Full-period operation data monitoring method and system applied to intelligent ring main unit

Abstract

The invention provides a full-period operation data monitoring method and system applied to an intelligent ring main unit, and the method comprises the steps: collecting multi-mode operation data in a full period in real time through operation data collection nodes disposed at all electrical nodes of the intelligent ring main unit; based on a preset time sequence feature extraction rule, performing segmentation alignment processing on the multi-mode operation data, and generating a full-period state parameter corresponding to each electrical node; training an operation monitoring model matched with the intelligent ring main unit according to the full-period state parameters; performing layer-by-layer analysis on the multi-modal operation data acquired in real time through the operation monitoring model, and generating a health degree score and an abnormal confidence interval corresponding to the current operation stage; and triggering an early warning signal or a model updating instruction matched with a preset threshold value, and dynamically adjusting the network parameters of the operation monitoring model. According to the method, multi-modal dynamic characteristics can be deeply fused, model parameters are optimized in real time, and a full-period monitoring system with effective feedback is formed.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method and system for monitoring the full-cycle operation data of intelligent ring network cabinets. Background Technology

[0002] With the rapid development of smart grids, intelligent ring main units (RMMs), as key equipment in distribution networks, directly impact power supply reliability and operation and maintenance efficiency through real-time monitoring and health management of their operating status. Current technologies for monitoring intelligent RMMs primarily employ data acquisition from fixed-type sensors and static model analysis. Examples include using current sensors to detect load fluctuations, temperature sensors to identify localized overheating, or vibration data to analyze mechanical wear. Some solutions trigger alarms using preset thresholds or rely on offline models to extrapolate trends from historical data to achieve basic fault early warning.

[0003] However, existing technologies have significant drawbacks: First, isolated data acquisition and analysis modes struggle to capture the dynamic correlation of multi-dimensional equipment states. For example, the coordinated changes in current harmonic distortion and mechanical vibration spectrum are not effectively correlated, resulting in insufficient ability to identify complex fault characteristics. Second, static models lack adaptability to parameter drift throughout the equipment's lifecycle, threshold settings rely on manual experience and are updated late, leading to frequent misjudgments and response delays. Finally, in traditional methods, the feedback link between monitoring data and operation and maintenance decisions is broken, and model optimization is disconnected from real-time operating status, failing to support proactive health management.

[0004] The aforementioned problems severely restrict the monitoring accuracy and operation and maintenance efficiency of intelligent ring main units, and there is a lack of reliable basis for predicting equipment degradation trends. There is an urgent need for a full-cycle monitoring system that can deeply integrate multimodal dynamic characteristics, optimize model parameters in real time, and form effective feedback. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring the full-cycle operation data of intelligent ring main units. This invention is achieved as follows: Firstly, this invention provides a method for monitoring the full-cycle operation data of intelligent ring main units, the method comprising: real-time acquisition of multi-modal operation data throughout the entire cycle through operation data acquisition nodes deployed at each electrical node of the intelligent ring main unit; the multi-modal operation data including voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra; based on preset time-series feature extraction rules, segmenting and aligning the multi-modal operation data to generate full-cycle state parameters corresponding to each electrical node; the full-cycle state parameters including steady-state feature vectors, transient feature matrices, and cumulative degradation factors; and so on. Based on the full-cycle state parameters, an operation monitoring model adapted to the intelligent ring network cabinet is trained. The operation monitoring model includes a state evaluation network and an anomaly detection network, and the output of the state evaluation network and the input of the anomaly detection network are connected through a dynamic weight allocation layer. The operation monitoring model analyzes the real-time collected multimodal operation data layer by layer to generate a health score and anomaly confidence interval corresponding to the current operation stage. Based on the health score and anomaly confidence interval, an early warning signal or model update instruction matching a preset threshold is triggered, and the network parameters of the operation monitoring model are dynamically adjusted based on the model update instruction.

[0006] In a second aspect, the present invention provides a data monitoring system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method described above.

[0007] The intelligent ring main unit full-cycle operation data monitoring method provided by this invention acquires multi-modal operation data such as voltage fluctuation sequences, current harmonic distribution, temperature change curves, and mechanical vibration spectra in real time through operation data acquisition nodes deployed at each electrical node. Based on time-series feature extraction rules, it generates steady-state feature vectors, transient feature matrices, and cumulative degradation factors covering the entire life cycle of the equipment. Combined with a two-layer operation monitoring model including a state assessment network and an anomaly detection network, and using a dynamic weight allocation layer to adaptively fuse spatiotemporal features, the real-time data is analyzed layer by layer to output a health score and anomaly confidence interval. This triggers graded early warnings or model update commands and dynamically adjusts network parameters. It can capture electrical stability through voltage fluctuation data, reflect load anomalies through current harmonic distribution, monitor contact point aging through temperature change curves, and identify mechanical vibration spectra. The system comprehensively assesses the electrical performance, thermodynamic state, and mechanical health of equipment by detecting loose mechanical structures. It utilizes steady-state feature vectors to quantify long-term operating trends, transient feature matrices to capture instantaneous event characteristics, and cumulative degradation factors to predict remaining equipment lifespan, forming a multi-dimensional condition assessment system. By leveraging spatial correlation analysis of the condition assessment network and temporal dependency modeling of the anomaly detection network, it collaboratively mines potential fault modes. A dynamic weight allocation layer balances the detection weights of long-term trends and short-term anomalies to improve decision-making accuracy under complex operating conditions. A combined judgment based on health scores and anomaly confidence intervals enables a response process from early warning to maintenance. Relying on incremental training datasets and local parameter adjustment mechanisms, the model achieves self-optimization, thereby comprehensively improving monitoring real-time performance, fault location accuracy, and the adaptability of operation and maintenance strategies, effectively reducing false alarms and missed alarms, and extending equipment lifespan. Attached Figure Description

[0008] Figure 1 This is a flowchart of a method for monitoring the full-cycle operation data of an intelligent ring network cabinet, provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the composition of an operational data monitoring system provided in an embodiment of the present invention.

[0009] Among them, 100 is the running data monitoring system; 101 is the processor; 102 is the bus; 103 is the memory; and 104 is the transceiver. Detailed Implementation

[0010] In this embodiment of the invention, the execution entity of the full-cycle operation data monitoring method applied to intelligent ring network cabinets is the operation data monitoring system, including but not limited to servers, personal computers, laptops, tablets, etc. For example... Figure 1 As shown, the method includes: Step S100: Real-time acquisition of multi-modal operation data throughout the entire cycle through operation data acquisition nodes deployed at each electrical node of the intelligent ring network cabinet; the multi-modal operation data includes voltage fluctuation sequence, current harmonic distribution, temperature change curve and mechanical vibration spectrum.

[0011] For example, the operational data acquisition node is a device used to collect operational data from each electrical node of the intelligent ring main unit. It can accurately sense and record changes in relevant electrical parameters. The entire cycle, for example, refers to the entire time period from the start of use of the intelligent ring main unit to its final decommissioning, covering various stages such as normal operation, potential anomalies, and equipment aging. Multimodal operational data, for example, is a data set containing multiple different types of characteristics. In this embodiment, the voltage fluctuation sequence is, for example, a series of values ​​showing the voltage changes of each electrical node of the intelligent ring main unit over time within the entire cycle. It reflects the voltage fluctuations at different times and is valuable for judging the stability of the power grid and the operating status of the equipment. Current harmonic distribution describes the content and distribution of each harmonic in the current, excluding the fundamental component. The presence of harmonics may lead to problems such as equipment overheating and increased losses. The temperature change curve records the temperature change trend of the electrical nodes over time. Excessively high temperatures may indicate potential faults such as overload or poor contact. The mechanical vibration spectrum is the result of spectral analysis of the vibration signals of the mechanical components of the intelligent ring main unit. It can reflect the operating status of the mechanical components, such as whether there are problems like loosening or wear.

[0012] Step S200: Based on the preset time-series feature extraction rules, the multimodal operation data is segmented and aligned to generate full-cycle state parameters corresponding to each electrical node; the full-cycle state parameters include steady-state feature vectors, transient feature matrices and cumulative degradation factors.

[0013] The preset time-series feature extraction rules are a series of pre-defined algorithms and criteria used to extract representative time-series features from multimodal operating data. Segmentation and alignment processing, for example, involves segmenting the multimodal operating data according to certain rules and aligning the segments to better analyze the data's characteristics and patterns. Full-cycle state parameters are a set of parameters used to comprehensively describe the operating state of each electrical node of the intelligent ring main unit throughout its entire lifecycle.

