeling item; obstruction density is the ratio of the number of obstructions per unit propagation path to the path length; the path complexity benchmark is an empirical threshold used to determine multipath environments, and a reflection count > 3 is usually considered a highly complex path. The initial cluster label generation module calls the millimeter-wave channel path damping segment information to obtain the loss factor, angle of arrival vector, and spectral trend category label of the path segment. The angle of arrival vector and loss factor are combined to form a two-dimensional feature vector group. Euclidean distance is calculated based on multiple feature vector groups with the same spectral trend category label. Feature groups with a distance less than the vector similarity threshold are aggregated into the same initial cluster, forming the initial cluster label set of the signal source. Angle of arrival (AHA) represents the angle of incidence of a signal from the source to the receiver, and is often used for source localization; Euclidean distance is the straight-line distance between two points, and is used to measure similarity in feature clustering; the vector similarity threshold is a specified upper limit of Euclidean distance, which is often set as the propagation angle difference corresponding to the shortest distinguishable path length in the system. The multi-solution path filtering module extracts the frequency domain amplitude response, phase profile point series, and mean square error residual value of the corresponding signal segment for each cluster based on the initial clustering label set of the signal source. It performs intra-frame point comparison between the frequency domain amplitude response and the phase profile point series, calculates the frequency response deviation between matching points, compares the mean square error residual value with the set mean square error residual tolerance threshold, filters out paths where both the deviation value and the residual value exceed the threshold, and obtains a set of blind source candidate reconstruction path numbers. The frequency domain amplitude response is the amplitude intensity distribution of the signal at different frequency points; the phase profile point list is the set of phase values ​​corresponding to each sampling point in the time or frequency domain, used to estimate frequency drift and path difference; the mean square error residual is the average squared difference between the fitted value and the original signal, used to judge the quality of the fitted path; the mean square error residual tolerance threshold is the upper limit of the acceptable residual. The fusion reconstruction execution module calls the blind source candidate reconstruction path number set, reads the time domain amplitude estimate, minimum mean square phase estimate, and frequency offset vector value under the corresponding path, resamples and adjusts the time domain amplitude estimate, maps the minimum mean square phase estimate to the frame-level time axis and superimposes it with the amplitude value, standardizes the frequency offset vector value in each path and weights and superimposes it to form a signal reconstruction set, and constructs and generates millimeter-wave blind source separation and reconstruction waveform information. The time-domain amplitude estimate is the time-axis amplitude sample point obtained from the signal conversion after frequency domain reconstruction; the minimum mean square phase estimate is the phase recovery value obtained under the minimum mean square error estimation; the frequency offset vector value is a vector sequence composed of the signal frequency offset in the path. The millimeter-wave spectral slope interval label information includes frequency domain partition number, fitting slope level identifier, and power spectrum trend type. The millimeter-wave channel path damping segment information includes propagation segment path number, path damping factor grouping label, and multipath propagation segment marker. The signal source initial clustering label set includes signal source cluster number, feature vector aggregation number, and initial cluster center vector identifier. The blind source candidate reconstruction path number set includes reconstruction path sequence number, path amplitude and phase error label, and path screening marker. The millimeter-wave blind source separation and reconstruction waveform information includes waveform amplitude sequence set, phase recovery function set, and frequency offset fusion vector set.

[0024] Specifically, such as Figure 2 , 3 As shown, the spectrum trend division module includes: The data frame receiving submodule acquires the power spectral density sequence in the millimeter-wave frequency domain data frame, extracts the frequency distribution information and corresponding power spectral density sequence within the frequency domain bandwidth, serializes each data frame according to the center frequency arrangement, extracts continuous sampling point data groups within a fixed interval length based on the frequency point index, and generates a frequency band power sampling sequence set. Millimeter-wave frequency domain data frames are snapshots of signal energy distribution within a specific frequency band, such as a data frame covering the 57 GHz to 64 GHz band. From the complete power spectral density sequence of this data frame, a preset frequency domain bandwidth is extracted, for example, specifying the extraction of frequency distribution information and the corresponding power spectral density value for each frequency point within the range of 60.0 GHz to 61.2 GHz. In practice, the system traverses the entire power spectral density sequence, filtering out all data points with frequencies between 60.0 GHz and 61.2 GHz. Next, this extracted data frame is serialized according to the order of its center frequency, meaning all frequency points are arranged in ascending order starting from 60.0 GHz, forming an ordered frequency sequence. Simultaneously, the corresponding power spectral density values ​​are arranged in the same order, ensuring a one-to-one correspondence between frequency and power. Based on the frequency point index, continuous sampling point data groups within a fixed interval length are extracted. This fixed interval length is set based on empirical values ​​of the channel coherence bandwidth, for example, 10 MHz. The basis for setting this length is that statistical analysis of a large amount of millimeter-wave indoor channel measurement data revealed that the channel's frequency response characteristics exhibit high consistency within a 10 MHz bandwidth, and the trend of frequency-selective fading is relatively gentle. The extraction process involves starting at 60.0 GHz and extracting all frequency points and their power values ​​within the range of 60.0 GHz to 60.01 GHz to form the first data set. Then, the starting point is moved by one step, for example, setting the step size to 5 MHz, extracting from 60.005 GHz to 60.015 GHz to form the second data set. This process continues until the window covers the entire frequency domain bandwidth from 60.0 GHz to 61.2 GHz, ultimately generating a frequency band power sampling sequence set composed of multiple consecutive sampling point data sets.

