Clock frequency offset information fast compensation method based on neural network model

Abstract

This application relates to the fields of time frequency and time synchronization technology, specifically, to a method for rapid compensation of clock frequency offset information based on a neural network model, including the following steps: Time calibration stage: Connecting a time reference source, using the same phase-locked unit to perform phase modulation on the main oscillator and the sub-oscillator respectively, generating the main reference voltage information of the main oscillator and the sub-reference voltage information of the sub-oscillator; monitoring the temperature and vibration information of the main oscillator during phase modulation; This application realizes real-time compensation of frequency offset caused by interference factors such as temperature and vibration during the absence of satellite correction window through dynamic learning of neural network; Utilizing the parallel compensation and confidence evaluation mechanism of dual oscillators, the optimal time source is dynamically selected, significantly improving the autonomous timekeeping accuracy and signal output stability of the system when an external time reference is missing.

Description

Technical Field

[0001] This application relates to the fields of time frequency and time synchronization technology, and more specifically, to a method for fast compensation of clock frequency offset information based on a neural network model. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] The clock signal is the reference signal for timing control in electronic systems. It provides the core rhythm for data synchronization, task scheduling, and precise triggering, and its accuracy directly determines the accuracy and stability of the entire system. If the clock signal deviates, it will lead to serious consequences such as data sampling errors, communication synchronization failures, and inaccurate positioning.

[0004] High-precision clock systems often rely on Global Navigation Satellite Systems (GNSS) such as BeiDou for periodic standard time correction, for example, by receiving 1PPS pulse signals to align with the UTC national time standard. Because satellite signals are susceptible to obstruction, weather interference, or equipment failure, there will inevitably be a time window of several hours to several days without external correction between two effective time synchronizations.

[0005] During periods without satellite signal coverage, frequency skew caused by factors such as temperature, voltage, and aging in local clock sources (e.g., crystal oscillators, atomic clocks) cannot be detected and corrected in a timely manner, accumulating over time to form significant deviations. Existing compensation methods based on fixed models or preset parameters struggle to dynamically adapt to the continuous changes in these frequency skews within the window, resulting in the inability to effectively suppress clock errors accumulated during this period. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a fast compensation method for clock frequency offset information based on a neural network model to solve the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a method for fast compensation of clock frequency offset information based on a neural network model, including the following steps: Time calibration phase: By connecting to a time reference source, the same phase-locked unit is used to perform phase modulation on the main oscillator and the sub-oscillator respectively, generating the main reference voltage information of the main oscillator and the sub-reference voltage information of the sub-oscillator. The temperature and vibration information of the master oscillator during phase modulation are monitored to generate master monitoring information; The temperature and vibration information of the secondary oscillator are monitored during phase modulation to generate secondary monitoring information; The main monitoring voltage information and main monitoring information are used as training data to train the neural network model and obtain the first model parameters. The secondary monitoring voltage information and secondary monitoring information are used as training data to train the neural network model and obtain the second model parameters. Adaptive calibration phase: The temperature and vibration information of the master oscillator are monitored in real time to generate master offset information. The master offset information is input into a neural network model loaded with the first model parameters to obtain the master voltage information of the master oscillator. The master voltage information is input into the master oscillator to generate the master signal. The temperature and vibration information of the sub-oscillator are monitored in real time to generate sub-offset information. The sub-offset information is input into a neural network model loaded with the second model parameters to obtain the sub-voltage information of the sub-oscillator. The sub-voltage information is input into the sub-oscillator to generate the sub-signal. The first feature information is obtained by extracting the signal features of the main signal, and the second feature is obtained by extracting the signal features of the sub-signal. The first feature signal, the second feature signal, the main offset information, and the secondary offset information are input into the confidence neural network model to generate the confidence of the main signal for the corresponding time period and the confidence of the secondary signal for the corresponding time period. For each time period, the signal with high confidence is used as the time source of the system.

[0009] This application achieves real-time compensation for frequency offsets caused by interference factors such as temperature and vibration during the period without satellite correction by using dynamic learning of neural networks; by using a dual-oscillator parallel compensation and confidence evaluation mechanism, the optimal time source is dynamically selected, which significantly improves the autonomous timekeeping accuracy and signal output stability of the system when an external time reference is missing.

[0010] Furthermore, during the adaptive calibration phase, the phase difference between the main signal and the secondary signal is compared in real time. When the phase difference exceeds the threshold, the signal with higher reliability built into the time period is used as the time source, and the other signal is calibrated.

