A circuit board failure prediction method and system

Abstract

This invention relates to the field of circuit board monitoring and fault diagnosis technology, and discloses a circuit board fault prediction method and system. The method includes collecting voltage and current signal data of the circuit board during operation, extracting time-series features to obtain a dynamic state change sequence; inputting the dynamic state change sequence into a preset support vector machine to obtain a state evolution trajectory, calculating the time point when the evolution rate changes abruptly, as the starting point of instability; calculating the amplitude intensity value of the starting point and judging the stability qualitative change trend; if a stability qualitative change trend exists, performing spectral decomposition on heterogeneous data sources to obtain a risk score for overall system instability; extracting state fluctuation features based on the risk score to determine the system safety boundary; if the state evolution sequence exceeds the system safety boundary, a potential threshold crossing occurs, generating a warning signal; extracting subtle feature sequences based on the warning signal to obtain the final stability assessment. This method can achieve accurate prediction of circuit board faults.

Description

Technical Field

[0001] This invention relates to the field of circuit board monitoring and fault diagnosis technology, and in particular to a circuit board fault prediction method and system. Background Technology

[0002] Circuit board failures can lead to equipment downtime, data loss, and even safety incidents. Therefore, integrating the concepts of pre-diagnosis and health management into the circuit board maintenance system to achieve early prediction and proactive warning of failures has become a core requirement for promoting the transformation of electronic equipment from "passive maintenance" to "proactive maintenance".

[0003] In one existing technology, the operation of a circuit board is a dynamic nonlinear process. Subtle signals in the early stages of a fault (such as the accumulation of small fluctuations caused by component aging) are often hidden in normal data. Existing technologies lack the ability to deeply characterize this dynamic evolution process, making it impossible to accurately track the gradual deterioration of the health status, resulting in delayed pre-diagnosis and difficulty in avoiding faults in advance. On the other hand, in complex environments, the operation of circuit boards is affected by various factors such as temperature, vibration, and electromagnetic interference, resulting in multi-source heterogeneous data (electrical quantities, environmental quantities). Existing health management solutions are unable to effectively integrate and eliminate interference from this type of data. Relying solely on surface data is prone to false alarms or missed alarms, failing to accurately reflect the true health status of the circuit board. Furthermore, existing technologies mostly use fixed thresholds or static models, which cannot adapt to the dynamic changes in the operating status of the circuit board, resulting in insufficient accuracy and reliability of pre-diagnosis and health management, making it difficult to meet the high availability requirements of equipment in critical fields.

[0004] In summary, existing technologies suffer from low accuracy in predicting circuit board faults. Summary of the Invention

[0005] This invention provides a circuit board fault prediction method and system to solve the problem of low accuracy in circuit board fault prediction in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a circuit board fault prediction method, comprising: Real-time acquisition of voltage and current signal data during the operation of the circuit board; extraction of time-series features from the voltage and current signal data to obtain a dynamic state change sequence; The dynamic state change sequence is input into a preset support vector machine model to obtain the state evolution trajectory. The time point when the evolution rate changes abruptly in the state evolution trajectory is calculated as the starting point of the instability sign. Calculate the amplitude intensity value at the starting point. If the amplitude intensity value exceeds a preset amplitude intensity threshold, extract the state fluctuation accumulation index from the voltage and current signal data to determine the stability change trend of the starting point. If the aforementioned trend of qualitative change in stability exists, heterogeneous data sources are acquired and spectral decomposition and interference removal are performed to obtain purified effective feature vectors. The effective feature vectors are then fused with the cumulative state fluctuation index to obtain a risk score for overall system instability. Based on the risk score, a time sliding window model is constructed to extract state fluctuation features. The state fluctuation features are fitted to obtain a state evolution sequence. The state evolution sequence is iteratively updated to obtain the system safety boundary. If the state evolution sequence exceeds the system safety boundary, a potential threshold crossing is determined, and instantaneous data at the time of the potential threshold crossing is extracted to generate an early warning signal. Based on the warning signal, a subtle feature sequence is extracted, and the feature value of the subtle feature sequence is calculated to obtain the final stability assessment.

[0007] Secondly, the present invention provides a circuit board fault prediction system, comprising: The data acquisition module is used to collect voltage and current signal data during the operation of the circuit board in real time, extract time-series features from the voltage and current signal data, and obtain a dynamic state change sequence. The starting point determination module is used to input the dynamic state change sequence into a preset support vector machine model to obtain the state evolution trajectory, and calculate the time point when the evolution rate changes abruptly in the state evolution trajectory as the starting point of the instability sign. The trend judgment module is used to calculate the amplitude intensity value of the starting point. If the amplitude intensity value exceeds the preset amplitude intensity threshold, the state fluctuation accumulation index is extracted from the voltage and current signal data to judge the stability change trend of the starting point. The risk calculation module is used to obtain heterogeneous data sources and perform spectrum decomposition and interference removal if the stability change trend exists, to obtain the purified effective feature vector, and to perform multi-dimensional data fusion of the effective feature vector and the state fluctuation accumulation index to obtain the overall system instability risk score. The safety boundary update module is used to construct a time sliding window model based on the risk score to extract state fluctuation features, fit the state fluctuation features to obtain a state evolution sequence, and iteratively update the state evolution sequence to obtain the system safety boundary. The warning signal generation module is used to determine a potential threshold crossing if the state evolution sequence exceeds the system safety boundary, extract instantaneous data when the potential threshold crossing occurs, and generate a warning signal. The stability assessment module is used to extract subtle feature sequences based on the warning signal, calculate the feature values ​​of the subtle feature sequences, and obtain the final stability assessment.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts the dynamic state change sequence of voltage and current signals from a long short-term memory network, combines it with a support vector machine to locate the starting point of instability signs, and then tracks the trend of qualitative change in stability based on the cumulative index of state fluctuations. Since subtle signals in the early stages of circuit board faults are hidden in normal data, existing technologies lack in-depth characterization of the dynamic evolution process and cannot track gradual degradation, leading to delayed pre-diagnosis. This invention, through the combination of temporal feature extraction and nonlinear pattern classification, accurately captures state transition points and abrupt changes in evolution rate, achieving accurate identification of early fault signs. This effectively solves the problem that existing technologies struggle to detect subtle signals in the early stages of faults, significantly improving the timeliness of fault prediction and early warning capabilities.

[0009] (2) This invention obtains a risk score by performing spectral decomposition and interference removal on heterogeneous data sources, and then fusing the data with the cumulative state fluctuation index. Addressing the difficulty in removing interference from multi-source heterogeneous data in complex environments, existing technologies rely solely on surface data, which easily leads to false alarms and missed alarms, failing to reflect the true health status. This invention utilizes an adaptive filtering operator to remove environmental interference, achieving effective fusion of multi-dimensional data and highlighting core features related to faults. This effectively solves the technical challenges of difficult multi-source heterogeneous data fusion and incomplete interference removal, reducing the false alarm and missed alarm rates caused by environmental factors and improving the accuracy of fault prediction.

[0010] (3) This invention dynamically adjusts the time window length based on the risk score, iteratively updates the system safety boundary, and generates an early warning signal by combining the judgment of potential threshold crossing. Traditional technologies often use fixed thresholds or static models, which cannot adapt to the dynamic changes in the circuit board's operating state, resulting in insufficient accuracy. This invention optimizes the system safety boundary in real time based on risk scores and multi-scale entropy analysis, allowing the fault judgment criteria to be dynamically adjusted according to the operating state. This effectively solves the problem of poor adaptability of fixed thresholds, achieves accurate adaptation to the dynamic operating state of the circuit board, and further improves the reliability and practicality of fault prediction. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of the circuit board fault prediction method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the circuit board fault prediction system provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a circuit board fault prediction method, including the following steps: S11, Real-time acquisition of voltage and current signal data during the operation of the circuit board, extraction of time-series features from the voltage and current signal data to obtain a dynamic state change sequence; S12, input the dynamic state change sequence into a preset support vector machine model to obtain the state evolution trajectory, and calculate the time point when the evolution rate changes abruptly in the state evolution trajectory as the starting point of the instability sign. S13, calculate the amplitude intensity value of the starting point. If the amplitude intensity value exceeds the preset amplitude intensity threshold, extract the state fluctuation accumulation index from the voltage and current signal data to determine the stability change trend of the starting point. S14, if the stability change trend exists, then obtain the heterogeneous data source and perform spectrum decomposition and interference removal to obtain the purified effective feature vector. Then, perform multi-dimensional data fusion of the effective feature vector and the state fluctuation accumulation index to obtain the risk score of overall system instability. S15, construct a time sliding window model based on the risk score to extract state fluctuation features, fit the state fluctuation features to obtain a state evolution sequence, iteratively update the state evolution sequence to obtain the system safety boundary; S16, if the state evolution sequence exceeds the system safety boundary, then a potential threshold crossing is determined, the instantaneous data at the time of the potential threshold crossing is extracted, and an early warning signal is generated; S17, extract subtle feature sequences based on the warning signal, calculate the feature values ​​of the subtle feature sequences, and obtain the final stability assessment.

