A neural network-based pulse signal intelligent switching control method and system

Abstract

The application provides a kind of neural network-based pulse signal intelligent switching control method and system, it is related to data processing technical field, the method includes: obtaining pulse output sequence and carrying out boundary fitting, generate periodic stable region list, construct heterodyne interference matrix and measure the independent response capability of channel in local interference area, get state separation index, filter out switching sensitive section, calculate the compressibility of local waveform fitting, get channel mapping compression rate, construct feature binary set to generate training sample set, update the parameters of neural network model, real-time acquisition data sequence to be judged, get the predicted input set input into neural prediction model, output switching signal sequence, and according to its identification whether it is currently in the best signal switching window, generate control instruction to switch the signal between channels.The application can accurately switch the signal between channels under different physical disturbances.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a neural network-based pulse signal intelligent switching control method and system. BACKGROUND

[0002] In a precision rotary measurement device, in order to improve the continuity and accuracy of angle detection, some systems have adopted a neural network-based pulse signal switching control technology. The neural network is used to predict and control the switching timing between the master reading head and the slave reading head. This method collects historical angle data, reading head signal strength, magnetic strip gap, electrical noise and other multi-dimensional input variables, constructs a feedforward neural network model, realizes dynamic discrimination of signal switching nodes, and avoids switching delay or misjudgment caused by threshold judgment.

[0003] However, the existing neural network-based signal switching scheme relies heavily on the consistency of neural network training data distribution and may lack the generalization ability for unseen scenes under magnetic field disturbance. For example, in some heavy load devices, due to temperature rise, the magnetic strip micro-deformation causes nonlinear changes in the magnetic flux density of the joint area, and the neural network model may fail to identify, causing the signal switching point to shift, causing measurement error accumulation and affecting the control accuracy of downstream devices. SUMMARY

[0004] The purpose of the present application is to provide a neural network-based pulse signal intelligent switching control method and system to solve the problems mentioned in the background.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] In a first aspect, a neural network-based pulse signal intelligent switching control method is provided, which comprises:

[0007] Obtaining the pulse output sequence of the main channel and the auxiliary channel in multiple cycles to construct a cycle boundary map, and fitting the cycle boundary map, generating a cycle stable region list based on statistical interval;

[0008] In the cycle stable region list, according to the relative fluctuation between the main channel and the auxiliary channel, an interference matrix of different frequencies is constructed, and the state separation index is obtained according to the interference matrix of different frequencies to measure the independent response ability of the channels in the local interference region;

[0009] According to the state separation index, the transition is identified, the switching sensitive section is selected, the signal mutation mapping table is constructed, the section in the signal mutation mapping table is extracted for path anomaly, the compressibility of local waveform fitting is calculated, and the channel mapping compression rate is obtained;

[0010] A set of feature pairs is constructed based on the state separation index and the channel mapping compression ratio, and the set of feature pairs is normalized to generate a training sample set.

[0011] The training sample set is input into the neural network model for training, the parameters of the neural network model are updated, and a neural prediction model is obtained.

[0012] The system acquires the data sequence to be judged in real time and performs the same feature processing procedure as the training sample set on the data sequence to be judged to obtain the prediction input set.

[0013] The predicted input set is fed into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it is determined whether the current signal switching window is in the optimal window, and control commands are generated to switch the signals between channels.

[0014] Furthermore, pulse output sequences of the main and secondary channels within multiple cycles are obtained to construct a periodic boundary map. Boundary fitting is then performed on the periodic boundary map, and a list of periodically stable regions is generated based on statistical intervals, including:

[0015] Time synchronization is performed based on the pulse output sequences of the main channel and the secondary channel, a correspondence is established between the sampling points of different channels, the difference in sampling starting point is eliminated, and an aligned signal sequence is obtained.

[0016] Based on the aligned signal sequence, the peak and valley values ​​in the aligned signal sequence are identified to mark the start and end positions of a complete pulse cycle, thus obtaining a set of candidate points for the cycle boundary;

[0017] Based on the set of candidate points for the periodic boundary, the time interval between adjacent candidate points is calculated one by one and the distribution of the intervals is statistically analyzed to obtain the boundary interval sequence.

[0018] The average interval value and standard deviation are calculated based on the boundary interval sequence. Interval values ​​with a difference from the average interval value greater than twice the standard deviation are excluded, and the remaining boundaries are compensated and corrected to obtain the periodic boundary map.

[0019] Based on the periodic boundary map, the fluctuation amplitude of the boundary interval of each segment is compared and segments below the preset fluctuation threshold are selected to obtain a list of periodic stable regions.

[0020] Furthermore, within the list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main and secondary channels. This hetero-frequency interference matrix is ​​then used to measure the independent response capabilities of the channels within local interference regions, yielding a state separation index, including:

[0021] Based on the list of periodically stable regions, the signal sequences of the main channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the main channel.

[0022] Based on the list of periodically stable regions, the signal sequences of the sub-channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the sub-channel.

[0023] Based on the statistical feature sets of the main channel and the sub-channel, the difference in mean, the difference in variance, and the covariance of each segment are calculated to obtain three types of difference index sequences.

[0024] By combining the three types of difference index sequences into a matrix, the difference in mean, difference in variance, and covariance of the segments are mapped to the same matrix structure in rows and columns to obtain the heterofrequency interference matrix.

[0025] Furthermore, in the list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel. The independent response capability between channels within local interference regions is measured using this hetero-frequency interference matrix to obtain the state separation index, which also includes:

[0026] Based on the difference in the average value of the segments between the main channel and the secondary channel, an average offset sequence for each segment is established, and the relative difference in the signal center position of different channels within the stable segment is calculated to obtain the mean difference anisotropy.

[0027] Based on the difference in segment variance between the main channel and the secondary channel, a segment fluctuation comparison sequence is formed. The relative difference in signal fluctuation amplitude between different channels within the same segment is calculated to obtain the fluctuation difference term.

[0028] Based on the segment covariance between the main channel and the secondary channel, a segment cooperative change sequence is constructed, and the degree of mutual dependence of the signals of different channels in the same segment over time is calculated to obtain the relevant difference terms.

[0029] The state separation index is obtained by weighting and fusing the mean difference, fluctuation difference, and correlation difference, and then statistically aggregating all segments.

[0030] Furthermore, transition identification is performed based on the state separation index to screen out switching sensitive segments, a signal mutation mapping table is constructed, and path anomalies are extracted from the segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio, including:

[0031] High-value intervals are identified based on the state separation index. Local average values ​​are calculated using a sliding window, and windows exceeding a preset average threshold are filtered out to obtain an initial set of sensitive segments.

[0032] Based on the initial set of sensitive segments, the original waveforms within the segments are compared point by point with the fitted waveforms at the periodic boundaries to obtain the residual sequence.

[0033] Based on the residual sequence, calculate the magnitude of change between adjacent residuals and mark the points that exceed a preset change threshold to obtain a set of mutation candidate points;

[0034] Based on the set of mutation candidate points, adjacent mutation candidate points are merged into mutation segments, and the start and end positions, average residual amplitude, and waveform offset of each mutation segment are recorded to obtain a signal mutation mapping table.

