Railway radio interference analysis method based on machine learning

Abstract

The invention discloses a railway radio interference analysis method based on machine learning. The method comprises the following steps: step 1, acquiring radio signal data in a railway communication scene; 2, performing signal preprocessing on the radio signal data; 3, extracting the interference structure characteristics of railway radio through the improved RepLKNet network, and generating an interference structure characteristic pattern; 4, constructing a training sample set based on the interference structure feature map; 5, inputting the training sample set into the improved TimesNet network to execute supervised training, and obtaining a railway radio interference analysis model; step 6, loading the railway radio interference analysis model, performing interference analysis on the radio signal data to be analyzed, and outputting an interference analysis result; and 7, outputting an interference analysis result. According to the invention, through the improved RepLKNet network and the improved Times Net network, the accuracy and stability of railway radio interference analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway wireless communication and intelligent signal analysis, and particularly relates to a railway radio interference analysis method based on machine learning. BACKGROUND

[0002] With the wide deployment of railway wireless communication systems in train dispatching, operation control and emergency command, the complexity of the radio environment along the railway has significantly increased, and radio interference analysis and interference situation recognition in the railway scenario have become one of the key technologies to ensure communication safety and operation reliability. The existing interference analysis methods mainly rely on fixed frequency band power detection, interference source signal template comparison or statistical discrimination based on empirical threshold, and still have the following problems in actual application:

[0003] The collected railway radio signals are significantly affected by factors such as multipath propagation, electrical equipment radiation, and train operation state changes, and have complex time-frequency structures, so it is difficult for traditional feature extraction methods to accurately capture the periodicity and disturbance characteristics of the interference signals; the existing modeling methods generally rely on artificial feature construction or static threshold judgment, lack the ability to model the evolution trend of non-steady-state interference, and are difficult to effectively deal with the mixed distribution of sudden interference and periodic interference; in the sample construction and time series analysis process, the modulation effect of the railway operation state is not fully considered, resulting in poor robustness or high false alarm rate of the model under complex operation conditions, and it is difficult to support the practical needs of interference evolution visualization, grade determination and risk warning of the actual railway communication monitoring system.

[0004] Therefore, how to provide a railway radio interference analysis method based on machine learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a railway radio interference analysis method based on machine learning, which fully combines time-frequency domain signal processing and deep neural network modeling, and describes in detail the extraction of interference structure features by an improved RepLKNet network and the periodicity and disturbance modeling by an improved TimesNet network. The present application has the advantages of high interference recognition accuracy, good adaptability to non-steady-state interference, and strong robustness in engineering application.

[0006] According to the railway radio interference analysis method based on machine learning of the present application, the following steps are included:

[0007] Step 1: Obtain radio signal data in a railway communication scenario;

[0008] Step 2: Perform signal preprocessing on the radio signal data to generate a time-frequency spectrum feature tensor;

[0009] Step three: input the time-frequency spectrum feature tensor into the improved RepLKNet network to extract the interference structure features of the railway radio and generate an interference structure feature map; the improved RepLKNet network comprises a time-frequency tensor encoding module, a multi-scale large kernel convolution module, a spectral prior channel modulation module, and an interference feature aggregation module;

[0010] Step four: based on the interference structure feature map, construct a railway interference feature sample to form a training sample set;

[0011] Step five: input the training sample set into the improved TimesNet network to perform supervised training and obtain a railway radio interference analysis model; the improved TimesNet network comprises a candidate period extraction module, a period selection gate module, a steady-state period modeling branch, a non-steady-state disturbance modeling branch, and a running state modulation module;

[0012] Step six: load the railway radio interference analysis model to perform interference analysis on the to-be-analyzed radio signal data and output an interference analysis result;

[0013] Step seven: output the interference analysis result to the railway communication monitoring system.

[0014] Optionally, the step one specifically comprises: collecting radio signal data from railway lines, station areas, or train running sections, wherein the radio signal data comprises a railway wireless communication signal sequence and a radio interference signal sequence superimposed on the railway wireless communication signal.

[0015] Optionally, the step two specifically comprises:

[0016] The signal preprocessing comprises denoising, normalization, time segmentation, and time-frequency transformation;

[0017] The radio signal data is denoised by a band-pass filter to generate a filtered signal sequence;

[0018] The maximum value normalization method is used to map the signal amplitude in the filtered signal sequence to the [0, 1] interval to obtain a normalized signal sequence;

[0019] The normalized signal sequence is time segmented by setting a sliding window length and a sliding step to obtain a plurality of sub-signal sequence blocks;

[0020] The one-dimensional time domain signal of each sub-signal sequence block is mapped to a two-dimensional time-frequency domain signal by using a short-time Fourier transform to obtain a time-frequency spectrum feature matrix; the time-frequency spectrum feature matrix structure is time step dimension x frequency resolution dimension; the time step dimension is the sliding window length;

[0021] Stacking time-frequency spectrum tensors corresponding to all sub-signal sequence blocks, a time-frequency spectrum feature tensor is obtained; the time-frequency spectrum feature tensor has a structure of sub-signal sequence block dimension x time step dimension x frequency resolution dimension.