[0014] In step S11, voltage and current signal data during the operation of the circuit board are acquired in real time, and time-series feature extraction is performed on the voltage and current signal data to obtain a dynamic state change sequence, including: The analog voltage and analog current signals of the circuit board are acquired in real time during operation. Continuous data points are obtained through high-precision sensors. The analog voltage and analog current signals are converted into a multi-dimensional original waveform matrix containing voltage and current dimensions to obtain voltage and current signal data. The original multidimensional waveform matrix is ​​sliced ​​to obtain multiple sub-matrices, and the sub-matrices are reorganized into a temporal feature tensor containing time dimension and feature dimension; The temporal feature tensor is input into a long short-term memory network, the data of each time window in the long short-term memory network is calculated, the hidden layer state vector is obtained, the transient feature descriptor is generated by mapping the hidden layer state vector, and the Euclidean distance between the transient feature descriptors corresponding to adjacent time windows is calculated. If the Euclidean distance exceeds a preset distance threshold, it is marked as a state transition point. All the state transition points are connected in chronological order to form a dynamic state change sequence.

[0015] In one implementation, this embodiment constructs a synchronous data acquisition link by deploying high-precision Hall sensors and voltage divider acquisition modules at key power supply circuits and signal interfaces on the circuit board. The sensor sampling frequency is set to 1kHz, i.e., acquiring 1000 continuous data points per second. This frequency setting is determined based on the typical duration (usually milliseconds) of transient signals from circuit board faults and the Nyquist sampling theorem, ensuring complete capture of the details and fluctuation characteristics of voltage and current abrupt changes. The acquired analog voltage signals (range 0.5-5.0V) and analog current signals (range 0.1-2.0A) are converted from analog to digital by a 16-bit ADC and organized into a multi-dimensional raw waveform matrix in timestamp order. The number of rows in the matrix equals the number of sampling points, and the number of columns is 2 (corresponding to the voltage and current dimensions, respectively). For example, the data acquired within 1 second corresponds to a 1000-row, 2-column matrix structure, providing structured basic data for subsequent time series analysis.

[0016] In another implementation, for slicing the multidimensional original waveform matrix, this embodiment adopts a fixed time window partitioning strategy, setting the window length to 100 sampling points (corresponding to 100 milliseconds) and the window sliding step size to 50 sampling points. Multiple consecutive overlapping sub-matrices are obtained through sliding truncation. Each sub-matrix has a dimension of 100 rows and 2 columns. These sub-matrices are then reorganized in chronological order into a three-dimensional temporal feature tensor with dimensions (number of windows, 100, 2). When the temporal feature tensor is input into a pre-set long short-term memory network, the model adopts a 3-layer hidden layer structure, with 50 neurons in each layer. The network analyzes the sub-matrices within each time window point by point, filtering effective information through a gating mechanism, and outputting a hidden layer state vector with a dimension of 64. This vector integrates the amplitude changes, fluctuation frequency, and trend characteristics of voltage and current within the window, accurately representing the current transient operating state of the circuit.

[0017] It should be noted that, based on the hidden layer state vector, a 10-dimensional transient feature descriptor is generated through mapping using a fully connected layer. This descriptor undergoes normalization to eliminate the influence of dimensions, ensuring that features of different dimensions are within the same numerical range. Subsequently, the Euclidean distance between descriptors corresponding to two adjacent time windows is calculated. In this embodiment, the preset distance threshold is determined statistically. During the offline phase, 1000 sets of feature descriptors under normal operating conditions are collected, and the 99.5th percentile of the Euclidean distance between adjacent descriptors is set as the preset distance threshold. When the real-time calculated Euclidean distance exceeds this threshold, it indicates a significant change in the circuit state, which is marked as a state transition point. Finally, all state transition points are connected sequentially in chronological order to obtain a dynamic state change sequence.

[0018] In step S12, the dynamic state change sequence is input into a preset support vector machine model to obtain a state evolution trajectory. The time point at which the evolution rate abruptly changes in the state evolution trajectory is calculated as the starting point of the instability indicator, including: The dynamic state change sequence is mapped to a high-dimensional state feature vector set using a radial basis kernel function. The high-dimensional state feature vector set is then input into a preset support vector machine model to construct the optimal decision hyperplane. Calculate the normal distance from each vector point in the high-dimensional state feature vector set to the optimal decision hyperplane, arrange the normal distances in chronological order, and generate the state evolution trajectory; The evolution rate characteristics are obtained by taking the second derivative of the normal distance sequence of the state evolution trajectory. The time index of the absolute value of the evolution rate feature that exceeds a preset mutation threshold is tracked, and the time corresponding to the time index is determined as the starting point of the unstable sign.

[0019] In one implementation, this embodiment addresses the low-dimensional nonlinearity of dynamic state change sequences by employing a radial basis function (RBF) kernel to increase the dimensionality of the feature space. The kernel width parameter undergoes rigorous offline optimization, traversing multiple candidate values ​​from 0.01 to 1.0 via grid search. The optimal value is determined to be 0.1, based on the classification accuracy on the validation set. This maps the original low-dimensional state transition point sequence (each state transition point being a 10-dimensional feature vector) into a 256-dimensional high-dimensional state feature vector set. In the high-dimensional space, feature points that originally overlapped or were close in the low-dimensional scene are effectively separated, and subtle state differences are amplified, resulting in the high-dimensional state feature vector set.

[0020] It should be noted that the preset support vector machine model adopts a binary classification architecture, and the training process undergoes multiple rounds of optimization to ensure generalization ability. During the offline training phase, the training data is divided into two categories: normal state evolution samples and abnormal state evolution samples. Normal samples are derived from the dynamic state change sequences of 30 identical circuit boards running stably for 72 consecutive hours, covering different load conditions and ambient temperature ranges to ensure sample diversity. Abnormal samples are generated by simulating more than 10 typical fault scenarios, such as component aging, instantaneous voltage fluctuations, and minor short-circuit precursors. 500 sets of data are collected for each scenario, resulting in a training set of 5000 sets for each category. During model training, a convex quadratic programming problem is solved to find the optimal decision hyperplane that maximizes the margin between the two classes of samples. Simultaneously, a regularization parameter is introduced to control model complexity and avoid overfitting.

[0021] In one implementation, this embodiment calculates the normal distance from the high-dimensional state feature vector to the optimal decision hyperplane by projecting the distance onto the hyperplane's normal vector and bias term. This distance value has a clear physical meaning: positive and negative values ​​represent the two sides of the hyperplane where the vector lies, corresponding to normal trends (negative values) and abnormal trends (positive values), respectively. The absolute value directly reflects the degree to which the vector deviates from the normal state baseline; the larger the absolute value, the more significant the difference between the current state and the normal operating mode. The normal distances of all vector points are arranged sequentially in chronological order to form a continuous state evolution trajectory. A stable trajectory close to zero indicates a stable circuit board operating state; a trajectory continuously drifting towards positive values ​​indicates an accumulating unstable trend; and a trajectory exhibiting large fluctuations suggests the possible presence of instantaneous interference or sudden anomalies. This visualized trajectory intuitively presents the dynamic changes in the circuit board's operating state, providing a clear basis for subsequent analysis.