[0025] The power sequence processing submodule calls the frequency band power sampling sequence set, extracts the power value sequence and frequency point sequence in each segment, and performs linear regression fitting with the power value sequence as the dependent variable and the frequency value sequence as the independent variable according to the order of the frequency point sequence. It extracts the absolute value of the regression slope of each frequency segment and assigns a corresponding index to generate a set of segmented fitting slope values. The system calls upon a set of frequency band power sampling sequences and performs independent linear regression analysis on each data group. The first data group is extracted from the sequence set, containing a frequency point sequence and a corresponding power value sequence. For example, a data group might contain a frequency point sequence of [60.000 GHz, 60.001 GHz, ..., 60.010 GHz] and a corresponding power value sequence of [-55.1 dBmW, -55.3 dBmW, ..., -55.8 dBmW]. Based on the inherent order of the frequency point sequences, the power value sequence is defined as the dependent variable, and the frequency value sequence is defined as the independent variable. The specific process of performing linear regression fitting involves calculating the sum of frequency values, the sum of power values, the sum of the products of each pair of frequency and power values, and the sum of the squares of each frequency value for all data points within the current data group. A simplified example with three data points is provided: frequency points [f1, f2, f3] and power values ​​[p1, p2, p3]. The calculation process first obtains the total frequency (f1+f2+f3), the total power (p1+p2+p3), and the total product (f1). p1+f2 p2+f3 p3), and the sum of squared frequencies (f1^2 + f2^2 + f3^2). Then, using these sums, the regression slope for that frequency band is calculated using the standard least squares method: (number of data points multiplied by the sum of their products minus the product of the sum of frequencies and the sum of power) divided by (number of data points multiplied by the sum of squared frequencies minus the square of the sum of frequencies). After calculation, the absolute value of the regression slope is extracted; for example, if the calculated slope is -0.07, its absolute value is extracted as 0.07. Simultaneously, a unique index is associated with this slope value, consistent with the order of the data group in the frequency band power sampling sequence set; for example, the slope value of the first data group is numbered 1. This extraction, calculation, and numbering operation is repeated for all data groups in the frequency band power sampling sequence set, ultimately generating a set of piecewise fitted slope values ​​containing the absolute values ​​of the regression slopes for all frequency bands and their corresponding numbers.

[0026] The trend interval labeling submodule uses the segmented fitted slope value set as a reference, calls the power spectrum slope segment threshold, and judges the slope value of each frequency segment according to the range of the value and the threshold. It sets the interval number and binds the frequency segment sequence number to generate millimeter wave spectrum slope interval label information. The process involves retrieving a set of segmented fitted slope values ​​and applying preset power spectrum slope segmentation thresholds. Each slope value is then categorized into intervals. The power spectrum slope segmentation thresholds are set based on the analysis of a large amount of prior millimeter-wave channel data. Tens of thousands of millimeter-wave signals were collected in standard laboratory environments, open corridor environments, and complex indoor environments filled with office furniture. The aforementioned slope extraction process is performed on these data frames to obtain the slope value distribution for various scenarios. Statistical analysis reveals that in scenarios with flat channels and near-line-of-sight propagation, the absolute slope value is generally below 0.05; in scenarios with some reflection or diffraction, the absolute slope value is concentrated between 0.05 and 0.2; while in complex multipath scenarios with multiple reflections and severe obstruction, the absolute slope value is typically greater than 0.2. Based on the boundaries of this statistical distribution, two threshold levels are set: the first threshold is 0.05, and the second threshold is 0.2. The specific interval judgment process involves reading a slope value from the set of segmented fitted slope values; for example, the slope value corresponding to frequency segment number 5 is 0.12. First, 0.12 is compared with the first threshold of 0.05. Since 0.12 is greater than 0.05, 0.12 is then compared with the second threshold of 0.2. Since 0.12 is less than 0.2, the slope value is determined to be in the interval between 0.05 and 0.2. Based on this determination, an interval number is assigned to this frequency segment. For example, values ​​less than 0.05 are assigned interval number 1 (representing a flat channel), values ​​between 0.05 and 0.2 are assigned interval number 2 (representing a moderately fluctuating channel), and values ​​greater than 0.2 are assigned interval number 3 (representing a highly fluctuating channel). Therefore, the segment with a slope value of 0.12 is assigned interval number 2. Next, this generated interval number 2 is bound to the original sequence number (i.e., 5) of the frequency segment to form a structured label. This determination, numbering, and binding process is repeated for all slope values ​​in the segmented fitted slope value set, ultimately generating millimeter-wave spectral slope interval label information containing trend classification information for all frequency segments.

