A self-calibration method and system based on a resistance simulator

Abstract

The application relates to the technical field of resistance simulator calibration, and discloses a self-calibration method and system based on a resistance simulator. The method comprises the following steps: collecting real-time operation parameters through a sensor array to form a data stream, extracting multi-scale feature components and residual components through adaptive signal decomposition, jointly analyzing and identifying abnormal fluctuation modes, and combining causal reasoning to mine key influencing factors. A historical calibration database and key factors are integrated, a reference performance trajectory is generated through a regression model, the deviation degree of an actual trajectory is calculated to evaluate calibration urgency, multi-dimensional information is fused, a calibration strategy is generated through a time series prediction model. The method constructs a closed-loop system from data collection to strategy generation, realizes dynamic monitoring and accurate evaluation, improves calibration automation and decision scientificity, provides a systematic solution for reliable operation of the resistance simulator under complex working conditions, and is suitable for the fields of high-precision electronic testing and measurement.

Description

Technical Field

[0001] This invention relates to the field of resistance simulator calibration technology, specifically to a self-calibration method and system based on a resistance simulator. Background Technology

[0002] As a core device in the field of electronic measurement and metrology, the performance stability of resistance simulators directly affects the accuracy of key processes such as circuit simulation and equipment calibration. With the development of electronic technology towards higher frequencies and greater integration, the application scenarios of resistance simulators in aerospace, power electronics, and precision instruments are becoming increasingly complex. They face operational challenges due to the superposition of multiple factors such as temperature drift, signal noise, and component aging, which places higher demands on the real-time performance, comprehensiveness, and adaptability of calibration methods.

[0003] Traditional calibration methods primarily rely on periodic manual inspection and static parameter comparison, using standard resistors to perform single-point calibration of simulator output values. This makes it difficult to capture the dynamic fluctuations during equipment operation. Furthermore, traditional methods typically calibrate only a single physical quantity (such as resistance value), ignoring the coupling effect between environmental parameters like temperature and voltage fluctuations and internal equipment parameters, thus failing to effectively identify abnormal states caused by multi-parameter interactions.

[0004] In assessing measurement uncertainty, existing technologies often employ a single method to evaluate calibration results, lacking in-depth analysis of the multi-scale characteristics of the data and making it difficult to distinguish between periodic noise and true performance degradation signals. Furthermore, calibration strategies are often based on fixed periods or threshold triggers, failing to fully integrate historical equipment operating data and real-time operating conditions. This leads to inappropriate calibration timing, increased maintenance costs due to over-calibration, and potential accumulation of measurement errors due to delayed calibration.

[0005] With the increasing demand for intelligent technology, resistance simulators are gradually integrating sensor arrays and digital control modules. The multi-dimensional monitoring data generated provides a data foundation for accurate calibration, but also poses challenges to data processing technology. How to extract effective features from complex data streams and establish dynamic calibration models has become a key technical challenge that the industry needs to address. Existing technologies for analog resistance calibration mostly focus on hardware circuit optimization or single-parameter calibration, lacking systematic analysis of equipment operating status and intelligent generation of calibration strategies. The self-calibration method proposed in this invention effectively solves the shortcomings of traditional methods in dynamic monitoring, anomaly identification, and strategy generation through adaptive signal decomposition, multi-factor joint analysis, and data-driven model construction, providing an innovative solution for the reliable operation of high-precision resistance simulators. Summary of the Invention

[0006] The purpose of this invention is to provide a self-calibration method and system based on a resistance simulator to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a self-calibration method based on a resistance simulator, the method comprising:

[0008] The sensor array collects the real-time operating parameters of the resistance simulator to form a real-time monitoring data stream;

[0009] Adaptive signal decomposition is performed on the real-time monitoring data stream to extract multi-scale feature components and residual components;

[0010] Joint analysis of the multi-scale feature components and the residual components is performed to identify abnormal fluctuation patterns.

[0011] Causal reasoning was performed on the aforementioned abnormal fluctuation patterns to identify key influencing factors;

[0012] The historical calibration database of the integrated resistor simulator and the aforementioned key influencing factors are used to generate a reference performance trajectory through a regression model.

[0013] Calculate the deviation between the actual operating trajectory and the reference performance trajectory to assess the urgency of calibrating the resistance simulator;

[0014] By integrating the aforementioned abnormal fluctuation patterns, key influencing factors, and calibration urgency, a calibration strategy is generated using a time series prediction model.

[0015] Preferably, the adaptive signal decomposition of the real-time monitoring data stream to extract multi-scale feature components and residual components includes:

[0016] The real-time monitoring data stream is processed using the empirical mode decomposition algorithm to obtain a set of intrinsic mode functions;

[0017] Perform a Hilbert transform on each intrinsic mode function and calculate the instantaneous frequency and amplitude;

[0018] Significant intrinsic mode functions are selected based on a preset threshold and used as multi-scale feature components.

[0019] The remaining unfiltered intrinsic mode functions are merged into residual components.

[0020] Preferably, joint analysis of the multi-scale feature components and the residual components is performed to identify anomalous fluctuation patterns, including:

[0021] The multi-scale feature components and the residual components are combined into a high-dimensional feature vector;

[0022] The isolation forest algorithm is used to perform unsupervised learning on the high-dimensional feature vector to divide normal data points into abnormal data points;

[0023] Density peak clustering is performed on abnormal data points to identify abnormal fluctuation patterns.

[0024] Preferably, the step of performing causal reasoning on the abnormal fluctuation pattern to uncover key influencing factors includes:

[0025] The abnormal fluctuation patterns are mapped to nodes to construct an initial graph structure;

[0026] Use a Bayesian network learning algorithm to analyze the conditional dependencies between nodes and add directed edges;

[0027] Edge weights are calculated using the maximum information coefficient to identify strong causal paths;

[0028] Frequently occurring nodes are extracted from strong causal paths as key influencing factors.

[0029] Preferably, the historical calibration database of the integrated resistor simulator and the key influencing factors are used to generate a reference performance trajectory through a regression model, including:

[0030] Load multiple sets of calibration event records from the historical calibration database, including environmental conditions, load parameters, and calibration results;

[0031] The key influencing factors are feature-aligned with the calibration event records;

[0032] A Gaussian process regression model is used to perform nonlinear fitting on the aligned features to generate a reference performance trajectory.

[0033] Adjust the hyperparameters of the Gaussian process regression model to minimize the prediction error.

[0034] Preferably, the calculation of the deviation between the actual operating trajectory and the reference performance trajectory, and the assessment of the urgency of calibration of the resistance simulator, includes:

[0035] A dynamic time warping algorithm is used to align the actual running trajectory and the reference performance trajectory, and the distance between the trajectories is calculated.

[0036] The distance between trajectories is input into a convolutional neural network to extract spatial features;

[0037] By weighting spatial features through an attention mechanism, a calibration urgency score is output.

[0038] Preferably, the step of weighting spatial features through an attention mechanism to output a calibration urgency score includes:

[0039] Add a self-attention module after the fully connected layer of the convolutional neural network;

[0040] Calculate the attention weight distribution of the feature vectors;

[0041] The feature vectors are weighted and summed according to the attention weights to generate aggregated features;

[0042] The aggregated features are input into the softmax classifier, which outputs a calibration urgency score.

[0043] Preferably, the step of integrating the abnormal fluctuation pattern, the key influencing factors, and the calibration urgency to generate a calibration strategy using a time series prediction model includes:

[0044] The abnormal fluctuation patterns, the key influencing factors, and the urgency of calibration are encoded as time series tensors;

[0045] Gated recurrent unit networks are used to process time series tensors and capture long-term dependencies.

[0046] Output the calibration probability distribution at multiple future time points;

[0047] Based on the calibration probability distribution, a threshold is set to generate a calibration schedule and operating parameters.