[0011] This application ensures the synchronization and consistency of the two time sources by real-time monitoring of the phase difference between the main and secondary signals and selecting the time source with higher confidence when the threshold is exceeded, while forcibly calibrating the other signal path, thus avoiding the accumulation of system time base divergence and persistent deviation caused by sudden drift of a single path.

[0012] Furthermore, the main oscillator is located in the main signal circuit, and the secondary oscillator is located in the secondary signal circuit. The main signal circuit and the secondary signal circuit are connected to the system's time source interface through a high-speed switching module. When the high-speed switching module is in the first state, the main signal circuit is connected to the system's time source interface. When the high-speed switching module is in the second state, the sub-oscillator is connected to the system's time source interface.

[0013] This application achieves physical isolation and lossless switching of the main and auxiliary signal circuits through a high-speed switching module. When the main oscillator or auxiliary oscillator experiences a momentary anomaly in a single path, nanosecond-level time source switching can be realized, avoiding the signal-free period during the switching time window and ensuring the physical continuity of the system's time output and its resistance to single-point failure.

[0014] Furthermore, the temperature information includes the current temperature h. t and the temperature rise coefficient H t ; ; in, Indicates time t before the current time. temperature, This indicates the time interval for temperature measurements.

[0015] This application introduces a temperature rise coefficient as a parameter of temperature change rate, enabling the neural network model to directly correlate the temperature dynamic characteristics of oscillator frequency drift and eliminate prediction lag bias caused by the accumulation of temperature gradient.

[0016] Furthermore, vibration information refers to vibration characteristics that characterize the current vibration energy and vibration frequency. These vibration characteristics include peak energy intensity, energy rise slope, and energy duration.

[0017] This application expands the vibration characteristics to three dimensions: peak energy intensity, energy rise slope, and energy duration. This enables the neural network model to fully capture the differentiated response mechanisms of vibration impact intensity, energy mutation rate, and decay duration to the frequency deviation characteristics of the oscillator, thus avoiding the failure of compensation in high-frequency oscillation scenarios caused by a single amplitude index.

[0018] Furthermore, the vibration features are extracted as follows: S1: Real-time acquisition of vibration data x[n] of the monitored area within the monitoring period, n=0,1,2...N-1, where N represents the signal length; S2: Calculate the energy sequence of vibration data ; ; S3: Extract vibration features; ; ; ; in, Peak energy intensity, The slope of energy rise. For energy duration, k represents the relative position offset. Let k represent the mean of k, and p represent the index of the peak position. This represents the energy value at position p offset by k. Let S represent the average energy of a local region, and let S represent the set of high-energy sample points. This represents the total energy of the high-energy region, and m represents the index of the high-energy sample point.

[0019] Furthermore, neural network models include: The input layer is used to standardize the main offset information and generate standard features; The feature encoding layer uses a fully connected network to extract nonlinear features from standard features and generate high-dimensional feature vectors. The spatiotemporal feature extraction layer uses dilated convolution and LSTM modules to extract the temporal information of high-dimensional feature vectors to obtain high-dimensional spatiotemporal features; The physical feature interaction layer obtains the main offset information, crosses the sub-features in the main offset information to generate cross features, and concatenates the cross features with high-dimensional spatiotemporal features to generate concatenated features. The residual compensation output layer concatenates the spliced ​​features and high-dimensional feature vectors, generates the compensation amount using a fully connected network, and then outputs the main voltage adjustment value using a linear output function.

[0020] This application achieves synergistic effects in the input standardization, feature encoding, dilated convolution and spatiotemporal feature extraction, physical feature crossing and residual output stages of the LSTM module through the hierarchical structure design of the neural network model. It efficiently models the nonlinear temporal relationship between environmental parameters and frequency offset, improves the prediction accuracy of voltage adjustment values, and thus reduces the cumulative error of clock frequency compensation.

[0021] Furthermore, the sub-features include: peak energy intensity, energy rise slope, energy duration, current temperature, and temperature rise coefficient; Cross features include at least: The cross term between current temperature and peak energy intensity; The interaction term between the current temperature and the slope of energy increase; The interaction between current temperature and energy duration; The interaction term between the temperature rise coefficient and peak energy intensity; The interaction term between the temperature rise coefficient and the energy rise slope; The interaction term between the temperature rise coefficient and the energy duration.