[0027] Specifically, such as Figure 2 , 4 As shown, the path fading construction module includes: The environmental element extraction submodule extracts the propagation path number corresponding to the millimeter wave spectrum slope interval label information, and collects the corresponding link propagation distance, reflection count, and obstruction density three types of environmental parameter values. It performs unified conversion and unit standardization on the data types of link propagation distance and obstruction density, establishes the number mapping relationship between propagation path and environmental parameters, and generates a link environmental parameter index table. Associating each labeled frequency band with specific physical environment parameters begins with extracting the propagation path number corresponding to each band from the tag information. This number, pre-established during the channel detection phase, uniquely identifies a physical path of a signal from the transmitter to the receiver. Subsequently, based on the extracted propagation path number, three types of environmental parameter values ​​matching that path number are collected from a pre-set environmental parameter database: link propagation distance, number of reflections, and obstruction density. Link propagation distance is a directly measured physical quantity, measured in meters. The number of reflections is an integer value obtained through ray tracing simulation or on-site survey. Obstruction density is a quantified parameter that requires standardized conversion. The quantification standards for obstruction density are as follows: unobstructed air is defined as 0.0; sparse vegetation obstruction is defined as 0.2; ordinary glass curtain walls are defined as 0.4; wooden furniture or plasterboard walls are defined as 0.6; and load-bearing concrete walls or metal cabinets are defined as 0.9. For example, a path numbered P-007 has a propagation distance of 15.5 meters, passes through a glass wall and a large potted plant, and its original environmental parameters are {distance: 15.5 meters, number of reflections: 1, obstructions: [glass wall, potted plant]}. The data type conversion and unit standardization process is as follows: the propagation distance is kept in meters and the value is 15.5. The obstruction density is calculated according to a quantification standard; in this example, it is 0.4 for the glass wall and 0.2 for the potted plant. If the path passes through multiple objects simultaneously, the maximum value is taken, so the obstruction density is standardized to 0.4. After parameter collection and standardization, a strict mapping relationship is established between the propagation path number and the processed environmental parameters. For example, the path number P-007 is mapped to the parameter set {propagation distance: 15.5, number of reflections: 1, obstruction density: 0.4}. This process is repeated for all involved propagation path numbers, ultimately generating a structured link environmental parameter index table.

[0028] The path damping calculation submodule calls the link environment parameter index table to obtain the link propagation distance and obstruction density corresponding to each path number, performs product calculation on the two parameters, and binds the product value with the path number to form a path factor structure. Based on the path factor structure, the path number is mapped and updated to obtain the path loss factor dataset. A comprehensive loss assessment index is calculated for each propagation path. Path information is read sequentially from the index table, obtaining two key environmental parameters corresponding to each path number: link propagation distance and obstruction density. For example, for path number P-007, its link propagation distance is 15.5 meters and obstruction density is 0.4. Next, these two parameters are multiplied, with the link propagation distance multiplied by the obstruction density. In this example, the result is 15.5 multiplied by 0.4, yielding 6.2. This product value is defined as the "path factor" for that path, comprehensively reflecting the combined impact of free space loss and penetration loss on signal strength. After calculation, the obtained path factor value is bound and stored with the corresponding path number, forming a "path factor structure," for example, {path number: P-007, path factor: 6.2}. Then, based on this newly generated path factor structure, the original path numbers are mapped and updated, essentially adding a quantified damping assessment value to each path. This process iterates through all path entries in the link environment parameter index table, performing the same parameter acquisition, product calculation, and binding storage operations for each entry. For example, for another path P-008, with a propagation distance of 8.0 meters and no obstructions (density of 0.0), its path factor is 8.0 multiplied by 0.0, which equals 0.0. After calculating all paths, all path factor structures are aggregated to obtain a path loss factor dataset containing all propagation paths and their corresponding quantized loss factors.

[0029] The multipath segment labeling submodule reads the reflection count value corresponding to each path based on the path loss factor dataset, and compares the reflection count value with the set path complexity benchmark value. The path number with a reflection count greater than the path complexity benchmark value is marked as a multipath propagation segment. The path number is then marked with multipath attributes and archived with damping values ​​to generate millimeter-wave channel path damping segment information. The process of determining the relationship between the reflection count value and the set path complexity benchmark value is as follows: obtain the reflection count value corresponding to each path, use the greater than or equal to judgment logic, and mark the path number with the reflection count value greater than or equal to the path complexity benchmark value as the path number with multipath propagation characteristics. The number of reflections is an integer value representing the number of relay points in a single-path propagation, the path complexity baseline is a fixed integer threshold, and the path complexity baseline is three. Based on the information in the path loss factor dataset, the complexity of the propagation path is classified and labeled. Each path number and its associated complete environmental parameters are read from the dataset, especially the recorded reflection count. The reflection count is an integer-order single-path propagation relay point count, representing the number of specular or non-spectral scattering events the signal experiences before reaching the receiver. Simultaneously, the module invokes a preset path complexity benchmark value. This benchmark value is set to a fixed integer threshold of three. This value is based on the following: extensive channel propagation experiments were conducted in various indoor scenarios, collecting and analyzing channel impulse response data under different reflection counts. Experimental results show that when the reflection count is zero, one, or two, the multipath components of the channel are relatively simple, and the complexity of channel estimation and equalization is low. However, when the reflection count reaches or exceeds three, the multipath effect significantly intensifies, leading to severe inter-symbol interference and signal fading; the deterioration of the channel state exhibits a sharp, inflection-point jump. Therefore, "three" is used as the critical benchmark for distinguishing between simple and complex multipath environments. The specific judgment process is as follows: The reflection count value corresponding to each path is obtained, and a greater than or equal to 3 value is used to compare the reflection count value with the path complexity benchmark value. For example, for path P-007, its reflection count is 1, which is less than 3, so it is not marked. For another path P-015, its reflection count is 3, which is determined to be greater than or equal to 3. Therefore, path number P-015 is marked as a path number with multipath propagation characteristics. For all marked paths, the system will mark its path number with multipath attributes and archive the corresponding damping value (i.e., path loss factor) in its data entry, ultimately generating millimeter-wave channel path damping segmentation information.