[0035] Furthermore, transition identification is performed based on the state separation index to screen out switching sensitive segments, a signal mutation mapping table is constructed, and path anomalies are extracted from segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio. This also includes:

[0036] Based on the original path length and fitted path length of each mutation segment, the relative shortening degree of the two is calculated, and the relative shortening degree is merged according to the original path length to obtain the path compression term;

[0037] Based on the average residual magnitude of each mutation segment, the proportion of the residual magnitude to the fitted path is calculated, and the proportion is constrained according to the length of the fitted path to obtain the residual weight term.

[0038] Based on the waveform offset of each abrupt change segment, the degree of coupling influence of the waveform offset on the fitted path is calculated, and the degree of coupling influence is suppressed according to the original path length to obtain the offset correction term.

[0039] The path compression term, residual weight term, and offset correction term are weighted and fused to obtain the segment compression value;

[0040] Calculate the degree of aggregation of the compression values ​​of each mutation segment and merge the aggregation degrees to obtain the channel mapping compression ratio.

[0041] Furthermore, the predicted input set is fed into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it is determined whether the current signal switching window is optimal, and control commands are generated to switch the signals between channels, including:

[0042] The switching signal sequence is filtered, and when the marker value at a certain moment is greater than the preset threshold, that moment is recorded as a candidate switching point, thus obtaining a set of candidate switching points;

[0043] Aggregate the candidate switching point set. When the time interval between adjacent candidate switching points is less than a preset interval threshold, merge them into a segment to obtain a candidate switching segment sequence.

[0044] Based on the candidate switching segment sequence, the variance of the switching signal sequence in each segment is calculated. When the variance is lower than the preset stability threshold, the segment is determined to be a stable segment, and a set of stable switching segments is obtained.

[0045] Based on the stable switching segment set, when the duration of a certain segment is greater than the preset minimum duration and there are no breakpoints in the segment that exceed the preset maximum interval, the segment is marked as the optimal signal switching window, and the optimal window index set is obtained.

[0046] Based on the optimal window index set, its start and end positions are bound to the channel switching logic to obtain control commands to control the signal switching between the main channel and the secondary channel.

[0047] Secondly, a pulse signal intelligent switching control system based on a neural network, the system comprising:

[0048] The basic module is used to acquire the pulse output sequences of the main channel and the sub-channel within multiple cycles to construct a periodic boundary map, and to perform boundary fitting on the periodic boundary map to generate a list of periodic stable regions based on statistical intervals.

[0049] The first analysis module is used to construct a hetero-frequency interference matrix based on the relative fluctuations between the main channel and the secondary channel in the list of periodically stable regions, and to measure the independent response capability between channels in the local interference region based on the hetero-frequency interference matrix to obtain the state separation index.

[0050] The second analysis module is used to identify transitions based on the state separation index, filter out switching sensitive sections, construct a signal mutation mapping table, extract path anomalies from the sections in the signal mutation mapping table, calculate the compressibility of local waveform fitting, and obtain the channel mapping compression rate.

[0051] The sample module is used to construct a set of feature pairs based on the state separation index and the channel mapping compression ratio, and to normalize the set of feature pairs to generate a training sample set.

[0052] The training module is used to input the training sample set into the neural network model for training, update the parameters of the neural network model, and obtain the neural prediction model.

[0053] The prediction module is used to acquire the data sequence to be judged in real time and perform the same feature processing process as the training sample set on the data sequence to be judged to obtain the prediction input set;

[0054] The instruction module is used to input the predicted input set into the neural prediction model, output the switching signal sequence, identify whether it is currently in the optimal signal switching window based on the switching signal sequence, and generate control instructions to switch the signals between channels.

[0055] The above-described solution of the present invention has at least the following beneficial effects:

[0056] This invention constructs a periodic boundary map by collecting pulse signal output sequences from the main and secondary channels within multiple periods. Based on the positions of the start and end points of the periods in the time series, the boundaries of each period are calibrated. This allows for the identification of time-series data segments with repetitive structures, stable waveforms, and clear variation patterns in the original signal. The periodic variation characteristics of the signal are easier to extract and analyze. Through the division of periodic boundaries and stability testing, the distortion of feature extraction caused by data inconsistency or excessive volatility in subsequent steps is significantly reduced, providing a high-confidence data foundation for subsequent construction of statistical feature matrices and training samples.

[0057] This invention expresses the statistical differences in the horizontal and vertical time segments of two channel signals in multiple periodic stable regions in a structured manner. It encapsulates the mean difference, variance difference, and covariance into a two-dimensional matrix. Subsequent calculations directly perform mathematical operations, feature extraction, or principal component analysis on this matrix, thereby enabling the system to handle high-dimensional statistical differences.

[0058] This invention generates a state separation index by weighted integration of mean difference anomalies, fluctuation differences, and correlation differences. This index quantifies the response separation of different channels under the same stable period, providing a basis for measuring the independence and decoupling of channel signals. It helps identify which regions exhibit signal behavior differences, making it suitable for subsequent switching timing determination. By fusing and compressing multiple statistical feature terms into a single value, while maintaining the integrity of segment features, subsequent sample construction and model input are no longer limited by the complexity of high-dimensional feature combinations, which is beneficial for model convergence.

[0059] This invention compares the original path and the fitted path of abrupt change segments, calculates the path compression term, residual weight term, and offset correction term respectively, and then weights and fuses them to obtain the channel mapping compression ratio. This achieves a unified characterization of signal structural complexity and anomalous coupling, and evaluates whether the segment possesses good compressibility during encoding or modeling. It not only reflects the signal change rate but also comprehensively embodies the fitting difficulty and anomalous distribution density, compressing complex morphological behaviors into a unified metric value. This provides the model with a directly applicable input for judging fitting difficulty and optimizes the learning performance of training samples on boundary transition states.

[0060] This invention identifies high-fluctuation segments through the state separation index, extracts abrupt change points, and constructs abrupt change segments. The start and end positions, average residual amplitude, and offset information of all abrupt change segments are summarized to generate a signal abrupt change mapping table. This not only identifies path segments with significant signal changes but also records multiple quantifiable indicators, providing structured data for subsequent modeling, compression, and prediction of these regions. Attached Figure Description

[0061] Figure 1This is a flowchart of a pulse signal intelligent switching control method based on a neural network, provided by an embodiment of the present invention. Detailed Implementation

[0062] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0063] like Figure 1 As shown, an embodiment of the present invention proposes a pulse signal intelligent switching control method based on a neural network, the method comprising:

[0064] The pulse output sequences of the main channel and the sub-channel within multiple cycles are obtained to construct a periodic boundary map. The periodic boundary map is then fitted to the boundary, and a list of periodic stable regions is generated based on statistical intervals.

[0065] In the list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel. The independent response capability between channels in the local interference region is measured based on the hetero-frequency interference matrix to obtain the state separation index.

[0066] Transition identification is performed based on the state separation index to screen out switching sensitive sections, a signal mutation mapping table is constructed, path anomalies are extracted from the sections in the signal mutation mapping table, the compressibility of local waveform fitting is calculated, and the channel mapping compression rate is obtained.

[0067] A set of feature pairs is constructed based on the state separation index and the channel mapping compression ratio, and the set of feature pairs is normalized to generate a training sample set.

[0068] The training sample set is input into the neural network model for training, the parameters of the neural network model are updated, and a neural prediction model is obtained.

[0069] The system acquires the data sequence to be judged in real time and performs the same feature processing procedure as the training sample set on the data sequence to be judged to obtain the prediction input set.

[0070] The predicted input set is fed into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it is determined whether the current signal switching window is in the optimal window, and control commands are generated to switch the signals between channels.