[0022] This application constructs six sets of temperature-vibration coupling characteristic cross terms to explicitly establish a multi-dimensional dynamic coupling relationship between temperature and vibration in terms of amplitude, rate, and duration. This enables the neural network compensation model to accurately match the physical mechanism of frequency drift caused by both temperature gradient and vibration shock, thus solving the problem of model inaccuracy caused by neglecting the interaction in single environmental parameter compensation.

[0023] Furthermore, confidence neural network models include: The signal feature input module uses dual channels to input the first signal feature and the second signal feature, and processes them separately into a first feature vector and a second feature vector. The environmental feature input module uses dual-channel input of main offset information and secondary offset information, and uses an environmental encoder to output an environmental confidence vector. The feature comparison module calculates signal quality differences, environmental disturbance differences, and signal alignment differences to generate difference information. The information fusion layer uses a gating attention mechanism to weight key features and generate confidence information. The confidence output layer uses a primary confidence predictor and a secondary confidence predictor to independently process the confidence information and generate the confidence of the first signal feature and the confidence of the second signal feature, respectively.

[0024] This application, through the structured design of a confidence neural network model, achieves decoupled evaluation and anti-interference decision-making of the reliability of the main and auxiliary signals through the synergistic effects of independent encoding of dual-channel signal features, quantization mapping of environmental disturbances, multi-dimensional difference calculation, and gating feature fusion. It effectively distinguishes between instantaneous frequency deviation caused by environmental mutations and systemic failures caused by hardware degradation.

[0025] Furthermore, signal characteristics include time-domain jitter, frequency shift, and phase noise. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a clock system.

[0027] Figure 2 This is a flowchart of a method for fast compensation of clock frequency offset information based on a neural network model. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0029] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0030] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0031] In a system, a master signal is required as a timing reference. The master signal directly ensures the logical consistency of data synchronous acquisition, transmission protocol parsing and timing-driven task triggering of each unit in the system under the control of clock tick, and prevents data misalignment, state machine disorder and execution sequence conflict caused by clock deviation.

[0032] The master oscillator generates the master signal, but during operation, temperature and vibration can cause frequency shifts, leading to clock deviations. Existing technology provides a UTC standard time reference through a reference time source (such as the 1PPS signal from a BeiDou receiver). The phase-locked loop (PLL) unit uses a phase detector to compare the phase difference Δφ between the reference signal and the oscillator output signal, outputting an error voltage signal. After loop filtering to suppress high-frequency noise, a control voltage Vc is generated and applied to the oscillator's voltage-controlled terminal. Frequency correction is achieved by adjusting the varactor diode bias until a phase synchronization lock state is reached, establishing a linear relationship between the standard control voltage and the oscillation frequency. This is a general overview of how the oscillator corrects its frequency; specific correction methods are existing technologies and will not be discussed further here.

[0033] The acquisition of a reference time source is limited by environmental interference and equipment availability, and can only be connected during the periodic correction window; the clock offset continues to accumulate during the gap between two effective corrections, leading to a deterioration in system accuracy. Embodiment 1 of this application provides a clock system.

[0034] like Figure 1 As shown, the clock system includes a main oscillator, a first phase-locked loop, a high-speed switching module, a system time source, a secondary oscillator, a second phase-locked loop, an environmental sensing module, a neural network compensation engine, a confidence decision unit, and a crystal oscillator voltage control terminal.

[0035] The main oscillator is located in the main signal circuit, and the secondary oscillator is located in the secondary signal circuit. The main signal circuit and the secondary signal circuit are connected to the system's time source interface through a high-speed switching module. When the high-speed switching module is in the first state, the main signal circuit is connected to the system's time source interface. When the high-speed switching module is in the second state, the sub-oscillator is connected to the system's time source interface.

[0036] Specifically, the voltage control terminal of the main oscillator is connected to the main output terminal (main DAC) of the crystal oscillator voltage control terminal, and its signal output terminal is connected in sequence to the reference input terminal of the first phase-locked loop, the port P1 of the high-speed switching module, and finally connected to the system time source interface through the COM common terminal; The voltage control terminal of the secondary oscillator is connected to the secondary output terminal (secondary DAC) of the crystal oscillator voltage control terminal. Its signal output terminal is connected in sequence to the reference input terminal of the second phase-locked loop, the port P2 of the high-speed switching module, and connected to the system time source interface through the COM common terminal. The environmental sensing module includes two sets of temperature and vibration probes, which can collect temperature and vibration data, and are respectively mounted on the main oscillator and the auxiliary oscillator. The environmental sensing module is connected to the neural network compensation engine. The neural network compensation engine has a built-in neural network model, which, after training, can generate compensation signals to the crystal oscillator voltage control terminal, and the crystal oscillator voltage control terminal sends the compensation signals to the corresponding oscillator.