[0030] Specifically, such as Figure 2 , 5 As shown, the initial cluster label generation module includes: The feature vector construction submodule calls the millimeter-wave channel path damping segmentation information to obtain the loss factor value, angle of arrival vector value and spectrum trend category label number corresponding to each path segment. It combines the loss factor value as the first dimension component and the angle of arrival vector value as the second dimension component to construct a two-dimensional feature vector group for the path and generate a set of path feature vectors. Each analyzed propagation path is assigned a numerical feature representation. Three core data points are extracted from each path segment: path loss factor, angle of arrival vector, and spectral trend category label. The path loss factor is calculated in previous steps; for example, path P-015 has a loss factor of 12.8. The angle of arrival vector is measured using antenna array signal processing technology and describes the spatial direction of the signal reaching the receiver, typically expressed as a combination of azimuth and elevation angles, such as 35 degrees azimuth and -10 degrees elevation. The spectral trend category label is assigned based on the power spectrum slope; for example, label number 3. These three data points are then combined. Specifically, the loss factor is used as the first dimension, and the angle of arrival vector is used as the second dimension, together constructing a two-dimensional path feature vector. Here, the angle of arrival vector needs to be converted into a scalar or a fixed-dimensional vector before it can be used for construction. One approach is to combine the azimuth and elevation angles. For example, the azimuth (0-360 degrees) and elevation (-90 to 90 degrees) can be normalized to the [0, 1] interval and used as a two-dimensional coordinate [normalized azimuth, normalized elevation], or fused into a scalar using a specific function. For simplicity, the two-dimensional angle of arrival vector value is directly used as the second dimension component. Therefore, the two-dimensional feature vector of path P-015 is constructed as [12.8, (35, -10)]. This process ignores the spectral trend category label number, as it will be used in subsequent filtering steps, not in the vector construction itself. This extraction and combination operation is repeated for each path segment contained in the millimeter-wave channel path damping segmentation information, ultimately generating a path feature vector set consisting of the two-dimensional feature vectors of all paths.

[0031] The label conditional filtering submodule extracts a subset of feature vectors with the same spectral trend category label number based on the path feature vector set. It then groups the feature set according to the label number as the conditional filtering parameter and retains the number index information for the feature vectors in each group to obtain a label mapping vector group set. The system groups vectors based on predefined labels. The process begins by extracting the spectral trend category label number associated with each feature vector in the set, and the filtering operation uses the label number as the sole filtering parameter. Specifically, the system iterates through the entire set of path feature vectors. When processing the first feature vector, the system reads its label number, for example, "2", and creates a subset specifically for storing vectors with the label "2". This feature vector is then stored in this subset. The next feature vector is processed; if its label number is also "2", it is added to the same subset; if its label number is "3", the system creates a new subset for storing vectors with the label "3" and stores that vector. During this process, the original path number index information of each feature vector is fully preserved and stored along with the vector in the corresponding subset. For example, a set may contain five vectors with path numbers and label numbers {P01: 2, P02: 3, P03: 2, P04: 1, P05: 3}. After filtering, three groups will be generated: the first group contains the feature vectors of paths P01 and P03, because their label numbers are both 2; the second group contains the feature vectors of paths P02 and P05, because their label numbers are both 3; and the third group contains only the feature vector of path P04, whose label number is 1. This process continues until all vectors in the path feature vector set have been assigned to their corresponding groups, ultimately resulting in a label mapping vector group set consisting of multiple feature vector subsets.

[0032] The Euclidean clustering labeling submodule obtains the feature vector coordinates of each group according to the label mapping vector group set, calculates the Euclidean distance between each pair of vectors, and judges the relationship between them with the set vector similarity threshold. It aggregates the vector number groups with distance values ​​less than the threshold to construct a set of the same cluster category, forming the initial clustering label set of the signal source. Clustering is performed within each group to identify the similarity of signal sources. The process proceeds sequentially through each group, obtaining all feature vectors within the first group. These vectors are initially categorized based on their shared spectral trend category labels. For any two feature vectors within a group, their coordinates are extracted. For example, vector A has coordinates (x1, y1), and vector B has coordinates (x2, y2). Next, the Euclidean distance between these two vectors is calculated. Specifically, the square of the difference between the first and second dimensions of the two vectors is added, and the square root of the result is taken. For example, if vector A has coordinates (12.8, 0.4) and vector B has coordinates (13.1, 0.5), the Euclidean distance between them is the square root of ((12.8 - 13.1)² + (0.4 - 0.5)²), which is approximately 0.316. The calculated distance value is then compared to a preset vector similarity threshold. The threshold is set based on the following: by performing extensive statistical analysis on the feature vectors of known homologous and heterologous signals, calculating the distribution of their intra-class and inter-class distances, and selecting a value that maximizes the distinction between these two types of distances. For example, experiments show that intra-class distances are usually less than 0.5, while inter-class distances are generally greater than 0.8; therefore, a similarity threshold of 0.5 can be set. In the above example, the calculated distance of 0.316 is less than the threshold of 0.5, therefore vectors A and B are determined to be highly similar. The system aggregates the indices of these two vectors (e.g., P01 and P03) together to construct an initial core of a clustered category set. Then, other vectors within the group are taken, and the distance calculation and comparison with both clustered and unclustered vectors are repeated. All vectors with distances less than the threshold are continuously aggregated until all vectors within the group have been compared. This clustering process is repeated for all groups, ultimately forming an initial cluster label set for the signal source consisting of multiple clusters.