[0071] In this embodiment of the invention, pulse output sequences of the main channel and sub-channel within multiple cycles are acquired to construct a periodic boundary map. Boundary fitting is performed on the periodic boundary map, and a list of periodically stable regions is generated based on statistical intervals. This achieves periodic structural processing of the original pulse signal sequence, eliminating the risk of feature extraction distortion caused by irregular signal changes and providing data information blocks for subsequent processing. In the list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and sub-channel. The independent response capability between channels within local interference regions is measured using the hetero-frequency interference matrix, resulting in a state separation index. A structured matrix of multi-dimensional statistical quantities is constructed, and the state separation index quantifies the differences in channel signal behavior, providing an indicator basis for identifying key regions with prominent differences in channel behavior. Transition identification is performed based on the state separation index to screen out switching sensitive segments. A signal mutation mapping table is constructed, and path anomalies are extracted from segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio, enabling the labeling of abnormal signal paths. Numerical values ​​reflecting path compressibility and waveform distortion are constructed, providing a metric benchmark for describing the complexity of the signal structure for subsequent training.

[0072] A feature binary set is constructed based on the state separation index and channel mapping compression ratio, and then normalized to generate a training sample set. This eliminates the influence of differences in physical quantity dimensions on model learning, ensuring that the evaluation of each feature dimension by the neural network during training is not affected by numerical scale differences. The training sample set is input into the neural network model for training, updating the parameters of the neural network model to obtain a neural prediction model. Supervised training is performed on the structured sample set, enabling prediction of whether to enter the switching window and achieving generalization decision-making under complex conditions. The data sequence to be judged is acquired in real time, and the same feature processing process as the training sample set is applied to the data sequence to be judged to obtain a prediction input set. This ensures that the training model receives the equivalent feature structure during runtime as during training, effectively avoiding the problem of declining model generalization performance and providing an input basis for prediction. The prediction input set is input into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it is identified whether the current signal switching window is in the optimal window, generating control commands to switch the signals between channels. This achieves quantitative positioning of the channel switching timing, providing data support for the switching control commands and meeting the signal switching requirements of high-precision rotating equipment.

[0073] Specifically, a feature binary set is constructed based on the state separation index and channel mapping compression ratio, and the feature binary set is normalized to generate a training sample set, which includes:

[0074] In multiple periodic stable regions, the state separation index reflects the degree of difference in signal response between the main channel and the sub-channel within the periodic segment, while the channel mapping compression ratio measures the expressibility of abrupt segments in terms of structural fitting and signal compression. These two factors are used to describe the characteristics of periodic segments in both statistical and signal structural dimensions.

[0075] The system sequentially pairs the state separation index of each period segment with the corresponding channel mapping compression ratio, forming the following structure: The binary pair, in which Indicates the state separation index. This represents the channel mapping compression ratio. The system sequentially scans all stable periodic segments according to the periodic order, constructing a complete set of feature pairs based on this structure. Each sample corresponds to a composite feature representation of a periodic segment. Since the two feature terms may originate from different computational dimensions and numerical ranges, it is necessary to normalize the two components in the set of pairs separately to eliminate the bias caused by the difference in dimensions to subsequent neural network training. The normalization process adopts a minimum-maximum normalization strategy: first, the minimum value of all state separation indices is calculated. and maximum value And separate the index for each state. according to Normalization is performed, where This serves as the index for the segment; similarly, the minimum compression ratio of all channel mappings is calculated. With the maximum value For each compression ratio according to After normalization, the two feature values ​​of all samples are mapped to the interval [0,1]. The normalized feature pairs are organized into a standardized training sample set, which can be directly input into the neural network model for training.

[0076] This process involves inputting a training sample set into a neural network model for training, updating the parameters of the neural network model, and obtaining a neural prediction model. Specifically, this includes:

[0077] First, a neural network model is constructed to handle this task. The model architecture adopts a feedforward neural network structure, including an input layer, at least one hidden layer, and an output layer. The number of neurons in the input layer is consistent with the dimension of the feature pairs, that is, each sample contains two input values: the state separation index and the channel mapping compression ratio. The number of hidden layers and the number of neurons in each layer can be set according to the sample size and fitting complexity. Typically, 2 to 4 hidden layers are selected, and the number of neurons in each layer can be set to an integer value between 8 and 64. Each hidden layer node uses the ReLU activation function to enhance the non-linear expressive power of the network; the output layer uses the Sigmoid activation function to output a probability value between 0 and 1, representing the probability of whether the corresponding input period segment is the optimal signal switching window.

[0078] The model training employs a supervised learning strategy, requiring the assignment of corresponding labels to the training sample set. Labels can be generated through historical switching records, baseline logic control strategies, or manual annotation. For each training sample, if its corresponding period segment has historically been identified as a valid switching window, the sample's label is set to 1; otherwise, it is set to 0. After constructing the training set, the cross-entropy loss function is used as the objective function to calculate the error between the neural network output and the true labels. The backpropagation algorithm combined with a gradient descent optimizer is then used to iteratively update the weight parameters of each layer in the network.

[0079] During training, the system groups the training sample set into batches according to a set batch size, typically 32 or 64, and inputs each batch into the neural network model in each training round. Each batch of training data is used to calculate the predicted output through forward propagation, then the current error is calculated using the loss function, and the gradient is calculated and the weights of each layer are updated using the chain rule during backpropagation. The number of training rounds is set according to the model's convergence performance, ranging from 100 to 1000 rounds. To avoid overfitting, the system uses a validation set to monitor model performance changes during training and introduces a Dropout mechanism to randomly deactivate hidden layer neurons when necessary. Finally, after all training rounds are completed or the model performance reaches the preset convergence condition, the trained neural network model is saved.

[0080] This includes real-time acquisition of the data sequence to be judged, and performing the same feature processing procedure on the data sequence as on the training sample set to obtain the prediction input set, specifically including:

[0081] After the neural network model is trained and deployed, the system enters the online operation phase. During this phase, the real-time acquired pulse signal data needs to be processed to determine whether channel switching is currently possible. To ensure the accuracy and consistency of the prediction results, the real-time input data must strictly follow the same feature extraction and normalization process as in the training phase, constructing a prediction input set with the same structure as the training samples.

[0082] Specifically, the system continuously acquires real-time pulse signal output sequences from the main and secondary channels, aligns them by timestamps to ensure the comparability of the data structure in the time domain. Within each fixed sampling window, the system extracts signal segments of multiple cycles and performs cycle boundary detection, identifying candidate cycle points consistent with the training phase and generating a cycle boundary map. Based on standard deviation analysis of the boundary time intervals, cycle segments that meet the stability threshold are selected to constitute the real-time periodic stable region.

[0083] Within the selected stable region, the system calculates the intra-period mean, variance, and covariance of the signal data from the main and secondary channels, respectively. Then, based on the previously trained process, it constructs a hetero-frequency interference matrix and calculates the corresponding state separation index. Simultaneously, it detects any significant signal abrupt changes within these regions. By fitting the periodic boundary curve and calculating the point-to-point residual sequence, it extracts abrupt change points and generates residual statistics to produce the channel mapping compression rate for the current period segment.

