A synchronous timing system fault diagnosis method based on unsupervised learning
By employing unsupervised learning methods, utilizing multidimensional data and dynamic control entropy features, and combining long short-term memory-autoencoder networks and Mahalanobis distance, early warning of metastable faults in synchronous timing systems is achieved. This solves the problems of fault identification lag and sample scarcity in existing technologies, and improves the diagnostic reliability and robustness of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAN RITRONTEK ELECTRONICS TECH
- Filing Date
- 2026-03-02
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies cannot identify metastable faults in synchronous timing systems in their early stages, resulting in alarms being triggered only after the system has completely lost lock, causing irreparable experimental losses. Furthermore, the lack of sufficient fault samples makes it difficult to train supervised learning models.
An unsupervised learning method is adopted to collect multi-dimensional operating state data, calculate dynamic control entropy features and reconstruct error vectors, use a long short-term memory-autoencoder network to learn the normal mode of the system, and combine Mahalanobis distance to make real-time anomaly score judgment, so as to realize early warning of metastable faults.
It enables early fault warning for synchronous timing systems, reduces false alarm rate, improves the reliability and robustness of diagnosis, avoids experimental losses, and is suitable for scenarios where fault samples are scarce in large-scale physical experimental devices.
Smart Images
Figure CN121765502B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of distributed control and intelligent operation and maintenance, specifically relating to a fault diagnosis method for synchronous timing systems based on unsupervised learning. Background Technology
[0002] In large-scale physics experimental setups, the control system relies on a high-precision synchronization timing system based on the White Rabbit protocol to achieve nanosecond-level synchronization of hundreds of devices distributed over a range of several kilometers. Terminal nodes achieve phase locking between their local clock and the network master clock through an internal soft phase-locked loop (SoftPLL, also known as a clock servo system). The core components of the SoftPLL include a digital dual-mixer phase detector and a voltage-controlled oscillator (VCO). The servo algorithm within the FPGA dynamically adjusts the control voltage of the digital-to-analog converter (DAC) output to the VCO based on the phase error measured by the phase detector, thereby regulating the local clock frequency to eliminate phase deviation.
[0003] In actual operation, due to the complex electromagnetic interference, temperature fluctuations and device aging in the accelerator tunnel, SoftPLL often faces a metastable critical fault state. This state is manifested as follows: although the phase error has not yet triggered the hard threshold of lockout, the control loop has already experienced high-frequency micro-oscillations, or the DAC control voltage has approached the range limit, and the synchronization timing system is on the dangerous edge of barely maintaining lockout, with extremely poor anti-interference ability.
[0004] Existing fault diagnosis methods mainly rely on monitoring systems to perform simple threshold judgments on locked state bits or phase error peaks. However, existing technologies typically only trigger alarms after the synchronization timing system has completely lost lock and synchronization has been interrupted. By this time, the time reference for the physical experiment has been lost, the data has been corrupted, and irreparable experimental losses have occurred. Furthermore, traditional univariate threshold methods cannot identify multidimensional parameter coupling anomalies during metastable states, leading to prolonged system operation with defects. Simultaneously, in large-scale physical experimental setups, actual lock-out faults are extremely rare, making it difficult to collect sufficient fault samples to train supervised machine learning classifiers. Therefore, there is an urgent need for a diagnostic method capable of early detection of metastable faults that does not rely on a large number of fault samples. Summary of the Invention
[0005] This invention provides a fault diagnosis method for synchronous timing systems based on unsupervised learning, in order to solve the technical problems of delayed fault warning, difficulty in identifying implicit coupling faults, and scarcity of fault samples in the prior art.