[0022] Optionally, the step three specifically includes:

[0023] The time-frequency tensor encoding module inputs the time-frequency spectrum feature tensor into a 1x1 convolution layer, maps channels for the frequency resolution dimension, and performs nonlinear feature transformation through a GELU activation function to obtain an initial mapping feature tensor;

[0024] The multi-scale large kernel convolution module includes three parallel large kernel convolution branches, each of which includes a two-dimensional convolution, batch normalization, ReLU activation and residual connection operation, specifically:

[0025] The initial mapping feature tensor is extracted for interference structure features using two-dimensional convolution with a convolution kernel size of 15x15, 31x31 and 51x51 respectively, and is normalized and nonlinearly transformed through batch normalization and ReLU activation to obtain a first scale convolution feature tensor, a second scale convolution feature tensor and a third scale convolution feature tensor respectively;

[0026] The first scale convolution feature tensor, the second scale convolution feature tensor and the third scale convolution feature tensor are respectively connected in residual connection with the initial mapping feature tensor to obtain a first scale structure feature tensor, a second scale structure feature tensor and a third scale structure feature tensor;

[0027] The first scale structure feature tensor, the second scale structure feature tensor and the third scale structure feature tensor are channel spliced to generate a fusion structure feature tensor;

[0028] In the spectral prior channel modulation module, the fusion structure feature tensor is subjected to two-dimensional average pooling operation in the sub-signal sequence block dimension and the time step dimension to obtain a channel energy mean vector;

[0029] Based on the channel energy mean vector, the maximum energy value and the minimum energy value of each channel are calculated, and each channel component of the channel energy mean vector is linearly normalized according to the maximum energy value and the minimum energy value to obtain a normalized channel energy vector;

[0030] A channel scaling coefficient is introduced to scale the normalized channel energy vector channel by channel to obtain a channel modulation weight vector, and the channel modulation weight vector is multiplied with the fusion structure feature tensor in the channel dimension element by element to obtain a modulated structure feature tensor;

[0031] The interference feature aggregation module performs average pooling on the modulation structure feature tensor in the time step dimension to generate an interference structure feature map, which has a structure of a sub-signal sequence block dimension x a channel dimension.

[0032] Optionally, the step four specifically includes:

[0033] The interference structure feature map is divided into a plurality of interference structure feature vectors according to the sub-signal sequence block dimension, and a starting timestamp of each interference structure feature vector corresponding to a sub-signal sequence block is obtained;

[0034] An interference judgment energy threshold is set, and an average energy value of a time-frequency spectrum feature matrix corresponding to each sub-signal sequence block is calculated;

[0035] If the average energy value is less than the interference judgment energy threshold, the interference state label of the corresponding sub-signal sequence block is set to 0: no interference is detected, otherwise, it is set to 1: there is railway radio interference;

[0036] The interference state label is used as a target label for supervised training;

[0037] The interference structure feature vectors are time-sequenced according to the starting timestamp, and the interference structure feature vectors and the interference state label are combined into an interference feature sample; each starting timestamp corresponds to a time step;

[0038] All interference feature samples are combined into a training sample set.

[0039] Optionally, the step five specifically includes:

[0040] Based on the training sample set, an interference structure feature vector sequence and an interference state label sequence are obtained;

[0041] The interference structure feature vector sequence is generated into a candidate period set through a candidate period extraction module;

[0042] In a period selection gating module, a matching score between each interference structure feature vector and each main period in the candidate period set is calculated, a matching score vector is obtained, and the matching score vector is subjected to Softmax normalization to obtain a period selection gating vector;

[0043] In a steady-state period modeling branch, the interference structure feature vector sequence is reconstructed into a period feature tensor according to the length of each main period in the candidate period set;

[0044] The period feature tensor is subjected to two-dimensional convolution to extract a stable period feature, a period modeling feature tensor is obtained, the period modeling feature tensor is reshaped to be consistent with the structure of the interference structure feature vector sequence, a period modeling feature vector sequence is obtained, and the period modeling feature vector sequence is consistent with the number of main periods in the candidate period set.

[0045] At the time step, based on the cycle selection gating vector, the cycle feature vectors of each main cycle in the candidate cycle set are weighted and fused to obtain a steady-state cycle feature vector, and the steady-state cycle feature vectors are grouped into a steady-state cycle feature vector sequence according to the time step;

[0046] In the non-steady-state disturbance modeling branch, the sliding window length is set, the sliding step is set to 1, the disturbance structure feature vector sequence is segmented by sliding to obtain a plurality of disturbance subsequences; the number of disturbance subsequences is consistent with the number of time steps, and when the sliding window length is greater than the number of remaining time steps, a forward padding method is used to fill in to generate the disturbance subsequence;

[0047] Each disturbance subsequence is extracted by one-dimensional convolution and ReLU activation to extract local mutation features to obtain a disturbance modeling feature vector;

[0048] All disturbance modeling feature vectors are spliced according to the time step sequence to obtain a disturbance modeling feature vector sequence;

[0049] The steady-state cycle feature vector sequence and the disturbance modeling feature vector sequence are spliced in the channel dimension to form a joint feature vector sequence;

[0050] In the running state modulation module, the train speed value sequence is obtained, and the train speed value at each time step is mapped by linear mapping to generate a running state modulation vector;

[0051] The running state modulation vector and the joint feature vector at the corresponding time step are executed by channel-by-channel modulation to obtain a running state feature vector, and the running state feature vector sequence is formed according to the time step;

[0052] Each running state feature vector is mapped by full connection to generate an interference state logic value, and the interference state logic value is mapped to the interval [0, 1] through the Sigmoid activation function to obtain an interference state prediction probability, and the interference state prediction probability sequence is formed according to the time step;

[0053] The binary cross entropy between the interference state prediction probability sequence and the interference state label sequence is calculated to obtain a supervised training loss function, and the improved TimesNet network is updated based on the supervised training loss function. The improved TimesNet network trained is used as a railway radio interference analysis model.

[0054] Optionally, the interference structure feature vector sequence is generated by the candidate cycle extraction module to form a candidate cycle set, which specifically includes:

[0055] The channel mean of each interference structure feature vector is calculated, and the interference feature sequence is formed according to the time step sequence;

[0056] mapping the interference feature sequence to a frequency domain based on a fast Fourier transform to obtain a complex frequency domain feature sequence;

[0057] calculating an amplitude spectrum of the complex frequency domain feature sequence, sorting the amplitude spectrum in descending order according to amplitudes, and selecting the first K frequencies with the largest amplitudes as main frequencies;

[0058] calculating a ratio of a sampling frequency to each main frequency to obtain K main periods;

[0059] setting a minimum period and a maximum period, removing duplicate K main periods, and retaining main periods that meet the minimum period and the maximum period to obtain a candidate period set.