[0033] Specifically, such as Figure 2 , 6 As shown, the multi-solution path filtering module includes: The spectrum feature extraction submodule extracts the frequency domain amplitude response sequence, phase profile point series and mean square error residual data corresponding to the signal segments in the clusters based on the initial clustering label set of the signal source. It performs time axis alignment on the signal segment frame sequence and synchronously numbers and indexes it to establish a set of spectrum response structure of the signal segments under each cluster and generate a clustered signal spectrum feature set. Detailed spectral information is compiled for each clustered signal cluster, extracting a cluster containing a set of path segment numbers considered to originate from the same type of signal. Based on these path segment numbers, three key data types corresponding to each signal segment are retrieved and extracted from the original frequency domain measurement data: frequency domain amplitude response sequence, phase profile point series, and mean square error residual value calculated during the initial fitting stage. The frequency domain amplitude response sequence describes the energy distribution of the signal within the frequency band, while the phase profile point series records the phase information of the signal at each frequency point. Subsequently, time axis alignment is performed on the frame sequences of all signal segments within the same cluster. Since signals from different paths may be captured at different time frames, this step calibrates the start time of all signal segments to the same reference time point by matching the timestamps or synchronization sequences in the frame headers. During alignment, the original number index is synchronized for each aligned data frame to ensure the continuity and traceability of the data in the time domain. After alignment and index synchronization are completed, the frequency domain amplitude response sequence, phase profile point series, and mean square error residual value data of all signal segments belonging to the same cluster are structured together. This organizational method establishes a clear correspondence between a cluster (representing a class of signal sources) and the detailed spectral characteristics of all its subordinate signal segments. The above extraction, alignment, indexing, and organization process is repeated for each cluster in the initial cluster label set of signal sources, ultimately generating a structured set containing the spectral response characteristics of the signal segments under each cluster, i.e., the clustered signal spectral feature set.

[0034] The error index calculation submodule calls the clustered signal spectrum feature set, extracts the frequency domain amplitude response value and phase profile point pair in each signal segment, calculates the frequency deviation value and phase difference value between each pair of points synchronously, and performs normalization processing. The residual value and frequency response deviation value in each signal segment are numerically superimposed to obtain the multipath error evaluation value group. The quality of each signal segment is evaluated in detail. A specific signal segment is extracted from the feature set, and its internal frequency domain amplitude response value sequence and phase profile point pair sequence are obtained. For each phase profile point pair, the system simultaneously calculates the frequency deviation and phase difference between each pair of adjacent points. For example, for two adjacent phase points (frequency f1, phase p1) and (frequency f2, phase p2), the frequency deviation is f2 minus f1, and the phase difference is p2 minus p1. After calculating the differences for all point pairs, these difference sequences are normalized. The normalization method is to divide each difference by the maximum absolute value of its sequence, so that all deviation values ​​fall within the range of -1 to 1. Next, for each signal segment, the mean square error residual value generated during the fitting process is numerically superimposed with the previously calculated normalized frequency response deviation value (which can be some combination of frequency deviation and phase difference, such as a weighted sum). This superposition operation combines the residual value representing the amplitude fitting goodness of fit and the deviation value representing phase stability to form a comprehensive error index. For example, if the residual value of a signal segment is 0.02 and its normalized frequency response deviation is 0.05, then the superimposed multipath error evaluation value is 0.07. This calculation process is repeated for each signal segment in the clustered signal spectral feature set, ultimately yielding a multipath error evaluation value set containing the evaluation values ​​corresponding to all signal segments.