[0006] A fault diagnosis method for synchronous timing systems based on unsupervised learning, comprising:
[0007] S1, collect multi-dimensional operating status data of the terminal nodes of the synchronous timing system, perform sliding window truncation and standardization processing on the multi-dimensional operating status data, and construct a time series of multi-dimensional state space;
[0008] S2, Based on the changes in phase error and DAC control word in the time series, calculate the dynamic control entropy characteristic that characterizes the disorder of the control loop;
[0009] S3, the time series and the dynamic control entropy feature are concatenated along the feature dimension and used as the input sequence. The input is then fed into a pre-trained long short-term memory-autoencoder network to obtain the reconstructed sequence of the input sequence and to calculate the reconstruction error vector.
[0010] S4. Based on the statistical distance of the reconstructed error vector and the gain adjustment of the dynamic control entropy feature, calculate the real-time anomaly score of the synchronous timing system, and determine whether the synchronous timing system has entered a metastable fault state based on the real-time anomaly score.
[0011] By collecting multidimensional data, the system's operating status can be comprehensively reflected, providing a rich information foundation for subsequent feature extraction. The dynamic control entropy feature aims to evaluate the disorder of the SoftPLL control loop. By nonlinearly amplifying signal-to-noise ratio anomalies and operating point deviations, weak metastable features become numerically significant. By utilizing a long short-term memory-autoencoder network to learn the normal mode of the system, the dependence on fault samples is avoided, solving the problem of scarce fault samples in large-scale physical experimental devices. At the same time, it can capture the long-term and short-term dependencies of time series. The real-time anomaly score combines the Mahalanobis distance's ability to measure the correlation between variables with the dynamic control entropy's sensitivity to weak oscillations. Even tiny metastable precursors that the neural network has not yet seen can be captured by physical mechanism indicators and trigger early warnings.
[0012] Furthermore, the multi-dimensional operating status data includes: the phase error between the local clock and the master clock, the DAC control word output to the voltage-controlled oscillator, the auxiliary clock frequency deviation, and the core temperature of the board.
[0013] Furthermore, the Z-Score standardization method is used to standardize the multidimensional operational status data.
[0014] Furthermore, the dynamic control entropy feature satisfies the following relationship:
[0015] ;
[0016] In the formula, This represents the dynamic control entropy characteristic. Indicates the current moment. Indicates the offset of a historical moment. The cumulative time window length, The first-order difference of the standardized phase error. Representing historical moments The dimensionless value of the DAC control word after standardization. For the first-order difference of the standardized DAC control word, This represents the optimal linear working center value for the DAC in the normalization domain. =0, To control the exponent of the sensitivity factor. The penalty coefficient for deviation from the working point. is the numerical stability constant.
[0017] Furthermore, the long short-term memory-autoencoder network includes an encoder and a decoder;
[0018] The encoder receives a multidimensional expanded sequence containing the time series and the dynamic control entropy features, extracts features through multiple long short-term memory units, and compresses the multidimensional expanded sequence into a latent variable vector.
[0019] Furthermore, the decoder uses the latent variable vector to reconstruct the input sequence in reverse and outputs the reconstructed sequence.
[0020] Furthermore, the training strategy of the Long Short-Term Memory-Autoencoder Network is as follows: select multi-dimensional operating state data of the synchronous timing system during the stable locked state as the training set, use mean squared error as the loss function, minimize the difference between the input sequence and the reconstructed sequence, so that the Long Short-Term Memory-Autoencoder Network can learn the normal steady-state mode of the synchronous timing system.
[0021] The strategy of training with only normal samples matches the reality that fault samples of large devices are extremely scarce, thus lowering the threshold for algorithm deployment.
[0022] Furthermore, the reconstruction error vector satisfies the following relationship: ;
[0023] In the formula, This represents the reconstruction error vector. For the input sequence at time... Multidimensional vectors, including time series and dynamic control entropy, For reconstructing the sequence.
[0024] Furthermore, the real-time anomaly score satisfies the following relationship:
[0025] ;
[0026] In the formula, For real-time anomaly scoring, This is the reconstruction error vector at the current moment. Let the mean vector of reconstruction errors on the validation set be _____. The reconstruction error covariance matrix on the validation set, The regularization coefficient is . It is the identity matrix. The average control entropy reference value for the synchronous timing system during steady-state operation. For adjustment coefficients, This represents the dynamic control entropy characteristic. is the numerical stability constant.