[0048] Preferably, the step of acquiring real-time operating parameters of the resistance simulator through a sensor array to form a real-time monitoring data stream includes:

[0049] The collected data is subjected to sliding window averaging filtering, and the filtered data is sorted by timestamp to form a real-time monitoring data stream.

[0050] The real-time monitoring data stream includes resistance drift, power consumption, and ambient temperature.

[0051] Preferably, the present invention also includes a self-calibration system based on a resistance simulator, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the self-calibration method based on a resistance simulator as described above.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] By acquiring multi-dimensional operating parameters of the resistance simulator in real time through a sensor array, a real-time monitoring data stream containing time and frequency domain characteristics is generated, overcoming the limitations of traditional single-point measurement. This mechanism can continuously capture subtle changes in equipment operation, providing a high-density, high-completeness data foundation for subsequent analysis. This ensures that the calibration process is based on the actual operating state of the equipment and avoids evaluation bias caused by missing data.

[0054] An adaptive signal decomposition technique is employed to process real-time data streams, effectively separating multi-scale feature components from residual components. This method can adaptively match signal characteristics under different operating conditions, exhibiting stronger characterization capabilities for nonlinear and non-stationary signals. It can accurately extract feature components at different time scales, such as periodic fluctuations and trend drifts, overcoming the limitations of traditional Fourier transform and other methods in complex signal analysis. By distinguishing between effective features and noise residuals, it provides clear feature input for anomaly pattern recognition, improving the reliability of subsequent analysis.

[0055] In the abnormal fluctuation pattern identification stage, a multi-scale feature fusion diagnostic model was constructed by combining multi-scale feature components and residual components for comprehensive analysis. Compared with single feature analysis, this method can capture the coupling relationship and synergistic change law between parameters, and effectively identify complex abnormal patterns caused by multiple factors such as component aging, temperature drift, and contact impedance changes.

[0056] The introduction of a causal reasoning mechanism enables in-depth analysis from abnormal phenomena to key influencing factors. By constructing a causal relationship model between parameters, the system can trace the root cause of abnormal fluctuations, such as identifying that changes in the contact resistance of a certain channel switch are the main cause of output resistance deviations, rather than the superficial fluctuations in resistance values.

[0057] In the reference performance trajectory generation stage, a dynamic performance benchmark for the equipment is constructed by integrating historical calibration databases with key influencing factors identified in real time and using a regression model. This model not only includes the typical parameter range during normal equipment operation but also incorporates the influence weights of key factors on performance, enabling dynamic adjustment of the reference trajectory based on current operating conditions.

[0058] The calibration urgency assessment module achieves dynamic quantitative assessment of calibration needs by quantifying the deviation between the actual operating trajectory and the reference trajectory. This mechanism changes the traditional, fixed-cycle calibration approach, intelligently determining the calibration timing based on the rate of equipment performance degradation. It extends the calibration cycle for stable equipment and triggers emergency calibration for rapidly degrading equipment, effectively reducing calibration costs and equipment downtime while ensuring measurement accuracy. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the working principle of the self-calibration method based on a resistance simulator described in this invention.

[0060] Figure 2 This is a flowchart of adaptive signal decomposition and feature extraction.

[0061] Figure 3 Flowchart for identifying abnormal fluctuation patterns;

[0062] Figure 4The graph shows the results of abnormal parameter detection for the resistance simulator.

[0063] Figure 5 This is a comparison chart of the actual running trajectory of the resistance simulator and the reference performance trajectory, along with a deviation analysis. Detailed Implementation

[0064] 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.

[0065] Please see Figure 1 This invention provides a self-calibration method based on a resistance simulator, the method comprising:

[0066] Intelligent self-calibration of a resistance simulator is achieved by integrating sensor technology, signal processing, machine learning, and time series analysis. The overall implementation scheme includes the following steps: Real-time operating parameters of the resistance simulator, such as resistance drift, power consumption, and ambient temperature, are collected via a sensor array. The collected data is then subjected to sliding window averaging filtering, and the filtered data is sorted by timestamp to form a real-time monitoring data stream. Adaptive signal decomposition is performed on the real-time monitoring data stream. An empirical mode decomposition algorithm is used to process the data stream to obtain a set of intrinsic mode functions (IMFs). Hilbert transform is applied to each IMF to calculate the instantaneous frequency and amplitude. Significant IMFs are selected as multi-scale feature components based on a preset threshold. The remaining unselected IMFs are merged into residual components. The multi-scale feature components are then further processed. Joint analysis of the feature components and residual components is performed, combining the multi-scale feature components and residual components into a high-dimensional feature vector. An isolated forest algorithm is used for unsupervised learning of the high-dimensional feature vector to classify normal and abnormal data points. Density peak clustering is performed on the abnormal data points to identify abnormal fluctuation patterns. Causal inference is then performed on the abnormal fluctuation patterns, mapping them to nodes to construct an initial graph structure. A Bayesian network learning algorithm is used to analyze the conditional dependencies between nodes and add directed edges. Edge weights are calculated using the maximum information coefficient to identify strong causal paths, and frequently occurring nodes are extracted from these strong causal paths as key influencing factors. The history of the integrated resistor simulator is also discussed. The calibration database and key influencing factors are determined by loading multiple sets of calibration event records, including environmental conditions, load parameters, and calibration results, from the historical calibration database. Key influencing factors are then aligned with the calibration event records. A Gaussian process regression model is used to perform nonlinear fitting on the aligned features to generate a reference performance trajectory. The hyperparameters of the Gaussian process regression model are adjusted to minimize prediction error. The deviation between the actual operating trajectory and the reference performance trajectory is calculated. A dynamic time warping algorithm is used to align the actual operating trajectory and the reference performance trajectory, calculating the distance between the trajectories. This distance is then input into a convolutional neural network to extract spatial features. An attention mechanism is used to weight the spatial features, and the resulting calibration trajectory is output. The quasi-urgency score is generated by adding a self-attention module after the fully connected layer of the convolutional neural network, calculating the attention weight distribution of the feature vectors, and weighting and summing the feature vectors according to the attention weights to generate aggregated features. The aggregated features are then input into a softmax classifier to output the calibration urgency score. Abnormal fluctuation patterns, key influencing factors, and calibration urgency are fused together and encoded into time series tensors. A gated recurrent unit network is used to process the time series tensors to capture long-term dependencies and output the calibration probability distribution at multiple future time points. Based on the calibration probability distribution, a threshold is set to generate a calibration schedule and operating parameters.

[0067] Example 1: See Figure 2In practical implementation, real-time operating parameters of the resistance simulator are collected through a sensor array. This array contains multiple high-precision digital sensors, distributed across the core functional modules of the simulator. Voltage and current sensors are installed on the power supply module, resistance sampling sensors on the load interface, and temperature sensors on the heat dissipation unit. Real-time monitoring data streams include resistance drift, power consumption, and ambient temperature. Resistance drift is calculated by measuring the voltage drop across the simulated resistor and the current flowing through it. Power consumption is obtained by multiplying the real-time voltage and current. Ambient temperature is read directly from a digital temperature sensor. Data acquisition is performed at a fixed frequency of 100 times per second. This frequency is set to ensure the capture of rapid transient changes that may occur during operation, such as instantaneous fluctuations caused by sudden load increases or decreases. The collected raw data is temporarily stored in a first-in, first-out (FIFO) data buffer.

[0068] In practice, the collected data undergoes a sliding window averaging filter, with the sliding window size set to 10 data points, corresponding to a 100-millisecond time window. The filtering algorithm employs a simple moving average method, summing the 10 data points within the window and dividing by 10 to obtain the average value, which serves as the filtered output. When a new data point arrives, the window slides forward one position, discarding the oldest data point, incorporating the newest, and recalculating the average. This method effectively smooths random white noise and occasional spike interference. The filtered data is then sorted according to timestamps provided by the high-precision clock of the acquisition system, forming a continuous and time-aligned real-time monitoring data stream. This data stream is transmitted to the central processing unit for subsequent analysis and is also written to non-volatile memory for historical data backtracking.