[0022] It should be noted that, to accurately capture the accelerated change characteristics of state evolution, the second derivative of the normal distance sequence of the state evolution trajectory is calculated using the central difference method to obtain the evolution rate characteristics. Solving by the normal distance at three adjacent time points effectively reduces the impact of single-point noise on the results, ensuring the reliability of the evolution rate characteristics. The absolute value of the second derivative directly characterizes the acceleration of state evolution; the larger the absolute value, the faster the state change and the more rapidly the unstable trend develops. Positive values ​​indicate that the unstable trend is intensifying, while negative values ​​indicate that the unstable trend is slowing down or stabilizing. During the calculation, the time interval is kept consistent with the window sliding step size in S11, set to 50 milliseconds. This ensures both the continuity of the time series data and accurate matching of the time scale of state changes, avoiding feature distortion caused by mismatched time granularity.

[0023] It is important to note that the preset mutation threshold requires collecting 1000 sets of evolution rate characteristic data under normal operating conditions during the offline phase, covering different operating conditions and individual differences of different equipment. The 99.8th percentile of the absolute values ​​of all data is calculated, and this percentile is set as the preset mutation threshold. The threshold setting ensures that the system will not falsely trigger mutation judgments during random fluctuations in normal operation, while also being able to accurately capture true unstable trend mutations. During real-time calculation, the system monitors the evolution rate characteristics hour by hour. When the absolute value of the characteristic first exceeds the threshold at a certain moment, the reverse tracking mechanism is immediately activated to find the time index corresponding to the characteristic. Combined with the timestamp of the state transition point in S11, the system accurately matches the actual operating time and determines the corresponding timestamp as the starting point of the unstable sign.

[0024] In step S13, the amplitude intensity value of the starting point is calculated. If the amplitude intensity value exceeds a preset amplitude intensity threshold, a state fluctuation accumulation index is extracted from the voltage and current signal data to determine the stability change trend of the starting point, including: Based on the timestamp of the starting point, extract the state vector data of the timestamp and calculate the amplitude intensity value; The amplitude intensity value is compared with a preset amplitude intensity threshold. If the amplitude intensity value exceeds the threshold, the full-dimensional running data set within a preset time range before and after the starting point is locked. The least squares method is used to fit the full-dimensional operational data set, and the deviation between the fitting result and the actual data is extracted to form a residual matrix containing information about subtle changes in the system, which serves as hidden information in the operational data. The energy spectral density is obtained by performing a Fourier transform on the residual matrix, and the energy spectral density is integrated within a preset frequency range to obtain the cumulative index of state fluctuation. Calculate the first-order difference vector of the cumulative index of state fluctuations over consecutive time steps. If the magnitude of the first-order difference vector of the cumulative index ...

[0025] In one implementation, this embodiment extracts state vector data based on the timestamp of the starting point, precisely extracting voltage and current data from five sampling points before and after the timestamp to form a state vector containing 11 data points (each data point includes both voltage and current dimensions). The calculation of the amplitude intensity value focuses on the fluctuation range of this vector, and is obtained by comprehensively analyzing the deviation of all data points within the vector from the mean. Specifically, the absolute value of the deviation of each data point from the mean is calculated, and the average value is taken as the final amplitude intensity value. This value can intuitively reflect the severity of the state fluctuation at the starting point. The determination of the preset amplitude intensity threshold has been verified in multiple scenarios. In the offline stage, normal operation data of 50 circuit boards of the same model under different loads and ambient temperatures were collected, and the amplitude intensity values ​​of the starting point (normal state transition point without actual unstable signs) under each operating condition were extracted, collecting a total of 1000 sets of valid data. The 99.5th percentile of this dataset was calculated through statistical analysis, and the quantile was set as the preset amplitude intensity threshold.

[0026] In another implementation, this embodiment, when locking the full-dimensional operational data set, expands the complete operational data for 30 seconds before and after the starting timestamp as the center, covering the original voltage and current data of all sampling points within the operational time period, ensuring the integrity and continuity of the data set. When fitting the full-dimensional operational data set using the least squares method, linear fitting models are constructed for both voltage and current dimensions. With time as the independent variable and voltage (or current) value as the dependent variable, the optimal fitting curve is obtained by minimizing the sum of squares error between the fitted value and the actual data. The fitted curve reflects the overall trend of data change, while the deviation between the fitted result and the actual data (i.e., the residual) contains hidden information such as subtle fluctuations and local anomalies that are masked by the overall trend. All residuals are organized according to time order and data dimensions to form a residual matrix.

[0027] It should be noted that a Fourier transform is performed on the residual matrix to convert the residual signal in the time domain into an energy spectral density in the frequency domain. The energy spectral density is then integrated within a preset frequency range to obtain the cumulative index of state fluctuations. The integration operation follows the formula... ,in Let be the energy spectral density function. and These are the lower and upper limits of the integration interval, respectively. The integration interval... The characteristic frequency range of common circuit board fault fluctuations is preset to 10Hz to 100Hz. This frequency band mainly reflects the characteristic fluctuations in the early stages of faults such as component aging and poor contact. The cumulative state fluctuation index... The physical meaning of is the total energy of the signal within a specified frequency range. Its value is positively correlated with the degree of accumulation of unstable energy inside the circuit board. The quantization range is from 0 to positive infinity, and the larger the value, the higher the potential fault risk. When calculating the first-order difference vector of the state fluctuation accumulation index under continuous time steps, each time step corresponds to 50 milliseconds. The difference vector reflects the change amplitude of the fluctuation accumulation index between adjacent time steps. Then, the magnitude of the difference vector is calculated, and the magnitude of the magnitude represents the rate of change of fluctuation accumulation. In this embodiment, the number of continuous time steps is preset to 5. If the magnitude of the difference vector shows a continuous increasing trend within 5 consecutive time steps (i.e., 250 milliseconds), it indicates that the speed of fluctuation accumulation is constantly accelerating, and the system state is changing from slight fluctuation to significant instability. Based on this, it can be clearly judged that there is a trend of qualitative change in stability. The trend of qualitative change in stability is defined as the critical stage of the system state from quantitative change to qualitative change. Specifically, it is manifested as the characteristic parameter reflecting the degree of system instability (the state fluctuation accumulation index) growing at a continuously accelerating rate within the continuous time step, that is, the magnitude of the first-order difference vector continuously increases. This trend indicates that the system is evolving from a state of controllable slight fluctuations to a state of uncontrollable failure, which is a key basis for triggering subsequent in-depth fault analysis.

[0028] In step S14, if the stability qualitative change trend exists, heterogeneous data sources are acquired and spectral decomposition and interference removal are performed to obtain purified effective feature vectors. These effective feature vectors are then fused with the accumulated state fluctuation index using multi-dimensional data to obtain a risk score for overall system instability, including: Based on the stability change trend, temperature data, vibration frequency data, and humidity data at the starting point are obtained as heterogeneous data sources. Fast Fourier Transform is performed on the heterogeneous data sources respectively to extract the frequency domain response features of the heterogeneous data sources in the characteristic frequency range, which is preset according to the common interference frequency range of the circuit board. An adaptive filtering operator is constructed based on the frequency domain response characteristics. The filtering operator is used to suppress interference components in the original running data and generate an interference removal vector that removes environmental interference. The interference removal vector and the state fluctuation accumulation index are jointly mapped to a high-dimensional state space by a kernel mapping function, and a multi-dimensional fusion matrix is ​​constructed by arranging them in time sequence. The multi-dimensional fusion matrix is ​​then subjected to manifold projection to reduce the dimensionality and obtain the evolution trajectory that reflects the overall state change of the system. Principal component analysis is used to determine the unstable principal axis that characterizes the unstable direction of the system, and the decay value of the projection energy of the evolution trajectory on the unstable principal axis over time is calculated. The attenuation value is normalized and mapped to a risk score of overall system instability between 0 and 10, thus obtaining the risk score of overall system instability.