[0027] Furthermore, determining whether the synchronous timing system has entered a metastable fault state based on the real-time anomaly score includes:
[0028] Determine whether the real-time anomaly score exceeds the upper limit of the preset confidence interval for multiple consecutive periods; if so, determine that the clock servo system of the terminal node has entered a metastable fault state and issue a warning signal.
[0029] By using a decision mechanism that operates over multiple consecutive cycles, the false alarm rate caused by a single threshold can be reduced, thereby improving the reliability of the diagnosis.
[0030] The beneficial effects are as follows: This invention achieves early warning of faults. By capturing temporal anomalies through a long short-term memory-autoencoder network and combining the sensitivity of dynamic control entropy to weak oscillations, it can issue an early warning as soon as the SoftPLL enters a metastable state before losing lock, thus providing maintenance personnel with a valuable repair window and preventing accidental loss of accelerator beams. This invention solves the problem of training with fault-free samples. It adopts an unsupervised learning architecture, requiring only data collected during normal system operation to complete model training, avoiding the problem of traditional supervised learning requiring a large amount of labeled fault data. Furthermore, by introducing Mahalanobis distance and multi-dimensional feature fusion, this invention can intelligently distinguish between normal fluctuations caused by changes in ambient temperature and abnormal oscillations caused by equipment aging or interference, suppressing false alarms and improving diagnostic robustness in complex environments. Attached Figure Description
[0031] Figure 1 A schematic diagram illustrating the actual evolution process of SoftPLL multidimensional state data;
[0032] Figure 2 A schematic diagram illustrating the nonlinear amplification effect of dynamically controlled entropy on weak oscillation characteristics;
[0033] Figure 3 This is a schematic diagram comparing the timeliness of fault diagnosis and early warning. Detailed Implementation
[0034] An embodiment of the fault diagnosis method for synchronous timing systems based on unsupervised learning provided by this invention:
[0035] A fault diagnosis method for a synchronous timing system based on unsupervised learning includes steps S1 to S4:
[0036] S1. Collect multi-dimensional operating status data of the terminal nodes of the synchronous timing system, perform sliding window truncation and standardization processing on the multi-dimensional operating status data, and construct a time series of multi-dimensional state space.
[0037] Specifically, at the edge of the synchronous timing system (such as a White Rabbit switch) or in the aggregation layer server, multi-dimensional operational status data of the terminal node SoftPLL is collected in real time at a frequency of 10Hz to 100Hz via SNMP protocol or UDP data stream. For each sampling moment, a multi-dimensional state vector is constructed.
[0038] The multidimensional state vector includes the following key components:
[0039] 1) Phase error: The phase difference between the local clock and the recovered master clock, measured in real time by the DDMTD module inside the FPGA, in picoseconds (ps). For example, the phase error acquired at a certain moment is 100ps.
[0040] 2) DAC Control Word: This is a digital value calculated and output to the auxiliary DAC chip by the FPGA's internal servo algorithm to control the frequency of the voltage-controlled oscillator (VCO). This control word can be linearly mapped to an equivalent control voltage value based on the full-scale range of the DAC chip. For example, if the DAC range is 0-2.5V, the DAC control word 32768 corresponds to an equivalent voltage of 1.25V.
[0041] 3) Auxiliary clock frequency deviation: reflects the stability of the digital dual-mixer phase detector's own reference reference.
[0042] 4) Core temperature of the board: The junction temperature collected by the temperature sensor on the FPGA chip, in degrees Celsius (°C). For example, the temperature collected at a certain moment is 45.