[0014] The steady-state feature vector is a vector composed of steady-state features from multimodal operating data, reflecting the characteristic information of electrical nodes under stable operating conditions. The transient feature matrix is ​​a matrix containing the characteristic information of electrical nodes during transient processes, such as the transition of electrical equipment from one stable state to another when subjected to external disturbances or failures. The cumulative degradation factor is an indicator used to measure the degree of performance degradation of electrical nodes over the entire lifecycle; it is obtained through analysis and calculation of multimodal operating data. By segmenting and aligning the multimodal operating data and generating full-cycle state parameters according to preset time-series feature extraction rules, the operating status of each electrical node in the intelligent ring main unit can be described more accurately, providing a more reliable basis for subsequent operation monitoring and fault diagnosis.

[0015] As one implementation method, step S200 involves segmenting and aligning the multimodal operation data based on preset time-series feature extraction rules to generate full-cycle state parameters corresponding to each electrical node. For example, it may include the following steps S210 to S250: Step S210: Extract periodic amplitude offset and phase change points from the voltage fluctuation sequence, and divide the voltage segment interval according to the phase change points.

[0016] Periodic amplitude offsets, for example, refer to the deviation of the voltage amplitude relative to the normal amplitude within a certain period in a voltage fluctuation sequence, reflecting the periodic changes in voltage amplitude. Phase abrupt changes are points in a voltage fluctuation sequence where the phase changes abruptly; these points are, for example, related to faults or abnormal operation of electrical equipment. Voltage segmentation intervals are different time periods divided into the voltage fluctuation sequence based on phase abrupt changes; the voltage changes within each time period have certain characteristics and patterns.

[0017] In practice, by analyzing and calculating voltage fluctuation sequences, periodic amplitude shifts and phase abrupt change points are extracted. Based on these phase abrupt change points, the voltage fluctuation sequence is divided into different voltage segment intervals. This allows for a more detailed analysis of voltage changes over different time periods, helping to identify voltage anomalies and potential faults.

[0018] Step S220: Remove the fundamental component from the current harmonic distribution, obtain the energy proportion and cumulative distortion of each harmonic, and screen key harmonic components based on the preset harmonic energy threshold.

[0019] The fundamental component is the component of the current with the same frequency as the power grid frequency; it is the main component of the current. Removing the fundamental component from the current harmonic distribution allows for a clearer analysis of each harmonic. The energy proportion of each harmonic is, for example, the percentage of each harmonic's energy in the total harmonic energy, reflecting the relative importance of each harmonic. Cumulative distortion is, for example, the cumulative value of the distortion of the current waveform by current harmonics over a certain period, reflecting the degree of impact of current harmonics on electrical equipment. A preset harmonic energy threshold is a pre-defined energy value used to screen key harmonic components. Key harmonic components are, for example, those with a large energy proportion and a significant impact on electrical equipment. By removing the fundamental component from the current harmonic distribution, obtaining the energy proportion and cumulative distortion of each harmonic, and screening key harmonic components based on the preset harmonic energy threshold, current harmonics can be analyzed and processed more effectively, reducing the harm caused by harmonics to intelligent ring main units.

[0020] Step S230: Calculate the gradient rise rate and thermal equilibrium stabilization time based on the temperature change curve, and determine the mechanical loosening index and contact resistance deterioration trend by combining the frequency domain energy distribution of the mechanical vibration spectrum.

[0021] The rate of temperature rise is the speed at which temperature increases over time in a temperature change curve, reflecting how fast or slow the temperature rises. The duration of thermal equilibrium is, for example, the length of time a temperature remains stable after reaching thermal equilibrium in a temperature change curve. Thermal equilibrium is, for example, the state where heat transfer between different parts of an object is balanced, and the temperature no longer changes significantly.

[0022] The frequency domain energy distribution of a mechanical vibration spectrum, for example, describes the energy distribution of a mechanical vibration signal at different frequencies, reflecting the vibration characteristics of mechanical components. The mechanical looseness index is an indicator used to measure the degree of looseness in mechanical components; it is obtained through analysis and calculation of the frequency domain energy distribution of the mechanical vibration spectrum. Contact resistance degradation trends, for example, describe the change in contact resistance of electrical nodes over time; increased contact resistance can lead to problems such as equipment overheating and increased losses.

[0023] For example, for the mechanical loosening index, vibration sensors (such as accelerometers) can be installed at key locations of mechanical components to collect vibration signals from the mechanical components of the intelligent ring main unit. The collected time-domain vibration signals are preprocessed (e.g., using low-pass, high-pass, or band-pass filters) to remove noise and interference components to improve signal quality. Sampling is performed at a certain sampling frequency (according to the Nyquist sampling theorem, greater than twice the highest frequency of the signal). The preprocessed time-domain signal is converted into a frequency-domain signal using a Fast Fourier Transform (FFT) to obtain a spectrum. Then, based on the structural characteristics and vibration properties of the mechanical components, the entire spectrum is divided into several frequency intervals (e.g., low-frequency band 0-100Hz, mid-frequency band 100-500Hz, high-frequency band 500-1000Hz, etc.). Within each frequency interval, the signal energy is calculated by integrating the square of the spectral amplitude (in actual discrete sampling, summation is used instead). Then, features related to mechanical loosening are extracted from the frequency-domain energy distribution, such as calculating the proportion of energy in certain specific frequency intervals to the total energy and the standard deviation of the energy distribution. Finally, the mechanical loosening index is calculated based on the extracted features using a linear combination algorithm or machine learning model. The trend of contact resistance degradation can be determined based on the multimodal operation data of the intelligent ring main unit. For example, sensors can be used to collect data related to contact resistance in real time, such as temperature change curves and current values. Temperature is closely related to contact resistance; when contact resistance increases, according to Joule's law, it will lead to increased heat generation and a rise in temperature. Analyzing the rate of rise and fluctuation of the temperature change curve, if the rate of rise is accelerated or the fluctuation is abnormally large, it may indicate that the contact resistance is deteriorating. At the same time, combined with current data, when the current is stable, the contact resistance value at different times can be calculated according to Ohm's law. The contact resistance values ​​calculated at different time points are arranged to form a sequence of resistance values ​​changing over time. By applying mathematical methods such as linear or nonlinear fitting to this sequence, its trend can be analyzed. If the resistance value shows an upward trend and the increase is gradually increasing, it can be determined that the contact resistance is deteriorating.

[0024] By calculating the gradient rise rate and thermal equilibrium stabilization time based on the temperature change curve, and combining this with the frequency domain energy distribution of the mechanical vibration spectrum, the mechanical loosening index and contact resistance degradation trend can be determined. This allows for a comprehensive evaluation of the mechanical and electrical performance of the intelligent ring main unit and timely detection of potential faults.

[0025] Step S240: Integrate the periodic amplitude offset, key harmonic components, mechanical loosening index, and contact resistance degradation trend into a steady-state feature vector.

[0026] The steady-state eigenvector is a vector used to describe the characteristic information of an electrical node under stable operating conditions. The periodic amplitude offset reflects the periodic changes in voltage amplitude, the critical harmonic component is the harmonic component that has a significant impact on electrical equipment, the mechanical loosening index measures the degree of looseness of mechanical components, and the contact resistance degradation trend reflects the change of contact resistance over time.

[0027] By fusing these features into a steady-state feature vector, multiple different types of feature information can be integrated to more comprehensively describe the steady-state operating state of electrical nodes. Through the analysis and processing of the steady-state feature vector, it is possible to more accurately determine whether the operating state of electrical nodes is normal, providing a more reliable basis for subsequent operation monitoring and fault diagnosis.

[0028] As one implementation method, step S240 integrates the periodic amplitude offset, key harmonic components, mechanical loosening index and contact resistance deterioration trend into a steady-state feature vector. For example, it may include the following steps S241 to S244: Step S241: Normalize the periodic amplitude offset to obtain a standardized voltage offset coefficient.

[0029] The standardized voltage offset coefficient is a coefficient obtained after normalization, reflecting the relative degree of voltage amplitude offset. By normalizing the periodic amplitude offset, the standardized voltage offset coefficient can be obtained, making the voltage offset of different electrical nodes comparable and providing a more accurate data foundation for subsequent feature fusion and analysis.

[0030] Step S242: Calculate the harmonic influence weight based on the energy proportion and cumulative distortion of the key harmonic components, and then perform weighted superposition of the key harmonic components based on the harmonic influence weight to obtain the weighted superposition of the key harmonic components.

[0031] The energy percentage of critical harmonic components reflects the proportion of each harmonic in the total harmonic energy, while the cumulative distortion reflects the cumulative degree of distortion of the current waveform by current harmonics. The harmonic impact weight is a weight value calculated based on the energy percentage of critical harmonic components and the cumulative distortion, used to measure the degree of influence of each critical harmonic component on electrical equipment.