[0060] Optionally, the step six specifically comprises:

[0061] The interference analysis result includes an interference evolution diagram, an interference judgment state and an interference risk level.

[0062] obtaining an interference structure feature vector sequence of the to-be-analyzed radio signal data, generating a running state feature vector sequence through a railway radio interference analysis model based on the interference structure feature vector sequence;

[0063] mapping each time-step running state feature vector to generate an interference state logic value through full connection, mapping the interference state logic value to an interference state prediction probability through a Sigmoid activation function, combining interference state prediction probabilities of each time step to form an interference state prediction probability sequence, and drawing the interference state prediction probability sequence into an interference evolution diagram with time steps as the horizontal axis and the interference state prediction probability as the vertical axis;

[0064] setting an interference judgment threshold, if the interference state prediction probability is greater than or equal to the interference judgment threshold, marking the interference judgment state of the corresponding time step as existing railway radio interference, otherwise marking it as not detecting railway radio interference;

[0065] setting interference level thresholds T1 and T2, and the interference level threshold T1 is greater than or equal to the interference judgment threshold; for the time step with railway radio interference, the interference risk level is determined based on the interference state prediction probability, specifically:

[0066] if the interference state prediction probability is less than the interference level threshold T1, the interference risk level is 0, indicating low risk; if the interference state prediction probability is less than the interference level threshold T2 and greater than or equal to the interference level threshold T1, the interference risk level is 1, indicating medium risk; if the interference state prediction probability is greater than or equal to the interference level threshold T2, the interference risk level is 2, indicating high risk.

[0067] The beneficial effects of the present application are:

[0068] The application provides a railway radio interference analysis method based on machine learning, which can effectively solve the problems of low interference feature extraction precision, weak time sequence modeling capability and lack of dynamic perception capability in interference judgment in the prior art. First, an improved RepLKNet network including a time-frequency tensor encoding module, a multi-scale large kernel convolution module, a spectrum prior channel modulation module and an interference feature aggregation module is constructed to extract multi-scale interference structure features for the complex time-frequency structure of railway radio signals, and the feature expression capability of the key frequency band enhanced by the spectrum energy is introduced to improve the feature precision of the interference structure feature map.

[0069] Secondly, the improved TimesNet network is used to deeply model the time sequence changes of railway radio interference, and through a candidate period extraction module, a period selection gating module, a steady-state period modeling branch, a non-steady-state disturbance modeling branch and an operating state modulation module, the feature modeling needs of periodic interference and burst interference are considered, and the train operating state information is introduced to dynamically modulate the time sequence features, thereby enhancing the adaptability of the improved TimesNet network to interference evolution patterns under different working conditions. The improved TimesNet network is supervised and trained through a supervision training module, and the railway radio interference analysis model obtained has strong robustness and time sequence reasoning capability.

[0070] In addition, by loading the railway radio interference analysis model, online inference can be performed on the radio signal data to be analyzed, and multi-dimensional interference analysis results including interference evolution graph, interference judgment state and interference risk level can be output, effectively supporting the application requirements of the railway communication monitoring system in interference early warning, risk identification and operation scheduling.

[0071] To sum up, the improved RepLKNet network and the improved TimesNet network solve the deficiencies of traditional methods in multi-source interference, non-steady-state evolution and dynamic working condition modeling, and realize high-precision identification and risk perception of railway radio interference. BRIEF DESCRIPTION OF DRAWINGS

[0072] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application are used to explain the application, and do not constitute a limitation on the application. In the drawings:

[0073] Figure 1 is a schematic diagram of a railway radio interference analysis method based on machine learning proposed by the application;

[0074] Figure 2 is a flow chart of an improved RepLKNet network structure in a railway radio interference analysis method based on machine learning proposed by the application;

[0075] Figure 3 is an improved TimesNet modeling flowchart in a machine learning-based railway radio interference analysis method. DETAILED DESCRIPTION

[0076] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show the basic structure of the application in a schematic manner only, and therefore only show those components which are relevant to the application.

[0077] REFERENCE Figures 1-3 A machine learning-based railway radio interference analysis method, comprising the following steps:

[0078] Step one: obtaining radio signal data under a railway communication scenario;

[0079] Step two: signal preprocessing of the radio signal data to generate a time-frequency spectrum feature tensor;

[0080] Step three: inputting the time-frequency spectrum feature tensor into an improved RepLKNet network to extract interference structure features of railway radio and generate an interference structure feature map; the improved RepLKNet network comprises a time-frequency tensor encoding module, a multi-scale large kernel convolution module, a spectrum prior channel modulation module, and an interference feature aggregation module;

[0081] Step four: based on the interference structure feature map, constructing railway interference feature samples to form a training sample set;

[0082] Step five: inputting the training sample set into an improved TimesNet network to perform supervised training and obtain a railway radio interference analysis model; the improved TimesNet network comprises a candidate period extraction module, a period selection gate module, a steady-state period modeling branch, a non-steady-state disturbance modeling branch, and a running state modulation module;

[0083] Step six: loading the railway radio interference analysis model to perform interference analysis on the radio signal data to be analyzed and output the interference analysis result;

[0084] Step seven: outputting the interference analysis result to a railway communication monitoring system.