[0043] In an optional embodiment, the acquired multidimensional operational status data is preprocessed: a fixed-length sliding window is used to divide the continuous time stream into several subsequences to capture temporal features; for example, the fixed length of the sliding window is 128 sampling points. Subsequently, the Z-Score normalization method is independently applied to each dimension of data to eliminate the physical dimension differences between phase error, DAC control word, frequency deviation, and temperature, so that the data of all dimensions are converted into dimensionless values with a mean of 0 and a variance of 1, forming the input tensor of the model.
[0044] By constructing a time series of a multidimensional state space and performing preprocessing, the influence of different physical dimensions is eliminated, providing a standardized data foundation for subsequent feature extraction and model training, and enabling a more comprehensive reflection of the SoftPLL's operating status.
[0045] S2, Based on the changes in phase error and DAC control word in the time series, calculate the dynamic control entropy characteristic that characterizes the disorder of the control loop.
[0046] Specifically, a dynamic control entropy is constructed based on automatic control theory. This index is used to evaluate the disorder of the SoftPLL control loop, thereby solving the problem that the general neural network is not good at extracting the weak oscillation features of the control loop.
[0047] Among them, for the current moment The dynamic control entropy satisfies the following relationship:
[0048] ;
[0049] In the formula, Represents dynamic control entropy; This represents the offset of a historical moment, used to iterate through the data within the cumulative time window; For example, to calculate the cumulative time window length, =50; This is the first-order difference of the standardized phase error; Representing historical moments The dimensionless value of the DAC control word after standardization; This is the first-order difference of the standardized DAC control word; This represents the optimal linear working center value of the DAC in the normalization domain. Since the normalization process makes the mean of the steady-state data zero, therefore... =0; To control the exponent of the sensitivity factor, for example, =2.0; For example, the penalty coefficient for deviation from the working point. =10; For example, the numerical stability constant is... = This is to prevent the denominator from being zero.
[0050] For example, suppose at a certain current moment At a certain historical moment within the cumulative time window The standardized data are as follows: the first-order differential change of the phase error is 10, the first-order differential of the DAC control word is 0.001, and the current DAC control word is 0.1.
[0051] The first term in the summation part of the relation is the control sensitivity factor: = The control sensitivity factor measures the phase change benefit per unit control input. This ratio increases significantly when the system exhibits abnormal behavior such as drastic control but poor phase convergence or large phase fluctuations caused by minor disturbances.
[0052] The second term in the summation part of the relation is the operating point deviation penalty factor: = The working point deviation penalty factor imposes a severe weight penalty on working behaviors that deviate from the optimal linear working center point.
[0053] Multiply the two terms, sum them, and take the logarithm to get the current time. Dynamic control entropy. For example... Figure 2 As shown, during the steady-state period of 0-50s, the entropy value of the dynamic control entropy remains at an extremely low level; once it enters the metastable oscillation period after 50s, the curve suddenly jumps and remains at a high level, proving that the dynamic control entropy can transform weak oscillations into clearly discernible numerical characteristics.
[0054] By calculating the dynamic control entropy, subtle oscillations that are difficult to distinguish with the naked eye can be transformed into significant numerical features, solving the problem that a single variable is difficult to identify metastable states and improving the sensitivity of the features.
[0055] S3, the time series and the dynamic control entropy feature are concatenated along the feature dimension and used as the input sequence. The input is then fed into a pre-trained long short-term memory-autoencoder network to obtain the reconstructed sequence of the input sequence and to calculate the reconstruction error vector.
[0056] Specifically, a Long Short-Term Memory-Autoencoder Network (hereinafter referred to as the Network) based on a Long Short-Term Memory (LSTM) network is constructed to learn the normal steady-state mode of SoftPLL.
[0057] The network consists of an encoder and a decoder. The encoder receives a multidimensional augmented sequence containing time series data and dynamic control entropy. Long- and short-term dependencies of the time series are extracted using multi-layer LSTM units and compressed into a low-dimensional latent variable vector. This process forces the network to discard noise, retaining only the core operational rules of the synchronous timing system. The decoder structure is symmetrical to the encoder, using the latent variables to reconstruct the input sequence and output the reconstructed sequence.