[0029] In one implementation, this embodiment uses the starting point timestamp corresponding to the trend of qualitative change in stability as a benchmark when acquiring heterogeneous data sources. Temperature, vibration frequency, and humidity data are simultaneously collected within one minute before and after this timestamp to ensure precise alignment of heterogeneous data with voltage and current data in the time dimension. Temperature data is acquired using a surface-mount temperature sensor deployed on the surface of the circuit board chip, with a range of -40℃ to 125℃ and an accuracy of ±0.1℃. Vibration frequency data is acquired by a miniature accelerometer installed in the housing of the device containing the circuit board, with a sampling frequency of 1kHz, consistent with the voltage and current data, capable of capturing vibration signals from 5Hz to 1000Hz. Humidity data is acquired using an environmental humidity sensor, with a range of 0%-100%RH and an accuracy of ±2%RH. All three types of data are stored and marked with a unified timestamp.

[0030] In this process, when performing Fast Fourier Transform (FFT) on heterogeneous data sources, transformation operations are performed separately for three types of data: temperature, vibration frequency, and humidity. This converts the time-domain signals into frequency-domain signals, highlighting the characteristic differences across different frequency bands. The preset characteristic frequency bands are determined based on common circuit board faults and environmental interference patterns. The characteristic frequency band for temperature data is 0.1Hz-1Hz, corresponding to the slow change trend of component heating, which is related to aging faults. The characteristic frequency band for vibration frequency is 10Hz-50Hz, covering characteristic fluctuations caused by equipment operation vibration and poor contact. The characteristic frequency band for humidity data is 0.01Hz-0.1Hz, focusing on the impact of slow humidity drift on circuit stability. By extracting the amplitude, phase, and other frequency-domain response characteristics of each heterogeneous data source in the corresponding characteristic frequency bands, irrelevant frequency components are filtered out, accurately identifying effective information related to faults.

[0031] In another implementation, this embodiment, when constructing the adaptive filtering operator, dynamically adjusts the filtering parameters based on the frequency domain response characteristics of each heterogeneous data source, using the minimum mean square error criterion. The filtering operator designs suppression logic for the frequency characteristics of different interference types. For power frequency interference (50Hz) and harmonics, it accurately attenuates the signal in this frequency band through notch filtering characteristics; for random noise interference, it uses a smoothing filtering mechanism to reduce its amplitude; for slow drift interference caused by environmental factors, it achieves suppression through adaptive threshold filtering. The filtering operator is applied to the original voltage and current operating data, processing the data point by point and removing various interference components to generate an interference removal vector. This interference removal vector retains the core signals related to the circuit board's own faults, significantly reducing the impact of environmental interference on subsequent analysis.

[0032] It should be noted that when performing high-dimensional space mapping using kernel mapping functions, a Gaussian kernel function is selected to achieve the joint mapping of the disturbance removal vector and the state fluctuation accumulation index. The disturbance removal vector (including voltage and current dimensions) and the state fluctuation accumulation index (single dimension) together constitute the input vector, which is transformed into a 128-dimensional high-dimensional feature vector after kernel mapping. This allows for the exploration of nonlinear correlations between data of different dimensions, the coupling relationship between voltage fluctuations and vibration characteristics, and the potential connection between the fluctuation accumulation index and temperature changes. The high-dimensional feature vectors of each time step are arranged in chronological order to construct a multi-dimensional fusion matrix with a dimension of (time step number × 128). When performing manifold projection on the matrix, a local linear embedding algorithm is used to reduce the 128-dimensional high-dimensional data to 3 dimensions while preserving the local topological structure of the data. This forms an evolution trajectory that intuitively reflects the overall state changes of the system. The direction, curvature, and distance from the origin of the trajectory all characterize the stability of the state. When determining the unstable principal axis through principal component analysis, the covariance matrix is ​​calculated on the dimensionality-reduced evolution trajectory data to solve for the eigenvalues ​​and eigenvectors of the covariance matrix.

[0033] The direction corresponding to the eigenvector with the largest eigenvalue is the unstable principal axis, which reflects the core trend of the system state evolving towards instability. When calculating the projected energy of the evolution trajectory on the unstable principal axis, each point on the trajectory is projected onto this principal axis, and the projected energy at each time step is obtained by summing the squares of the projected values. A larger energy value indicates that the system deviates further from a stable state. As time progresses, if the projected energy gradually decreases, it indicates that the system state is tending towards stability; if the energy continues to increase, the instability trend intensifies, and the difference in energy over time is the energy decay value (a negative value indicates energy growth, i.e., intensified instability).

[0034] It is important to note that when normalizing the energy decay value, a maximum-minimum normalization method is used to map the decay value to a numerical range of 0-10, resulting in a risk score for overall system instability. The mapping rule is set so that the minimum energy decay value (most severe instability trend) corresponds to a risk score of 10, and the maximum energy decay value (most stable system) corresponds to a risk score of 0. The numerical range of the risk score has clear warning significance: 0-3 points indicate stable system operation with no obvious failure risk; 4-6 points indicate a slight instability trend requiring continuous monitoring; and 7-10 points indicate a high failure risk requiring timely intervention. The risk score quantifies the overall instability of the system, providing an intuitive and quantifiable basis for subsequent safety boundary updates and early warning signal generation.

[0035] In step S15, a time sliding window model is constructed based on the risk score to extract state fluctuation features. The state fluctuation features are fitted to obtain a state evolution sequence. The state evolution sequence is iteratively updated to obtain the system safety boundary, including: The length of the time window is adjusted according to the numerical range of the risk score to construct a time window model. The running data within the continuous window is slidably extracted from the time window model to extract the state fluctuation characteristics of voltage fluctuation frequency, current fluctuation amplitude, and characteristic value change rate. The state fluctuation characteristics are input into a weighted accumulator, and weights are assigned to the degree of influence on system stability. The fluctuation accumulation increment is generated by weighted summation, and the fluctuation accumulation increment is mapped to a nonlinear state space to obtain a continuous and smooth state evolution sequence. Multi-scale entropy calculation is performed on the state evolution sequence to obtain the uncertainty component in the sequence. The fault alarm baseline is adjusted according to the magnitude of the uncertainty component by a preset ratio. The adjusted baseline is used as the iteratively updated system safety boundary to obtain the system safety boundary.

[0036] In one implementation, the adjustment of the time window length is strongly tied to the risk score range, dynamically adapting to the risk level to ensure the accuracy and efficiency of feature extraction under different operating conditions. Specifically, the range division and corresponding window lengths are as follows: for risk scores of 0-3 (low risk, stable system operation), the window length is set to 10 seconds; this long window design effectively filters random noise and instantaneous interference, focusing on long-term stable trends. For scores of 4-6 (medium risk, slight signs of instability), the window length is adjusted to 5 seconds, balancing sensitivity and anti-interference capabilities, capturing potential fluctuations while avoiding overreaction. For scores of 7-10 (high risk, significant instability), the window length is shortened to 2 seconds, quickly capturing high-frequency fluctuation features and reserving time for emergency intervention. The sliding step size is uniformly set to half the window length (5 seconds for low risk, 2.5 seconds for medium risk, and 1 second for high risk) to ensure data overlap and continuity, avoiding feature omissions due to excessively large step sizes, while controlling computational load and ensuring real-time performance.

[0037] It is worth noting that the length of the time window is dynamically adjusted based on the risk score, which aims to optimize the granularity of feature extraction for different risk levels. A short window is used for high risk to improve the sensitivity to capture rapid fluctuations, while a long window is used for low risk to smooth random noise and observe long-term trends. Although the window length changes, the three types of state fluctuation features extracted—voltage fluctuation frequency, current fluctuation amplitude, and feature value change rate—are all designed as normalized or rate indicators that are relatively independent of the time window length.

[0038] For example, voltage fluctuation frequency is the number of zero-crossings per unit time, current fluctuation amplitude is the ratio of peak difference to mean, and eigenvalue change rate is the change per unit time. Therefore, even when calculated under different time windows, these eigenvalues ​​are still physically comparable. The fixed weights assigned to each feature in the subsequent weighted accumulator are parameters that reflect the inherent influence of each feature on system stability, determined based on a large amount of historical data and fault case analysis. The effectiveness of these weights is not sensitive to the absolute scale of the eigenvalues, but mainly depends on the relative importance between features. In addition, the step of mapping the fluctuation accumulation increment to the nonlinear state space itself has a certain scale adaptation and normalization capability, which can further ensure that the final generated state evolution sequence remains continuous, smooth, and comparable under different sliding window lengths, thereby supporting the reliable updating of the dynamic system safety boundary. Specifically, in the offline stage, operating data of multiple circuit boards of the same model under various fault modes are collected. Through feature importance assessment, the correlation coefficients of three types of features—voltage fluctuation frequency, current fluctuation amplitude, and eigenvalue change rate—with the fault occurrence time are calculated. After normalization, the weight assignment is obtained, and its effectiveness is verified using validation set data.