[0069] In practical implementation, adaptive signal decomposition is performed on the real-time monitoring data stream, employing the Empirical Mode Decomposition (EMD) algorithm. EMD is a decomposition method based on local data features, processing the aforementioned real-time monitoring data stream. Through an iterative sieving process, EMD decomposes the nonlinear, non-stationary data stream into a finite number of fundamental modal components, i.e., intrinsic mode functions (EMFs). The sieving process begins by identifying all local extrema in the data stream, including local maxima and local minima. Cubic spline interpolation is used to connect all local maxima to form an upper envelope, and all local minima to form a lower envelope. The mean curves of the upper and lower envelopes are calculated, and the original data stream is subtracted from these mean curves to obtain a candidate component. This candidate component is then checked to ensure it meets the conditions for an EMF: the number of extrema and zero-crossings within the entire data segment is equal or at most differs by one, and the mean of the upper envelope defined by local maxima and the lower envelope defined by local minima is zero at any point. If the conditions are not met, the candidate component is treated as a new data stream, and the above screening process is repeated until the conditions are met. The candidate component obtained at this point is an intrinsic mode function (EMF). This EMF is extracted from the original data stream, and the entire screening process is repeated for the remaining data until the remaining components become monotonic functions or constants. Finally, a set of EMFs is obtained, arranged from high frequency to low frequency.

[0070] In practice, significant intrinsic mode functions (IMFs) are selected based on a preset threshold, which is set according to the calculated instantaneous amplitude. The threshold calculation is based on the statistical characteristics of the instantaneous amplitudes of all IMFs at all time points, specifically set to twice the global average instantaneous amplitude. Each IMF is iterated over, and its average instantaneous amplitude over the entire time range is calculated and compared with the preset threshold. If the average instantaneous amplitude of an IMF is greater than or equal to the preset threshold, then that IMF is determined to be a significant IMF, serving as a multi-scale feature component. These multi-scale feature components typically correspond to strong, physically meaningful oscillation patterns in the data stream. IMFs with average instantaneous amplitudes below the preset threshold are considered insignificant components. All insignificant IMFs are algebraically summed and merged into a single residual component. The residual component represents the long-term trend term and very low-frequency fluctuation components in the data stream. The multi-scale feature components and residual components are used together as the final output of adaptive signal decomposition, providing input for the subsequent joint analysis stage.

[0071] In some embodiments, the sensor array configuration may include more types of sensors, such as humidity sensors or vibration sensors, to collect more comprehensive operating parameters. The data acquisition frequency can be adjusted according to the specific application scenario of the resistance simulator; for scenarios with higher dynamic response requirements, a higher acquisition frequency can be used. The window size of the sliding window averaging filter can also be optimized based on noise characteristics and system response speed; a larger window provides a stronger smoothing effect but reduces time resolution.

[0072] It is understandable that the specific implementation of the Empirical Mode Decomposition (EMD) algorithm can employ different stopping criteria to control the number of screening iterations, such as setting an upper limit on the number of screening iterations or a standard deviation threshold between two adjacent screening results. The Hilbert transform can also be calculated using different numerical methods, such as the Discrete Hilbert Transform. The setting of the preset threshold can be diversified, for example, using robust statistics based on the absolute deviation of the median, or dynamically adjusting the threshold level based on prior knowledge. Optionally, a data quality check step can be added during the data acquisition phase, such as checking whether sensor readings are within a reasonable range and marking or removing abnormal readings. During the signal decomposition phase, the EMD algorithm can be improved, for example, by integrating the overall EMD to suppress mode aliasing.

[0073] It is understandable that the formation of real-time monitoring data streams forms the data foundation of the entire self-calibration method, and its quality and reliability directly affect the effectiveness of all subsequent processing steps. Adaptive signal decomposition transforms complex operating parameter data into a set of components with clear physical meaning, facilitating subsequent feature extraction and pattern recognition. The extraction of multi-scale feature components and residual components provides a multi-resolution perspective for a deeper understanding of the operating state of the resistance simulator.

[0074] In practical implementation, the cubic spline interpolation method used in the Empirical Mode Decomposition (EMD) algorithm to construct the upper and lower envelopes is a method of constructing smooth curves through local extrema. This method begins with identifying all local maxima and local minima. Local maxima are used as interpolation nodes, each containing its timestamp and magnitude. For adjacent nodes, a cubic polynomial of the form S(x) = ax³ + bx² + cx + d is used to connect them. To ensure the smoothness and continuity of the entire envelope, adjacent polynomial segments at each node are required to have the same first and second derivative values, forming a system of linear equations. For boundary nodes, natural boundary conditions are typically used, i.e., the second derivative values ​​of the second and penultimate nodes are set to zero. By solving this tridiagonal linear system of equations composed of all node conditions, the coefficients of each polynomial segment can be obtained, thus fitting the complete upper envelope. The construction process of the lower envelope is exactly the same as that of the upper envelope, the only difference being the use of local minima as interpolation nodes. The entire solution process is completed using an efficient chasing method, ultimately yielding a smooth envelope curve that passes through all specified extreme points.

[0075] In practical implementation, one of the iteration termination conditions of the Empirical Mode Decomposition (EMD) algorithm, "until the remaining components become monotonic functions," requires a clear criterion. The determination of a monotonic function is achieved by calculating the difference values ​​of the remaining component sequence. For the residual components remaining after multiple screenings, the numerical difference between adjacent data points is calculated, i.e., the value of the later data point minus the value of the earlier data point, yielding the first-order difference of the entire sequence. If all first-order difference values ​​are greater than or equal to zero, the residual component is determined to be a monotonically increasing function; if all first-order difference values ​​are less than or equal to zero, the residual component is determined to be a monotonically decreasing function. In practice, considering numerical calculation errors and minor fluctuations, a small positive number is set as a tolerance threshold. When the signs of all difference values ​​remain consistent and their absolute values ​​are greater than this tolerance threshold, the monotonicity condition is considered satisfied. Once a residual component is determined to be a monotonic function, the screening process of the EMD algorithm immediately terminates, and the residual component is no longer further decomposed but directly output as the final residual component.

[0076] Example 2: See Figure 3In practice, multi-scale feature components and residual components are combined into a high-dimensional feature vector. The construction of this high-dimensional feature vector is based on the principle of time alignment. Each multi-scale feature component is a time series obtained after Empirical Mode Decomposition (EMD) and Hilbert Transform (HLD), with each time point containing instantaneous frequency and amplitude information. The residual component is also a time series, representing the trend of the data. The combination operation is performed at each unified timestamp. For a given time point, the instantaneous frequency and amplitude values ​​of all multi-scale feature components at that time point are extracted, and the values ​​of the residual components at that time point are also extracted. These values ​​are then concatenated in a fixed order to form a vector, the dimension of which is equal to the number of multi-scale feature components multiplied by two plus one. For example, if there are five multi-scale feature components, the dimension of each feature vector is five times two plus one, which equals eleven dimensions. The feature vectors at all time points are arranged in chronological order to form a complete set of high-dimensional feature vectors for subsequent anomaly detection analysis.

[0077] In its implementation, the Isolation Forest algorithm is used for unsupervised learning of high-dimensional feature vectors. The Isolation Forest algorithm is an anomaly detection method based on ensemble learning. It works by constructing multiple isolation trees, each constructed randomly. When constructing a single isolation tree, a subsample is randomly selected from the set of high-dimensional feature vectors, typically with 256 vectors. Starting from the root node, a dimension of the feature vector is randomly selected as the splitting feature, and a splitting value is randomly chosen between the maximum and minimum values ​​of that feature. The sample is then assigned to either the left or right subtree based on whether the value of each feature vector in that dimension is less than the splitting value. This process is recursively repeated at each child node until the following termination conditions are met: the node contains only one feature vector, all feature vectors are identical, or the maximum depth of the tree is reached. The maximum depth of the tree is limited by a logarithmic function to base 2, where the input parameter is the subsample size. The Isolation Forest algorithm assesses the anomaly degree of each feature vector by calculating the path length from the root node to a leaf node. A shorter path length indicates that the feature vector is more easily isolated and therefore more likely to be an anomaly. Finally, normal and abnormal data points are distinguished by the average path length of the entire isolated forest. Data points with an average path length that is one standard deviation lower than the overall average path length of the forest are marked as abnormal data points.