[0084] The state separation index, calculated in real time, is paired with the channel mapping compression ratio to form feature pairs for prediction, and then standardized using the same normalization formula as during training. The maximum and minimum values ​​used in normalization should be consistent with the training sample set to avoid shifts in the model input feature distribution due to inconsistent normalization standards. For example, the system should save the data recorded during training. , , , The parameters are normalized, and these constants are reused during the online inference phase to normalize new predicted features, ensuring consistency of the input space between training and inference. All standardized feature pairs constitute the prediction input set, whose data structure strictly matches the input interface of the neural network model and can be directly fed into the pre-trained neural network model for online prediction.

[0085] In a preferred embodiment of the present invention, pulse output sequences of the main channel and sub-channel within multiple periods are obtained to construct a periodic boundary map, and boundary fitting is performed on the periodic boundary map to generate a list of periodically stable regions based on statistical intervals, including:

[0086] Time synchronization is performed based on the pulse output sequences of the main channel and the secondary channel, a correspondence is established between the sampling points of different channels, the difference in sampling starting point is eliminated, and an aligned signal sequence is obtained.

[0087] Based on the aligned signal sequence, the peak and valley values ​​in the aligned signal sequence are identified to mark the start and end positions of a complete pulse cycle, thus obtaining a set of candidate points for the cycle boundary;

[0088] Based on the set of candidate points for the periodic boundary, the time interval between adjacent candidate points is calculated one by one and the distribution of the intervals is statistically analyzed to obtain the boundary interval sequence.

[0089] The average interval value and standard deviation are calculated based on the boundary interval sequence. Interval values ​​with a difference from the average interval value greater than twice the standard deviation are excluded, and the remaining boundaries are compensated and corrected to obtain the periodic boundary map.

[0090] Based on the periodic boundary map, the fluctuation amplitude of the boundary interval of each segment is compared and segments below the preset fluctuation threshold are selected to obtain a list of periodic stable regions.

[0091] In this embodiment of the invention, time synchronization is performed based on the pulse output sequences of the main channel and the secondary channel. A correspondence is established between the sampling points of different channels to eliminate differences in sampling start points and obtain an aligned signal sequence. This avoids timing mismatch caused by sampling delays or start time offsets between channels, ensuring that subsequent period detection and boundary determination have a unified reference standard. Based on the aligned signal sequence, the peak and valley values ​​in the aligned signal sequence are identified to mark the start and end positions of a complete pulse cycle, obtaining a set of candidate points for the period boundary. This captures the local trend extrema of the signal and effectively extracts the structural boundary of a period pulse. Based on the set of candidate points for the period boundary, the time interval between adjacent candidate points is calculated one by one, and the distribution of the interval is statistically analyzed to obtain... Boundary interval sequences are used to model the global time interval of the periodic structure, providing a foundation for subsequent stable region screening and abnormal period removal. The average interval value and standard deviation are calculated based on the boundary interval sequence. Interval values ​​with a difference greater than twice the standard deviation from the average interval value are excluded, and the remaining boundaries are compensated and corrected to obtain a periodic boundary map. Abnormal periods deviating from the main distribution are removed, maintaining a unified analysis window for subsequent extraction of periodic features. Based on the periodic boundary map, the fluctuation amplitude of the boundary intervals in each segment is compared, and segments below a preset fluctuation threshold are selected to obtain a list of periodic stable regions. This ensures that the selected periodic segments are consistent in the time dimension and effectively eliminates unstable segments in the pulse waveform caused by electromagnetic disturbances or mechanical errors.

[0092] The process involves calculating the average interval and standard deviation based on the boundary interval sequence, excluding intervals whose difference from the average interval is greater than twice the standard deviation, and compensating and correcting the remaining boundaries to obtain the periodic boundary map. Specifically, this includes:

[0093] First, the time length between each period segment needs to be quantified and statistically analyzed. The system sequentially extracts the start timestamps of two adjacent period segments from the set of period boundary points, and calculates the time interval sequence, which can be represented as follows: ,in Indicates the first The starting boundary time of each period segment Indicates the first The starting boundary time of each periodic segment.

[0094] To remove potential disturbances and anomalous segments from periodic waveforms, such as non-periodic oscillations caused by signal jitter, electrical interference, or sensor malfunction, statistical analysis should be performed on the boundary interval sequence. Specifically, the mean of the sequence is first calculated. and standard deviation Then set filtering conditions to include all conditions that meet the criteria. Periodic segments are considered statistical outliers and removed to ensure that the retained periodic segments have concentrated distribution and typical periodic characteristics, effectively avoiding interference from atypical segments in periodic modeling. When reconstructing the sequence of the remaining periodic boundary points after removal, to avoid periodic breaks due to discontinuous intervals, the system adopts linear extrapolation and local smoothing compensation to perform temporal interpolation and boundary time axis correction on some missing intervals, ultimately forming a periodic boundary map that satisfies statistical consistency constraints.

[0095] Specifically, based on the periodic boundary map, the fluctuation amplitude of the boundary intervals of each segment is compared and segments below a preset fluctuation threshold are selected to obtain a list of periodically stable regions, which includes:

[0096] First, the system needs to perform regional stability analysis on the periodic spectrum based on the fluctuation amplitude between periodic segments, and then recalculate the time interval for all adjacent periodic segments in the spectrum. And define the relative variation amplitude between each periodic segment and its neighboring periodic segments on this sequence, denoted as . The system has a pre-set fluctuation threshold. This value, which can be obtained based on empirical data or adjusted to the target accuracy, defines the allowable range of fluctuations during the cycle. When the fluctuation amplitude corresponding to a certain cycle segment is less than this threshold, it indicates that its cycle behavior is relatively stable and has not been significantly affected by sudden disturbances or slow trends. The system includes this cycle segment in the set of candidate stable cycle segments. All cycle segments that meet this condition are sequentially combined on the time axis to form a list of non-overlapping stable cycle regions. Each stable segment has a clear start and end time boundary to ensure that its internal signal behavior meets the requirements of small fluctuations and good cycle consistency.

[0097] In a preferred embodiment of the present invention, in a list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel. The independent response capability between channels within a local interference region is measured using this hetero-frequency interference matrix to obtain a state separation index, including:

[0098] Based on the list of periodically stable regions, the signal sequences of the main channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the main channel.

[0099] Based on the list of periodically stable regions, the signal sequences of the sub-channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the sub-channel.

[0100] Based on the statistical feature sets of the main channel and the sub-channel, the difference in mean, the difference in variance, and the covariance of each segment are calculated to obtain three types of difference index sequences.

[0101] By combining the three types of difference index sequences into a matrix, the difference in mean, difference in variance, and covariance of the segments are mapped to the same matrix structure in rows and columns to obtain the heterofrequency interference matrix.

[0102] In this embodiment of the invention, based on the list of periodically stable regions, the signal sequences of the main channel within each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the main channel. This accurately reflects the output characteristics of the main channel under different time windows, laying a quantitative foundation for subsequent comparison with the secondary channel and providing a standard reference for measuring the differences between the two channels. Similarly, based on the list of periodically stable regions, the signal sequences of the secondary channel within each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the secondary channel. This ensures that the data processing methods are consistent across different channels, eliminates systematic errors caused by differences in calculation methods, and provides a data set for the construction of the hetero-frequency interference matrix. Consistency foundation: Based on the statistical feature sets of the main channel and the secondary channel, the difference in mean, difference in variance, and covariance of each segment are calculated one by one to obtain three types of difference index sequences. This quantifies the difference in signal behavior between the two channels in each stable segment, captures the offset of the average output, reflects the consistency of fluctuations and the degree of coordination of responses, and provides basic data for matrix structure construction. Based on the matrix combination of the three types of difference index sequences, the difference in mean, difference in variance, and covariance of each segment are mapped to the same matrix structure according to rows and columns to obtain the hetero-frequency interference matrix, which provides a two-dimensional expression structure for the signal differences between channels and provides a unified input source for the calculation of the state separation index.