[0039] It should be noted that after constructing the time sliding window model, voltage and current operating data within a continuous window are extracted according to the adjusted length. Three types of core state fluctuation characteristics are extracted for each window. The voltage fluctuation frequency is obtained by counting the number of times the voltage signal crosses the zero point of the mean line within the window and dividing it by the window duration. This reflects the frequency of voltage changes. An abnormally high frequency may indicate poor component contact or circuit load fluctuations. The current fluctuation amplitude is calculated by dividing the difference between the maximum and minimum current values ​​within the window by the mean current within the window to obtain the normalized fluctuation amplitude (dimensionless). This avoids the distortion of the index caused by the difference in the mean current under different operating conditions and intuitively reflects the relative severity of current fluctuations. The characteristic value change rate is obtained by selecting the state fluctuation cumulative index generated in S14 as the core characteristic value, calculating the difference between the first and last characteristic values ​​within the window, and then dividing it by the window duration.

[0040] It should be noted that the three types of extracted state fluctuation features are simultaneously input into a weighted accumulator. The weight allocation is based on the degree of influence of each feature on system stability. Voltage fluctuation frequency is most strongly correlated with faults such as circuit insulation aging and poor contact, and its weight is set to 0.4; current fluctuation amplitude directly reflects load anomalies and short-circuit precursors, and its weight is set to 0.35; the rate of change of feature values ​​can detect trend changes in advance, and its weight is set to 0.25. The calculation logic of the weighted accumulator is to multiply the three types of features at each time step by their corresponding weights and then sum them to obtain the fluctuation accumulation increment of the current window. The increment directly reflects the comprehensive intensity of fluctuations within the current window; a positive increment indicates that the fluctuation accumulation is intensifying, and a negative increment indicates that the fluctuation is becoming smoother. To obtain a continuous and smooth state evolution sequence, a nonlinear mapping function is used to map the fluctuation accumulation increment to a two-dimensional nonlinear state space. During the mapping process, the increment value is used as input, converted into two-dimensional coordinate points by a function, and all two-dimensional coordinate points are connected sequentially in time to form a state evolution sequence.

[0041] In another implementation, when calculating the multi-scale entropy of the state evolution sequence, three typical time scales are selected (corresponding to 2 seconds, 5 seconds, and 10 seconds). For each time scale, the entropy value of the sequence is calculated using the sample entropy algorithm (entropy value range 0-3). The larger the entropy value, the higher the complexity of the sequence at that scale, and the stronger the uncertainty of the system state. To comprehensively reflect multi-dimensional uncertainty, the entropy values ​​of the three scales are weighted equally (1 / 3 each) to obtain a comprehensive uncertainty component, with the component value ranging from 0-3. 0-1 indicates low uncertainty (system state is predictable), 1-2 indicates moderate uncertainty (some fluctuations exist but are controllable), and 2-3 indicates high uncertainty (system state is complex and difficult to predict).

[0042] Specifically, the calculation of the multi-scale entropy involves coarsening the original state evolution sequence, specifically for the scale factor. Construct coarse-grained sequences ,in The original sequence, , Let the sequence length be ; then, for each scale factor, calculate the sample entropy value of the coarse-grained sequence. This value is the multi-scale entropy at that scale. The formula for calculating sample entropy is: ,in Let be the embedding dimension, be the similarity tolerance, and be the dimension between the two sequences. The probability of a match. For two sequences in dimension The probability of matching. The uncertainty component is the comprehensive value obtained by weighted averaging of the sample entropy values ​​under multiple selected scale factors, used to quantify the complexity and unpredictability of the system at different time scales.

[0043] It is worth noting that the embedding dimension is empirically set to 2 based on the sequence length and the number of samples, and the similarity tolerance is set to 0.2 times the standard deviation of the sequence. This value can effectively distinguish between ordered and disordered states and is a commonly used setting in the field of signal processing. The time scale is selected as 2 seconds, 5 seconds, and 10 seconds, which correspond to the transient fluctuations, short-term trends, and medium-term evolution of the circuit board, respectively, and can comprehensively reflect the state complexity under different time scales.

[0044] It should be noted that the initial value of the fault alarm baseline is determined through offline training. 100 sets of state evolution sequences of the same model of circuit board under stable operating conditions are collected, the maximum fluctuation peak value of each sequence is calculated, and the 99th percentile is taken as the initial baseline (set to 5). The baseline is dynamically adjusted according to the magnitude of the uncertainty component at a preset ratio to ensure the adaptability of the safety boundary. When the uncertainty component is less than 1.0 (low uncertainty), the baseline is lowered by 5% (e.g., adjusted to 4.75) to enhance the system's anti-interference capability and avoid false alarms due to random fluctuations. When the uncertainty component is between 1.0 and 2.0 (medium uncertainty), the baseline remains unchanged to balance sensitivity and stability. When the uncertainty component is greater than 2.0 (high uncertainty), the baseline is raised by 10% (e.g., adjusted to 5.5) to improve system sensitivity and promptly capture fault symptoms under complex fluctuations. The iteratively updated system safety boundary is a dynamically changing curve, not a fixed threshold; its core characteristic is its adaptation to the fluctuation patterns of the state evolution sequences. The upper and lower limits of the safety boundary are generated based on the adjusted fault alarm baseline. The upper limit corresponds to the critical state requiring vigilance, and the lower limit corresponds to the critical state of safe and stable operation. When the state evolution sequence is between the upper and lower limits, the system is considered to be operating normally; when the sequence exceeds the upper limit, it indicates a fault risk; when the sequence falls below the lower limit, it indicates that the system state is approaching stability. The dynamic safety boundary accurately adapts to the dynamic changes in the circuit board's operating state, solving the problem of poor adaptability of traditional fixed thresholds, and significantly improving the flexibility and reliability of fault diagnosis.

[0045] In step S16, if the state evolution sequence exceeds the system safety boundary, a potential threshold crossing is determined, and instantaneous data at the time of the potential threshold crossing is extracted to generate a warning signal, including: The state evolution sequence is compared with the system safety boundary in real time. When the feature value of any moment in the state evolution sequence exceeds the system safety boundary for the first time, it is determined that a potential threshold crossing has occurred, and the timestamp of the corresponding moment is recorded. Based on the timestamp, the instantaneous values ​​of voltage, current, temperature and vibration within a preset time period are extracted and arranged in chronological order to construct a state sequence vector; The state sequence vector is input into the Long Short-Term Memory (LSTM) network to filter historical dependency information corresponding to the potential threshold crossing, filter irrelevant and redundant data, and update the memory units of the LTM network. Based on the deviation between the output of the memory unit and the actual instantaneous data, the error gradient is calculated, and the long short-term memory network weight matrix is ​​iteratively corrected using the error gradient until the error gradient is lower than a preset accuracy threshold. Then, a calibrated activation state prediction sequence is output. If any feature value in the activation state prediction sequence exceeds the preset safety tolerance, an early warning signal is generated. The early warning signal includes a timestamp and abnormal location information.

[0046] In one implementation, this embodiment employs a time-by-time traversal mechanism for real-time comparison of the state evolution sequence and the system safety boundary. The comparison frequency is consistent with the data sampling frequency (1kHz) to ensure zero-delay capture of boundary crossing events. The comparison logic focuses on the first crossing determination. When a feature value in the state evolution sequence first exceeds the upper safety limit or falls below the lower safety limit, a potential threshold crossing is immediately determined, and only the timestamp of this first crossing moment (accurate to the millisecond level) is recorded to avoid repeated triggering of the same crossing event and reduce redundant data. The timestamp recording uses a dual verification mechanism: on the one hand, the system's local clock time (synchronized to UTC time) is extracted; on the other hand, the hardware timestamp of the data acquisition module is correlated. The error between the two is controlled within 1 millisecond to ensure the accuracy of time information. The recorded timestamp includes year, month, day, hour, minute, and second, and is accurate to the millisecond, providing a precise time anchor for subsequent instantaneous data extraction and fault tracing.