[0032] The key harmonic components are weighted and summed based on their harmonic influence weights. This involves summing each key harmonic component according to its corresponding harmonic influence weight to obtain the weighted key harmonic components. This approach can more accurately reflect the comprehensive impact of key harmonic components on electrical equipment, providing a more reliable basis for subsequent condition assessment and fault diagnosis.

[0033] Step S243: Using a preset mechanical health mapping table, convert the mechanical loosening index into a first health score, and map the contact resistance deterioration trend into a second health score using linear interpolation.

[0034] The preset mechanical health mapping table is a pre-defined table that records the correspondence between the mechanical loosening index and the mechanical health score. The mechanical loosening index is an indicator that measures the degree of looseness of mechanical components. By querying the preset mechanical health mapping table, the mechanical loosening index can be converted into the corresponding first health score, which reflects the health status of the mechanical components. Linear interpolation is used to perform interpolation calculations between known data points. For example, the contact resistance degradation trend is the change trend of the contact resistance of electrical nodes over time. The contact resistance degradation trend can be mapped to a second health score using linear interpolation, which reflects the health status of the contact resistance. By converting the mechanical loosening index into the first health score and mapping the contact resistance degradation trend into the second health score, mechanical and electrical characteristics can be transformed into a unified health score, facilitating the assessment and comparison of the overall health status of the intelligent ring main unit.

[0035] Step S244: Concatenate the standardized voltage offset coefficient, the weighted superposition of key harmonic components, the first health score and the second health score into a steady-state feature vector according to preset dimensions.

[0036] The standardized voltage offset coefficient reflects the relative degree of voltage amplitude offset, and the weighted superposition of key harmonic components more accurately reflects the comprehensive impact of key harmonic components on electrical equipment. The first health score reflects the health status of mechanical components, and the second health score reflects the health status of contact resistance.

[0037] The preset dimension is a pre-defined dimensional requirement used to determine the dimension of the steady-state feature vector. By concatenating the standardized voltage offset coefficient, the weighted key harmonic components, the first health score, and the second health score according to the preset dimension to form the steady-state feature vector, multiple different types of feature information can be integrated to form a more comprehensive and accurate vector describing the steady-state operating state of the electrical node. Through the analysis and processing of the steady-state feature vector, the normal operating state of the electrical node can be determined more effectively, providing a more reliable basis for subsequent operation monitoring and fault diagnosis.

[0038] Step S250: Construct a transient feature matrix based on the maximum instantaneous deviation and thermal equilibrium stabilization time within the voltage segment interval, and generate a cumulative degradation factor by weighted accumulation of distortion accumulation.

[0039] The maximum instantaneous deviation within a voltage segment interval is, for example, the maximum instantaneous offset of the voltage amplitude relative to the normal amplitude within the voltage segment interval, reflecting the degree of voltage fluctuation during the transient process. The thermal equilibrium settling time is, for example, the length of time that the temperature remains stable after reaching thermal equilibrium in a temperature change curve.

[0040] The transient characteristic matrix is ​​a matrix used to describe the characteristic information of electrical nodes during transient processes. By constructing the transient characteristic matrix based on the maximum instantaneous deviation within the voltage segment interval and the thermal equilibrium stabilization time, the operating state of electrical nodes during transient processes can be reflected more comprehensively.

[0041] Accumulated distortion, for example, is the cumulative value of the distortion of the current waveform caused by current harmonics over a certain period of time. Weighted accumulated distortion, for example, is generated by summing the accumulated distortion values ​​from different time periods according to certain weights, thus generating a cumulative degradation factor. The cumulative degradation factor is an indicator used to measure the degree of performance degradation of an electrical node over its entire lifecycle. By generating the cumulative degradation factor through weighted accumulated distortion, the performance degradation of the electrical node can be more accurately reflected, providing a more reliable basis for subsequent condition assessment and fault diagnosis.

[0042] Step S300: Based on the full-cycle state parameters, train an operation monitoring model adapted to the intelligent ring network cabinet; the operation monitoring model includes a state assessment network and an anomaly detection network, and the output of the state assessment network and the input of the anomaly detection network are connected through a dynamic weight allocation layer.

[0043] The full-cycle state parameters are a set of parameters used to comprehensively describe the operating status of each electrical node of the intelligent ring main unit throughout its entire lifecycle. These parameters include the steady-state characteristic vector, transient characteristic matrix, and cumulative degradation factor. The operation monitoring model is a model used to monitor and evaluate the operating status of the intelligent ring main unit. It can analyze the operating data of the intelligent ring main unit in real time, determine whether it is operating normally, and promptly identify potential faults.

[0044] The status assessment network, part of the operation monitoring model, evaluates the operational status of the intelligent ring main unit and outputs characteristic information related to that status. The anomaly detection network, another part of the model, detects anomalies in the intelligent ring main unit and outputs the anomaly detection results. The dynamic weight allocation layer connects the output of the status assessment network and the input of the anomaly detection network. It dynamically adjusts the weights of the output features of the status assessment network based on different situations, enabling the anomaly detection network to more accurately utilize the output information of the status assessment network.

[0045] By training an operation monitoring model adapted to the intelligent ring main unit based on the full-cycle state parameters, the model can better adapt to the operation characteristics of the intelligent ring main unit, thereby improving the accuracy and reliability of operation monitoring.

[0046] As one implementation method, step S300, based on the full-cycle state parameters, trains an operation monitoring model adapted to the intelligent ring network cabinet, which may specifically include the following steps S310~S350: Step S310: Divide the full-cycle state parameters into a training set and a validation set, and inject noise to enhance the steady-state feature vector and transient feature matrix in the training set.

[0047] The training set is the dataset used to train and monitor the model, while the validation set is the dataset used to validate the performance of the trained model. Dividing the full-cycle state parameters into training and validation sets allows the model's performance to be evaluated using the validation set during training, thus preventing overfitting.

[0048] Noise injection augmentation is a data augmentation method that increases the diversity and complexity of data by injecting noise into the steady-state feature vectors and transient feature matrices in the training set, thereby improving the model's generalization ability. The noise can be random interference signals, which can simulate various noise and interference situations that may occur in actual operation.

[0049] By dividing the full-cycle state parameters into training and validation sets, and enhancing the steady-state feature vectors and transient feature matrices in the training set with noise injection, the training effect and performance of the operation monitoring model can be improved, enabling it to better adapt to the actual operating environment.

[0050] Step S320: Construct a multi-layer convolutional channel for the state evaluation network, wherein the first convolutional layer is used to extract the spatial correlation of the steady-state feature vector, and the second convolutional layer is used to capture the temporal dependence of the transient feature matrix.

[0051] Multi-layer convolutional channels are a crucial component of state evaluation networks, consisting of multiple convolutional layers, each capable of extracting and processing different features from the input data. The first convolutional layer, the first layer in the multi-layer convolutional channel, primarily extracts the spatial correlation of steady-state feature vectors. A steady-state feature vector is a vector containing multiple feature information; the first convolutional layer, through convolution operations, can discover the spatial relationships and patterns between these features, thereby extracting more representative features.

[0052] The second convolutional layer is the second layer in a multi-layer convolutional channel. Its main function is to capture the temporal dependence of the transient feature matrix. The transient feature matrix is ​​a matrix containing feature information of electrical nodes during transient processes. This feature information changes over time, and the second convolutional layer can capture the changing patterns and dependencies of these features in the time dimension through convolution operations.

[0053] By constructing a multi-layer convolutional network for the state assessment network, the first convolutional layer extracts the spatial correlation of steady-state feature vectors, and the second convolutional layer captures the temporal dependence of transient feature matrices. This allows for a more comprehensive and accurate extraction of the feature information of the intelligent ring network cabinet's operating state, providing a more reliable basis for subsequent state assessment and anomaly detection.

[0054] Step S330: Set up bidirectional long short-term memory units in the anomaly detection network to generate time-series prediction results of health scores based on the fusion features output by the dynamic weight allocation layer.

[0055] Bidirectional Long Short-Term Memory (BiLSTM) units can simultaneously consider both past and future information of the input sequence, thus better processing sequence data. In anomaly detection networks, BiLSTM units can be incorporated to analyze and process the fused features output by the dynamic weight allocation layer. These fused features are obtained by dynamically adjusting the feature information output by the state assessment network, integrating both steady-state and transient features. For example, the time-series prediction results of health scores, such as those for intelligent ring main units over a future period, can be more accurately predicted by using BiLSTM units to generate these predictions based on the fused features, allowing for timely detection of potential anomalies and more accurate prediction of the operating status trends of intelligent ring main units.