[0037] The confidence decision unit can receive temperature and vibration data from the environmental sensing module, as well as the signal quality generated by the main oscillator and the sub-oscillator, and make corresponding decisions to control the working state of the high-speed switch.

[0038] The principle of the entire system is as follows: both the master oscillator and the slave oscillator are used to generate timing signals, which are input to the system time source interface through a high-speed switching module. Under normal circumstances, the high-speed switching module (first state) inputs the signal generated by the master oscillator to the time source interface. When the signal generated by the master oscillator is distorted, the high-speed switching module (second state) inputs the signal generated by the slave oscillator to the time source interface.

[0039] The first and second phase-locked loops (PLLs) are used for timing adjustments of the master oscillator and the slave oscillator, respectively. That is, when the system is connected to an accurate external time source, the frequencies of the master and slave oscillators are calibrated separately. When the external time source is removed, the signals generated by the master and slave oscillators are compared, and the optimal signal is input to the time source interface. The specific operating logic is a fast clock frequency offset compensation method based on a neural network model.

[0040] refer to Figure 2 Fast compensation methods for clock frequency offset information based on neural network models include: Time calibration phase: By connecting to a time reference source, the same phase-locked unit is used to perform phase modulation on the main oscillator and the sub-oscillator respectively, generating the main reference voltage information of the main oscillator and the sub-reference voltage information of the sub-oscillator. The temperature and vibration information of the master oscillator during phase modulation are monitored to generate master monitoring information; The temperature and vibration information of the secondary oscillator are monitored during phase modulation to generate secondary monitoring information; The main monitoring voltage information and main monitoring information are used as training data to train the neural network model and obtain the first model parameters. The secondary monitoring voltage information and secondary monitoring information are used as training data to train the neural network model and obtain the second model parameters.

[0041] As independent physical devices, the master oscillator and slave oscillator exhibit inherent individual differences in their frequency response to environmental factors (temperature, vibration) and their own electrical characteristics (such as voltage control characteristics). Therefore, it is necessary to model the highly nonlinear relationship between the frequency deviation characteristics of the master and slave oscillators and their operating environment and required control voltage separately. By using "master monitoring voltage information + master monitoring information" and "slave monitoring voltage information + slave monitoring information" as training data, the trained first and second model parameters represent the weights of the neural network models established for the master and slave oscillators, respectively, which accurately map their unique physical response behaviors. This separate modeling method ensures that during the adaptive calibration phase, the environmental parameter input can be adapted to the corresponding individual oscillator characteristics to the greatest extent, outputting the most accurate compensation voltage and achieving high-frequency precision compensation.

[0042] Specifically: During the time calibration phase, when a time reference source is connected and the master oscillator is subjected to phase-locked loop (the time reference source is connected to the first phase-locked loop), master monitoring information (including temperature and vibration information of the master oscillator) and master reference voltage information applied to the master oscillator are simultaneously acquired. Using the acquired master monitoring information as input conditions for the model and the corresponding master reference voltage information as the desired output target, these paired data are used to train the initial neural network model (considered a black box). By learning a large number of such input-output mapping relationships, the model adjusts its internal parameters (i.e., the first model parameters) so that when receiving new, similar master monitoring information (temperature, vibration) input, it can accurately predict the master reference voltage information required to maintain or restore a stable frequency on the master oscillator. In other words, when the same master reference voltage information appears again, it can provide the closest possible master reference voltage information.

[0043] The generation method for the second model parameters is roughly the same: When a time reference source is connected (synchronously connected to the second phase-locked loop) and the sub-oscillator is subjected to phase-locked regulation, the system synchronously acquires sub-monitoring information and sub-reference voltage information. Using the acquired series of sub-monitoring information as input conditions for the model, and the corresponding sub-reference voltage information as the desired output target, this paired data is used to train another initial neural network model (also considered a black box). By learning a large number of such input-output mapping relationships, the model ultimately adjusts its internal parameters (i.e., the second model parameters) so that it can accurately predict the amount of compensation voltage required to maintain or restore a stable frequency when receiving new, similar sub-monitoring information (temperature, vibration) input.