[0047] In another implementation, this embodiment extracts instantaneous data and constructs a state sequence vector by using the recorded timestamp as the center and extracting instantaneous data for 2 seconds before and after it. The extracted data types cover four key physical quantities: voltage, current, temperature, and vibration instantaneous values. Real-time voltage data of the core power supply circuit of the circuit board is collected with an accuracy of 0.001V, covering subtle fluctuations within the normal operating voltage range (0.5-5.0V); current data of the corresponding circuit is collected simultaneously with an accuracy of 0.001A to capture abnormal features such as current abrupt changes and spikes; data from the surface-mount temperature sensor on the circuit board chip is extracted with an accuracy of ±0.1℃, reflecting sudden temperature rises or falls across time intervals; and data from the miniature accelerometer on the device housing is acquired with a sampling frequency of 1kHz to capture vibration and shock signals in the range of 5Hz-1000Hz. The four types of extracted instantaneous data are arranged in chronological order (from 2 seconds before to 2 seconds after the timestamp, a total of 4000 sampling points) to construct a state sequence vector. The vector dimension is (4000×4), where the four feature dimensions correspond to the instantaneous values ​​of voltage, current, temperature, and vibration, respectively. The values ​​of each dimension are normalized (mapped to the 0-1 interval) to eliminate the influence of dimensional differences on subsequent model inputs and ensure data semantic consistency.

[0048] It should be noted that the Long Short-Term Memory (LSTM) network used to process the state sequence vector adopts a deep optimization architecture. The LTM network contains four hidden layers, each with 128 neurons. The input layer dimension matches the state sequence vector dimension (4-dimensional), and the output layer dimension is 64-dimensional (corresponding to the core features of the activation state prediction). The network uses a gating mechanism (forget gate, input gate, and output gate) to accurately filter historical dependency information. Specifically, the forget gate filters redundant data in the state sequence vector (normal voltage and current fluctuations, temperature drift unrelated to the crossing event), outputting a weight value between 0 and 1 through the activation function. A weight close to 0 indicates discarding the information, while a weight close to 1 indicates retention. The input gate focuses on key information related to potential threshold crossings (voltage surges, vibration and shock data), transforming it into candidate state update values ​​and writing them into the memory unit. The output gate generates an output vector reflecting the core features based on the current memory unit state, completing the update of the memory unit.

[0049] The memory unit is updated using a dynamic iterative mechanism. For each time step of the state sequence vector, the memory unit’s storage content is adjusted according to the output of the gating mechanism. This continuously retains historical dependency information that is strongly correlated with the threshold crossing (accumulated fluctuation data before the crossing and instantaneous peak characteristics at the crossing moment), filters out irrelevant and redundant data, and ensures that the memory unit focuses on the core fault symptoms.

[0050] It is worth noting that the network was trained using data collected from multiple circuit boards of the same model as the circuit board to be predicted under normal operation and simulated fault conditions. The training samples covered four types of signals: voltage, current, temperature, and vibration, and thousands of samples were collected and divided into training and validation sets according to the proportion. During training, mean square error was used as the loss function, and the network weights were optimized through backpropagation until the validation set error converged.

[0051] In another implementation, this embodiment calculates the deviation between the output (predicted value) of the Long Short-Term Memory (LSTM) network memory unit and the actual instantaneous data (true value), using mean squared error as the deviation quantification index to highlight the significant difference between the predicted and true values. The error gradient is calculated based on the backpropagation algorithm, tracing back layer by layer from the output layer to the input layer, calculating the contribution of each weight parameter to the error, forming a complete error gradient matrix. When using the error gradient for iterative correction, the Adam optimizer controls the correction rhythm. During the correction process, the network weight matrix is ​​updated with each iteration, and a new error gradient is calculated simultaneously, until the error gradient falls below a preset accuracy threshold (0.0001), at which point iteration stops, and the activated state prediction sequence is output. The activated state prediction sequence is a 64-dimensional vector, with each dimension corresponding to a core fault characteristic. The temporal correlation and numerical magnitude of the sequence reflect the evolution trend of the system state in the near future (1-5 minutes).

[0052] In another implementation, this embodiment sets preset safety tolerances for each of the 64 feature dimensions of the active state prediction sequence. The tolerance range for each dimension is determined based on the normal operating threshold corresponding to that feature. 1000 sets of active state sequences under stable circuit board operating conditions are collected offline. The maximum and minimum values ​​of each feature dimension are calculated. 1.2 times the maximum value is taken as the upper safety tolerance limit for that dimension, and 0.8 times the minimum value is taken as the lower safety tolerance limit, forming a 64-dimensional safety tolerance matrix. During judgment, the feature values ​​of the active state prediction sequence are compared with the safety tolerance matrix dimension by dimension. If the feature value of any dimension exceeds the corresponding tolerance range (higher than the upper tolerance limit or lower than the lower tolerance limit), it is determined that the system has a clear fault risk, and an early warning signal is immediately triggered.

[0053] In step S17, a subtle feature sequence is extracted based on the warning signal, and the feature values ​​of the subtle feature sequence are calculated to obtain the final stability assessment, including: Based on the timestamp and abnormal location information in the warning signal, the corresponding original monitoring data stream is locked, and wavelet transform is used for multi-resolution analysis to decompose the original monitoring data stream into signal components of different frequency scales. Feature sequences reflecting subtle fluctuations in the high-frequency band are extracted to obtain subtle feature sequences. Calculate the topological distance between the subtle feature sequences at adjacent time steps, construct a transformation trajectory vector based on the topological distance, input the transformation trajectory vector into a preset fault mode space, calculate the degree of matching with various fault mode vectors in the library through cosine similarity, and output a state observation matrix, where the matrix elements reflect the degree of fit between the current state and the fault mode. The eigenvalues ​​of the state observation matrix are calculated, and the degree of deviation of the system state is quantified by the degree of divergence of the eigenvalues. If the degree of divergence exceeds a preset divergence threshold, the offset between the current state and the normal state is calculated. Based on the offset and the risk score of overall system instability, a final stability assessment report is generated.

[0054] In one implementation, this embodiment constructs a dual-screening mechanism based on the precise timestamp and abnormal location information in the early warning signal to lock the original monitoring data stream. In the time dimension, centered on the timestamp, the complete data stream is expanded to include 5 seconds before and after the locked data, ensuring coverage of subtle fluctuations in the data before and after the threshold. In the spatial dimension, the focus is on the monitoring loop corresponding to the abnormal location, extracting only the four types of original data—voltage, current, temperature, and vibration—from that loop, filtering out data from other irrelevant loops. When using wavelet transform for multi-resolution analysis, the db4 wavelet basis function is selected. During the decomposition process, the original monitoring data stream is decomposed into five scales: Scale 1 (high frequency band, 125Hz-250Hz), Scale 2 (high frequency band, 62.5Hz-125Hz), Scale 3 (mid frequency band, 31.25Hz-62.5Hz), Scale 4 (mid-low frequency band, 15.625Hz-31.25Hz), and Scale 5 (low frequency band, 0Hz-15.625Hz). Among them, the high-frequency signals corresponding to scale 1 and scale 2 concentrate the subtle fluctuations in the early stage of faults and are the core frequency bands for extracting subtle features.

[0055] It should be noted that when extracting the high-frequency feature sequence, threshold denoising was performed on the signal components at scales 1 and 2 to filter out high-frequency noise interference. Subsequently, four key features—peak value, valley value, number of zero crossings, and waveform slope—were extracted and arranged in chronological order to form a subtle feature sequence. The sequence dimension is (10000×4), where 10000 time steps correspond to a 10-second data stream (1kHz sampling frequency), and the four feature dimensions correspond to the four types of high-frequency features, thus fully preserving the subtle fluctuation information in the early stages of the fault, resulting in the subtle feature sequence.