[0103] Specifically, based on the statistical feature sets of the main channel and the secondary channel, the difference in mean, the difference in variance, and the covariance of each segment are calculated to obtain three types of difference index sequences, including:

[0104] First, iterate through each stable segment in the list of periodically stable regions. For each segment, extract the signal sequence within the corresponding time range from both the main channel and the secondary channel. Denote the signal from the main channel within that segment as... The signal of the secondary channel is denoted as ,in Indicates the first A stable periodic segment, This indicates the number of sampling points contained in the period segment, with a one-to-one correspondence between the two sequences in time. For each pair of signal sequences... and Calculate their statistical characteristics separately to obtain the mean values ​​of the main channel and the secondary channel within this section. , and variance , Based on this, the covariance of the two signal sequences within this segment is further calculated. The calculation formula is as follows: Then, three dissimilarity metrics are constructed segment by segment: segment mean difference, segment variance difference, and covariance. The above calculation is repeated in all stable segments to obtain three index sequences that correspond one-to-one with the periodic segments: mean difference sequence, variance difference sequence, and covariance sequence.

[0105] Specifically, based on the matrix combination of three types of difference index sequences, the segment mean difference, variance difference, and covariance value are mapped to the same matrix structure according to rows and columns to obtain the heterofrequency interference matrix, which specifically includes...

[0106] For each periodically stable segment, row vectors are constructed. These row vectors are then stacked row by row to form an M×3 two-dimensional matrix, denoted as the heterofrequency interference matrix, where M is the total number of periodically stable regions. Each row in this matrix represents the channel difference characteristics of a periodically stable segment. The three columns of the matrix correspond to three indicators: the difference in segment mean, the difference in segment variance, and the segment covariance. To prevent matrix skew caused by differences in the dimensions of different indicators, the matrix can be standardized by performing zero-mean unit variance normalization on each column, ensuring that the three types of indicators have equivalent weights in subsequent aggregation modeling.

[0107] In a preferred embodiment of the present invention, in the list of periodically stable regions, a hetero-frequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel, and the independent response capability between channels in the local interference region is measured based on the hetero-frequency interference matrix to obtain the state separation index, further comprising:

[0108] Based on the difference in the average value of the segments between the main channel and the secondary channel, an average offset sequence for each segment is established, and the relative difference in the signal center position of different channels within the stable segment is calculated to obtain the mean difference anisotropy.

[0109] Based on the difference in segment variance between the main channel and the secondary channel, a segment fluctuation comparison sequence is formed. The relative difference in signal fluctuation amplitude between different channels within the same segment is calculated to obtain the fluctuation difference term.

[0110] Based on the segment covariance between the main channel and the secondary channel, a segment cooperative change sequence is constructed, and the degree of mutual dependence of the signals of different channels in the same segment over time is calculated to obtain the relevant difference terms.

[0111] The state separation index is obtained by weighting and fusing the mean difference, fluctuation difference, and correlation difference, and then statistically aggregating all segments.

[0112] In this embodiment of the invention, based on the difference in the segment mean values ​​between the main channel and the secondary channel, an average offset sequence for each segment is established. The relative difference in the signal center position of different channels within the stable segment is calculated, yielding the mean difference term. Using the segment mean as a benchmark, high-frequency noise interference is eliminated, accurately characterizing the difference in the signal center position of the main and secondary channels within the same segment. Based on the difference in the segment variance between the main channel and the secondary channel, a segment fluctuation comparison sequence is formed. The relative difference in the signal fluctuation amplitude of different channels within the same segment is calculated, yielding the fluctuation difference term. The numerical differences in signal amplitude and interference response amplitude between the main and secondary channels are extracted, quantifying the two channels. The system measures the elasticity of the channel under noise interference or vibration response. Based on the segment covariance between the main channel and the secondary channel, a segment cooperative change sequence is constructed. The interdependence of the signals of different channels in the same segment over time is calculated to obtain the relevant difference term. This captures the degree of cooperative response of the main and secondary channels in each period segment, effectively identifying whether there are strong coupling regions in the signal. The mean difference term, fluctuation difference term, and relevant difference term are weighted and fused, and statistically aggregated for all segments to obtain the state separation index. The three different dimensions of channel response differences are unified and fused to construct a single comparable index for subsequent judgment of which segments have higher channel response separation.

[0113] In a preferred embodiment of the present invention, transition identification is performed based on the state separation index to screen out switching sensitive segments, a signal mutation mapping table is constructed, and path anomaly extraction is performed on the segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio, including:

[0114] High-value intervals are identified based on the state separation index. Local average values ​​are calculated using a sliding window, and windows exceeding a preset average threshold are filtered out to obtain an initial set of sensitive segments.

[0115] Based on the initial set of sensitive segments, the original waveforms within the segments are compared point by point with the fitted waveforms at the periodic boundaries to obtain the residual sequence.

[0116] Based on the residual sequence, calculate the magnitude of change between adjacent residuals and mark the points that exceed a preset change threshold to obtain a set of mutation candidate points;

[0117] Based on the set of mutation candidate points, adjacent mutation candidate points are merged into mutation segments, and the start and end positions, average residual amplitude, and waveform offset of each mutation segment are recorded to obtain a signal mutation mapping table.

[0118] In this embodiment of the invention, high-value intervals are identified based on the state separation index. A sliding window is used to statistically analyze local average values ​​and filter windows exceeding a preset average threshold to obtain an initial set of sensitive segments. This enables the location of significantly abnormal local regions in time-series data and automatically identifies time intervals with prominent differences in the responses of the main and secondary channels, providing a regional range for subsequent mutation point location. Based on the initial set of sensitive segments, the original waveforms within each segment are compared point-by-point with the fitted waveforms at the periodic boundaries to obtain residual sequences. This constructs a numerical expression reflecting signal shift behavior, providing data for subsequent mutation identification and structural complexity assessment. The system is based on the residual sequence. It calculates the change amplitude between adjacent residuals and marks the points that exceed the preset change threshold to obtain a set of mutation candidate points. This can extract local jump regions in continuous change trends and avoid noise points being mistaken for valid mutations. Based on the set of mutation candidate points, adjacent mutation candidate points are merged into mutation segments, and the start and end positions, average residual amplitude, and waveform offset of each mutation segment are recorded to obtain a signal mutation mapping table. This table uses structured data to characterize the path shape, anomaly degree, and waveform offset direction of each mutation segment, providing a data foundation for subsequent channel mapping compression ratio calculations.

[0119] Specifically, based on the initial set of sensitive segments, the original waveform within each segment is compared point-by-point with the fitted waveform at the periodic boundary to obtain the residual sequence, which includes:

[0120] For each identified initial sensitive segment, the original signal sequence of the main channel corresponding to that segment is extracted. This sequence covers all sampling points from the start time to the end time of the sensitive segment. Simultaneously, based on the standard period fitting model generated in the period boundary map, a fitted waveform sequence within the same time range is extracted. First, the original and fitted waveforms are time-synchronized, i.e., the time axes of the two sequences are aligned to ensure that each sampling point has a consistent timestamp index, thus ensuring point-to-point matching consistency in subsequent comparison operations. After time alignment, the difference between the original and fitted signals is calculated point-by-point to form a residual sequence. In practice, the residual values ​​can be in absolute form to reflect the waveform deviation magnitude, or the positive and negative directions can be retained to record the offset trend, depending on the subsequent processing requirements. The residual sequence is arranged sequentially with time as the index, forming a one-dimensional residual signal stream, used for subsequent abrupt change point identification and abnormal segment extraction.