[0044] The time calibration phase primarily involves using a high-precision external time reference source (such as a GNSS signal) to stabilize the main oscillator and the secondary oscillator through a standard phase-locked loop (PLL) control process, while simultaneously acquiring the core "environment-control voltage" mapping data under stable PLL conditions. This establishes and solidifies a precise prediction model (i.e., first model parameters and second model parameters) for the "environmental disturbance characteristics -> optimal compensation voltage" when the main and secondary oscillators operate independently, enabling adaptive compensation in the absence of an external reference.

[0045] The master oscillator and the slave oscillator are two identical devices, but they are installed in different locations and are from different production batches. Therefore, they can be trained using the same neural network model. However, different model parameters are needed for compensation.

[0046] Adaptive calibration phase: The temperature and vibration information of the master oscillator are monitored in real time to generate master offset information. The master offset information is input into a neural network model loaded with the first model parameters to obtain the master voltage information of the master oscillator. The master voltage information is input into the master oscillator to generate the master signal. The temperature and vibration information of the sub-oscillator are monitored in real time to generate sub-offset information. The sub-offset information is input into a neural network model loaded with the second model parameters to obtain the sub-voltage information of the sub-oscillator. The sub-voltage information is input into the sub-oscillator to generate the sub-signal. The first feature information is obtained by extracting the signal features of the main signal, and the second feature is obtained by extracting the signal features of the sub-signal. The first feature signal, the second feature signal, the main offset information, and the secondary offset information are input into the confidence neural network model to generate the confidence of the main signal for the corresponding time period and the confidence of the secondary signal for the corresponding time period. For each time period, the signal with high confidence is used as the time source of the system.

[0047] Specifically: During the adaptive calibration phase (i.e., the window period without an external time reference), the environmental physical quantities (temperature information, vibration information) of the master and slave oscillators are monitored in real time to generate master offset information and slave offset information respectively. Then, the neural network models loaded with their respective dedicated model parameters (first model parameters and second model parameters) receive the corresponding offset information as input, independently predict and output the main voltage adjustment value and the secondary voltage adjustment value, and immediately apply them to the corresponding oscillators to offset the influence of environmental disturbances on their frequency in real time. Then the main oscillator and the secondary oscillator each output timing signals.

[0048] The system extracts the time-frequency characteristics (such as time-domain jitter, frequency offset, and phase noise) of the output signals of the main and auxiliary oscillators in real time as the first and second feature information. These two signal features, along with their corresponding environmental offset information, are fed into the confidence neural network model. The model comprehensively evaluates and outputs the independent confidence scores of the main and auxiliary signals within the current time period. During each decision time period, the system automatically selects and seamlessly switches the oscillation signal with the higher confidence (main signal or auxiliary signal) as the system's time reference output through a high-speed switching module, thereby ensuring that a high-precision and stable local time signal can be maintained even without an external reference.

[0049] Furthermore, during the adaptive calibration phase, the phase difference between the main signal and the secondary signal is compared in real time. When the phase difference exceeds the threshold, the signal with higher reliability built into the time period is used as the time source, and the other signal is calibrated.

[0050] Specifically, when real-time comparison detects that the phase difference between the primary and secondary signals exceeds a preset threshold: a control command is immediately sent to the high-speed switching module to switch the system's time source interface to the output signal of the phase-locked loop with higher confidence (e.g., primary signal: first phase-locked loop → high-speed switch → system time source interface). A forced phase-locked calibration command is sent to the second phase-locked loop of another path (e.g., the secondary signal path) (flow: forced correction command of the "core processing unit" → second phase-locked loop), using the phase / frequency of the current high-confidence signal as the reference target for strong phase-locked correction (essentially temporarily changing the reference source of the second phase-locked loop), thereby quickly pulling the secondary signal that is identified as having significantly deviated back to synchronization.

[0051] Example 2: Example 2 provides a neural network model based on Example 1. The neural network model is deployed in the neural network compensation engine. The key requirement of this application is a neural network model used to process the main offset signal and the secondary offset signal. For the processing of each, the neural network model is trained separately. Therefore, this application uses the generation of the main offset signal as an example to explain in detail the structure of the neural network model.

[0052] The master offset information is actually the master monitoring information. The master monitoring information is the name during training, while the master offset information is the name during usage.

[0053] The main offset information includes temperature information and vibration information.

[0054] Temperature information includes the current temperature h t and the temperature rise coefficient H t ; ; in, Indicates time t before the current time. temperature, This indicates the time interval for temperature measurements.