[0035] The path number filtering submodule performs interval judgment on the evaluation value group of each path segment and the set mean square error residual tolerance threshold and frequency response error tolerance threshold, based on the multi-path error evaluation value group. It filters out the path segment numbers that exceed both thresholds at the same time, retains the path numbers that do not exceed the thresholds, and performs number classification and sorting to obtain the blind source candidate reconstruction path number set. The mean square error residual tolerance threshold is set by identifying the upper limit of the deviation range after statistical normalization of the residual distribution corresponding to each sampling point in the frequency domain amplitude response sequence of the signal segment. The frequency response error tolerance threshold is set based on the maximum allowable offset range of the frequency drift amplitude change rate of the phase profile point column in adjacent frames. The evaluation values ​​for each path segment are compared with two independent tolerance thresholds. The first is the mean square error residual tolerance threshold. This threshold is set by statistically analyzing the residual distribution corresponding to each sampling point in the frequency domain amplitude response sequence of a large number of signal segments, normalizing these residual values, and identifying the upper limit of the deviation range of the distribution. For example, the 95th percentile of the statistical distribution is taken as the threshold, set to 0.08. The second is the frequency response error tolerance threshold. This threshold is set by analyzing the rate of change of frequency drift amplitude of the phase profile points in the stable channel within adjacent frames, setting a maximum allowable offset range. For example, if the standard deviation of the normal drift rate is 0.02, the threshold is set to three times its standard deviation, i.e., 0.06. The specific screening process is as follows: for each path segment, it is checked whether its mean square error residual value exceeds the mean square error residual tolerance threshold and whether its frequency response error value exceeds the frequency response error tolerance threshold. Only when both error values ​​of a path segment exceed their respective thresholds will the path segment be screened out. For example, if path A has a residual of 0.09 (exceeding 0.08) and a frequency response error of 0.07 (exceeding 0.06), then path A is eliminated. Path B has a residual of 0.10 (exceeding 0.08) but a frequency response error of 0.05 (not exceeding 0.06), so path B is retained. Path C also retains because both errors do not exceed the threshold. After performing this dual conditional judgment on the path segments in all evaluation value groups, all path numbers that were not eliminated are retained. These retained numbers are then categorized and sorted to ultimately obtain a set of blind source candidate reconstruction path numbers that can be used for subsequent reconstruction.

[0036] Specifically, such as Figure 2 , 7 As shown, the fusion refactoring execution module includes: The amplitude data correction submodule calls the blind source candidate reconstruction path number set, reads the time domain amplitude estimation value sequence under the corresponding path, performs time axis resampling on each amplitude value according to the frame index order, performs equal-interval interpolation on the sampling point spacing and generates a resampled amplitude data structure of uniform length to obtain the path amplitude resampling information. The time-domain amplitude data corresponding to the candidate paths is preprocessed. Based on a number in the path number set, the estimated time-domain amplitude value sequence associated with that path is read. This sequence is arranged in the order of the time frames at the time of acquisition. Since the signal acquisition process of different paths may have inconsistent sampling rates or non-uniform distribution of sampling points, time-axis resampling is required for each amplitude value sequence. The resampling operation is based on a unified, high-density standard time axis. Specifically, the sampling point spacing in the original amplitude value sequence is interpolated at equal intervals. For example, using linear interpolation, multiple new, equally spaced sampling points are linearly calculated between two original sampling points based on their time interval and amplitude difference. Let the amplitude of the original sequence at time t1 be A1 and at time t2 be A2. If the standard time axis requires inserting a point t_new between t1 and t2, then the amplitude A_new of the new point is calculated as A1 + (A2 - A1). (t_new-t1) / (t2-t1). This method converts the amplitude sequences of all candidate paths into sequences with the same number of sampling points and a uniform sampling interval, generating a resampled amplitude data structure of uniform length. This reading, resampling, and interpolation process is repeated for all path numbers in the blind source candidate reconstruction path number set, ultimately yielding a path amplitude resampled information containing the normalized amplitude information of all candidate paths.

[0037] The phase data mapping submodule obtains the minimum mean square phase estimate sequence corresponding to the path number based on the path amplitude resampling information, performs frame-level index mapping processing on the sequence with the resampling time axis, maps the phase estimate of each frame to the amplitude value at the same frame position, performs point-to-point superposition, and generates a phase amplitude fusion frame sequence group. The phase data is precisely aligned with the calibrated amplitude data. Based on the path number contained in the path amplitude resampling information, a sequence of minimum mean square phase estimates corresponding to that path number is obtained. This phase sequence, obtained during the channel estimation stage, represents the optimal estimate of the signal phase at each time frame. Next, this phase sequence undergoes frame-level index mapping with the unified time axis used by the resampled amplitude data. Specifically, each phase estimate in the phase sequence and its corresponding time frame index are read, and then the resampled amplitude value with the same time frame index in the path amplitude resampling information is found. The phase estimate of that frame is precisely mapped onto the amplitude value at the same frame position, performing point-to-point overlay. Conceptually, this operation is equivalent to attaching a simultaneous phase value to the amplitude value at each time point, thereby merging the originally independent amplitude and phase sequences into a complex sequence or a sequence containing (amplitude, phase) pairs. For example, in time frame 100, the resampled amplitude value is 0.8, while the corresponding phase estimate is 1.2 radians. These are then superimposed into a single data point (0.8, 1.2). This mapping and superposition operation is performed on every frame of data for all candidate paths, ultimately combining the data from all paths into a single phase-amplitude fused frame sequence.