[0121] Specifically, based on the set of candidate mutation points, adjacent candidate mutation points are merged into mutation segments, and the start and end positions, average residual amplitude, and waveform offset of each mutation segment are recorded to obtain a signal mutation mapping table, which specifically includes:

[0122] The residual sequence is subjected to first-order differencing to calculate the rate of change of the residuals between consecutive time points, thus constructing a rate of change sequence. A threshold comparison is then performed on this rate of change sequence; when any time point satisfies… (in When the preset mutation detection threshold is used, the time point is marked as a mutation candidate point, and all time points that meet the mutation conditions constitute the mutation candidate point set.

[0123] After obtaining the set of mutation candidate points, iterate through them in time index order, and select all adjacent time differences less than or equal to a set time interval threshold. The mutation points are merged into a single mutation segment. The merging operation employs a segment clustering method. If the time difference between two candidate points is within a threshold range, they are determined to belong to the same mutation event. The time boundary of the mutation segment is constructed based on the start and end points, ultimately forming several non-overlapping mutation segments. Subsequently, feature calculation operations are performed on each mutation segment: first, the average value of the residual sequence within the segment is calculated to describe the average residual amplitude of the mutation segment; then, the arithmetic mean of the residual sequence is calculated as the waveform offset of the segment, indicating whether the mutation behavior is predominantly upward or downward. Finally, the start and end times, average residual amplitude, and waveform offset of each mutation segment are summarized to generate structured record entries, resulting in a signal mutation mapping table.

[0124] In a preferred embodiment of the present invention, transition identification is performed based on the state separation index to screen out switching sensitive sections, a signal mutation mapping table is constructed, and path anomalies are extracted from the sections in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio. The method further includes:

[0125] Based on the original path length and fitted path length of each mutation segment, the relative shortening degree of the two is calculated, and the relative shortening degree is merged according to the original path length to obtain the path compression term;

[0126] Based on the average residual magnitude of each mutation segment, the proportion of the residual magnitude to the fitted path is calculated, and the proportion is constrained according to the length of the fitted path to obtain the residual weight term.

[0127] Based on the waveform offset of each abrupt change segment, the degree of coupling influence of the waveform offset on the fitted path is calculated, and the degree of coupling influence is suppressed according to the original path length to obtain the offset correction term.

[0128] The path compression term, residual weight term, and offset correction term are weighted and fused to obtain the segment compression value;

[0129] Calculate the degree of aggregation of the compression values ​​of each mutation segment and merge the aggregation degrees to obtain the channel mapping compression ratio.

[0130] In this embodiment of the invention, based on the original path length and the fitted path length of each abrupt change segment, the relative shortening degree of the two is calculated, and the relative shortening degree is merged according to the original path length to obtain a path compression term. This quantitatively evaluates the difference between the original signal and the fitted signal, reflecting the compressibility characteristics of the abrupt change segment in terms of morphological structure. Based on the average residual amplitude of each abrupt change segment, the proportion of the residual amplitude to the fitted path is calculated, and the proportion is constrained according to the fitted path length to obtain a residual weight term. This evaluates the proportion of the residual signal in the fitted path, reflecting the fitting difficulty and local anomaly density of the fitting model in a specific abrupt change segment. Based on the waveform offset of each abrupt change segment, the waveform offset is calculated. The coupling effect of the quantity on the fitted path is measured and the coupling effect is suppressed according to the original path length to obtain the offset correction term, which models the offset of the signal path center and reflects the degree of morphological distortion of the local fitted curve. The path compression term, residual weight term and offset correction term are weighted and fused to obtain the segment compression value, realizing the transition from single-dimensional structural index to multi-factor structural constraint index and measuring the complex shape of the signal. The aggregation degree of segment compression values ​​between each abrupt segment is calculated and the aggregation degree is merged to obtain the channel mapping compression ratio, which not only reflects the compressibility of the local abrupt state, but also reflects the distribution density of the abrupt phenomenon between channels in the whole cycle, providing a global evaluation standard for feature tuples.

[0131] In a preferred embodiment of the present invention, a predicted input set is input to a neural prediction model, a switching signal sequence is output, and the model identifies whether it is currently in the optimal signal switching window based on the switching signal sequence, generating control commands to switch the signals between channels, including:

[0132] The switching signal sequence is filtered, and when the marker value at a certain moment is greater than the preset threshold, that moment is recorded as a candidate switching point, thus obtaining a set of candidate switching points;

[0133] Aggregate the candidate switching point set. When the time interval between adjacent candidate switching points is less than a preset interval threshold, merge them into a segment to obtain a candidate switching segment sequence.

[0134] Based on the candidate switching segment sequence, the variance of the switching signal sequence in each segment is calculated. When the variance is lower than the preset stability threshold, the segment is determined to be a stable segment, and a set of stable switching segments is obtained.

[0135] Based on the stable switching segment set, when the duration of a certain segment is greater than the preset minimum duration and there are no breakpoints in the segment that exceed the preset maximum interval, the segment is marked as the optimal signal switching window, and the optimal window index set is obtained.

[0136] Based on the optimal window index set, its start and end positions are bound to the channel switching logic to obtain control commands to control the signal switching between the main channel and the secondary channel.

[0137] In this embodiment of the invention, the switching signal sequence is filtered. When the marker value at a certain moment is greater than a preset confidence threshold, that moment is recorded as a candidate switching point, resulting in a candidate switching point set. High-confidence switching moments are extracted from the continuous probability output, while low-confidence signals are eliminated to avoid false triggering of switching behavior. The candidate switching point set is aggregated. When the time interval between adjacent candidate switching points is less than a preset interval threshold, they are merged into a segment, resulting in a candidate switching segment sequence. The time-discrete candidate points are organized into continuous switching candidate segments to avoid switching responses to occasional spike signals and reduce misjudgments. Based on the candidate switching segment sequence, the variance of the switching signal sequence within each segment is calculated. When the variance is lower than a preset stability threshold, the segment is determined to be a stable segment, resulting in a stable switching segment set. This further filters out segments with relatively low confidence value fluctuations. For short time periods, ensure consistent switching decisions by the model to avoid switching operations in areas of drastic fluctuation in model output confidence values. Based on a set of stable switching segments, when the duration of a segment exceeds a preset minimum duration and there are no breakpoints exceeding a preset maximum interval within that segment, mark that segment as the optimal signal switching window, obtaining an optimal window index set. This ensures that signal switching operations are only performed within areas of stable confidence output without abnormal interruptions, avoiding pseudo-stable segments that are momentarily stable but discontinuous overall. Based on the optimal window index set, bind its start and end positions to the channel switching logic to obtain control commands to control signal switching between the main and secondary channels. After ensuring that the optimal switching window has been clearly defined, automatically issue control commands and complete the switching of physical signal channels, realizing a closed loop between feature-driven, probability output, and action control.