[0055] Vibration information refers to vibration characteristics that characterize the current vibration energy and vibration frequency. Vibration characteristics include peak energy intensity, energy rise slope, and energy duration.

[0056] The vibration features are extracted as follows: S1: Real-time acquisition of vibration data x[n] of the monitored area within the monitoring period, n=0,1,2...N-1, where N represents the signal length; S2: Calculate the energy sequence of vibration data ; ; S3: Extract vibration features; ; ; ; in, Peak energy intensity, The slope of energy rise. For energy duration, k represents the relative position offset. Let k represent the mean of k, and p represent the index of the peak position. This represents the energy value at position p offset by k. Let S represent the average energy of a local region, and let S represent the set of high-energy sample points. This represents the total energy of the high-energy region, and m represents the index of the high-energy sample point.

[0057] The neural network model takes principal offset information as input and outputs principal voltage information. The structure of the neural network model is as follows: Neural network models include: The input layer is used to standardize the main offset information and generate standard features. ; ; The symbol represents the transpose, and D represents the total number of dimensions of the main offset information. This represents the standardized value of the feature in the D-th dimension of the main offset information. In this scheme, D=5, which means: the current temperature h t and the temperature rise coefficient H t Peak energy intensity, energy rise slope, and energy duration.

[0058] The feature encoding layer uses a fully connected network to extract nonlinear features from standard features and generate high-dimensional feature vectors. ; in: Represents a high-dimensional feature vector. Represents the weight matrix. N represents the number of hidden units in the feature coding layer. Represents the bias vector; Indicates the activation function; The spatiotemporal feature extraction layer uses dilated convolution and LSTM modules to extract the temporal information of high-dimensional feature vectors to obtain high-dimensional spatiotemporal features; Specifically, the spatiotemporal feature extraction layer includes a dilated convolution module and an LSTM module. The dilated convolution module is used to extract convolutional features from high-dimensional feature vectors, and the LSTM module is used to extract the temporal information of the convolutional features to obtain high-dimensional spatiotemporal features. ; in, This represents the gating parameters of the LSTM module, used to control the input gate, forget gate, and output gate. This is the output of the LSTM module from the previous time step. This is the output of the LSTM module at the current time step. As a high-dimensional spatiotemporal feature, The convolutional features at time step t; ; in, The parameter 'd' represents the dilation kernel parameter, 'd' represents the dilation coefficient, and 'C' represents the convolution output image. This represents dilated convolution; Where C is a two-dimensional matrix, the column index is the time, and the row index is the number of channels; That is to say Let be the feature vector of all channels in column t of C.

[0059] The physical feature interaction layer obtains the main offset information, crosses the sub-features in the main offset information to generate cross features, and concatenates the cross features with high-dimensional spatiotemporal features to generate concatenated features. Sub-features include: peak energy intensity, energy rise slope, energy duration, current temperature, and temperature rise coefficient; Cross features include: The cross term between current temperature and peak energy intensity; The interaction term between the current temperature and the slope of energy increase; The interaction between current temperature and energy duration; The interaction term between the temperature rise coefficient and peak energy intensity; The interaction term between the temperature rise coefficient and the energy rise slope; The interaction term between the temperature rise coefficient and the energy duration.

[0060] Obtain all cross features to get the cross feature vector. ; ; ; Where i and j are both indices of the sub-features, i ≠ j. Let T be the sub-feature set, and T be the matrix transformation symbol. This represents the k-th cross feature, where k denotes the index of the cross feature. Feature splicing is as follows: ; in, Indicates splicing characteristics, Represents high-dimensional spatiotemporal features. Represents the cross feature vector; The residual compensation output layer concatenates the spliced ​​features and high-dimensional feature vectors, generates the compensation amount using a fully connected network, outputs the main voltage adjustment value using a linear output function, and adds the main voltage adjustment value to the current main voltage to obtain the main voltage information. Specifically: ; The features are the concatenated features and the high-dimensional feature vectors. Indicates splicing characteristics, Represents a high-dimensional feature vector; ; This represents the first residual weight matrix. This represents the second weight matrix. This represents the first residual bias term. This represents the second residual bias term. This is the residual voltage; ; This is the output main voltage information. This is the current voltage.

[0061] Example 3: Example 3 provides a convolutional network model, which is the core component of the confidence decision unit.