[0078] In practical implementation, density peak clustering is used to cluster the outlier data points identified by the Isolation Forest algorithm. Density peak clustering is a clustering algorithm based on density and distance. The first step of density peak clustering is to calculate the local density of each outlier data point. The local density is estimated using a Gaussian kernel function. For each outlier data point, the Euclidean distance between it and all other outlier data points is calculated. A cutoff distance is used as the bandwidth parameter of the kernel function, and the local density is defined as the sum of all distance values ​​after being transformed by the Gaussian kernel function. The second step is to calculate the minimum distance between each outlier data point and points with higher local densities. For the point with the highest local density, this distance is defined as its maximum distance to all other points. By calculating the product of the local density value and the distance value of each point, points with significantly higher product values ​​than their surrounding points are identified as cluster centers. The number of cluster centers is automatically determined by the algorithm, usually selecting those points whose product values ​​are significantly higher than their neighbors. After determining the cluster centers, each non-center point is assigned to the category of its nearest point with a higher density, thus completing the clustering of all outlier data points. Each cluster category is identified as an outlier fluctuation pattern.

[0079] In practice, identified abnormal fluctuation patterns are mapped to nodes to construct an initial graph structure, with each abnormal fluctuation pattern corresponding to one node. The attributes of a node include the characteristic statistics of all data points within that abnormal fluctuation pattern, such as mean, variance, skewness, and kurtosis. The initial graph structure is an undirected graph, where edges between nodes represent potential correlations between abnormal fluctuation patterns. The initial graph contains all nodes, but the edge connections are not yet determined. Typically, fully connected or distance-threshold-based methods are used to establish initial edge connections. In a fully connected scheme, each node is connected to all other nodes; in a distance-threshold-based scheme, an edge is established only when the Euclidean distance between the feature vectors of the abnormal fluctuation patterns represented by two nodes is less than a predetermined threshold.

[0080] In practical implementation, Bayesian network learning algorithms are used to analyze the conditional dependencies between nodes. A Bayesian network is a directed acyclic graph (DAG) model. Bayesian network learning includes two parts: structure learning and parameter learning. The goal of structure learning is to determine the directed edge connections between nodes. A score search method is used for structure learning, with the scoring function employing the Bayesian information criterion, which comprehensively considers both the model's goodness of fit to the data and its complexity. The search algorithm uses a hill-climbing algorithm, starting from an initial graph structure and generating candidate graph structures by adding, deleting, or reversing directed edges. The candidate graph structure with the highest score is selected as the current graph structure. This process is iterated until the score can no longer be improved. The final DAG structure reflects the conditional dependencies between variables in the abnormal fluctuation pattern, and the direction of the directed edges represents the potential causal relationship between the variables.

[0081] In practical implementation, the weights of directed edges in a Bayesian network are calculated using the maximum information coefficient (MIC), an information-theoretic correlation metric. The MIC measures the strength of the association between two variables, including linear and nonlinear associations. For each directed edge connecting two node variables in a Bayesian network, their MIC values ​​are calculated as edge weights. The calculation of the MIC is based on the concept of mutual information, which measures the degree of interdependence between two variables. The MIC is calculated by dividing the variable's value range into different grids, calculating the mutual information value for each grid division, taking the maximum value, and then normalizing it. The final MIC value is between zero and one; a larger value indicates a stronger association between the two variables. After calculating the MIC weights of all directed edges, strong causal paths are identified. A strong causal path is defined as a sequence of consecutive directed edges where the MIC weight of all edges is greater than 0.7.

[0082] In practice, frequently occurring nodes are extracted from the identified strong causal paths. Frequently occurring nodes are defined as those appearing more than 50% of the total number of paths within a strong causal path. The frequency of each node across all strong causal paths is statistically analyzed, including instances where it is the start, middle, or end point of a path. Nodes with a frequency exceeding a set threshold are marked as key influencing factors. These key influencing factors represent variables at the core of the causal network of anomalous fluctuation patterns, significantly impacting the abnormal behavior of the resistance simulator. Information on key influencing factors is then passed to the subsequent reference performance trajectory generation stage to guide the formulation of calibration strategies. In some embodiments, the combination of high-dimensional feature vectors can employ different feature encoding schemes, such as normalization or the inclusion of multinomial features. The parameter settings of the isolation forest algorithm can be adjusted according to data characteristics, such as the number of subtrees and subsample size. The cutoff distance in density peak clustering can be automatically determined by analyzing the distance distribution, without requiring manual specification.

[0083] Optionally, Bayesian network learning can employ different scoring functions, such as Bayesian Dirichlet equivalence scoring. The calculation of the maximum information coefficient can utilize different grid partitioning strategies, such as adaptive grid partitioning. The threshold for identifying strong causal paths can be adjusted based on network density; a lower threshold can be used in sparse networks. It is understandable that the identification of anomalous fluctuation patterns provides a foundation for understanding the abnormal behavior of resistance simulators, and the discovery of key influencing factors helps to pinpoint the root cause of the anomalies. Causal reasoning-based analysis methods can surpass simple correlation analysis, providing more interpretable results. Density peak clustering can automatically determine the number of clusters, avoiding the limitation of traditional clustering algorithms requiring pre-specified numbers of categories.

[0084] In some embodiments, Bayesian network structure learning can incorporate prior knowledge, such as prohibiting or enforcing certain edge connections. The calculation of the maximum information coefficient can employ different estimation methods for continuous and discrete variables. The extraction of key influencing factors can consider the centrality of nodes in the network, rather than just their frequency of occurrence. Optionally, handling outliers can add visualization analysis steps to aid in understanding clustering results. The causal inference process can introduce time lag effects to analyze time-delayed causal relationships between variables. Validation of key influencing factors can be performed through intervention analysis or sensitivity analysis.

[0085] It is understandable that joint analysis of multi-scale feature components and residual components can capture system behavior characteristics from different scales. Unsupervised learning methods are suitable for scenarios without labeled data and can automatically discover anomalous patterns in the data. Causal inference methods provide a theoretical basis for the formulation of calibration strategies, helping to achieve more targeted calibration operations. In specific implementation, the Gaussian kernel function used in the density peak clustering algorithm to estimate the local density of each outlier data point is a distance-based smoothing weight function. The application of the Gaussian kernel function begins with calculating the Euclidean distance between any two outlier data points, which is obtained by taking the square root of the sum of the squares of the differences in each dimension of the corresponding feature vector. A local neighborhood is defined for each outlier data point, and this range is defined by the cutoff distance parameter. For a given data point, the distance values ​​between it and all other points in the dataset are calculated, and these distance values ​​are substituted into the Gaussian kernel function for calculation. The standard form of the Gaussian kernel function is... ,in Represents the distance between two points. This is the bandwidth parameter of the kernel function. (Bandwidth parameter) The value of is usually set to the median of the distances between all points or a fixed percentile. The local density of a data point is ultimately defined as the sum of the Gaussian kernel function values ​​of that point and all other points. This sum reflects the density of the data distribution around that point; a larger value indicates that the point is located in a denser region.