[0058] During the training phase, a strategy of training only on normal samples was adopted, selecting data from the historical database during periods when the synchronization timing system was in a stable locked state as the training set. The mean squared error loss function was used, and the training objective was to minimize the difference between the input sequence and the reconstructed sequence. Through training, the model learned what normal SoftPLL behavior is. During the online execution phase, the reconstruction error vector was calculated, and the reconstruction error vector satisfies the following relationship: ;
[0059] in, This represents the reconstruction error vector. The input sequences include time series and dynamic control entropy. For reconstructing the sequence.
[0060] By utilizing a long short-term memory-autoencoder network to learn the normal patterns of a synchronous timing system, the dependence on fault samples is avoided, solving the problem of scarce fault samples in large-scale devices, while also capturing the long- and short-term dependencies of time series.
[0061] S4. Based on the statistical distance of the reconstructed error vector and the gain adjustment of the dynamic control entropy feature, calculate the real-time anomaly score of the synchronous timing system, and determine whether the synchronous timing system has entered a metastable fault state based on the real-time anomaly score.
[0062] Specifically, to reduce the false alarm rate caused by a single threshold, a dynamic decision mechanism based on Mahalanobis distance is introduced. The real-time anomaly score of the synchronous timing system satisfies the following relationship: ;
[0063] In the formula, For real-time anomaly scoring, This is the reconstruction error vector at the current moment. Let the mean vector of reconstruction errors on the validation set be _____. The reconstruction error covariance matrix on the validation set, For example, the regularization coefficient is... Take the smallest positive number, The identity matrix is used to prevent the covariance matrix from being singular and thus impossible to invert. This is the baseline value of the average control entropy of the system during steady-state operation on the verification set. This is an adjustment coefficient used to control the weighting of physical features in the final score. is the numerical stability constant.
[0064] The square root term in the relation uses Mahalanobis distance to measure the degree to which the reconstruction error vector at the current moment deviates from the normal error distribution. Compared to Euclidean distance, Mahalanobis distance takes into account the correlation between state variables. For example, under normal circumstances, an increase in temperature leads to a slight increase in phase error, and the Mahalanobis distance will not increase significantly; however, if the temperature remains constant but the phase error increases abnormally, violating statistical laws, the Mahalanobis distance will increase significantly.
[0065] Product terms in the relation Using the physical layer's feature metrics as amplifiers for the neural network output, when the reconstruction error vector is small but the dynamic control entropy of the physical layer shows an abnormal increase, The value will be greater than 1, thus forcibly amplifying the final real-time anomaly score, ensuring that even tiny metastable precursors that the network has not yet seen can be captured by the physical mechanism index and trigger an early warning.
[0066] When executing a judgment, if the real-time anomaly score exceeds the upper limit of the preset confidence interval for N consecutive cycles, the clock servo system of the terminal node is determined to have entered a metastable fault state, and a yellow warning signal is issued to the control center. For example, when the real-time anomaly score exceeds the upper limit of the preset confidence interval for 3 consecutive cycles, a yellow warning signal is issued to the control center.
[0067] In an optional embodiment, such as Figure 1 As shown, the operation of SoftPLL includes a steady-state operation period, a metastable oscillation period, and a lockout failure period. For example... Figure 3 As shown, the anomaly judgment score rapidly exceeds the warning threshold standard in about 50 seconds, which is the metastable starting point, thus presenting a three-level ladder pattern of low steady state, high metastable state, and extremely high loss of lock. Compared with the existing technology that only alarms after complete loss of lock in 82 seconds, this invention issues a warning about 30 seconds earlier, buying maintenance personnel time for emergency repairs.
[0068] By integrating Mahalanobis distance and dynamic control entropy to calculate anomaly scores, not only is the correlation between variables considered, but weak fault signals are also amplified using physical characteristics, enabling accurate judgment and early warning of metastable faults and improving the robustness of the system.