[0062] Convolutional network models play the role of efficient compression and deep representation extraction of environmental disturbance features in confidence neural networks: they perform nonlinear transformation and dimensionality reduction mapping on environmental features through a shared environmental encoder (generating a low-dimensional environmental confidence vector from high-dimensional main / secondary offset information), and use convolution operations to capture the local spatial correlation and cross-dimensional dependency between multi-dimensional environmental parameters (temperature, vibration intensity, temperature rise rate, vibration energy slope, etc.), thereby generating environmental disturbance intensity and decoupled feature vectors as the basis for confidence decisions.

[0063] Confidence neural network models include: The signal feature input module uses dual channels to input the first signal feature and the second signal feature, and processes them separately into a first feature vector and a second feature vector. In the first channel: ; ; In the second channel: ; ; in, Let denot be the first feature vector, and GELU be the GELU activation function. Indicates the first signal encoding weight. This represents the first signal encoding bias term. Indicates the characteristics of the first signal. This indicates time-domain jitter of the main signal. Indicates the frequency offset of the main signal. The main signal phase noise is represented by T, which represents the matrix transpose. Indicates the second signal encoding weight. This represents the second signal encoding bias term. Indicates the characteristics of the second signal. These represent the sub-signal time-domain jitter, sub-signal frequency offset, and sub-signal phase noise, respectively. The environmental feature input module uses dual-channel input of main offset information and secondary offset information, and uses an environmental encoder to output an environmental confidence vector. ; ; in, Indicates the environmental coding weight. This indicates the environment coding bias term. Indicates the main offset information. This represents the environmental credibility vector of the main signal. Indicates secondary offset information. This represents the environmental credibility vector of the secondary signal; The feature comparison module calculates signal quality differences, environmental disturbance differences, and signal alignment differences to generate difference information. ; ; ; ; in, Indicates differences in signal quality. Represents the hyperbolic tangent function. Represents the first eigenvector. Represents the second eigenvector. This indicates differences in environmental disturbances. This represents the environmental credibility vector of the main signal. This represents the environmental credibility vector of the secondary signal. Indicates the signal difference amplification factor. This represents the amplification factor for environmental differences. Indicates the minimum value (to prevent division by zero errors); Indicates signal alignment; Indicates differences; The information fusion layer uses a gating attention mechanism to weight key features and generate confidence information. The information fusion process involves the information fusion layer concatenating the first feature vector, the second feature vector, the environmental credibility vector of the main signal, the environmental credibility vector of the sub-signal, and the difference information to obtain the aggregated feature. ; ; in, Represents the first eigenvector. Represents the second eigenvector. This represents the environmental credibility vector of the main signal. This represents the environmental credibility vector of the secondary signal. Indicates differences; Temporal feature extraction process: The information fusion layer extracts the temporal features of the aggregated features to obtain confidence information. ; ; ; ; in, This represents the activation function. Indicates element-wise multiplication. , These represent the first confidence weight and the second confidence weight of the information fusion layer, respectively. , These represent the first confidence bias term and the second confidence bias term of the information fusion layer, respectively. This represents the Sigmoid activation function.

[0064] The confidence output layer uses a primary confidence predictor and a secondary confidence predictor to independently process the confidence information and generate the confidence of the first signal feature and the confidence of the second signal feature, respectively.

[0065] ; ; in, The confidence level of the first signal characteristic is indicated. This represents the Sigmoid activation function. Indicates the confidence weight of the main signal. This represents the confidence bias term of the main signal. The confidence level representing the second signal characteristic. Indicates the confidence weight of the secondary signal. This represents the confidence bias term for the secondary signal.

[0066] The methods for extracting signal characteristics, including time-domain jitter, frequency shift, and phase noise, are existing technologies and will not be described further here.

[0067] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for fast compensation of clock frequency offset information based on a neural network model, characterized in that, include: Time calibration phase: By connecting to a time reference source, the same phase-locked unit is used to perform phase modulation on the main oscillator and the sub-oscillator respectively, generating the main reference voltage information of the main oscillator and the sub-reference voltage information of the sub-oscillator; The temperature and vibration information of the master oscillator during phase modulation are monitored to generate master monitoring information; The temperature and vibration information of the secondary oscillator are monitored during phase modulation to generate secondary monitoring information; The main monitoring voltage information and main monitoring information are used as training data to train the neural network model and obtain the first model parameters. The secondary monitoring voltage information and secondary monitoring information are used as training data to train the neural network model and obtain the second model parameters. Adaptive calibration phase: The temperature and vibration information of the master oscillator are monitored in real time to generate master offset information. The master offset information is input into a neural network model loaded with the first model parameters to obtain the master voltage information of the master oscillator. The master voltage information is input into the master oscillator to generate the master signal. The temperature and vibration information of the sub-oscillator are monitored in real time to generate sub-offset information. The sub-offset information is input into a neural network model loaded with the second model parameters to obtain the sub-voltage information of the sub-oscillator. The sub-voltage information is input into the sub-oscillator to generate the sub-signal. The first feature information is obtained by extracting the signal features of the main signal, and the second feature is obtained by extracting the signal features of the sub-signal. The first feature signal, the second feature signal, the main offset information, and the secondary offset information are input into the confidence neural network model to generate the confidence of the main signal for the corresponding time period and the confidence of the secondary signal for the corresponding time period. For each time period, the signal with high confidence is used as the time source of the system.

2. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, During the adaptive calibration phase, the phase difference between the main signal and the secondary signal is compared in real time. When the phase difference exceeds the threshold, the signal with higher reliability built in that time period is used as the time source, and the other signal is calibrated.

3. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, The main oscillator is located in the main signal circuit, and the secondary oscillator is located in the secondary signal circuit. The main signal circuit and the secondary signal circuit are connected to the system's time source interface through a high-speed switching module. When the high-speed switching module is in the first state, the main signal circuit is connected to the system's time source interface. When the high-speed switching module is in the second state, the sub-oscillator is connected to the system's time source interface.

4. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, Temperature information includes the current temperature h t and the temperature rise coefficient H t ; ； in, Indicates time t before the current time. temperature, This indicates the time interval for temperature measurements.

5. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, Vibration information refers to vibration characteristics that characterize the current vibration energy and vibration frequency. Vibration characteristics include peak energy intensity, energy rise slope, and energy duration.

6. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 5, characterized in that, The vibration features are extracted as follows: S1: Real-time acquisition of vibration data x[n] of the monitored area within the monitoring period, n=0,1,2...N-1, where N represents the signal length; S2: Calculate the energy sequence of vibration data ; ; S3: Extract vibration features; ； ; ; in, Peak energy intensity The slope of energy rise. For energy duration, k represents the relative position offset. Let k represent the mean of k, and p represent the index of the peak position. This represents the energy value at position p offset by k. Let S represent the average energy of a local region, and let S represent the set of high-energy sample points. This represents the total energy of the high-energy region, and m represents the index of the high-energy sample point.

7. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, Neural network models include: The input layer is used to standardize the main offset information and generate standard features; The feature encoding layer uses a fully connected network to extract nonlinear features from standard features and generate high-dimensional feature vectors. The spatiotemporal feature extraction layer uses dilated convolution and LSTM modules to extract the temporal information of high-dimensional feature vectors to obtain high-dimensional spatiotemporal features; The physical feature interaction layer obtains the main offset information, crosses the sub-features in the main offset information to generate cross features, and concatenates the cross features with high-dimensional spatiotemporal features to generate concatenated features. The residual compensation output layer concatenates the spliced ​​features and high-dimensional feature vectors, generates the compensation amount using a fully connected network, and then outputs the main voltage adjustment value using a linear output function.

8. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 7, characterized in that, Sub-features include: peak energy intensity, energy rise slope, energy duration, current temperature, and temperature rise coefficient; Cross features include at least: The cross term between current temperature and peak energy intensity; The interaction term between the current temperature and the slope of energy increase; The interaction between current temperature and energy duration; The interaction term between the temperature rise coefficient and peak energy intensity; The interaction term between the temperature rise coefficient and the energy rise slope; The interaction term between the temperature rise coefficient and the energy duration.

9. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 1, characterized in that, Confidence neural network models include: The signal feature input module uses dual channels to input the first signal feature and the second signal feature, and processes them separately into a first feature vector and a second feature vector. The environmental feature input module uses dual-channel input of main offset information and secondary offset information, and uses an environmental encoder to output an environmental confidence vector. The feature comparison module calculates signal quality differences, environmental disturbance differences, and signal alignment differences to generate difference information. The information fusion layer uses a gating attention mechanism to weight key features and generate confidence information. The confidence output layer uses a primary confidence predictor and a secondary confidence predictor to independently process the confidence information and generate the confidence of the first signal feature and the confidence of the second signal feature, respectively.

10. The method for fast compensation of clock frequency offset information based on a neural network model according to claim 9, characterized in that, Signal characteristics include time-domain jitter, frequency shift, and phase noise.

Images