[0138] The preset confidence threshold is set between 0.7 and 0.95, with a recommended value of 0.8. This value originates from the probability distribution characteristics of the neural network output; the label value essentially represents the neural network's confidence in predicting whether a given moment represents a switching point. Considering the inherent fluctuations in neural network false positives, a threshold below 0.7 can easily introduce false switching points, while a threshold above 0.95 may cause missed switching points that should be identified. Therefore, a value around 0.8 strikes a balance between false positive and false negative rates and is an empirical threshold widely used in engineering practice.

[0139] The preset interval threshold depends on the sampling period of the pulse signal and the response time of the device, and is typically set in the range of 5ms to 50ms, with a recommended value of 20ms. This value is used to determine whether two candidate points belong to the same event segment. If the value is set too small, it will cause a large number of consecutive switching points to be split into multiple small segments; if it is set too large, it may merge unrelated events, reducing the accuracy of local judgment. Therefore, the threshold value needs to be set in conjunction with the signal sampling frequency (e.g., 1 kHz corresponds to a sampling interval of 1 ms).

[0140] The preset stability threshold is used to determine whether the switching signal sequence maintains a stable output within a certain segment. Its value ranges from 0.005 to 0.05, with a recommended value of 0.02. This value represents the fluctuation range of the neural network output confidence value within a local segment. When the neural network gives a high consistency in its judgment of whether a switching point is reached (i.e., the variance is below the threshold) within a certain segment, its judgment can be considered relatively reliable. This value should be set based on statistical analysis of the fluctuation of the neural network output, and can be obtained during the training set evaluation phase by taking the confidence variance of the output in the actual switching window.

[0141] The preset minimum duration is used to ensure that the switching window has a sufficiently long effective length in time, preventing occasional high-confidence fluctuations within a short period from being mistaken for a stable segment. This value is set between 100ms and 500ms, with a recommended value of 200ms. The rationale for this setting is that in industrial automation or high-precision rotary measurement scenarios, the switching action between the main and auxiliary channels requires at least one complete signal cycle to provide stable support. If the system's signal cycle is 100ms, then the minimum duration should at least include one complete cycle. 200ms can cover two cycles, helping to filter out transient abnormal responses while preserving an effective switching window.

[0142] The preset maximum interval represents the maximum time span during which a low confidence value can exist, typically set between 30ms and 100ms, with 50ms being the recommended value. This threshold is used to determine whether there is a significant signal interruption or confidence collapse within a stable handover segment. If the predicted value is lower than the preset confidence threshold for a certain period, and this low value persists beyond this interval, the segment is considered to no longer be stable. In system handover determination, if a signal confidence breakpoint occurs, it is usually caused by sudden interference or model hesitation; this segment should not trigger a real handover and should therefore be explicitly excluded.

[0143] Specifically, based on the optimal window index set, its start and end positions are bound to the channel switching logic to obtain control commands that control the signal switching between the main channel and the secondary channel, including:

[0144] Each entry in the optimal window index set contains a start timestamp and an end timestamp, indicating the optimal time for signal channel switching as determined by the system within a certain time period. In implementation, the system first iterates through each entry in the index set, extracts the start and end timestamps for each optimal switching window segment, and establishes a binding relationship between this time period and the system's internal channel control logic, marking this segment as allowed to switch. This binding operation can be represented by a mapping table structure, for example, using time intervals as keys and the bound channel switching operation commands as values, thus establishing a mapping table between timing indexes and control commands.

[0145] During actual operation, the system receives signal outputs from the main channel and the secondary channel in real time, and timestamps the current moment. This timestamp is then used as a query index to retrieve the aforementioned switching control mapping table. If the current timestamp falls within any of the pre-defined optimal switching window segments, and the current signal source of the system is the main channel, then the channel switching operation is automatically executed according to the switching logic, i.e., switching to the secondary channel. Conversely, if the current channel is the secondary channel, then the system switches to the main channel. After this operation is completed, the current channel is locked as the new main channel until the next valid switching window appears.

[0146] When binding channel switching logic, the system can set the response delay time and minimum channel hold time for the switching action to avoid frequent back-and-forth switching of signals at the edge of the optimal switching window. In embedded implementation, this binding process can be controlled by a state machine, setting multiple system states such as "waiting to switch," "switching completed," and "holding." Based on the relationship between the current time and the window interval, as well as changes in the signal source, the corresponding state transition event is triggered, driving the switching execution module to output the corresponding low-level or high-level control signal to the hardware execution circuit, thereby completing the actual physical channel switching.

[0147] To ensure data continuity and system synchronization during the handover process, synchronization commands must be sent to the signal acquisition and processing module before and after the handover action. This ensures that the data processing link updates the current signal source identifier, guaranteeing that subsequent data streams remain consistent with the physical channel logic at the algorithm layer. This completes the closed loop from the prediction model to control command generation and then to physical signal handover.

[0148] Embodiments of the present invention also provide a pulse signal intelligent switching control system based on a neural network, the system comprising:

[0149] The basic module is used to acquire the pulse output sequences of the main channel and the sub-channel within multiple cycles to construct a periodic boundary map, and to perform boundary fitting on the periodic boundary map to generate a list of periodic stable regions based on statistical intervals.

[0150] The first analysis module is used to construct a hetero-frequency interference matrix based on the relative fluctuations between the main channel and the secondary channel in the list of periodically stable regions, and to measure the independent response capability between channels in the local interference region based on the hetero-frequency interference matrix to obtain the state separation index.

[0151] The second analysis module is used to identify transitions based on the state separation index, filter out switching sensitive sections, construct a signal mutation mapping table, extract path anomalies from the sections in the signal mutation mapping table, calculate the compressibility of local waveform fitting, and obtain the channel mapping compression rate.

[0152] The sample module is used to construct a set of feature pairs based on the state separation index and the channel mapping compression ratio, and to normalize the set of feature pairs to generate a training sample set.

[0153] The training module is used to input the training sample set into the neural network model for training, update the parameters of the neural network model, and obtain the neural prediction model.

[0154] The prediction module is used to acquire the data sequence to be judged in real time and perform the same feature processing process as the training sample set on the data sequence to be judged to obtain the prediction input set;

[0155] The instruction module is used to input the predicted input set into the neural prediction model, output the switching signal sequence, identify whether it is currently in the optimal signal switching window based on the switching signal sequence, and generate control instructions to switch the signals between channels.