[0086] In practice, the method of dividing the variable value range into different grids during the calculation of the maximum information coefficient employs a dynamic programming optimization strategy. For two variables whose correlation strength needs to be calculated, their observed values ​​are arranged in ascending order, forming two one-dimensional data sequences. The goal of grid partitioning is to find a partitioning method that maximizes the mutual information value calculated under that grid. Mutual information measures the difference between the product of the joint distribution and the marginal distribution of two variables under a specific grid partition. The partitioning process starts with the smallest grid, such as a 2x2 grid, dividing the variable value range into two intervals. The mutual information value under the current grid partition is calculated, and all possible grid partitioning methods are tried, including changing the position of the interval boundaries. The number of rows and columns of the grid is gradually increased, for example, trying grid partitioning with different resolutions such as 3x3 and 4x4, and searching for the optimal partitioning scheme for each resolution. Since exhaustively searching for all partitioning methods is computationally intensive, a dynamic programming algorithm is actually used to efficiently search for the optimal grid partitioning. Dynamic programming avoids redundant calculations by storing the optimal solutions to subproblems. Finally, the partition scheme with the largest mutual information value is selected from all possible grid partitions. This maximum mutual information value is then normalized so that its value falls between zero and one, thus obtaining the final result of the maximum information coefficient.

[0087] See Figure 4 This chart is a key visualization result of the abnormal fluctuation pattern identification stage in the resistance simulator self-calibration method. With time on the horizontal axis and resistance value on the vertical axis, it visually presents the abnormal distribution characteristics of resistance value drift during real-time operation of the resistance simulator through the visual distinction of blue dots and red crosses. The value of this chart lies in providing intuitive data support for subsequent density peak clustering to identify abnormal fluctuation patterns and Bayesian network causal inference to mine key influencing factors. It is a core step in achieving early detection and diagnosis of anomalies in the resistance simulator, helping technicians quickly locate the time periods and characteristics of equipment anomalies, and laying a reliable analytical foundation for the generation of accurate calibration strategies.

[0088] Example 3: In a specific implementation, the historical calibration database and key influencing factors of the integrated resistance simulator are used to generate a reference performance trajectory. This process begins by loading multiple sets of calibration event records from the historical calibration database, which is typically stored and managed using a relational database management system, such as SQLite or MySQL. The calibration event records contain multiple fields, including timestamps, environmental conditions such as temperature and humidity, load parameters such as current and voltage, and calibration results such as adjusted resistance values. Database query operations are performed using Structured Query Language (SCL), retrieving the most recent records in chronological order using SELECT statements, for example, loading the most recent 1000 calibration event records to ensure data timeliness and representativeness. The loaded data is converted into a format suitable for machine learning models, typically a two-dimensional array or data frame structure, where each row represents a calibration event and each column represents a feature variable.

[0089] In practice, key influencing factors are aligned with calibration event records. Key influencing factors are variables extracted from causal inferences of abnormal fluctuation patterns, such as changes in ambient temperature or load fluctuations. Feature alignment is based on timestamp matching principles, aligning the time series of key influencing factors with the timestamps of calibration event records. The alignment process employs a sliding time window method, setting an alignment window size, for example, 1 minute. For each calibration event record's timestamp, key influencing factor data points within the window before and after that timestamp are identified. If multiple key influencing factor data points exist within the window, their average or median is calculated as the representative value; if no data points exist within the window, interpolation methods such as linear interpolation are used to fill in missing values. The aligned dataset forms a unified feature matrix, where each row corresponds to a calibration event. Features include environmental conditions, load parameters, and aligned key influencing factor values ​​from the calibration event records.

[0090] In practice, a Gaussian process regression model is used to nonlinearly fit the aligned features to generate a reference performance trajectory. Gaussian process regression is a nonparametric regression method based on Bayesian theory. The Gaussian process regression model uses the calibration results as the target variable and the aligned features as the input variables. The model is defined as a Gaussian stochastic process, whose characteristics are fully specified by the mean function and covariance function. The mean function is usually set as a zero function or a constant function, and the covariance function is chosen from the radial basis function kernel. Its mathematical expression is:

[0091]

[0092] in: and This represents two input feature vectors. It is the amplitude of the signal variance parameter control function. It is a diagonal matrix whose elements are the reciprocals of the length scale parameter, which controls the influence range of each feature dimension. It is the noise variance parameter, representing the observed noise. This is the Kronecker delta function, which is 1 when i=j and 0 otherwise. The fitting process involves calculating the covariance matrix of the training data and optimizing the model parameters by maximizing the marginal likelihood function, which is the logarithmic form of a multivariate Gaussian distribution. When generating the reference performance trajectory, for new input features, Gaussian process regression provides the predicted mean and variance. The predicted mean serves as a point estimate of the reference performance trajectory, which is a time series representation of the performance curve of the resistor simulator under ideal conditions.

[0093] In practice, the hyperparameters of the Gaussian process regression model are adjusted to minimize the prediction error. These hyperparameters include the length scaling parameter, signal variance parameter, and noise variance parameter in the covariance function. The adjustment process employs gradient-based optimization algorithms, such as the conjugate gradient method or the quasi-Newton method, with the optimization objective being to minimize the negative log-marginal likelihood function. The negative log-marginal likelihood function is a function of the model parameters, and its gradient is obtained analytically. The optimization algorithm iteratively updates the hyperparameter values ​​until it converges to a local minimum. The prediction error is evaluated using cross-validation methods, such as dividing the data into training and validation sets, calculating the mean squared error or mean absolute error on the validation set as the error metric, and repeating the optimization process until the error falls below a preset threshold or the maximum number of iterations is reached.

[0094] In some embodiments, loading a historical calibration database may include data preprocessing steps, such as removing outliers or standardizing feature values. The time window size for feature alignment can be dynamically adjusted according to the data acquisition frequency, using a smaller window for high-frequency data to improve alignment accuracy. The covariance function of the Gaussian process regression model can be selected from other kernel functions, such as the Matérn kernel or the periodic kernel, to capture different data characteristics.

[0095] It is understandable that the nonparametric nature of Gaussian process regression models allows them to flexibly adapt to complex data patterns, while the Bayesian framework provides quantification of uncertainty. Feature alignment ensures that key influencing factors are synchronized with calibration events in time, avoiding biases caused by phase differences. Hyperparameter tuning is performed automatically through optimization algorithms, reducing the workload of manual parameter tuning. Optionally, dynamic time warping algorithms can be used in the feature alignment stage to handle time series alignment and non-uniformly sampled data. Gaussian process regression predictions can be combined with active learning strategies to select points with high information content for updates to improve model efficiency. In some embodiments, queries to the historical calibration database can incorporate conditional filtering, such as loading only records under specific environmental conditions. The implementation of Gaussian process regression models can employ sparse approximation methods to handle large-scale datasets, such as using induced point sets to reduce computational complexity. Hyperparameter optimization can use global optimization algorithms such as Bayesian optimization to avoid getting trapped in local optima.

[0096] It is understandable that the generation of the reference performance trajectory depends on the quality and quantity of historical data; sufficient data helps improve the accuracy of the trajectory. Gaussian process regression provides probabilistic predictions, offering risk estimates for calibration decisions. Integrating key influencing factors enhances the model's responsiveness to anomalies. In some embodiments, calibration event logs can be expanded to include more contextual information, such as equipment uptime or maintenance history. Feature alignment can introduce a weighting mechanism, assigning different weights to different key influencing factors. The mean function of Gaussian process regression can be set as a linear function to incorporate prior knowledge. It is understood that the entire implementation process is a data-driven modeling workflow; from data preparation to model fitting and optimization, each step affects the reliability of the final trajectory. The processing power of Gaussian process regression makes it suitable for nonlinear dynamic modeling of resistance simulators, while hyperparameter tuning ensures optimal model-data matching. The reference performance trajectory serves as a benchmark for subsequent calibration urgency assessment.

[0097] In its implementation, the feature alignment process employs a sliding time window method, a data matching technique based on fixed time intervals. This method achieves data association by defining a time interval centered on the timestamp of the calibration event record. The time window size is set to one minute, i.e., a range of thirty seconds before and after the timestamp. For each calibration event record requiring alignment, the system retrieves all key influencing factor data points, each with a precise timestamp. The retrieval operation iterates through the time series of key influencing factors, identifying all data points whose timestamps fall within the thirty-second range before and after the current calibration event record's timestamp, forming a temporary data subset. If this subset contains multiple data points, the system calculates the arithmetic mean of these data points as a representative value. The arithmetic mean is calculated by summing the values ​​of all data points and dividing by the number of data points.