[0056] Step S340: By comparing the time-series prediction results with the cumulative degradation factor in the validation set, calculate the mean squared error loss function, and use the adaptive momentum optimization algorithm to update the weight parameters of the multi-layer convolutional channels and bidirectional long short-term memory units.

[0057] The time-series prediction result is the health score generated by the bidirectional long short-term memory units in the anomaly detection network based on the fusion features output by the dynamic weight allocation layer. It reflects the predicted health score of the intelligent ring main unit over a future period. The cumulative degradation factor in the validation set is an indicator of the performance degradation degree of electrical nodes recorded in the validation set over the entire cycle.

[0058] The mean squared error loss function (MSE) measures the error between the predicted result and the true value. By comparing the time-series prediction results with the cumulative degradation factor in the validation set, the MSE loss function can be calculated to evaluate the model's prediction accuracy. The adaptive momentum optimization algorithm automatically adjusts the learning rate based on the model's training progress, accelerating the model's convergence speed. This algorithm updates the weight parameters of multi-layer convolutional channels and bidirectional long short-term memory units (LSTMs) based on the MSE loss function calculation, making the model's predictions closer to the true values ​​and improving the model's performance and accuracy.

[0059] As one implementation method, step S340, by comparing the time series prediction results with the cumulative degradation factors in the validation set, calculates the mean squared error loss function, which may include the following steps S341~S345: Step S341: Extract the true value sequence of the cumulative degradation factors corresponding to the current training batch from the validation set.

[0060] The validation set is a dataset used to validate the performance of the trained model. It contains information such as the cumulative degradation factor of electrical nodes over the entire lifecycle. The current training batch is, for example, a specific training batch during the model training process. The ground truth sequence of cumulative degradation factors is, for example, the sequence of true cumulative degradation factor values ​​recorded in the validation set corresponding to the current training batch, which reflects the performance degradation of electrical nodes during actual operation.

[0061] Extracting the true value sequence of cumulative degradation factors corresponding to the current training batch from the validation set can provide accurate true value data for subsequent calculation of the mean squared error loss function, so as to more accurately evaluate the prediction accuracy of the model.

[0062] Step S342: Obtain the health score prediction sequence output by the anomaly detection network, and perform trend alignment on the health score prediction sequence using the exponential smoothing method.

[0063] The health score prediction sequence output by the anomaly detection network is, for example, the time-series prediction result of the health score generated by the bidirectional long short-term memory unit in the anomaly detection network based on the fusion features output by the dynamic weight allocation layer. It reflects the health score prediction of the intelligent ring network cabinet in the future.

[0064] Exponential smoothing predicts future values ​​by weighting historical data. By aligning the health score prediction sequence with its trend using exponential smoothing, short-term fluctuations and noise can be eliminated, resulting in a smoother trend that better reflects actual changes. This improves the accuracy of subsequent calculations of the mean squared error loss function, allowing for a more precise evaluation of the model's performance.

[0065] Step S343: Calculate the sum of squared differences between the trend-aligned health score prediction sequence and the cumulative degradation factor true value sequence.

[0066] The trend-aligned health score prediction sequence is a health score prediction sequence with a smoother trend after exponential smoothing. The cumulative degradation factor ground value sequence is the sequence of true cumulative degradation factors corresponding to the current training batch, extracted from the validation set.

[0067] The pointwise sum of squared differences is calculated by subtracting the corresponding values ​​in the trend-aligned health score prediction sequence from the true cumulative degradation factor sequence, squaring the difference, and then summing all the squared values. Calculating the pointwise sum of squared differences measures the degree of difference between the trend-aligned health score prediction sequence and the true cumulative degradation factor sequence, providing a basis for subsequent calculations of the mean squared error loss.

[0068] Step S344: Determine the mean squared error loss value for the current training batch based on the sum of squared differences at each point and the total number of samples in the validation set.

[0069] The sum of squared differences is an indicator that measures the degree of difference between the trend-aligned health score prediction sequence and the cumulative degradation factor true value sequence. The total number of samples in the validation set is, for example, the number of samples contained in the validation set.

[0070] The mean squared error loss (MSE) is obtained by dividing the sum of squared point-by-point differences by the total number of samples in the validation set. It reflects the average prediction error of the model in the current training batch. Determining the MSE for the current training batch based on the sum of squared point-by-point differences and the total number of samples in the validation set allows for a more accurate assessment of the model's performance in the current training phase, providing a basis for subsequent adjustments to the model's weight parameters.

[0071] Step S345: Accumulate the mean squared error loss values ​​of all training batches to obtain the mean squared error loss function.

[0072] Each training batch generates a mean squared error loss value, which reflects the model's prediction error at different training stages. The mean squared error loss function is obtained by accumulating the mean squared error losses of all training batches, and it comprehensively reflects the model's prediction error throughout the entire training process.

[0073] By accumulating the mean squared error loss value of all training batches, the mean squared error loss function can be obtained, which can more comprehensively evaluate the training effect and performance of the model, and provide an important basis for judging whether the model has converged and whether further adjustment of weight parameters is needed.

[0074] Step S350: When the mean squared error loss function is lower than the preset convergence threshold, freeze the parameters of the state evaluation network and the anomaly detection network to complete the training of the running monitoring model.

[0075] The mean squared error loss function is a function that comprehensively reflects the prediction error of the model throughout the entire training process. The preset convergence threshold is a pre-set error threshold used to determine whether the model has converged. When the mean squared error loss function is lower than the preset convergence threshold, it indicates that the model's prediction error has reached an acceptable level, and the model has converged.

[0076] Freezing the parameters of the state evaluation network and the anomaly detection network involves fixing the weight parameters of these networks and refraining from further adjustment. Training the operational monitoring model is completed by stopping the training process after the model converges, resulting in a model suitable for practical operational monitoring. By freezing the parameters of the state evaluation network and the anomaly detection network when the mean squared error loss function falls below a preset convergence threshold, the training of the operational monitoring model can be completed, ensuring that the model has good performance and stability and can accurately monitor and evaluate the operational status of the intelligent ring network cabinet.

[0077] Step S400: The real-time collected multimodal operation data is analyzed layer by layer through the operation monitoring model to generate a health score and anomaly confidence interval corresponding to the current operation stage.

[0078] Real-time multimodal operational data is collected in real time, such as voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra, obtained through operational data acquisition nodes deployed at each electrical node of the intelligent ring main unit. The operational monitoring model is a trained model used to monitor and evaluate the operational status of the intelligent ring main unit, and it includes a status assessment network and an anomaly detection network.

[0079] Layer-by-layer analysis, for example, involves the operation monitoring model analyzing and processing real-time collected multimodal operation data step by step according to a certain hierarchical structure. The health score is an indicator used to measure the current operational health of the intelligent ring main unit, comprehensively considering various characteristic information in the multimodal operation data. The anomaly confidence interval is a range used to represent the probability of anomalies in the intelligent ring main unit, helping maintenance personnel determine whether there are potential anomaly risks. By performing layer-by-layer analysis of the real-time collected multimodal operation data through the operation monitoring model, generating health scores and anomaly confidence intervals corresponding to the current operational stage, the operational status of the intelligent ring main unit can be understood in a timely and accurate manner, providing a basis for subsequent early warning and decision-making.

[0080] As one implementation method, step S400 involves analyzing the real-time collected multimodal operation data layer by layer through the running monitoring model to generate a health score and anomaly confidence interval corresponding to the current operation stage. Specifically, this may include the following steps S410~S450: Step S410: Input the real-time collected voltage fluctuation sequence and current harmonic distribution into the first convolutional layer of the state assessment network to extract the real-time steady-state feature vector.

[0081] The real-time voltage fluctuation sequence is, for example, a series of values ​​of the voltage change of each electrical node of the intelligent ring main unit over time, collected in real time at the current moment. The current harmonic distribution is, for example, the content and distribution of each harmonic in the current, excluding the fundamental component, collected in real time at the current moment.

[0082] The first convolutional layer of the state assessment network is the first layer in a multi-layer convolutional channel, and its main function is to extract the spatial correlation of steady-state feature vectors. Real-time acquired voltage fluctuation sequences and current harmonic distributions are input into the first convolutional layer of the state assessment network. Through convolution operations, the first convolutional layer can discover the spatial relationships and patterns between features in these data, thereby extracting real-time steady-state feature vectors. These real-time steady-state feature vectors reflect the steady-state operating characteristics of the intelligent ring main unit at the current moment, providing important basic data for subsequent state assessment and anomaly detection.

[0083] Step S420: Input the real-time temperature change curve and mechanical vibration spectrum into the second convolutional layer of the state assessment network to generate a real-time transient feature matrix.