[0085] In the present application, based on the dual complexity of the structural characteristics and the time evolution characteristics of the railway radio interference signal, the improved RepLKNet network and the improved TimesNet network are selected, and the combination of the two can realize comprehensive modeling from spatial structure to time sequence dynamics, significantly improving the accuracy and practicality of railway radio interference analysis. Among them, the improved RepLKNet network can effectively capture interference patterns at different spatial scales through a multi-scale large kernel convolution structure, and at the same time, combined with the spectrum prior channel modulation mechanism, the response capability to the key frequency area is enhanced, so that the interference structure characteristics can be more accurately extracted. The improved TimesNet network is designed for the time sequence characteristics of the interference signal, which has periodicity and burstiness, and a dual-branch structure of steady-state cycle modeling and non-steady-state disturbance modeling is designed, and a running state modulation mechanism is introduced, so that the model can dynamically adapt to the change of train operation conditions, and the modeling ability and risk perception level of the interference evolution trend are improved.

[0086] In the present embodiment, the step one specifically comprises: collecting radio signal data from railway lines, station areas or train operation sections, wherein the radio signal data includes a railway wireless communication signal sequence and a radio interference signal sequence superimposed on the railway wireless communication signal.

[0087] In the present embodiment, the step two specifically comprises:

[0088] The signal preprocessing includes denoising, normalization, time segmentation and time-frequency transformation;

[0089] The radio signal data is denoised by a band-pass filter to generate a filtered signal sequence;

[0090] The maximum normalization method is used to map the signal amplitude in the filtered signal sequence to the interval [0, 1] to obtain a normalized signal sequence; the maximum normalization method specifically refers to the ratio of the signal amplitude of each sampling point in the filtered signal sequence to the maximum absolute signal amplitude in the filtered signal sequence;

[0091] The normalized signal sequence is time segmented by setting a sliding window length and a sliding step to obtain a plurality of sub-signal sequence blocks;

[0092] The one-dimensional time domain signal of each sub-signal sequence block is mapped to a two-dimensional time-frequency domain signal by using short-time Fourier transform to obtain a time-frequency spectrum feature matrix; the time-frequency spectrum feature matrix structure is time step dimension x frequency resolution dimension; the time step dimension is the sliding window length;

[0093] The time-frequency spectrum tensors corresponding to all sub-signal sequence blocks are stacked to obtain a time-frequency spectrum feature tensor; the time-frequency spectrum feature tensor structure is sub-signal sequence block dimension x time step dimension x frequency resolution dimension.

[0094] In this embodiment, step three specifically includes:

[0095] The time-frequency tensor encoding module inputs the time-frequency spectrum feature tensor into a 1x1 convolution layer, performs channel mapping on the frequency resolution dimension, and performs nonlinear feature transformation through a GELU activation function to obtain an initial mapping feature tensor;

[0096] The multi-scale large kernel convolution module includes three parallel large kernel convolution branches, each of which includes two-dimensional convolution, batch normalization, ReLU activation, and residual connection operations, specifically:

[0097] The initial mapping feature tensor is extracted using two-dimensional convolution with a convolution kernel size of 15x15, 31x31, and 51x51 to extract interference structure features, and normalized and nonlinearly transformed through batch normalization and ReLU activation to obtain a first scale convolution feature tensor, a second scale convolution feature tensor, and a third scale convolution feature tensor, respectively;

[0098] The first scale convolution feature tensor, the second scale convolution feature tensor, and the third scale convolution feature tensor are respectively connected in residual connection with the initial mapping feature tensor to obtain a first scale structure feature tensor, a second scale structure feature tensor, and a third scale structure feature tensor;

[0099] The first scale structure feature tensor, the second scale structure feature tensor, and the third scale structure feature tensor are channel spliced to generate a fused structure feature tensor;

[0100] In the spectrum prior channel modulation module, the fused structure feature tensor is subjected to two-dimensional average pooling operation in the sub-signal sequence block dimension and the time step dimension to obtain a channel energy mean vector;

[0101] Based on the channel energy mean vector, the maximum energy value and the minimum energy value of each channel are calculated, and each channel component of the channel energy mean vector is linearly normalized according to the maximum energy value and the minimum energy value to obtain a normalized channel energy vector;

[0102] The normalized channel energy vector is scaled channel by channel by introducing a channel scaling coefficient to obtain a channel modulation weight vector, and the channel modulation weight vector is multiplied element by element with the fused structure feature tensor in the channel dimension to obtain a modulated structure feature tensor;

[0103] The interference feature aggregation module performs average pooling on the modulated structure feature tensor in the time step dimension to generate an interference structure feature map, and the interference structure feature map has a structure of sub-signal sequence block dimension x channel dimension.

[0104] In the embodiment, the improved RepLKNet network is capable of perceiving different scale interference structure features such as wideband disturbance, narrowband modulation and local mutation at the same time in a single forward propagation process by setting different size large kernel convolution branches in parallel, and combining the residual connection mechanism to maintain the complete transmission of the original structure information, thereby enhancing the expression ability of complex interference patterns; the normalization and scaling mechanism based on channel energy distribution is introduced in the spectrum prior channel modulation module to adaptively adjust the response strength of different spectrum channels, so that the network pays more attention to the high-energy and strong interference related spectrum channels in the feature aggregation process, and suppresses redundant or low correlation features. Through the above improvements, the improved RepLKNet network can not only maintain the calculation stability, but also effectively improve the accuracy and discriminability of railway radio interference structure feature extraction.

[0105] In the embodiment, the step four specifically includes:

[0106] The interference structure feature map is divided into a plurality of interference structure feature vectors according to the sub-signal sequence block dimension, and the starting timestamp of each interference structure feature vector corresponding to the sub-signal sequence block is obtained;

[0107] An interference judgment energy threshold is set, the average energy value of the time-frequency spectrum feature matrix corresponding to each sub-signal sequence block is calculated, if the average energy value is less than the interference judgment energy threshold, the interference state label of the corresponding sub-signal sequence block is set to 0: no interference is detected, otherwise, it is set to 1: there is railway radio interference;

[0108] The interference state label is used as the target label of supervised training;

[0109] The interference structure feature vectors are time-sequenced according to the starting timestamp, and the interference structure feature vectors and the interference state label form an interference feature sample; each starting timestamp corresponds to a time step;

[0110] All interference feature samples form a training sample set.