[0098] In practice, the conjugate gradient method used to adjust the hyperparameters of a Gaussian process regression model is an iterative optimization algorithm used to find the combination of hyperparameters that minimizes the negative log-marginal likelihood function. The conjugate gradient method iterates from a set of initial hyperparameter values, which are usually set empirically or randomly. In each iteration, the algorithm calculates the gradient vector of the negative log-marginal likelihood function at the current hyperparameter value. The gradient vector represents the rate of change of the function in the steepest ascending direction. Based on the current gradient and the information from the previous search direction, the conjugate gradient method calculates a new conjugate direction. The choice of the conjugate direction ensures that the algorithm's search efficiency in the parameter space is higher than that of the steepest descent method. Along this conjugate direction, the algorithm performs a one-dimensional linear search to determine the optimal step size. The linear search finds the step size that reduces the function value the most by trying different step sizes and evaluating the function value. After updating the hyperparameter values, the algorithm checks whether the convergence condition is met. The convergence condition can be that the magnitude of the gradient vector is less than a certain threshold, the change in the function value is less than a threshold, or the maximum number of iterations has been reached. If convergence is not achieved, the algorithm continues to the next iteration, using the new gradient and the old search direction to generate a new conjugate direction. The entire process is repeated until the convergence condition is met, and the final output hyperparameter values ​​are the optimization results, which minimize the prediction error of the Gaussian process regression model on the training data.

[0099] See Figure 5 This chart is a core visualization result of the reference performance trajectory generation and deviation assessment stages in the self-calibration method of the resistance simulator. With time as the horizontal axis and resistance value as the vertical axis, it visually presents the difference between the actual performance of the device and the ideal benchmark through solid blue lines, dashed red lines, and light blue fill. The reference performance trajectory is generated by a Gaussian process regression model, integrating environmental conditions, load parameters, and key influencing factors from the historical calibration database, serving as a dynamic benchmark for normal device operation. The actual operating trajectory is real-time data collected by the sensor array; the light blue filled area represents the deviation range, reflecting the degree of deviation between the actual performance and the ideal state. The value of this chart lies in providing a quantitative basis for subsequent calibration urgency assessment. It calculates the distance between trajectories using a dynamic time warping algorithm, and combines convolutional neural networks and attention mechanisms to output a calibration urgency score, ultimately supporting the generation of intelligent calibration strategies. This is a key analytical step in achieving on-demand and accurate calibration of the resistance simulator.

[0100] Example 4: In specific implementation, calculating the deviation between the actual running trajectory and the reference performance trajectory begins with aligning the two trajectories using a dynamic time warping algorithm. This algorithm can handle the stretching and distortion of time series along the time axis. The actual running trajectory comes from real-time data sequences collected by a sensor array, while the reference performance trajectory is a predicted sequence generated by a Gaussian process regression model. The dynamic time warping algorithm constructs a cumulative cost matrix, where rows correspond to data points on the actual running trajectory and columns correspond to data points on the reference performance trajectory. The value of each element in the matrix represents the local distance between the corresponding data points, typically calculated using Euclidean distance. Starting from the top left corner of the matrix, the dynamic time warping algorithm progressively calculates the minimum cumulative path cost at each position. The cumulative path cost equals the local distance at the current position plus the minimum cumulative cost among adjacent positions (above, left, and top left). Finally, the minimum cumulative distance between the two trajectories is obtained at the bottom right corner of the matrix. This distance value serves as a quantification of the distance between the trajectories. The dynamic time warping algorithm aligns the trajectories by finding the optimal curved path, thereby eliminating the influence of nonlinear deformation along the time axis.

[0101] In the specific implementation, referring to Table 1, the distance between trajectories is input into a convolutional neural network (CNN) to extract spatial features. A CNN is a deep learning model specifically designed for processing grid-like data. The distance between trajectories is input into the CNN in sequence form and processed through a one-dimensional convolutional layer. This one-dimensional convolutional layer contains multiple convolutional kernels, each sliding along the time dimension. The convolution operation calculates the dot product between the input sequence and the kernel weights, adds a bias term, and then generates a feature map through an activation function. Multiple convolutional kernels extract different local patterns, forming multiple feature maps. A pooling layer follows the convolutional layer. The pooling layer downsamples the feature maps, and max pooling selects the maximum value within a local region as the output, reducing data dimensionality while preserving salient features. Through the alternating stacking of multiple convolutional and pooling layers, the CNN progressively extracts spatial features from low to high levels. The final feature map is flattened into a feature vector for subsequent processing.

[0102] Table 1: Parameter Configuration Table for One-Dimensional Convolutional Layers in Convolutional Neural Networks

[0103]

[0104] In practice, an attention mechanism is used to weight spatial features to output a calibration urgency score. This attention mechanism automatically learns the importance weights of features. A self-attention module is added after the fully connected layer of the convolutional neural network (CNN). This module receives the spatial feature vector extracted by the CNN as input. The self-attention module maps the input feature vector into three different representations: a query vector, a key vector, and a value vector, through a linear transformation. The dot product score of the query vector and all key vectors is calculated, and the score is scaled by dividing by the square root of the key vector dimension to prevent gradient vanishing. The scaled score is then transformed into an attention weight distribution using a softmax function. The attention weights represent the relative importance of each feature in the context. The value vectors are weighted and summed according to the attention weights to generate an aggregated feature representation. This aggregated feature focuses on key information while ignoring irrelevant features. The aggregated feature is then input into a softmax classifier, which maps the features to a calibration urgency score. The score ranges from 0 to 1, with higher values ​​indicating greater calibration urgency.

[0105] In its implementation, the softmax classifier is based on a multinomial probability distribution model, mapping the aggregated feature vector to a probability output. The softmax classifier calculates a score for each category, obtained through a linear transformation of the weight matrix and the input features. The scores are then transformed using an exponential function to ensure all values ​​are positive. The exponentialized value is divided by the sum of the exponential values ​​for all categories to obtain a normalized probability distribution. In the calibration urgency assessment, the softmax classifier outputs probabilities for two categories, representing "low urgency" and "high urgency," respectively. The probability value of the "high urgency" category is used as the final calibration urgency score. This score reflects the severity of the deviation of the actual operating trajectory from the reference performance trajectory, providing a quantitative basis for calibration decisions.

[0106] In some embodiments, the local distance calculation of the dynamic time warping algorithm can employ other distance metrics such as Manhattan distance or cosine distance. The structure of the convolutional neural network can be adjusted according to the characteristics of the trajectory data, such as increasing or decreasing the number of convolutional layers. The self-attention module can be extended to a multi-head self-attention mechanism to capture feature information from different subspaces. Optionally, trajectory alignment can employ other methods such as the longest common subsequence algorithm. Regularization techniques such as dropout can be incorporated into the training of the convolutional neural network to prevent overfitting. The calculation of attention weights can introduce a local attention mechanism to limit the attention range and improve computational efficiency.

[0107] It is understandable that dynamic time warping algorithms can effectively handle phase differences in time series and improve the accuracy of trajectory comparison. Convolutional neural networks, through local perception and weight sharing characteristics, can effectively extract local fluctuation patterns in trajectories. Attention mechanisms enable the model to focus on key features, improving the reliability of calibration urgency assessment. Optionally, the calculation of distances between trajectories can be combined with multi-scale analysis to compare trajectory differences at different time granularities. The input of convolutional neural networks can include differential features of trajectories, emphasizing trends. Self-attention modules can be combined with gating mechanisms to control the direction of information flow.