[0156] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0157] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0158] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0159] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A pulse signal intelligent switching control method based on neural networks, characterized in that, The method includes: The pulse output sequences of the main channel and the sub-channel within multiple cycles are obtained to construct a periodic boundary map. The periodic boundary map is then fitted to the boundary, and a list of periodic stable regions is generated based on statistical intervals. In the list of periodically stable regions, a heterofrequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel. This heterofrequency interference matrix is ​​then used to measure the independent response capability between channels within local interference regions, yielding a state separation index, including: Based on the difference in the average value of the segments between the main channel and the secondary channel, an average offset sequence for each segment is established, and the relative difference in the signal center position of different channels within the stable segment is calculated to obtain the mean difference anisotropy. Based on the difference in segment variance between the main channel and the secondary channel, a segment fluctuation comparison sequence is formed. The relative difference in signal fluctuation amplitude between different channels within the same segment is calculated to obtain the fluctuation difference term. Based on the segment covariance between the main channel and the secondary channel, a segment cooperative change sequence is constructed, and the degree of mutual dependence of the signals of different channels in the same segment over time is calculated to obtain the relevant difference terms. The state separation index is obtained by weighting and fusing the mean difference anomaly, fluctuation difference, and correlation difference, and then statistically aggregating all segments. Transition identification is performed based on the state separation index to screen out switching-sensitive segments, a signal mutation mapping table is constructed, and path anomalies are extracted from segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio, including: Based on the original path length and fitted path length of each mutation segment, the relative shortening degree of the two is calculated, and the relative shortening degree is merged according to the original path length to obtain the path compression term; Based on the average residual magnitude of each mutation segment, the proportion of the residual magnitude to the fitted path is calculated, and the proportion is constrained according to the length of the fitted path to obtain the residual weight term. Based on the waveform offset of each abrupt change segment, the degree of coupling influence of the waveform offset on the fitted path is calculated, and the degree of coupling influence is suppressed according to the original path length to obtain the offset correction term. The path compression term, residual weight term, and offset correction term are weighted and fused to obtain the segment compression value; Calculate the degree of aggregation of the compression values ​​of each mutation segment and merge the aggregation degrees to obtain the channel mapping compression ratio; A set of feature pairs is constructed based on the state separation index and the channel mapping compression ratio, and the set of feature pairs is normalized to generate a training sample set. The training sample set is input into the neural network model for training, the parameters of the neural network model are updated, and a neural prediction model is obtained. The system acquires the data sequence to be judged in real time and performs the same feature processing procedure as the training sample set on the data sequence to be judged to obtain the prediction input set. The predicted input set is fed into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it is determined whether the current signal switching window is in the optimal window, and control commands are generated to switch the signals between channels.

2. The intelligent pulse signal switching control method based on neural networks according to claim 1, characterized in that, A periodic boundary map is constructed by acquiring pulse output sequences of the main and secondary channels within multiple periods. Boundary fitting is then performed on the periodic boundary map, and a list of periodically stable regions is generated based on statistical intervals, including: Time synchronization is performed based on the pulse output sequences of the main channel and the secondary channel, a correspondence is established between the sampling points of different channels, the difference in sampling starting point is eliminated, and an aligned signal sequence is obtained. Based on the aligned signal sequence, the peak and valley values ​​in the aligned signal sequence are identified to mark the start and end positions of a complete pulse cycle, thus obtaining a set of candidate points for the cycle boundary; Based on the set of candidate points for the periodic boundary, the time interval between adjacent candidate points is calculated one by one and the distribution of the intervals is statistically analyzed to obtain the boundary interval sequence. The average interval value and standard deviation are calculated based on the boundary interval sequence. Interval values ​​with a difference from the average interval value greater than twice the standard deviation are excluded, and the remaining boundaries are compensated and corrected to obtain the periodic boundary map. Based on the periodic boundary map, the fluctuation amplitude of the boundary interval of each segment is compared and segments below the preset fluctuation threshold are selected to obtain a list of periodic stable regions.

3. The intelligent pulse signal switching control method based on neural networks according to claim 2, characterized in that, In the list of periodically stable regions, a heterofrequency interference matrix is ​​constructed based on the relative fluctuations between the main channel and the secondary channel. This heterofrequency interference matrix is ​​then used to measure the independent response capability between channels within local interference regions, yielding a state separation index, including: Based on the list of periodically stable regions, the signal sequences of the main channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the main channel. Based on the list of periodically stable regions, the signal sequences of the sub-channel in each periodic segment are extracted, and the mean and variance are calculated to obtain the statistical feature set of the sub-channel. Based on the statistical feature sets of the main channel and the sub-channel, the difference in mean, the difference in variance, and the covariance of each segment are calculated to obtain three types of difference index sequences. By combining the three types of difference index sequences into a matrix, the difference in mean, difference in variance, and covariance of the segments are mapped to the same matrix structure in rows and columns to obtain the heterofrequency interference matrix.

4. The intelligent pulse signal switching control method based on neural networks according to claim 3, characterized in that, Transition identification is performed based on the state separation index to screen out switching-sensitive segments, a signal mutation mapping table is constructed, and path anomalies are extracted from segments in the signal mutation mapping table. The compressibility of local waveform fitting is calculated to obtain the channel mapping compression ratio, including: High-value intervals are identified based on the state separation index. Local average values ​​are calculated using a sliding window, and windows exceeding a preset average threshold are filtered out to obtain an initial set of sensitive segments. Based on the initial set of sensitive segments, the original waveforms within the segments are compared point by point with the fitted waveforms at the periodic boundaries to obtain the residual sequence. Based on the residual sequence, calculate the magnitude of change between adjacent residuals and mark the points that exceed a preset change threshold to obtain a set of mutation candidate points; Based on the set of mutation candidate points, adjacent mutation candidate points are merged into mutation segments, and the start and end positions, average residual amplitude, and waveform offset of each mutation segment are recorded to obtain a signal mutation mapping table.

5. The intelligent pulse signal switching control method based on neural networks according to claim 4, characterized in that, The predicted input set is fed into the neural prediction model, which outputs a switching signal sequence. Based on the switching signal sequence, it identifies whether the current signal switching window is optimal and generates control commands to switch the signals between channels, including: The switching signal sequence is filtered, and when the marker value at a certain moment is greater than the preset threshold, that moment is recorded as a candidate switching point, thus obtaining a set of candidate switching points; Aggregate the candidate switching point set. When the time interval between adjacent candidate switching points is less than a preset interval threshold, merge them into a segment to obtain a candidate switching segment sequence. Based on the candidate switching segment sequence, the variance of the switching signal sequence in each segment is calculated. When the variance is lower than the preset stability threshold, the segment is determined to be a stable segment, and a set of stable switching segments is obtained. Based on the stable switching segment set, when the duration of a certain segment is greater than the preset minimum duration and there are no breakpoints in the segment that exceed the preset maximum interval, the segment is marked as the optimal signal switching window, and the optimal window index set is obtained. Based on the optimal window index set, its start and end positions are bound to the channel switching logic to obtain control commands to control the signal switching between the main channel and the secondary channel.

6. A pulse signal intelligent switching control system based on neural networks, characterized in that, The system is used to perform the method as described in any one of claims 1 to 5, the system comprising: The basic module is used to acquire the pulse output sequences of the main channel and the sub-channel within multiple cycles to construct a periodic boundary map, and to perform boundary fitting on the periodic boundary map to generate a list of periodic stable regions based on statistical intervals. The first analysis module is used to construct a hetero-frequency interference matrix based on the relative fluctuations between the main channel and the secondary channel in the list of periodically stable regions, and to measure the independent response capability between channels in the local interference region based on the hetero-frequency interference matrix to obtain the state separation index. The second analysis module is used to identify transitions based on the state separation index, filter out switching sensitive sections, construct a signal mutation mapping table, extract path anomalies from the sections in the signal mutation mapping table, calculate the compressibility of local waveform fitting, and obtain the channel mapping compression rate. The sample module is used to construct a set of feature pairs based on the state separation index and the channel mapping compression ratio, and to normalize the set of feature pairs to generate a training sample set. The training module is used to input the training sample set into the neural network model for training, update the parameters of the neural network model, and obtain the neural prediction model. The prediction module is used to acquire the data sequence to be judged in real time and perform the same feature processing process as the training sample set on the data sequence to be judged to obtain the prediction input set; The instruction module is used to input the predicted input set into the neural prediction model, output the switching signal sequence, identify whether it is currently in the optimal signal switching window based on the switching signal sequence, and generate control instructions to switch the signals between channels.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images