[0111] In the embodiment, the step five specifically includes:

[0112] Based on the training sample set, the interference structure feature vector sequence and the interference state label sequence are obtained;

[0113] The interference structure feature vector sequence is generated into a candidate period set through a candidate period extraction module;

[0114] In the period selection gating module, the matching score between each interference structure feature vector and each main period in the candidate period set is calculated, a matching score vector is obtained, and the matching score vector is subjected to Softmax normalization to obtain a period selection gating vector;

[0115] In the steady-state cycle modeling branch, the interference structure feature vector sequence is reconstructed according to each main cycle length in the candidate cycle set to generate a cycle feature tensor, specifically: the ratio of the time step number of the interference structure feature vector sequence to the main cycle is calculated to obtain the cycle segment number, and the interference structure feature vector sequence is reconstructed according to the cycle segment number x main cycle length x channel dimension;

[0116] The cycle feature tensor is extracted through two-dimensional convolution to obtain a cycle modeling feature tensor, and the cycle modeling feature tensor is reshaped to be consistent with the structure of the interference structure feature vector sequence to obtain a cycle modeling feature vector sequence, which is consistent with the number of main cycles in the candidate cycle set;

[0117] In the present application, the reconstruction operation divides the original time sequence into multiple cycle segments, so that the interference changes inside each segment are aligned in the time dimension, which is conducive to the two-dimensional convolution kernel to extract the interference repetition pattern and local change trend in the cycle internal space and time dimension. After convolution, the feature is reshaped back to the same time step structure as the original sequence, which can ensure that the cycle modeling result is consistent with the disturbance modeling feature and the running state feature in the time step on the global timeline, thereby improving the modeling ability and expression accuracy of the entire time sequence model for periodic interference. If two-dimensional convolution is performed directly, the interference patterns between different cycle segments will be confused, and the stable structure within the cycle cannot be explicitly captured. Through the above processing steps, the cycle structure expression is retained, and the operability and semantic consistency of fusion with other module features are considered.

[0118] In the time step, based on the cycle selection gating vector, the cycle modeling feature vectors of each main cycle in the candidate cycle set are weighted and fused to obtain a steady-state cycle feature vector, and the steady-state cycle feature vector is composed of a steady-state cycle feature vector sequence according to the time step;

[0119] In the present application, it is assumed that the time step number of the interference structure feature vector sequence is 6, the channel dimension is 4, and the candidate cycle set contains two main cycles 2 and 3. In the cycle selection gating module, first, each interference structure feature vector is matched with the two main cycles respectively to obtain two matching score vectors; then, through the Softmax normalization operation, the cycle selection gating weights of the main cycle 3 and the main cycle 2 at each time step are obtained, for example, at time step 1, the weights of the main cycle 3 and the main cycle 2 are 0.6 and 0.4 respectively. In the steady-state cycle modeling branch, the interference structure feature vector sequence is reconstructed into a cycle feature tensor with the shape of cycle segment number × cycle length × channel dimension, for example, for the main cycle 3, a tensor with the shape of 2 × 3 × 4 is obtained; then, a two-dimensional convolution operation is performed on each cycle feature tensor respectively to obtain a cycle modeling feature tensor, and the cycle modeling feature tensor is reshaped into a cycle modeling feature vector sequence aligned with the time step, that is, the main cycle 3 and the main cycle 2 each output a cycle modeling feature vector sequence with a time step of 6 and a channel dimension of C, wherein C represents the channel dimension after performing the two-dimensional convolution operation. Finally, at each time step, the cycle modeling feature vectors from the main cycle 3 and the main cycle 2 are weighted and fused according to the cycle selection gating vector, for example, the fusion calculation at time step 1 is: 0.6 × cycle modeling feature vector of the main cycle 3 + 0.4 × cycle modeling feature vector of the main cycle 2, forming a steady-state cycle feature vector; the steady-state cycle feature vectors at all time steps together form a steady-state cycle feature vector sequence.

[0120] In the non-steady-state disturbance modeling branch, the sliding window length is set, the sliding step is set to 1, the interference structure feature vector sequence is segmented by sliding to obtain a plurality of disturbance sub-sequences; the number of disturbance sub-sequences is consistent with the time step number, and when the sliding window length is greater than the remaining time step number, a forward padding method is used to fill in to generate the disturbance sub-sequences;

[0121] Each disturbance sub-sequence is extracted through one-dimensional convolution and ReLU activation to extract local mutation features to obtain a disturbance modeling feature vector;

[0122] All disturbance modeling feature vectors are spliced in time step order to obtain a disturbance modeling feature vector sequence;

[0123] The steady-state cycle feature vector sequence and the disturbance modeling feature vector sequence are spliced in the channel dimension to form a joint feature vector sequence;

[0124] In the running state modulation module, the train speed value sequence is obtained, and the train speed value at each time step is generated through linear mapping to generate a running state modulation vector;

[0125] The running state modulation vector is performed channel-by-channel modulation with the joint feature vector of the corresponding time step to obtain a running state feature vector, and the running state feature vectors are grouped according to time steps to form a running state feature vector sequence;

[0126] In the present application, since the train speed can intuitively reflect the current running condition of the train, and different speed intervals are often accompanied by different types or intensities of radio interference, the train speed value sequence is selected for running state modulation, thereby improving the dynamic judgment ability of the interference risk.