[0084] Real-time temperature change curves, for example, are the temperature trends of each electrical node of the intelligent ring main unit collected in real time over time. Mechanical vibration spectra, for example, are the spectral analysis results of vibration signals of mechanical components of the intelligent ring main unit collected in real time.

[0085] The second convolutional layer of the state assessment network is the second layer in a multi-layer convolutional network. Its main function is to capture the time dependence of the transient feature matrix. Real-time temperature change curves and mechanical vibration spectra are input into the second convolutional layer of the state assessment network. Through convolution operations, the second convolutional layer can capture the changing patterns and dependencies of these data over time, thereby generating a real-time transient feature matrix. The real-time transient feature matrix reflects the transient operating characteristics of the intelligent ring main unit at the current moment. Together with the real-time steady-state feature vector, it provides richer information for comprehensively assessing the operating status of the intelligent ring main unit.

[0086] Step S430: The real-time steady-state feature vector and the real-time transient feature matrix are fused through a dynamic weight allocation layer to obtain the enhanced spatiotemporal feature tensor.

[0087] The dynamic weight allocation layer connects the output of the state assessment network and the input of the anomaly detection network. It can dynamically adjust the weights of the output features of the state assessment network according to different situations. The real-time steady-state feature vector reflects the steady-state operating characteristics of the intelligent ring main unit at the current moment, while the real-time transient feature matrix reflects the transient operating characteristics of the intelligent ring main unit at the current moment.

[0088] Feature fusion, for example, integrates real-time steady-state feature vectors and real-time transient feature matrices to extract more valuable feature information. Enhanced spatiotemporal feature tensors, obtained through feature fusion, are tensors containing richer spatiotemporal information. They combine the advantages of real-time steady-state feature vectors and real-time transient feature matrices, enabling a more comprehensive and accurate description of the operating status of intelligent ring network cabinets.

[0089] By fusing real-time steady-state feature vectors and real-time transient feature matrices through a dynamic weight allocation layer, an enhanced spatiotemporal feature tensor is obtained, which can provide higher-quality input information for anomaly detection networks and improve the accuracy and reliability of anomaly detection.

[0090] As one implementation method, step S430 involves fusing the real-time steady-state feature vector and the real-time transient feature matrix through a dynamic weight allocation layer to obtain an enhanced spatiotemporal feature tensor. This may include the following steps S431 to S434: Step S431: Expand the dimension of the real-time steady-state feature vector to generate a first extended feature that matches the dimension of the real-time transient feature matrix.

[0091] The real-time steady-state feature vector is extracted through the first convolutional layer of the state evaluation network, reflecting the steady-state operating characteristics of the intelligent ring main unit at the current moment. The real-time transient feature matrix is ​​generated through the second convolutional layer of the state evaluation network, reflecting the transient operating characteristics of the intelligent ring main unit at the current moment.

[0092] Dimension expansion, for example, involves increasing the dimension of the real-time steady-state feature vector to match the dimension of the real-time transient feature matrix. The first expanded feature is the feature obtained after dimensional expansion, which has the same dimension as the real-time transient feature matrix, facilitating subsequent feature fusion operations. By expanding the dimension of the real-time steady-state feature vector to generate the first expanded feature that matches the dimension of the real-time transient feature matrix, a more suitable data foundation can be provided for subsequent feature fusion, ensuring the effectiveness and accuracy of feature fusion.

[0093] Step S432: Calculate the feature weights of each time step in the real-time transient feature matrix using an attention mechanism, and resample the real-time transient feature matrix according to the feature weights.

[0094] Attention mechanisms are commonly used in deep learning. They automatically learn the importance of different parts of the input data and assign them appropriate weights. A real-time transient feature matrix is ​​a matrix containing feature information of electrical nodes during transient processes; this feature information changes over time.

[0095] The attention mechanism calculates the feature weights for each time step in the real-time transient feature matrix. This means the attention mechanism automatically learns and determines the importance of each time step based on its feature information, and then assigns it a corresponding weight. Resampling, for example, involves resampling the real-time transient feature matrix based on the calculated feature weights, highlighting important time step features and weakening less important ones.

[0096] By calculating the feature weights of each time step in the real-time transient feature matrix through an attention mechanism, and resampling the real-time transient feature matrix according to the feature weights, key information in the real-time transient feature matrix can be extracted more effectively, thereby improving the effect of feature fusion and the accuracy of anomaly detection.

[0097] Step S433: Concatenate the first extended feature with the resampled real-time transient feature matrix along the channel dimension to obtain the initial fused feature.

[0098] The first extended feature is obtained by expanding the dimensions of the real-time steady-state feature vector, and its dimension is consistent with that of the real-time transient feature matrix. The resampled real-time transient feature matrix is ​​obtained by calculating feature weights through an attention mechanism and then resampling, which highlights important time-step features.

[0099] Channel dimension, for example, refers to the dimension of the data along the channel direction. The first extended feature is concatenated with the resampled real-time transient feature matrix along the channel dimension, that is, these two features are connected along the channel direction to form a new feature. The initial fused feature is obtained through this concatenation operation, which integrates information from the real-time steady-state feature vector and the real-time transient feature matrix, providing a richer feature foundation for subsequent processing.

[0100] Step S434: The initial fused features are reduced in dimensionality by a fully connected layer, and the reduced features are nonlinearly activated by a gated linear unit to generate an enhanced spatiotemporal feature tensor.

[0101] Fully connected layers are a common type of layer in neural networks, connecting each input neuron to each output neuron. The initial fused features are obtained by concatenating the first expanded features with the resampled real-time transient feature matrix along the channel dimension; these features may contain significant dimensionality and redundant information.

[0102] The initial fused features are reduced in dimensionality by using fully connected layers, which means reducing the dimensionality of the initial fused features, removing redundant information, and extracting more critical features. A Gated Linear Unit (GLU) is a non-linear activation function that can perform non-linear transformations on the input features, increasing the model's expressive power.

[0103] Gated linear units are used to perform nonlinear activation on the dimensionality-reduced features. This involves inputting the dimensionality-reduced features into a gated linear unit and generating new features through nonlinear transformation. The enhanced spatiotemporal feature tensor, obtained after dimensionality reduction and nonlinear activation, integrates real-time steady-state and transient feature information and, having undergone nonlinear transformation, can more comprehensively and accurately describe the operating status of the intelligent ring network cabinet.

[0104] Step S440: Input the enhanced spatiotemporal feature tensor into the bidirectional long short-term memory unit of the anomaly detection network, and output the health score sequence for the next N running cycles, where N>2.

[0105] The enhanced spatiotemporal feature tensor is a tensor obtained by feature fusion, dimensionality reduction and nonlinear activation of real-time steady-state feature vectors and real-time transient feature matrices. It integrates real-time steady-state and transient feature information and can more comprehensively and accurately describe the operating status of intelligent ring network cabinets.

[0106] The Bidirectional Long Short-Term Memory (BiLSTM) unit in anomaly detection networks is a special type of recurrent neural network unit that can simultaneously consider past and future information of the input sequence, thus better processing sequence data. By inputting an enhanced spatiotemporal feature tensor into the BiLSTM unit of the anomaly detection network, the BiLSTM unit can analyze and process the enhanced spatiotemporal feature tensor, predicting the health score sequence for the next N operating cycles based on its internal memory and computation mechanism.

[0107] The health score sequence for the next N operating cycles reflects the predicted health score of the intelligent ring main unit over a period of time. By outputting this sequence, maintenance personnel can understand the trend of changes in the operating status of the intelligent ring main unit in advance and take corresponding measures in a timely manner to ensure the safe and stable operation of the intelligent ring main unit.

[0108] Step S450: Calculate the anomaly confidence interval based on the variance and the maximum-minimum interval of the health score sequence.

[0109] The health score sequence is the health score sequence for the next N operating cycles output by the bidirectional long short-term memory unit of the anomaly detection network, which reflects the predicted health score of the intelligent ring main unit in the future.

[0110] Variance is a statistic that measures the dispersion of data. The variance of a health score sequence reflects the fluctuation of each health score within the sequence. The maximum-minimum interval, for example, is the difference between the maximum and minimum values ​​in the health score sequence, reflecting the range of variation in the health scores.

[0111] An anomaly confidence interval is a range used to represent the probability of an anomaly in a smart ring main unit. By calculating the anomaly confidence interval based on the variance of the health score sequence and the range between the maximum and minimum values, the fluctuation and range of change in the health score can be comprehensively considered, allowing for a more accurate assessment of the probability of anomalies in the smart ring main unit. This provides a more reliable basis for subsequent early warning and decision-making.