[0056] In another implementation, this embodiment quantifies the dynamic changes of subtle feature sequences by calculating the differences between subtle feature sequences at adjacent time steps. Euclidean distance is used to calculate the distance between the 4-dimensional subtle feature vectors at time t and time t+1. The larger the distance value, the more significant the subtle feature changes between the two time steps. When constructing the transformation trajectory vector based on topological distance, the topological distance values ​​of 100 consecutive time steps are concatenated in chronological order to form a 100-dimensional transformation trajectory vector, which is then normalized (mapped to the 0-1 interval). The construction of the pre-defined fault mode space involves training with multi-source data. In the offline phase, subtle feature sequences of 15 typical circuit board faults are collected, including but not limited to component aging, power supply fluctuations, short circuit precursors, cold solder joints, and overheating. 500 samples are collected for each fault type, and the topological distance is calculated to construct the transformation trajectory vector, which is then stored in the mode library as a fault mode vector. Simultaneously, each fault mode vector is labeled with its corresponding fault type, severity, and typical feature information, forming a structured pre-defined fault mode space that supports rapid matching and querying.

[0057] It should be noted that after inputting the transformed trajectory vector into a preset fault mode space, the degree of matching with various fault mode vectors in the library is calculated using cosine similarity. The cosine similarity value ranges from -1 to 1, with values ​​closer to 1 indicating a higher degree of matching and values ​​closer to -1 indicating a lower degree of matching. After calculating the similarity for each type of fault mode vector, the data is organized into a state observation matrix according to the fault type. The matrix has a dimension of (15×1), and each element corresponds to the matching similarity of a fault mode type.

[0058] In another implementation, the eigenvalues ​​of the state observation matrix are calculated using eigenvalue decomposition. The eigenvalues ​​are obtained by solving the matrix's characteristic equation. The degree of divergence of the eigenvalues ​​quantifies the deviation of the system state, characterized by the coefficient of variation (the ratio of standard deviation to mean). A larger coefficient of variation indicates a more dispersed distribution of eigenvalues ​​and a more significant deviation of the system state from the normal mode; a smaller coefficient of variation indicates a more concentrated distribution of eigenvalues ​​and a clearer indication of the system state's fault. A preset divergence threshold is determined through offline experiments. One hundred sets of state observation matrices under normal operating conditions are collected, and their eigenvalue coefficients of variation are calculated. The 99.5th percentile is used as the preset divergence threshold. If the eigenvalue coefficient of variation exceeds the preset divergence threshold, it indicates that the system state has significantly deviated from the normal mode, and further calculation of the offset between the current state and the normal state is required. The reference vector for the normal state is taken from the mean of the transformation trajectory vectors of all normally operating samples in the preset fault mode space. The offset is obtained by calculating the Euclidean distance between the current transformation trajectory vector and the reference vector; a larger distance indicates a greater deviation from the normal state.

[0059] In summary, this invention discloses a circuit board fault prediction method, which constructs a dynamic prediction closed loop with the risk score of overall system instability as the pivot. This includes: real-time acquisition of voltage and current signal data during circuit board operation using high-precision sensors; extraction of dynamic state change sequences through time-series feature extraction; nonlinear classification using a support vector machine model to generate state evolution trajectories and locate the starting points of instability signs; identification of stability qualitative change trends through amplitude intensity threshold judgment and fluctuation accumulation index analysis; quantification of the risk score of overall system instability by fusing heterogeneous data sources of temperature, vibration, and humidity, followed by spectral decomposition, interference removal, and multi-dimensional data fusion; dynamic adjustment of the time sliding window model based on the risk score to extract state fluctuation features and iteratively update the system safety boundary; when the state evolution sequence exceeds the safety boundary, a potential threshold crossing is determined and a warning signal containing timestamps and abnormal locations is generated; finally, subtle feature sequences are extracted through wavelet transform, combined with pre-set fault mode space matching and feature value divergence analysis to generate a final evaluation report covering stability levels and maintenance recommendations.

[0060] Reference Figure 2 The second embodiment of the present invention provides a circuit board fault prediction system, comprising: The data acquisition module is used to collect voltage and current signal data during the operation of the circuit board in real time, extract time-series features from the voltage and current signal data, and obtain a dynamic state change sequence. The starting point determination module is used to input the dynamic state change sequence into a preset support vector machine model to obtain the state evolution trajectory, and calculate the time point when the evolution rate changes abruptly in the state evolution trajectory as the starting point of the instability sign. The trend judgment module is used to calculate the amplitude intensity value of the starting point. If the amplitude intensity value exceeds the preset amplitude intensity threshold, the state fluctuation accumulation index is extracted from the voltage and current signal data to judge the stability change trend of the starting point. The risk calculation module is used to obtain heterogeneous data sources and perform spectrum decomposition and interference removal if the stability change trend exists, to obtain the purified effective feature vector, and to perform multi-dimensional data fusion of the effective feature vector and the state fluctuation accumulation index to obtain the overall system instability risk score. The safety boundary update module is used to construct a time sliding window model based on the risk score to extract state fluctuation features, fit the state fluctuation features to obtain a state evolution sequence, and iteratively update the state evolution sequence to obtain the system safety boundary. The warning signal generation module is used to determine a potential threshold crossing if the state evolution sequence exceeds the system safety boundary, extract instantaneous data when the potential threshold crossing occurs, and generate a warning signal. The stability assessment module is used to extract subtle feature sequences based on the warning signal, calculate the feature values ​​of the subtle feature sequences, and obtain the final stability assessment.

[0061] It should be noted that the circuit board fault prediction system provided in this embodiment of the invention is used to execute all the process steps of the circuit board fault prediction method in the above embodiment. The working principle and beneficial effect of the two correspond one-to-one, so they will not be described again.

[0062] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0063] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A circuit board fault prediction method, comprising: Real-time acquisition of voltage and current signal data during the operation of the circuit board; extraction of time-series features from the voltage and current signal data to obtain a dynamic state change sequence; The dynamic state change sequence is input into a preset support vector machine model to obtain the state evolution trajectory. The time point when the evolution rate changes abruptly in the state evolution trajectory is calculated as the starting point of the instability sign. Calculate the amplitude intensity value at the starting point. If the amplitude intensity value exceeds a preset amplitude intensity threshold, extract the state fluctuation accumulation index from the voltage and current signal data to determine the stability change trend of the starting point. If the aforementioned trend of qualitative change in stability exists, heterogeneous data sources are acquired and spectral decomposition and interference removal are performed to obtain purified effective feature vectors. The effective feature vectors are then fused with the cumulative state fluctuation index to obtain a risk score for overall system instability. Based on the risk score, a time sliding window model is constructed to extract state fluctuation features. The state fluctuation features are fitted to obtain a state evolution sequence. The state evolution sequence is iteratively updated to obtain the system safety boundary. If the state evolution sequence exceeds the system safety boundary, a potential threshold crossing is determined, and instantaneous data at the time of the potential threshold crossing is extracted to generate an early warning signal. Based on the warning signal, a subtle feature sequence is extracted, and the feature value of the subtle feature sequence is calculated to obtain the final stability assessment. The step of constructing a time-sliding window model based on the risk score to extract state fluctuation features, fitting the state fluctuation features to obtain a state evolution sequence, and iteratively updating the state evolution sequence to obtain the system safety boundary includes: The length of the time window is adjusted according to the numerical range of the risk score to construct a time window model. The running data within the continuous window is slidably extracted from the time window model to extract the state fluctuation characteristics of voltage fluctuation frequency, current fluctuation amplitude, and characteristic value change rate. The state fluctuation characteristics are input into a weighted accumulator, and weights are assigned to the degree of influence on system stability. The fluctuation accumulation increment is generated by weighted summation, and the fluctuation accumulation increment is mapped to a nonlinear state space to obtain a continuous and smooth state evolution sequence. Multi-scale entropy calculation is performed on the state evolution sequence to obtain the uncertainty component in the sequence. The fault alarm baseline is adjusted according to the magnitude of the uncertainty component by a preset ratio. The adjusted baseline is used as the iteratively updated system safety boundary to obtain the system safety boundary.