[0127] Each running state feature vector is mapped through full connection to generate an interference state logic value, and the interference state logic value is mapped to the interval [0, 1] through a Sigmoid activation function to obtain an interference state prediction probability, and the interference state prediction probabilities are grouped according to time steps to form an interference state prediction probability sequence;

[0128] The binary cross entropy between the interference state prediction probability sequence and the interference state label sequence is calculated to obtain a supervised training loss function, and the improved TimesNet network is updated through back propagation based on the supervised training loss function, and the trained improved TimesNet network is used as a railway radio interference analysis model.

[0129] In the present embodiment, the interference structure feature vector sequence is generated through a candidate period extraction module to form a candidate period set, which specifically includes:

[0130] The channel mean of each interference structure feature vector is calculated, and the interference feature sequences are grouped according to time steps in sequence;

[0131] Based on fast Fourier transform, the interference feature sequence is mapped to the frequency domain to obtain a complex frequency domain feature sequence;

[0132] The amplitude spectrum of the complex frequency domain feature sequence is calculated, the amplitude spectrum is sorted in descending order according to the amplitude, and the first K frequencies with the largest amplitude are selected as the main frequencies;

[0133] The ratio of the sampling frequency to each main frequency is calculated to obtain K main periods;

[0134] The minimum period and the maximum period are set, the K main periods are de-duplicated, and the main periods that meet the minimum period and the maximum period are retained to obtain the candidate period set.

[0135] In the present application, the time scale with significant periodic characteristics is extracted from the interference structure feature vector to depict the main repeating mode and change rhythm of the interference signal. By performing frequency domain analysis on the interference feature sequence, selecting the main frequency with high energy proportion in the amplitude spectrum, further converting it into the corresponding main period, and screening out the candidate period set within the set period range, the potential periodic interference rule can be effectively identified, and the modeling accuracy and robustness of complex periodic interference characteristics are improved.

[0136] In the present embodiment, the step six specifically comprises:

[0137] The interference analysis result includes interference evolution graph, interference judgment state and interference risk level;

[0138] Obtain the interference structure feature vector sequence of the radio signal data to be analyzed, and generate the running state feature vector sequence through the railway radio interference analysis model;

[0139] Map each time step running state feature vector to generate interference state logic value, and map the interference state logic value to interference state prediction probability through Sigmoid activation function; combine the interference state prediction probability of each time step to form an interference state prediction probability sequence, and draw the interference state prediction probability sequence into an interference evolution graph with time step as the horizontal axis and interference state prediction probability as the vertical axis;

[0140] Set an interference judgment threshold, if the interference state prediction probability is greater than or equal to the interference judgment threshold, the interference judgment state of the corresponding time step is marked as existing railway radio interference, otherwise it is marked as no railway radio interference detected;

[0141] Set interference level thresholds T1 and T2, , and the interference level threshold T1 is greater than or equal to the interference judgment threshold; for the time step with railway radio interference, the interference risk level is determined based on the interference state prediction probability, specifically:

[0142] If the interference state prediction probability is less than the interference level threshold T1, the interference risk level is 0, indicating low risk; if the interference state prediction probability is less than the interference level threshold T2 and greater than or equal to the interference level threshold T1, the interference risk level is 1, indicating medium risk; if the interference state prediction probability is greater than or equal to the interference level threshold T2, the interference risk level is 2, indicating high risk.

[0143] Example 1

[0144] In order to verify the feasibility of the present application in implementation, the method of the present application is applied to a certain railway interference monitoring and communication guarantee system, and a wireless communication section along a certain railway and a certain station section is selected as a typical test scene, the test duration is 30 days, covering peak period, night, construction period and sudden dispatching period, etc. The typical GSM-R and LTE-R wireless frequency bands are detected, analyzed and judged for all period interference.

[0145] In the implementation process, first, the radio signal data in the railway communication environment is collected by the track side perception terminal, the low frequency and high frequency background noise is eliminated through the band pass filter, the sliding window segmentation is performed after the maximum value normalization processing, the short time Fourier transform is used to construct the three-dimensional structure time-frequency spectrum feature tensor, and the improved RepLKNet network deployed on the server is used to model the interference features, and the interference structure feature map is generated. By constructing the training sample set and inputting into the improved TimesNet network for time series modeling and supervised training, the railway radio interference analysis model is obtained. The real-time radio signal is analyzed by loading the railway radio interference analysis model.

[0146] In order to evaluate the actual effect of the method of the present application, the method of the present application is compared with two kinds of comparison schemes: comparison scheme A is the traditional energy threshold method, and comparison scheme B is the short time Fourier energy feature model based on SVM classifier. The three methods are deployed in the same hardware environment and test line, and the same data source is used for training and evaluation, and the statistical results are shown in Table 1.

[0147] Table 1 Performance comparison table of different schemes in railway radio interference detection

[0148] Scheme name Interference identification accuracy Interference type identification accuracy False alarm rate Missed alarm rate Average detection delay (ms) Risk level consistency rate Method of the invention 96.8% 94.3% 2.7% 1.5% 18.4 93.5% Comparative scheme A 78.2% 64.5% 11.6% 10.3% 36.9 52.7% Comparative scheme B 85.6% 78.9% 7.2% 6.5% 29.1 71.4%

[0149] As can be seen from the data in Table 1, the method of the present application is significantly better than the comparison schemes in multiple performance indicators. In terms of interference recognition accuracy, the method of the present application reaches 96.8%, which is improved by 18.6 percentage points compared with comparison scheme A and 11.2 percentage points compared with comparison scheme B, fully proving the high reliability of the method of the present application in recognizing interference signals in complex railway wireless environment. In terms of interference type recognition accuracy, the method of the present application reaches 94.3%, which is improved by nearly 30 percentage points compared with comparison scheme A and 15.4 percentage points compared with comparison scheme B. It shows that through the improved RepLKNet network, the multi-scale interference structure is accurately modeled, so that the present application can more effectively distinguish amplitude modulation type, narrowband pulse type and wideband disturbance type interference.