[0112] Step S500: Based on the health score and the abnormal confidence interval, trigger an early warning signal or model update command that matches the preset threshold, and dynamically adjust the network parameters of the running monitoring model based on the model update command.

[0113] The health score is an indicator used to measure the current operational health of the intelligent ring main unit. The anomaly confidence interval is a range representing the probability of an anomaly in the intelligent ring main unit. The preset thresholds are a series of pre-set thresholds used to determine whether the intelligent ring main unit is operating normally and whether corresponding measures need to be taken.

[0114] Early warning signals are issued when the operating status of the intelligent ring main unit (RMU) becomes abnormal. These signals alert maintenance personnel to take timely measures to ensure the safe and stable operation of the RMU. Model update commands are issued when the operating status of the RMU changes and the existing operation monitoring model is no longer adequately adapted to the new situation. These commands trigger an update of the network parameters in the operation monitoring model.

[0115] The network parameters of the operation monitoring model are dynamically adjusted based on model update commands. Specifically, the weight parameters of the state assessment network and anomaly detection network in the operation monitoring model are adjusted according to the model update commands, enabling the model to better adapt to the new operating states of the intelligent ring main unit. By triggering early warning signals or model update commands matching preset thresholds based on health scores and anomaly confidence intervals, and dynamically adjusting the network parameters of the operation monitoring model based on these commands, abnormal situations in the intelligent ring main unit can be detected in a timely manner, ensuring the operation monitoring model maintains good performance and accurately monitors and evaluates the operating status of the intelligent ring main unit.

[0116] As one implementation method, in step S500, based on the health score and the abnormal confidence interval, a warning signal or model update instruction matching a preset threshold is triggered. Specifically, this may include the following steps S510~S530: Step S510: When the health score is lower than the first threshold and the upper limit of the abnormal confidence interval exceeds the second threshold, a first-level warning signal is generated and an emergency shutdown instruction is triggered.

[0117] The health score is an indicator that measures the current operational health of the intelligent ring main unit. The first threshold is a pre-set threshold used to determine if the health score is too low. The upper limit of the anomaly confidence interval is the upper limit of the anomaly confidence interval, representing the upper limit of the probability of an anomaly in the intelligent ring main unit. The second threshold is a pre-set threshold used to determine if the upper limit of the anomaly confidence interval is too high. The Level 1 warning signal is a relatively serious warning signal, indicating that the intelligent ring main unit's operational status has encountered a serious anomaly, which may cause significant harm to the equipment and personnel. The emergency shutdown command is an instruction issued when a serious anomaly occurs in the intelligent ring main unit, used to immediately stop the operation of the intelligent ring main unit to avoid accidents. When the health score is below the first threshold and the upper limit of the anomaly confidence interval exceeds the second threshold, it indicates that the intelligent ring main unit's operational status is very dangerous, with a significant risk of anomalies. At this time, generating a Level 1 warning signal and triggering an emergency shutdown command can promptly remind maintenance personnel to take emergency measures to ensure the safety of the intelligent ring main unit.

[0118] Step S520: When the health score is between the first threshold and the third threshold and the width of the abnormal confidence interval exceeds the fourth threshold, a secondary warning signal is generated and a local maintenance command is triggered.

[0119] The health score is an indicator that measures the current operating health of the intelligent ring main unit. The first threshold and the third threshold are two pre-set thresholds used to divide the range of the health score. When the health score is between these two thresholds, it indicates that the operating status of the intelligent ring main unit is at a certain abnormal level, but it is not very serious.

[0120] The anomaly confidence interval width is the difference between the upper and lower limits of the anomaly confidence interval, reflecting the range of possibilities for anomalies in the intelligent ring main unit. The fourth threshold is a pre-set threshold used to determine whether the anomaly confidence interval width is too large.

[0121] A Level 2 warning signal is a milder warning signal than a Level 1 warning signal. It indicates that the intelligent ring main unit has experienced some abnormalities in its operation, requiring the attention of maintenance personnel. A partial maintenance command is issued when the intelligent ring main unit exhibits abnormalities. It is used to inspect certain components of the intelligent ring main unit to restore its normal operating condition.

[0122] When the health score falls between the first and third thresholds and the anomaly confidence interval exceeds the fourth threshold, it indicates that the intelligent ring main unit's operating status poses a certain risk of abnormality and requires partial maintenance. At this point, a level-two warning signal is generated and a partial maintenance command is triggered, promptly reminding maintenance personnel to perform maintenance on the intelligent ring main unit and preventing the abnormal situation from worsening.

[0123] Step S530: When the health score decline rate exceeds the fifth threshold for N consecutive running cycles, generate a model update instruction and start the incremental learning mode; wherein, the incremental learning mode includes: adding the multimodal running data collected in the current cycle to the end of the training set, and updating the weight parameters of the running monitoring model through the sliding window method.

[0124] The health score decline rate is, for example, the ratio of the decline in the health score over N consecutive operating cycles to the initial health score, reflecting the rate at which the intelligent ring network cabinet's operating status deteriorates. The fifth threshold is a pre-set threshold used to determine whether the health score decline rate is too high.

[0125] The model update command is issued when the operating status of the intelligent ring network cabinet changes and the original operation monitoring model can no longer adapt well to the new situation. It is used to trigger the update of the network parameters of the operation monitoring model. Incremental learning mode is a machine learning mode that allows new data to be continuously added to the original training data for training, in order to adapt to changes in data distribution.

[0126] The multimodal operational data collected in the current cycle is added to the end of the training set. This involves adding newly collected data such as voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra to the end of the existing training set, thus expanding the scale of the training data. The sliding window method is a data processing technique that uses a fixed-size window to slide across the data sequence, continuously selecting new data for training. The sliding window method is used to update the weight parameters of the operational monitoring model. Specifically, after adding new data, a portion of the data is selected using the sliding window method to update the weight parameters of the operational monitoring model, enabling the model to better adapt to the new operational state.

[0127] When the health score decline rate exceeds the fifth threshold for N consecutive operating cycles, it indicates a significant change in the operating status of the intelligent ring main unit, and the original operation monitoring model may no longer be applicable. At this point, generating a model update command and initiating incremental learning mode allows for timely updates to the operation monitoring model, improving its accuracy and adaptability.

[0128] In one implementation, the incremental learning mode is initiated, which may include the following steps S531 to S534: Step S531: Keep the parameters of the first K layers of the convolutional channels of the state evaluation network in the running monitoring model unchanged, and unlock the parameters of the last L layers of the bidirectional long short-term memory units of the anomaly detection network.

[0129] The operational monitoring model comprises a state evaluation network and an anomaly detection network. The first K convolutional layers of the state evaluation network are the first K layers in a multi-layer convolutional network. These layers are used to extract some basic features of the data and have been extensively trained, exhibiting good stability. The last L layers of the bidirectional long short-term memory units in the anomaly detection network are the last L layers in the bidirectional long short-term memory units. These layers have a significant impact on the final prediction results.

[0130] The parameters of the first K convolutional channels of the state evaluation network are kept unchanged, meaning that the weight parameters of these layers are not changed during incremental learning to maintain the model's ability to extract basic features. The parameters of the last L layers of the bidirectional long short-term memory units in the anomaly detection network are unlocked, allowing the weight parameters of these layers to be adjusted during incremental learning so that the model can better adapt to new data and changes.

[0131] By retaining the parameters of the first K convolutional channels of the state evaluation network in the runtime monitoring model unchanged, and unlocking the parameters of the last L layers of the bidirectional long short-term memory units in the anomaly detection network, the model's original basic feature extraction capabilities can be preserved during incremental learning, while also allowing the model to adjust according to new data, thus improving the model's adaptability and accuracy. N, K, and L are all positive integers greater than 0.

[0132] Step S532: Fine-tune the parameters of the last L layer that has been unlocked using a local training set, while fixing the parameters of the remaining layers.

[0133] The local training set is a subset of data selected from the incremental training dataset, used for local training of the unlocked parameters. In incremental learning mode, the parameters of the last L layers of the bidirectional long short-term memory units of the anomaly detection network have been unlocked, while the parameters of the remaining layers remain fixed.

[0134] Fine-tuning the parameters of the final L layers by using a local training set—that is, making small adjustments to the weight parameters of these final L layers using a local training set—allows the model to better adapt to new data and changes. Meanwhile, fixing the parameters of the remaining layers is to avoid unnecessary impact on other parts of the model during fine-tuning, thus maintaining model stability.

[0135] In this way, the model can be adjusted according to new data without destroying its original structure and performance, thereby improving the model's adaptability and accuracy.