2. The circuit board fault prediction method according to claim 1, characterized in that, The real-time acquisition of voltage and current signal data during the operation of the circuit board, and the extraction of time-series features from the voltage and current signal data to obtain a dynamic state change sequence, including: The analog voltage and analog current signals of the circuit board are acquired in real time during operation. Continuous data points are obtained through high-precision sensors. The analog voltage and analog current signals are converted into a multi-dimensional original waveform matrix containing voltage and current dimensions to obtain voltage and current signal data. The original multidimensional waveform matrix is ​​sliced ​​to obtain multiple sub-matrices, and the sub-matrices are reorganized into a temporal feature tensor containing time dimension and feature dimension; The temporal feature tensor is input into a long short-term memory network, the data of each time window in the long short-term memory network is calculated, the hidden layer state vector is obtained, the transient feature descriptor is generated by mapping the hidden layer state vector, and the Euclidean distance between the transient feature descriptors corresponding to adjacent time windows is calculated. If the Euclidean distance exceeds a preset distance threshold, it is marked as a state transition point. All the state transition points are connected in chronological order to form a dynamic state change sequence.

3. The circuit board fault prediction method according to claim 1, characterized in that, The step of inputting the dynamic state change sequence into a preset support vector machine model to obtain a state evolution trajectory, and calculating the time point at which the evolution rate in the state evolution trajectory changes abruptly as the starting point of the instability sign, includes: The dynamic state change sequence is mapped to a high-dimensional state feature vector set using a radial basis kernel function. The high-dimensional state feature vector set is then input into a preset support vector machine model to construct the optimal decision hyperplane. Calculate the normal distance from each vector point in the high-dimensional state feature vector set to the optimal decision hyperplane, arrange the normal distances in chronological order, and generate the state evolution trajectory; The evolution rate characteristics are obtained by taking the second derivative of the normal distance sequence of the state evolution trajectory. The time index of the absolute value of the evolution rate feature that exceeds a preset mutation threshold is tracked, and the time corresponding to the time index is determined as the starting point of the unstable sign.

4. The circuit board fault prediction method according to claim 1, characterized in that, The calculation of the amplitude intensity value at the starting point, if the amplitude intensity value exceeds a preset amplitude intensity threshold, involves extracting a state fluctuation accumulation index from the voltage and current signal data to determine the stability change trend at the starting point, including: Based on the timestamp of the starting point, extract the state vector data of the timestamp and calculate the amplitude intensity value; The amplitude intensity value is compared with a preset amplitude intensity threshold. If the amplitude intensity value exceeds the threshold, the full-dimensional running data set within a preset time range before and after the starting point is locked. The least squares method is used to fit the full-dimensional operational data set, and the deviation between the fitting result and the actual data is extracted to form a residual matrix containing information about subtle changes in the system, which serves as hidden information in the operational data. The energy spectral density is obtained by performing a Fourier transform on the residual matrix, and the energy spectral density is integrated within a preset frequency range to obtain the cumulative index of state fluctuation. Calculate the first-order difference vector of the cumulative index of state fluctuations over consecutive time steps. If the magnitude of the first-order difference vector of the cumulative index ...

5. The circuit board fault prediction method according to claim 1, characterized in that, If the aforementioned trend of qualitative change in stability exists, heterogeneous data sources are acquired and subjected to spectral decomposition and interference removal to obtain purified effective feature vectors. These effective feature vectors are then fused with the accumulated state fluctuation index using multi-dimensional data to obtain a risk score for overall system instability, including: Based on the stability change trend, temperature data, vibration frequency data, and humidity data at the starting point are obtained as heterogeneous data sources. Fast Fourier Transform is performed on the heterogeneous data sources respectively to extract the frequency domain response features of the heterogeneous data sources in the characteristic frequency range, which is preset according to the common interference frequency range of the circuit board. An adaptive filtering operator is constructed based on the frequency domain response characteristics. The filtering operator is used to suppress interference components in the original running data and generate an interference removal vector that removes environmental interference. The interference removal vector and the state fluctuation accumulation index are jointly mapped to a high-dimensional state space by a kernel mapping function, and a multi-dimensional fusion matrix is ​​constructed by arranging them in time sequence. The multi-dimensional fusion matrix is ​​then subjected to manifold projection to reduce the dimensionality and obtain the evolution trajectory that reflects the overall state change of the system. Principal component analysis is used to determine the unstable principal axis that characterizes the unstable direction of the system, and the decay value of the projection energy of the evolution trajectory on the unstable principal axis over time is calculated. The attenuation value is normalized and mapped to a risk score of overall system instability between 0 and 10, thus obtaining the risk score of overall system instability.

6. The circuit board fault prediction method according to claim 1, characterized in that, If the state evolution sequence exceeds the system safety boundary, a potential threshold crossing is determined, and instantaneous data at the time of the potential threshold crossing is extracted to generate a warning signal, including: The state evolution sequence is compared with the system safety boundary in real time. When the feature value of any moment in the state evolution sequence exceeds the system safety boundary for the first time, it is determined that a potential threshold crossing has occurred, and the timestamp of the corresponding moment is recorded. Based on the timestamp, the instantaneous values ​​of voltage, current, temperature and vibration within a preset time period are extracted and arranged in chronological order to construct a state sequence vector; The state sequence vector is input into the Long Short-Term Memory (LSTM) network to filter historical dependency information corresponding to the potential threshold crossing, filter irrelevant and redundant data, and update the memory units of the LTM network. Based on the deviation between the output of the memory unit and the actual instantaneous data, the error gradient is calculated, and the long short-term memory network weight matrix is ​​iteratively corrected using the error gradient until the error gradient is lower than a preset accuracy threshold. Then, a calibrated activation state prediction sequence is output. If any feature value in the activation state prediction sequence exceeds the preset safety tolerance, an early warning signal is generated. The early warning signal includes a timestamp and abnormal location information.

7. The circuit board fault prediction method according to claim 1, characterized in that, The step of extracting subtle feature sequences based on the warning signal, calculating the feature values ​​of the subtle feature sequences, and obtaining the final stability assessment includes: Based on the timestamp and abnormal location information in the warning signal, the corresponding original monitoring data stream is locked, and wavelet transform is used for multi-resolution analysis to decompose the original monitoring data stream into signal components of different frequency scales. Feature sequences reflecting subtle fluctuations in the high-frequency band are extracted to obtain subtle feature sequences. Calculate the topological distance between the subtle feature sequences at adjacent time steps, construct a transformation trajectory vector based on the topological distance, input the transformation trajectory vector into a preset fault mode space, calculate the degree of matching with various fault mode vectors in the library through cosine similarity, and output a state observation matrix, where the matrix elements reflect the degree of fit between the current state and the fault mode. The eigenvalues ​​of the state observation matrix are calculated, and the degree of deviation of the system state is quantified by the degree of divergence of the eigenvalues. If the degree of divergence exceeds a preset divergence threshold, the offset between the current state and the normal state is calculated. Based on the offset and the risk score of overall system instability, a final stability assessment report is generated.

8. A circuit board fault prediction system, characterized in that, For implementing the method as described in any one of claims 1-7, comprising: The data acquisition module is used to collect voltage and current signal data during the operation of the circuit board in real time, extract time-series features from the voltage and current signal data, and obtain a dynamic state change sequence. The starting point determination module is used to input the dynamic state change sequence into a preset support vector machine model to obtain the state evolution trajectory, and calculate the time point when the evolution rate changes abruptly in the state evolution trajectory as the starting point of the instability sign. The trend judgment module is used to calculate the amplitude intensity value of the starting point. If the amplitude intensity value exceeds the preset amplitude intensity threshold, the state fluctuation accumulation index is extracted from the voltage and current signal data to judge the stability change trend of the starting point. The risk calculation module is used to obtain heterogeneous data sources and perform spectrum decomposition and interference removal if the stability change trend exists, to obtain the purified effective feature vector, and to perform multi-dimensional data fusion of the effective feature vector and the state fluctuation accumulation index to obtain the overall system instability risk score. The safety boundary update module is used to construct a time sliding window model based on the risk score to extract state fluctuation features, fit the state fluctuation features to obtain a state evolution sequence, and iteratively update the state evolution sequence to obtain the system safety boundary. The warning signal generation module is used to determine a potential threshold crossing if the state evolution sequence exceeds the system safety boundary, extract instantaneous data when the potential threshold crossing occurs, and generate a warning signal. The stability assessment module is used to extract subtle feature sequences based on the warning signal, calculate the feature values ​​of the subtle feature sequences, and obtain the final stability assessment.

Images