[0150] In addition, in terms of false positive rate and false negative rate, the method of the present application is controlled at 2.7% and 1.5% respectively, which is significantly lower than 11.6% and 10.3% of comparative scheme A and 7.2% and 6.5% of comparative scheme B, indicating that the method of the present application can effectively inhibit the risk hidden danger caused by misjudgment or omission, and enhance the stability and practicality in the actual scene of railway communication monitoring. In terms of average detection delay, the average detection delay of the method of the present application is 18.4 milliseconds, which is far superior to 36.9 milliseconds of comparative scheme A and 29.1 milliseconds of scheme B, verifying its faster response capability, which can meet the high real-time requirement of high-speed railway and other systems. In terms of risk level consistency rate, the method of the present application reaches 93.5%, while comparative scheme A and comparative scheme B are 52.7% and 71.4% respectively, further verifying that the method of the present application has higher accuracy and intelligent degree in multi-source information fusion.

[0151] In the task of railway radio interference analysis, the method of the present application not only has higher recognition ability and classification accuracy, but also has lower misjudgment risk, faster detection response and more robust risk judgment ability, which comprehensively improves the intelligent perception and safety protection ability of railway communication system to radio interference.

[0152] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A machine learning based railway radio interference analysis method, characterized by, The method comprises the following steps: Step one: obtaining radio signal data in a railway communication scenario; Step two: signal preprocessing of the radio signal data to generate a time-frequency spectrum feature tensor; Step three: inputting the time-frequency spectrum feature tensor into an improved RepLKNet network to extract interference structure features of railway radio and generate an interference structure feature map; the improved RepLKNet network comprises a time-frequency tensor encoding module, a multi-scale large kernel convolution module, a spectrum prior channel modulation module and an interference feature aggregation module; Step four: based on the interference structure feature map, constructing a railway interference feature sample to form a training sample set; Step five: inputting the training sample set into an improved TimesNet network to perform supervised training and obtain a railway radio interference analysis model; the improved TimesNet network comprises a candidate period extraction module, a period selection gating module, a steady-state period modeling branch, a non-steady-state disturbance modeling branch and a running state modulation module; Step six: loading the railway radio interference analysis model to perform interference analysis on the radio signal data to be analyzed and output an interference analysis result; Step seven: outputting the interference analysis result to a railway communication monitoring system.

2. The method of claim 1, wherein, The step one specifically comprises: collecting radio signal data from railway lines, station areas or train running sections, wherein the radio signal data comprises a railway wireless communication signal sequence and a radio interference signal sequence superimposed on the railway wireless communication signal.

3. The method of claim 1, wherein the method further comprises: The step two specifically comprises: The signal preprocessing comprises denoising, normalization, time segmentation and time-frequency transformation; The radio signal data is denoised by a band-pass filter to generate a filtered signal sequence; The maximum value normalization method is used to map the signal amplitude in the filtered signal sequence to the [0, 1] interval to obtain a normalized signal sequence; The normalized signal sequence is time segmented by setting a sliding window length and a sliding step to obtain a plurality of sub-signal sequence blocks; A short-time Fourier transform is used to map one-dimensional time domain signals of each sub-signal sequence block to two-dimensional time-frequency domain signals to obtain a time-frequency spectrum feature matrix; the time-frequency spectrum feature matrix structure is time step dimension x frequency resolution dimension; the time step dimension is the sliding window length; All sub-signal sequence block corresponding time-frequency spectrum tensors are stacked to obtain a time-frequency spectrum feature tensor; the time-frequency spectrum feature tensor structure is sub-signal sequence block dimension x time step dimension x frequency resolution dimension.

4. The method of claim 1, wherein the method further comprises: The step three specifically comprises: The time-frequency tensor encoding module inputs the time-frequency spectrum feature tensor into a 1x1 convolution layer, maps the frequency resolution dimension to a channel, and performs nonlinear feature transformation through a GELU activation function to obtain an initial mapping feature tensor; The multi-scale large kernel convolution module comprises three parallel large kernel convolution branches, each of which comprises a two-dimensional convolution, a batch normalization, a ReLU activation and a residual connection operation, specifically: The initial mapping feature tensor is extracted by two-dimensional convolution with a convolution kernel size of 15*15, 31*31 and 51*51 respectively to obtain a first scale convolution feature tensor, a second scale convolution feature tensor and a third scale convolution feature tensor; The first scale convolution feature tensor, the second scale convolution feature tensor and the third scale convolution feature tensor are respectively connected with the initial mapping feature tensor to obtain a first scale structure feature tensor, a second scale structure feature tensor and a third scale structure feature tensor; The first scale structure feature tensor, the second scale structure feature tensor and the third scale structure feature tensor are channel spliced to generate a fusion structure feature tensor; In the spectral prior channel modulation module, the fusion structure feature tensor is subjected to two-dimensional average pooling operation in the sub-signal sequence block dimension and the time step dimension to obtain a channel energy mean vector; Based on the channel energy mean vector, the maximum energy value and the minimum energy value of each channel are calculated, and each channel component of the channel energy mean vector is linearly normalized according to the maximum energy value and the minimum energy value to obtain a normalized channel energy vector; The normalized channel energy vector is scaled channel by channel by introducing a channel scaling coefficient to obtain a channel modulation weight vector, and the channel modulation weight vector is multiplied with the fusion structure feature tensor in the channel dimension to obtain a modulation structure feature tensor; The interference feature aggregation module performs average pooling on the modulation structure feature tensor in the time step dimension to generate an interference structure feature map, and the interference structure feature map has a structure of sub-signal sequence block dimension*channel dimension.