[0038] The frequency offset superposition and reconstruction submodule is based on the phase amplitude fusion frame sequence group. It extracts the frequency offset vector value corresponding to each path segment and performs zero mean normalization. The frequency offset vector is weighted and superimposed according to the path number order. The fusion frame sequence and the frequency offset superposition vector are spliced ​​together in a one-dimensional structure to form a continuous reconstruction signal and construct millimeter wave blind source separation and reconstruction waveform information. The weighted average superposition process is as follows: the standardized vectors are weighted according to the path bandwidth ratio in the order of path number, and the weighted vectors are summed by component level to obtain the fused frequency offset superposition vector. The process of performing one-dimensional structure splicing of the fused frame sequence and the frequency offset superposition vector is as follows: according to the frame index, the frequency offset superposition vector component under the corresponding frame index is added to the end of the phase amplitude fused frame sequence, and the one-dimensional continuous output format of the data structure is maintained. From the relevant channel parameters, the frequency offset vector value corresponding to each candidate path segment is extracted. This frequency offset vector describes the offset of the signal carrier frequency caused by factors such as the Doppler effect. Each extracted frequency offset vector is then normalized to zero mean; that is, the average value of all elements in the vector is calculated, and then this average value is subtracted from each element, resulting in a vector with a mean of zero. This step eliminates the DC component deviation of the frequency offset of each path. Next, all normalized frequency offset vectors are weighted and superimposed according to their path numbers. The normalized vectors are weighted according to their path bandwidth proportions to highlight the frequency offset characteristics of the dominant signal. For example, if path 1 has a weight of 0.7 and a frequency offset vector of V1, and path 2 has a weight of 0.3 and a frequency offset vector of V2, then the superimposed frequency offset vector is 0.7. V1+0.3 V2. Then, the phase-amplitude fused frame sequence is concatenated with the final frequency offset superposition vector using a one-dimensional structure. The concatenation operation combines the fused frame sequence (a complex sequence that varies with time) representing the main information of the signal with the frequency offset vector representing the micro-dynamic changes in frequency, forming a more complete continuous reconstructed signal. This final signal structure not only contains the amplitude and phase at each moment but also the overall frequency drift trend, thus constructing the final millimeter-wave blind source separation and reconstruction waveform information.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A system for blind source separation and reconstruction of millimeter wave signals in complex electromagnetic environment, characterized in that, The system comprises: The spectrum trend division module obtains the power spectrum density sequence in the millimeter wave frequency domain data frame, calls the frequency division segment value, the sampling point sequence in the bandwidth, performs trend identification on the frequency segment, and generates millimeter wave spectrum slope interval label information; The path fading construction module calculates the product of the link propagation distance and the obstacle density based on the millimeter wave spectrum slope interval label information, obtains the segment path loss factor, and establishes millimeter wave channel path damping segmented information for the segment with a reflection number greater than a path complexity reference value; The initial cluster label generation module calls the millimeter wave channel path damping segmented information, forms a two-dimensional feature vector group with the angle of arrival vector and the loss factor, aggregates the feature groups with a distance less than a vector similarity threshold value as the same initial cluster, and forms a signal source initial clustering label set; The multi-solution path screening module calculates the frequency response deviation between matching points according to the signal source initial clustering label set, compares the mean square error residual value with a set mean square error residual tolerance threshold value, screens out the paths with deviation values and residual values exceeding the threshold value, and obtains a blind source candidate reconstruction path number set; The fusion reconstruction execution module calls the blind source candidate reconstruction path number set, resamples and adjusts the time domain amplitude estimation value, maps the minimum mean square phase estimation value to the frame level time axis after the amplitude value is superimposed, normalizes and weights the frequency offset vector value in each path to form a signal reconstruction set, and constructs and generates millimeter wave blind source separation reconstruction waveform information.

2. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The millimeter wave spectrum slope interval label information includes frequency domain partition number, fitting slope level identification and power spectrum trend type, the millimeter wave channel path damping segmented information includes propagation segment path number, path damping factor grouping label and multipath propagation paragraph mark, the signal source initial clustering label set includes signal source clustering number, feature vector aggregation number and initial cluster center vector identification, the blind source candidate reconstruction path number set includes reconstruction path sequence number, path amplitude and phase error label and path screening mark identification, and the millimeter wave blind source separation reconstruction waveform information includes waveform amplitude sequence set, phase recovery function set and frequency offset fusion vector set.

3. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The spectrum trend division module comprises: The data frame receiving submodule obtains the power spectrum density sequence in the millimeter wave frequency domain data frame, extracts the frequency distribution information and the corresponding power spectrum density sequence in the frequency domain bandwidth range, serializes each data frame according to the center frequency arrangement order, extracts the continuous sampling point data group in the fixed interval length according to the frequency point index, and generates a frequency segment power sampling sequence set; The power sequence processing submodule calls the frequency segment power sampling sequence set, extracts the power value sequence and the frequency point sequence in each segment, performs linear regression fitting with the power value sequence as the dependent variable and the frequency value sequence as the independent variable according to the order of the frequency point sequence, extracts the regression slope absolute value of each frequency segment and corresponds to the index, and generates a segmented fitting slope value set; The trend interval labeling submodule calls a power spectrum slope segment threshold according to the set of segment fitting slope values, performs interval judgment on the slope value of each frequency segment according to the range in which the value is located and the threshold, sets interval numbers and binds frequency segment serial numbers, and generates millimeter wave spectrum slope interval label information.

4. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The path fading construction module comprises: The environmental element extraction submodule extracts the interval corresponding propagation path number based on the millimeter wave spectrum slope interval label information, collects three types of environmental parameter values of corresponding link propagation distance, reflection times and obstacle density, uniformly converts and unit standardizes the data types of the link propagation distance and the obstacle density, establishes a number mapping relationship between the propagation path and the environmental parameters, and generates a link environmental parameter index table; The path damping calculation submodule calls the link environmental parameter index table, obtains the link propagation distance and the obstacle density corresponding to each path number, performs product calculation on the two parameters, and stores the product value and the path number to form a path factor structure, updates the path number according to the path factor structure, and obtains a path loss factor data set; The multipath paragraph labeling submodule reads the reflection times value corresponding to each path according to the path loss factor data set, judges the size relationship between the reflection times value and the set path complexity reference value, marks the path number whose reflection times is greater than the path complexity reference value as a multipath propagation paragraph, marks the path number with the multipath attribute and archives the damping value, and generates millimeter wave channel path damping segment information.

5. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 4, wherein, The process of judging the size relationship between the reflection times value and the set path complexity reference value specifically comprises: obtaining the reflection times value corresponding to each path, using greater than or equal judgment logic, and marking the path number whose reflection times value is greater than or equal to the path complexity reference value as a path number with multipath propagation characteristics; The reflection times value is an integer order single-path propagation relay point count value, the path complexity reference value is a fixed integer threshold, and the path complexity reference value is three.

6. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The initial cluster label generation module comprises: The feature vector construction submodule calls the millimeter wave channel path damping segment information, obtains the loss factor value, the angle of arrival vector value and the spectrum trend category label number corresponding to each path segment, combines the loss factor value as the first dimension component and the angle of arrival vector value as the second dimension component to construct a path two-dimensional feature vector group, and generates a path feature vector set; The label condition filtering submodule extracts a feature vector subset with the same spectrum trend category label number based on the path feature vector set, groups the feature set according to the label number as a condition filtering parameter, retains the index information of the feature vectors in each group, and obtains a label mapping vector grouping set; The Euclidean clustering labeling submodule obtains the feature vector coordinate values in each group in turn according to the label mapping vector grouping set, calculates the Euclidean distance values between two vectors, judges the size relationship between the distance values and a set vector similarity threshold, aggregates the vector number groups with distance values less than the threshold to construct a same cluster category set, and forms an initial clustering label set of the signal source.

7. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The multi-solution path screening module comprises: The spectrum feature extraction submodule extracts the frequency domain amplitude response sequence, the phase profile point column and the mean square error residual value data corresponding to the signal segment in the clustering cluster according to the signal source initial clustering label set, performs time axis alignment on the signal segment frame sequence, synchronously indexes, establishes the spectrum response structure set of the signal segment under each clustering cluster, and generates the clustering signal spectrum feature set; The error index calculation submodule calls the clustering signal spectrum feature set, extracts the frequency domain amplitude response value and the phase profile point pair in each signal segment, synchronously calculates the frequency deviation value and the phase difference value between each pair of points, performs normalization processing, and superimposes the residual value and the frequency response deviation value in each signal segment to obtain a multi-path error evaluation value group; The path number screening submodule performs interval judgment on the evaluation value group of each path segment and the set mean square error residual tolerance threshold and the frequency response error tolerance threshold respectively according to the multi-path error evaluation value group, screens out the path segment numbers that simultaneously exceed the two threshold values, retains the path numbers that do not exceed the threshold values and performs number classification sorting to obtain a blind source candidate reconstruction path number set.

8. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 7, wherein, The setting mode of the mean square error residual tolerance threshold is to identify the upper limit value of the deviation range based on the statistical normalization of the residual distribution corresponding to each sampling point in the frequency domain amplitude response sequence of the signal segment; The setting mode of the frequency response error tolerance threshold is to set the maximum allowed deviation range based on the frequency drift amplitude change rate of the phase profile point column in adjacent frames.

9. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 1, wherein, The fusion reconstruction execution module comprises: The amplitude data correction submodule calls the blind source candidate reconstruction path number set, reads the time domain amplitude estimation value sequence under the corresponding path, performs time axis resampling on each amplitude value in the order of frame index, performs equidistant interpolation processing on the sampling point interval and generates a uniform length resampled amplitude data structure to obtain path amplitude resampling information; The phase data mapping submodule obtains the minimum mean square phase estimation value sequence corresponding to the path number according to the path amplitude resampling information, performs frame level index mapping processing on the sequence with the resampled time axis, maps the phase estimation value of each frame to the amplitude value at the same frame position, implements point superposition, generates a phase amplitude fusion frame sequence group, and performs one-dimensional structure splicing on the fusion frame sequence and the frequency offset superposition vector to form a continuous reconstruction signal and construct millimeter wave blind source separation reconstruction waveform information. The process of the weighted average superposition is specifically that the standardized vectors are distributed with weights according to the path bandwidth proportion in the order of the path number, and the weighted vectors are summed at the component level to obtain the superimposed frequency offset vector; 10. The millimeter wave signal blind source separation and reconstruction system for complex electromagnetic environment according to claim 9, wherein, The process of performing one-dimensional structure splicing on the fusion frame sequence and the frequency offset superposition vector is specifically that the frame end of the phase amplitude fusion frame sequence is added with the frequency offset superposition vector component at the corresponding frame index according to the frame index, and the data structure is kept in one-dimensional continuous output format. ​

Images