[0108] In practice, the generation of calibration urgency scores requires a model training phase, using historical data and their corresponding calibration urgency labels for supervised learning. Training data includes the distance sequence between the actual running trajectory and the reference performance trajectory, as well as expert-annotated calibration urgency levels. Model training employs backpropagation to optimize network parameters, using cross-entropy loss as the loss function to measure the difference between the model's predicted probability distribution and the true label distribution. By iteratively optimizing to minimize the loss function, the trained model can accurately assess the calibration urgency of new trajectory data. An early stopping strategy is used during training to prevent overfitting; training terminates when performance on the validation set no longer improves, preserving the optimal model parameters for subsequent calibration urgency assessments.

[0109] In its implementation, the cumulative cost matrix constructed in the dynamic time warping algorithm is a two-dimensional array structure. The number of rows equals the number of data points on the actual running trajectory, and the number of columns equals the number of data points on the reference performance trajectory. The matrix is ​​initialized starting from the top-left element, whose value is the Euclidean distance between the first data point on the actual running trajectory and the first data point on the reference performance trajectory. The matrix is ​​traversed in row-major order. For each element (i,j), its corresponding local distance d(i,j) is calculated, which is the Euclidean distance between the i-th point on the actual running trajectory and the j-th point on the reference performance trajectory. Three possible path directions are considered: moving horizontally from the left element (i-1,j), moving vertically from the top element (i,j-1), or moving diagonally from the top-left element (i-1,j-1). The minimum cumulative cost among these three adjacent elements is taken and added to the current local distance d(i,j) to obtain the cumulative cost of the current element. This recursive process continues until the bottom-right element of the matrix is ​​reached. This element stores the minimum cumulative distance between the two trajectories after alignment via the optimal curved path. To find the specific optimal curved path, we need to backtrack from the bottom right element, selecting the neighboring element with the minimum cumulative cost at each position as the previous point on the path, until we backtrack to the top left corner of the matrix. The resulting path is the optimal alignment path found by dynamic time warping.

[0110] In practice, the process of mapping input feature vectors to query vectors, key vectors, and value vectors in the self-attention module is implemented through three independent linear transformation layers. Each linear transformation layer contains a learnable weight matrix and a bias vector, where the dimension of the weight matrix is ​​the input feature dimension multiplied by the projection dimension. For the input feature vector sequence, a linear transformation is performed using the weight matrix and bias vector of the query transformation layer to obtain the query vector sequence. The same input feature vector sequence is then transformed by the linear transformations of the key transformation layer and the value transformation layer to obtain the key vector sequence and the value vector sequence, respectively. These three transformations use different weight parameters, allowing the query vector, key vector, and value vector to represent the input features from different perspectives. The attention weights are calculated by multiplying the dot product of the query vector and all key vectors to obtain the original attention score matrix. The score matrix is ​​then scaled by dividing by the square root of the key vector dimension to prevent gradient explosion or vanishing. The scaled score matrix is ​​then row-normalized using the softmax function to obtain the attention weight matrix, where each element represents the association strength between the corresponding query and key. The attention weights are used to perform a weighted summation of the value vectors, that is, the attention weight matrix is ​​multiplied by the value vector sequence through matrix multiplication to obtain a weighted aggregate feature representation.

[0111] Example 5: In specific implementation, the process of integrating anomalous fluctuation patterns, key influencing factors, and calibration urgency to generate a calibration strategy begins by encoding these three factors into a time-series tensor. The anomalous fluctuation patterns are sequence of patterns identified from density peak clustering, each containing a type identifier and intensity value. Key influencing factors are sequences of variables extracted from causal inference, such as the rate of change of ambient temperature or the amplitude of load fluctuations. Calibration urgency is a sequence of scores output from a convolutional neural network and attention mechanism, ranging from 0 to 1. The encoding operation aligns these three data sources with a unified timestamp, with the timestamp interval set to a fixed value, such as one data point per minute. Missing data is supplemented using forward imputation or linear interpolation. Data from each time point is combined into a feature vector, which includes the type encoding of the anomalous fluctuation pattern, the intensity value, the value of the key influencing factor, and the calibration urgency score. The feature vectors from all time points are stacked chronologically to form a three-dimensional tensor. The dimension of the tensor is the time step multiplied by the number of features multiplied by 1, where the time step depends on the time range of the historical data; for example, data from the past 24 hours corresponds to 1440 time steps. This time series tensor serves as the input to a gated recurrent unit network to capture long-term temporal dependencies.

[0112] In practice, gated recurrent unit (GRU) networks are used to process time-series tensors to capture long-term dependencies. GRU is a variant of recurrent neural networks with update and reset gate mechanisms. The input layer of the GRU receives a three-dimensional tensor, and the network structure includes two hidden layers, each containing 64 GRU neurons. The processing begins at the first time step. For each time step, the input feature vector is computed through update and reset gates. The update gate controls the contribution of the previous hidden state to the current time step, while the reset gate controls the degree of forgetting of the previous hidden state. Specifically, the update gate signal is generated by a linear transformation of the input features and the previous hidden state using a sigmoid function; the reset gate signal is calculated similarly. Candidate hidden states are obtained through element-wise multiplication of the reset gate and the previous hidden state, along with a transformation of the input features, and then activated by a tanh function. The final hidden state is a weighted combination of the update gate-weighted previous hidden state and candidate hidden states. The gated recurrent unit network processes all time steps sequentially. The hidden state of the last time step contains a summary of the entire sequence and is mapped to the output dimension through the fully connected layer at the end of the network.

[0113] In practical implementation, the system outputs a calibration probability distribution for multiple future time points, with the output dimension set to the number of future time points, e.g., the next 24 hours, with each time point corresponding to a calibration probability value. The fully connected layer transforms the final hidden state of the gated recurrent unit network into a probability vector. This probability vector is normalized using a softmax function to ensure that the sum of all probability values ​​is 1. The softmax function calculates the calibration probability for each future time point by exponentializing each output value and dividing it by the sum of the exponents of all output values. This probability value represents the likelihood that calibration is required at that time point. The output of the probability distribution allows the system to predict the optimal timing for calibration; for example, the probability peak corresponds to the most probable calibration time point. This probability distribution is learned based on the temporal patterns of abnormal fluctuations, key influencing factors, and calibration urgency in historical data, and the prediction error is minimized by training the gated recurrent unit network.

[0114] In practice, a threshold is set based on the calibration probability distribution to generate a calibration schedule and operating parameters. The threshold is set to 0.8, meaning that time points with a probability value greater than 0.8 are marked as requiring calibration. The calibration schedule is a list of all time points exceeding the threshold, arranged chronologically, with each time point accompanied by a calibration priority determined by its probability value. Operating parameters are generated by referencing a historical calibration database. For each calibration time point, the system retrieves calibration records under similar historical conditions; for example, when the rate of change in ambient temperature exceeds a specific value, the corresponding resistance adjustment value in the historical record is extracted as a reference. Operating parameters include the type of calibration action, such as resistance adjustment, power compensation, or system restart, and specific parameter values, such as the target resistance value or adjustment range. The calibration schedule and operating parameters are combined into a complete calibration strategy document for use by the execution unit.

[0115] In some embodiments, the encoding of the time series tensor can incorporate a sliding window technique, using a fixed-length window to generate multiple training samples. The number of hidden layers in the gated recurrent unit network (GRU) can be adjusted according to data complexity, increasing the number of layers for longer-term dependencies. The output of the calibration probability distribution can be extended to multi-class probabilities to distinguish different types of calibration operations. Optionally, the features of the time series tensor can include difference features or moving average features to emphasize trends. The GRU can be replaced with a Long Short-Term Memory (LSTM) network to handle more complex time-series patterns. The probability distribution threshold can be dynamically adjusted, adapting to changes based on system load or historical calibration results. It is understood that the gating mechanism of the GRU effectively alleviates the gradient vanishing problem, making it suitable for processing long-sequence data. The calibration probability distribution provides probabilistic predictions, supporting decision-making under uncertainty. The generation of the calibration schedule is based on a data-driven approach, reducing the need for manual intervention.