[0069] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.
Claims
1. A fault diagnosis method for a synchronous timing system based on unsupervised learning, characterized in that, include: S1, collect multi-dimensional operating status data of the terminal nodes of the synchronous timing system, perform sliding window truncation and standardization processing on the multi-dimensional operating status data, and construct a time series of multi-dimensional state space; The multi-dimensional operating status data includes: the phase error between the local clock and the master clock, the DAC control word output to the voltage-controlled oscillator, the auxiliary clock frequency deviation, and the core temperature of the board. S2, based on the changes in phase error and DAC control word in the time series, calculate the dynamic control entropy characteristic representing the disorder of the control loop, satisfying the following relationship: ； In the formula, This represents the dynamic control entropy characteristic. Indicates the current moment. Indicates the offset of a historical moment. The cumulative time window length, The first-order difference of the standardized phase error. Representing historical moments The dimensionless value of the DAC control word after standardization. For the first-order difference of the standardized DAC control word, This represents the optimal linear working center value for the DAC in the normalization domain. =0, To control the exponent of the sensitivity factor. The penalty coefficient for deviation from the working point. It is the numerical stability constant; S3, the time series and the dynamic control entropy feature are concatenated along the feature dimension and used as the input sequence. The input is then fed into a pre-trained long short-term memory-autoencoder network to obtain the reconstructed sequence of the input sequence and to calculate the reconstruction error vector. S4. Based on the statistical distance of the reconstructed error vector and the gain adjustment of the dynamic control entropy feature, calculate the real-time anomaly score of the synchronous timing system, and determine whether the synchronous timing system has entered a metastable fault state based on the real-time anomaly score.
2. The fault diagnosis method for synchronous timing systems based on unsupervised learning according to claim 1, characterized in that, The Z-Score standardization method is used to standardize the multidimensional operational status data.
3. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 1 or 2, characterized in that, The long short-term memory-autoencoder network includes an encoder and a decoder; The encoder receives a multidimensional expanded sequence containing the time series and the dynamic control entropy features, extracts features through multiple long short-term memory units, and compresses the multidimensional expanded sequence into a latent variable vector.
4. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 3, characterized in that, The decoder uses the latent variable vector to reconstruct the input sequence in reverse and outputs the reconstructed sequence.
5. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 4, characterized in that, The training strategy of the Long Short-Term Memory-Autoencoder Network is as follows: select multi-dimensional operating state data of the synchronous timing system during the stable locked state as the training set, use mean squared error as the loss function, minimize the difference between the input sequence and the reconstructed sequence, and enable the Long Short-Term Memory-Autoencoder Network to learn the normal steady-state mode of the synchronous timing system.
6. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 4, characterized in that, The reconstruction error vector satisfies the following relationship: ; In the formula, This represents the reconstruction error vector. For the input sequence at time... Multidimensional vectors, including time series and dynamic control entropy, For reconstructing the sequence.
7. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 6, characterized in that, The real-time anomaly score satisfies the following relationship: ； In the formula, For real-time anomaly scoring, This is the reconstruction error vector at the current moment. Let the mean vector of reconstruction errors on the validation set be _____. The reconstruction error covariance matrix on the validation set, The regularization coefficient is . It is the identity matrix. The average control entropy reference value for the synchronous timing system during steady-state operation. For adjustment coefficients, This represents the dynamic control entropy characteristic. is the numerical stability constant.
8. The fault diagnosis method for a synchronous timing system based on unsupervised learning according to claim 7, characterized in that, Determining whether the synchronous timing system has entered a metastable fault state based on the real-time anomaly score includes: Determine whether the real-time anomaly score exceeds the upper limit of the preset confidence interval for multiple consecutive periods; if so, determine that the clock servo system of the terminal node has entered a metastable fault state and issue a warning signal.