5. The machine learning based railway radio interference analysis method of claim 1, wherein, The step four specifically includes: The interference structure feature map is divided into a plurality of interference structure feature vectors according to the sub-signal sequence block dimension, and the starting time stamp of each sub-signal sequence block corresponding to each interference structure feature vector is obtained; An interference judgment energy threshold is set, the average energy value of the time-frequency spectrum feature matrix corresponding to each sub-signal sequence block is calculated, if the average energy value is less than the interference judgment energy threshold, the interference state label of the corresponding sub-signal sequence block is set to 0: no interference is detected, otherwise, it is set to 1: there is railway radio interference; The interference state label is used as a target label for supervised training; The interference structure feature vectors are time-ordered according to the starting time stamp, and the interference structure feature vectors and the interference state label form an interference feature sample; each starting time stamp corresponds to a time step; All interference feature samples form a training sample set.

6. The machine learning based railway radio interference analysis method of claim 1, wherein, The step five specifically includes: Based on the training sample set, the interference structure feature vector sequence and the interference state label sequence are obtained; The interference structure feature vector sequence is generated into a candidate period set through a candidate period extraction module; In the period selection gating module, the matching score between each interference structure feature vector and each main period in the candidate period set is calculated to obtain a matching score vector, and the matching score vector is subjected to Softmax normalization to obtain a period selection gating vector; The period selection gating vector is input into the period selection gating module to obtain a final interference state label sequence. In the steady-state cycle modeling branch, reconstruct the interference structure feature vector sequence according to each main cycle length in the candidate cycle set to generate a cycle feature tensor; extract stable cycle features through two-dimensional convolution of the cycle feature tensor to obtain a cycle modeling feature tensor, reshape the cycle modeling feature tensor to be consistent with the structure of the interference structure feature vector sequence, obtain a cycle modeling feature vector sequence, and the cycle modeling feature vector sequence is consistent with the number of main cycles in the candidate cycle set; In the time step, based on the cycle selection gating vector, the cycle modeling feature vectors of each main cycle in the candidate cycle set are weighted and fused to obtain a steady-state cycle feature vector, and the steady-state cycle feature vectors are combined into a steady-state cycle feature vector sequence according to the time step; In the non-steady-state disturbance modeling branch, set the sliding window length, and set the sliding step to 1, segment the interference structure feature vector sequence to obtain a plurality of disturbance subsequences; the number of disturbance subsequences is consistent with the number of time steps, and when the sliding window length is greater than the number of remaining time steps, the forward padding method is used to fill in to generate the disturbance subsequence; extract local mutation features through one-dimensional convolution and ReLU activation of each disturbance subsequence to obtain a disturbance modeling feature vector; concatenate all disturbance modeling feature vectors in time step order to obtain a disturbance modeling feature vector sequence; concatenate the steady-state cycle feature vector sequence and the disturbance modeling feature vector sequence in the channel dimension to form a joint feature vector sequence; obtain a running state modulation vector by linearly mapping the train speed value sequence at each time step in the running state modulation module; perform channel-by-channel modulation on the running state modulation vector and the joint feature vector at the corresponding time step to obtain a running state feature vector, and combine the running state feature vectors into a running state feature vector sequence according to the time step; generate an interference state logical value by fully connected mapping of each running state feature vector, and map the interference state logical value to the interval [0, 1] through the Sigmoid activation function to obtain an interference state prediction probability, and combine the interference state prediction probabilities into an interference state prediction probability sequence according to the time step; calculate the binary cross-entropy between the interference state prediction probability sequence and the interference state label sequence to obtain a supervised training loss function, and update the improved TimesNet network based on the supervised training loss function, and use the trained improved TimesNet network as a railway radio interference analysis model.

7. The machine learning based railway radio interference analysis method of claim 6, wherein, The candidate cycle set generated by the interference structure feature vector sequence through the candidate cycle extraction module specifically includes: calculate the channel mean of each interference structure feature vector, and combine the interference feature sequences in time step order; map the interference feature sequence to the frequency domain based on the fast Fourier transform to obtain a complex frequency domain feature sequence; calculate the amplitude spectrum of the complex frequency domain feature sequence, sort the amplitude spectrum in descending order according to the amplitude, and select the top K frequencies with the largest amplitude as the main frequencies; calculate the ratio of the sampling frequency to each main frequency to obtain K main cycles; Set the minimum period and the maximum period, remove the K main periods, and retain the main periods between the minimum period and the maximum period to obtain a candidate period set.

8. The machine learning based railway radio interference analysis method of claim 1, wherein, The step six specifically comprises: The interference analysis result comprises an interference evolution diagram, an interference judgment state, and an interference risk level. Obtain an interference structure feature vector sequence of the to-be-analyzed radio signal data, pass the interference structure feature vector sequence through a railway radio interference analysis model, and generate an operation state feature vector sequence; Map each time step operation state feature vector to generate an interference state logic value, map the interference state logic value to an interference state prediction probability through a Sigmoid activation function, combine the interference state prediction probabilities of each time step to form an interference state prediction probability sequence, take a time step as a horizontal axis and an interference state prediction probability as a vertical axis, draw the interference state prediction probability sequence into an interference evolution diagram; Set an interference judgment threshold, if the interference state prediction probability is greater than or equal to the interference judgment threshold, mark the interference judgment state of the corresponding time step as existing railway radio interference, otherwise mark it as not detecting railway radio interference; The interference level threshold values T1 and T2 are set, The interference level threshold values T1 and T2 are set, For the time steps in which the railway radio interference exists, the interference risk level is determined based on the interference state prediction probability, specifically: If the interference state prediction probability is less than an interference level threshold T1, the interference risk level is 0, indicating low risk, if the interference state prediction probability is less than an interference level threshold T2 and greater than or equal to the interference level threshold T1, the interference risk level is 1, indicating medium risk, and if the interference state prediction probability is greater than or equal to the interference level threshold T2, the interference risk level is 2, indicating high risk.

Images