[0136] Step S533: Calculate the mean absolute percentage error of the fine-tuned anomaly detection network on the validation set. If it exceeds the preset fine-tuning threshold, roll back to the parameter state before unlocking.

[0137] Mean Absolute Percentage Error (MAPE) is a metric used to measure the error between predicted and true values. It calculates the average of the absolute percentage errors between the predicted and true values. After fine-tuning the parameters of the last L layers of the bidirectional long short-term memory units in the anomaly detection network, a validation set is needed to evaluate the performance of the fine-tuned model.

[0138] The preset fine-tuning threshold is a pre-defined error threshold used to determine whether the performance of the fine-tuned model meets the requirements. If the mean absolute percentage error of the fine-tuned anomaly detection network on the validation set exceeds the preset fine-tuning threshold, it indicates that the performance of the fine-tuned model is not as good as before fine-tuning, which may be due to overfitting or other problems during the fine-tuning process.

[0139] Rolling back to the parameter state before unlocking means restoring the parameters of the last L layers of the bidirectional long short-term memory units of the anomaly detection network to their pre-unlocked state, ensuring that the model's performance is not affected. By calculating the mean absolute percentage error of the fine-tuned anomaly detection network on the validation set and rolling back to the pre-unlocked parameter state when it exceeds a preset fine-tuning threshold, the stability and reliability of the model during incremental learning can be ensured.

[0140] Step S534: When the mean absolute percentage error does not exceed the preset fine-tuning threshold, synchronize the fine-tuned parameters to all online running monitoring model instances.

[0141] When the mean absolute percentage error of the fine-tuned anomaly detection network on the validation set does not exceed the preset fine-tuning threshold, it indicates that the performance of the fine-tuned model has been improved and it can better adapt to new data and changes.

[0142] All online operational monitoring model instances are, for example, multiple copies of the operational monitoring model running in a real-world application. These instances may be distributed across different devices or nodes for real-time monitoring of the operational status of the intelligent ring network cabinet. The fine-tuned parameters are synchronized to all online operational monitoring model instances; that is, the parameters of the last L layer of the bidirectional long short-term memory unit of the anomaly detection network, which has been fine-tuned and has improved performance, are updated in all online operational model instances. This ensures that these instances can use the latest parameters for operational monitoring, improving the overall monitoring accuracy and reliability.

[0143] As one implementation method, in step S500, the network parameters of the running monitoring model are dynamically adjusted based on the model update instruction, which may include the following steps S540~S5100: Step S540: Receive the model update instruction and parse the associated anomaly confidence interval type to determine the target network layer to be adjusted in the running monitoring model; the target network layer includes the convolution kernel parameters of the first K layers of convolution channels in the state evaluation network and the weight parameters of the last L layers of bidirectional long short-term memory units in the anomaly detection network.

[0144] The model update command is issued when the operating status of the intelligent ring main unit changes and the original operation monitoring model can no longer adapt well to the new situation. It is used to trigger the update of the network parameters of the operation monitoring model. The anomaly confidence interval type is a type of information associated with the model update command. It reflects the different types of possible anomalies in the intelligent ring main unit. Different anomaly confidence interval types may correspond to different network layers that need to be adjusted.

[0145] The operational monitoring model comprises a state evaluation network and an anomaly detection network. In the state evaluation network, the kernel parameters of the first K convolutional channels are the core parameters of the first K layers in a multi-layer convolutional network, determining the feature extraction capability of the convolutional layers from the input data. Similarly, in the anomaly detection network, the weight parameters of the last L bidirectional long short-term memory units are the parameters of the last L layers of bidirectional long short-term memory, significantly influencing the anomaly detection results.

[0146] The system receives model update instructions and parses the associated anomaly confidence interval types. It then determines the target network layer to be adjusted in the running monitoring model. Specifically, based on the model update instructions and anomaly confidence interval types, it determines that the kernel parameters of the first K layers of convolutional channels in the state evaluation network and the weight parameters of the last L layers of bidirectional long short-term memory units in the anomaly detection network need to be adjusted so that the model can better adapt to the new running state.

[0147] Step S550: Starting from the current running cycle, collect multimodal running data within a preset time window, and merge the collected data with the N sets of data at the end of the historical training set to generate an incremental training dataset.

[0148] The current operating cycle is, for example, the current operating cycle of the intelligent ring network cabinet. The preset time window is a pre-set time range. Multimodal operating data within the preset time window are collected back from the current operating cycle. That is, starting from the current moment, a certain time is traced back to collect data such as voltage fluctuation sequence, current harmonic distribution, temperature change curve and mechanical vibration spectrum during this period.

[0149] The historical training set is the collection of training data used during model training. For example, the last N data sets in the historical training set are the final N data sets. Merging the newly collected data with the last N data sets in the historical training set, that is, combining the newly collected data with a portion of the historical training set, forms a new dataset.

[0150] The incremental training dataset is a dataset used for incremental training of the operation monitoring model. It is generated by collecting multimodal operation data within a preset time window from the current operation cycle and merging the collected data with the N sets of data at the end of the historical training set. This provides richer and more up-to-date data for the incremental training of the model, enabling the model to better adapt to changes in the operating status of the intelligent ring network cabinet.

[0151] Step S560: Locally train the target network layers on the incremental training dataset, wherein the parameters of the remaining convolutional channels in the fixed-state evaluation network are excluding the first K layers, and the weight parameters of the bidirectional long short-term memory units in the last L layers of the anomaly detection network are unlocked.

[0152] The incremental training dataset is generated by merging newly collected data with the last N sets of data from the historical training set. It contains the latest multimodal operational data and can reflect the latest operational status of the intelligent ring network cabinet. The target network layer is the network layer that needs to be adjusted, determined according to the model update instructions and the anomaly confidence interval type. This includes the convolution kernel parameters of the first K layers of convolutional channels in the state evaluation network and the weight parameters of the last L layers of bidirectional long short-term memory units in the anomaly detection network.

[0153] Local training of the target network layers is performed on the incremental training dataset, meaning the parameters of the target network layers are trained and adjusted using the incremental training dataset. Fixing the parameters of the remaining convolutional channels in the evaluation network (excluding the first K layers) is to maintain the model's ability to extract basic features and avoid unnecessary impact on the parameters of other layers during training. Unlocking the weight parameters of the bidirectional long short-term memory units in the last L layers of the anomaly detection network allows the parameters of these layers to be adjusted according to new data, enabling the model to better adapt to new operating conditions.

[0154] By locally training the target network layers on the incremental training dataset, while fixing the parameters of the remaining convolutional channels in the state evaluation network except for the first K layers, and unlocking the weight parameters of the bidirectional long short-term memory units in the last L layers of the anomaly detection network, the model can be adjusted according to new data without destroying the original structure and performance of the model, thereby improving the model's adaptability and accuracy.

[0155] Step S570: Extract a local training set from the incremental training dataset using the sliding window method, and input the steady-state feature vectors in the local training set into the first K layers of convolution channels to output the intermediate feature tensor.

[0156] The sliding window method uses a fixed-size window to slide across a data sequence, continuously selecting new data for training. The incremental training dataset contains the latest multimodal operating data. The sliding window method extracts a local training set from the incremental training dataset, that is, selects a portion of the data as a local training set for local training of the model.

[0157] The steady-state feature vectors in the local training set are obtained after feature extraction and processing, reflecting the characteristic information of the intelligent ring network cabinet under stable operating conditions. The first K convolutional channels are the first K convolutional channels in the state evaluation network, and the kernel parameters of these layers are one of the target parameters that need to be adjusted. The steady-state feature vectors from the local training set are input into the first K convolutional channels, which perform convolution operations on the steady-state feature vectors, extracting the feature information and outputting intermediate feature tensors.

[0158] The intermediate feature tensor is the result obtained after processing the steady-state feature vector by the first K convolutional channels. It contains the feature information extracted by the first K convolutional channels, providing important basic data for subsequent anomaly detection.

[0159] Step S580: Input the intermediate feature tensor and the transient feature matrix in the local training set into the L-layer bidirectional long short-term memory unit, calculate the prediction error gradient of the L-layer bidirectional long short-term memory unit, and update the weight parameters of the L-layer based on the prediction error gradient.

[0160] The intermediate feature tensor is the result obtained after processing the steady-state feature vectors in the local training set through the first K convolutional channels, and it contains the feature information extracted by the first K convolutional channels. The transient feature matrix in the local training set is obtained after feature extraction and processing, and it reflects the feature information of the intelligent ring network cabinet during the transient process.