[0116] Optionally, post-processing steps, such as non-maximum suppression, can be added after outputting the probability distribution to avoid repeated calibration at adjacent time points. The generation of operating parameters can be combined with reinforcement learning to optimize the long-term benefits of calibration actions. The calibration strategy can be integrated into the control system to achieve closed-loop automatic calibration. In some embodiments, the time alignment of time series tensors can be managed using a more precise time series database. Regularization techniques, such as dropout or weight decay, can be incorporated into the training of the gated recurrent unit network. A visualization interface for the calibration probability distribution can assist in manual review and adjustment. Optionally, contextual information, such as duration or frequency, can be incorporated into the encoding of abnormal fluctuation patterns. The sequences of key influencing factors can be preprocessed into standardized values ​​to improve model convergence. The calibration urgency score can be smoothed to avoid abrupt changes affecting the probability distribution.

[0117] In practice, the gated recurrent unit network (GRU) uses historical data as input and future calibration events as labels during training. The loss function is cross-entropy loss, and the optimizer uses the Adam algorithm. Training data is divided into training, validation, and test sets. The validation set is used for early stopping to prevent overfitting, and the test set evaluates the model's generalization performance. After model deployment, the real-time data stream continuously updates the time-series tensor, and the GRU generates the latest calibration strategy online, ensuring the strategy's timeliness and adaptability. The entire process is automated, and the model is periodically retrained to maintain prediction accuracy.

[0118] In practice, the forward imputation method used to handle missing data during time series data alignment is a time-order-based data imputation technique. The implementation of forward imputation begins with identifying missing data points in the time series. The system checks whether there are numerical gaps in the feature vector corresponding to each unified timestamp. These gaps may be caused by sensor transmission interruptions or data acquisition failures. For each timestamp with a missing value, the system searches forward along the time axis for the nearest valid data point. This valid data point must belong to the same feature type as the missing point and have a timestamp earlier than the current point. Once the nearest valid data point is found, the system copies its value to the current missing position, completing the imputation of a single missing value. If multiple consecutive timestamps have missing data, forward imputation will continue to use the value of the same valid data point until a new valid data point is encountered. This method is based on the assumption that the system state remains relatively stable over a short period, and that the missing value is close to its previous valid value. After imputation is complete, the system records the imputation location and the source timestamp of the imputation value for subsequent quality assessment.

[0119] In practical implementation, non-maximum suppression (NMS) is used when generating calibration schedules based on calibration probability distributions. NMS extracts significant peaks from the probability sequence. The processing object of NMS is the calibration probability sequence of future time points output by a gated recurrent unit network (GRU). Each point in the sequence corresponds to a calibration probability value for a future time point. NMS slides a fixed-width window across the probability sequence, typically set to the number of time points corresponding to several hours. For each position within the window, the algorithm compares the probability value of the center point with the probability values ​​of other points within the window. If the probability value of the center point is greater than the probability values ​​of all other points within the window and exceeds a preset threshold of 0.8, then the center point is marked as a candidate calibration point. After completing a full sequence scan, multiple candidate calibration points are obtained. NMS further processes these candidate points to ensure that the minimum time interval is maintained between any two candidate points. If the time interval between two candidate points is too close, the point with the higher probability value is retained, and the point with the lower probability value is discarded. The list of candidate points processed by NMS constitutes the final determined calibration time points. The corresponding operating parameters for these time points are generated from the historical calibration database.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A self-calibration method based on a resistance simulator, characterized in that, The method includes the following steps: The sensor array collects the real-time operating parameters of the resistance simulator to form a real-time monitoring data stream; Adaptive signal decomposition is performed on the real-time monitoring data stream to extract multi-scale feature components and residual components; Joint analysis of the multi-scale feature components and the residual components is performed to identify abnormal fluctuation patterns. Causal reasoning was performed on the aforementioned abnormal fluctuation patterns to identify key influencing factors; The historical calibration database of the integrated resistor simulator and the aforementioned key influencing factors are used to generate a reference performance trajectory through a regression model. Calculate the deviation between the actual operating trajectory and the reference performance trajectory to assess the urgency of calibrating the resistance simulator; By integrating the aforementioned abnormal fluctuation patterns, key influencing factors, and calibration urgency, a calibration strategy is generated using a time series prediction model. The historical calibration database of the integrated resistor simulator and the key influencing factors are used to generate a reference performance trajectory through a regression model, including: Multiple sets of calibration event records, including environmental conditions, load parameters, and calibration results, are loaded from the historical calibration database; the key influencing factors are feature-aligned with the calibration event records; a Gaussian process regression model is used to perform nonlinear fitting on the aligned features to generate a reference performance trajectory; the hyperparameters of the Gaussian process regression model are adjusted to minimize the prediction error. The calculation of the deviation between the actual operating trajectory and the reference performance trajectory, and the assessment of the urgency of calibrating the resistance simulator, includes: A dynamic time warping algorithm is used to align the actual running trajectory and the reference performance trajectory, and the distance between the trajectories is calculated. The distance between the trajectories is input into a convolutional neural network to extract spatial features. The spatial features are weighted through an attention mechanism to output a calibration urgency score. The process of weighting spatial features through an attention mechanism to output a calibration urgency score includes: A self-attention module is added after the fully connected layer of the convolutional neural network; the attention weight distribution of the feature vectors is calculated; the feature vectors are weighted and summed according to the attention weights to generate aggregated features; the aggregated features are input into a softmax classifier to output a calibration urgency score. The process of integrating the abnormal fluctuation pattern, the key influencing factors, and the calibration urgency to generate a calibration strategy using a time series prediction model includes: The abnormal fluctuation pattern, the key influencing factors, and the calibration urgency are encoded as time series tensors; the time series tensors are processed using a gated recurrent unit network to capture long-term dependencies; the calibration probability distribution at multiple future time points is output; and a calibration schedule and operating parameters are generated based on the calibration probability distribution by setting thresholds.

2. The self-calibration method based on a resistance simulator as described in claim 1, characterized in that, The adaptive signal decomposition of the real-time monitoring data stream to extract multi-scale feature components and residual components includes: The real-time monitoring data stream is processed using the empirical mode decomposition algorithm to obtain a set of intrinsic mode functions; Perform a Hilbert transform on each intrinsic mode function and calculate the instantaneous frequency and amplitude; Significant intrinsic mode functions are selected based on a preset threshold and used as multi-scale feature components. The remaining unfiltered intrinsic mode functions are merged into residual components.

3. The self-calibration method based on a resistance simulator as described in claim 1, characterized in that, The joint analysis of the multi-scale feature components and the residual components to identify abnormal fluctuation patterns includes: The multi-scale feature components and the residual components are combined into a high-dimensional feature vector; The isolation forest algorithm is used to perform unsupervised learning on the high-dimensional feature vector to divide normal data points into abnormal data points; Density peak clustering is performed on abnormal data points to identify abnormal fluctuation patterns.

4. The self-calibration method based on a resistance simulator as described in claim 1, characterized in that, The causal reasoning of the abnormal fluctuation pattern to uncover key influencing factors includes: The abnormal fluctuation patterns are mapped to nodes to construct an initial graph structure; Use a Bayesian network learning algorithm to analyze the conditional dependencies between nodes and add directed edges; Edge weights are calculated using the maximum information coefficient to identify strong causal paths; Frequently occurring nodes are extracted from strong causal paths as key influencing factors.

5. The self-calibration method based on a resistance simulator as described in claim 1, characterized in that, The process of acquiring real-time operating parameters of the resistance simulator through a sensor array to form a real-time monitoring data stream includes: The collected data is subjected to sliding window averaging filtering, and the filtered data is sorted by timestamp to form a real-time monitoring data stream. The real-time monitoring data stream includes resistance drift, power consumption, and ambient temperature.

6. A self-calibration system based on a resistance simulator, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the self-calibration method based on a resistor simulator as described in any one of claims 1 to 5.

Images