[0161] The L-layer bidirectional long short-term memory (BSS) units are the core components of an anomaly detection network. The weight parameters of these layers are among the target parameters that need to be adjusted. The intermediate feature tensor and the transient feature matrix from the local training set are input into the L-layer BSS units. These units analyze and process the input features and output prediction results. The prediction error gradient, for example, is the gradient of the error between the predicted result and the true value over the weight parameters, reflecting the direction and magnitude of the weight parameter adjustment. Calculating the prediction error gradient of the L-layer BSS units involves calculating the gradient of the weight parameters based on the error between the predicted result and the true value. Updating the weight parameters of the L-layer BSS units based on the prediction error gradient further adjusts the weight parameters, making the model's predictions closer to the true values ​​and improving the model's performance and accuracy.

[0162] Step S590: After completing local training, the updated weight parameters of the last L layers are recombined with the unadjusted convolutional kernel parameters of the first K layers to generate the adjusted running monitoring model, and the mean absolute percentage error of the adjusted running monitoring model is calculated through the validation set.

[0163] After completing the local training of the target network layers, the weight parameters of the bidirectional long short-term memory units in the last L layers have been updated according to the prediction error gradient, while the kernel parameters of the convolutional channels in the first K layers remain unchanged during the local training process.

[0164] The updated weight parameters of the last L layers are recombined with the unadjusted kernel parameters of the first K layers. In other words, the updated weight parameters of the bidirectional long short-term memory units of the last L layers are combined with the kernel parameters of the unadjusted convolutional channels of the first K layers to form a new operation monitoring model, namely the adjusted operation monitoring model.

[0165] The validation set is a dataset used to validate the performance of the trained model. The mean absolute percentage error (MASE) of the adjusted monitoring model is calculated using the validation set. In other words, the adjusted monitoring model is tested using the validation set, and the average absolute percentage error between its predictions and the true values ​​is calculated. The mean absolute percentage error is an important metric for measuring the accuracy of model predictions; by calculating the mean absolute percentage error, the performance of the adjusted monitoring model can be evaluated.

[0166] Step S5100: If the mean absolute percentage error is lower than the preset error threshold, the adjusted parameters of the running monitoring model will be synchronized to all online instances; if it exceeds the preset error threshold, the parameters will be rolled back to the state before adjustment and the collection range of the preset time window will be expanded before regenerating the incremental training dataset.

[0167] Mean absolute percentage error is an indicator for measuring the prediction accuracy of the adjusted operation monitoring model. The preset error threshold is a pre-set error threshold used to determine whether the performance of the adjusted operation monitoring model meets the requirements.

[0168] If the mean absolute percentage error is lower than the preset error threshold, it indicates that the performance of the adjusted operation monitoring model has been improved, enabling more accurate monitoring and evaluation of the intelligent ring network cabinet's operating status. At this point, the parameters of the adjusted operation monitoring model are synchronized to all online instances. This means updating the parameters of the adjusted and improved operation monitoring model to all online running model instances, allowing these instances to use the latest parameters for operation monitoring, thereby improving the overall monitoring accuracy and reliability.

[0169] If the mean absolute percentage error exceeds the preset error threshold, it indicates that the performance of the adjusted monitoring model has not been effectively improved, possibly due to insufficient quality or quantity of the incremental training dataset. In this case, roll back to the parameter state before adjustment, restoring the parameters of the monitoring model to their pre-adjustment state to ensure that the model's performance is not affected. Expand the collection range of the preset time window and regenerate the incremental training dataset, i.e., increase the time range of data collection to obtain more multimodal operational data, and regenerate the incremental training dataset for further local training and adjustment to improve model performance.

[0170] As an optional derivative implementation, the method provided by the present invention may further include the following steps S600~S1100: Step S600: During the decommissioning stage of the intelligent ring network cabinet, collect historical full-cycle status parameters and corresponding equipment fault record tables, and divide the historical full-cycle status parameters into initial operating parameters, stable operating parameters and degradation critical parameters according to the operating stage.

[0171] The decommissioning phase of an intelligent ring main unit (RNB) refers to the stage when it reaches the end of its service life or needs to be taken out of service for other reasons. Historical full-cycle state parameters include the RNB's state parameters throughout its entire service life, such as steady-state characteristic vectors, transient characteristic matrices, and cumulative degradation factors. These parameters record the RNB's state information at different operational stages. The equipment fault log table records fault information that occurs during the operation of the intelligent RNB, including detailed information such as the time and type of fault.

[0172] Initial operating parameters, for example, are the full-cycle status parameters of a smart ring main unit when it is first put into use, reflecting the performance and characteristics of the equipment in its initial state. Stable operating parameters, for example, are the full-cycle status parameters of a smart ring main unit during its normal and stable operation phase, during which the equipment's performance is relatively stable. Degradation critical parameters, for example, are the full-cycle status parameters of a smart ring main unit when it is nearing decommissioning, and its performance begins to show significant degradation, indicating that the equipment is about to enter an unusable state.

[0173] During the decommissioning phase of intelligent ring main units, collecting historical full-cycle status parameters and corresponding equipment fault record tables, and dividing the historical full-cycle status parameters into initial operating parameters, stable operating parameters, and degradation critical parameters according to the operating stage, can provide an important data foundation for subsequent life prediction and fault analysis during the decommissioning phase.

[0174] Step S700: Based on the difference between the initial operating parameters and the stable operating parameters, a baseline degradation trajectory is generated, and key degradation inflection points are marked on the baseline degradation trajectory by superimposing abnormal fluctuation events in the degradation critical parameters.

[0175] Initial operating parameters are the full-cycle state parameters of the intelligent ring main unit when it is first put into use, while stable operating parameters are the full-cycle state parameters of the intelligent ring main unit during its normal and stable operation phase. The difference between the initial operating parameters and the stable operating parameters reflects the performance changes of the intelligent ring main unit from its initial state to its stable operating state.

[0176] The baseline degradation trajectory is a trajectory that reflects the performance degradation trend of the intelligent ring main unit, generated based on the difference between the initial operating parameters and the stable operating parameters. It can be used as a reference standard to evaluate the degradation of the intelligent ring main unit in subsequent operation.

[0177] Degradation critical parameters are the full-cycle state parameters of an intelligent ring main unit when its performance begins to degrade significantly as it approaches retirement. Abnormal fluctuation events, such as unusual performance fluctuations occurring during the degradation critical stage, are included. By overlaying these abnormal fluctuation events from the degradation critical parameters, key degradation inflection points are marked on the baseline degradation trajectory. This means marking the time points and characteristic information corresponding to these abnormal fluctuation events on the baseline degradation trajectory. These key degradation inflection points can help more accurately identify the critical stages and characteristics of intelligent ring main unit performance degradation, providing important basis for subsequent lifespan prediction and fault analysis.

[0178] Step S800: Associate and map key degradation inflection points with fault types in the equipment fault record table to generate a labeled degradation trajectory map, wherein each degradation inflection point is associated with at least one fault type and the corresponding fault occurrence timestamp.

[0179] Critical degradation inflection points are points marked on the baseline degradation trajectory that reflect key stages of performance degradation in the intelligent ring main unit. The equipment fault record table is a table that records fault information occurring during the operation of the intelligent ring main unit, including detailed information such as the time and type of the fault. Association mapping, for example, establishes a correspondence between critical degradation inflection points and fault types in the equipment fault record table. This method clarifies the types of faults that each critical degradation inflection point may trigger.

[0180] The labeled degradation trajectory map is obtained by annotating key degradation inflection points with fault types and corresponding fault occurrence timestamps based on the baseline degradation trajectory. Each degradation inflection point is associated with at least one fault type and its corresponding fault occurrence timestamp, which makes it possible to intuitively see the types of faults that may occur at different stages of intelligent ring network cabinet performance degradation and the approximate time of fault occurrence.

[0181] Step S900: Input the labeled degradation trajectory map into the pre-trained transfer learning framework, perform structural pruning on the anomaly detection network of the running monitoring model, and retain only the network branches related to the most frequent fault types in the degradation trajectory map to form a decommissioning stage life prediction model.

[0182] Pre-trained transfer learning frameworks are machine learning frameworks that have been pre-trained and can quickly adapt to new tasks by leveraging knowledge and features learned from other related tasks. Labeled degradation trajectory maps contain important information such as fault types and timestamps corresponding to different performance degradation stages of intelligent ring main units.

[0183] Labeled degradation trajectory maps are input into a pre-trained transfer learning framework, which can analyze and learn from this data to extract key features and patterns. The anomaly detection network of the monitoring model is used to detect abnormalities in the intelligent ring network cabinet. Structural pruning, for example, optimizes the structure of the anomaly detection network by removing unnecessary network branches and parameters to improve the model's efficiency and performance.

[0184] By retaining only network branches associated with the most frequent fault types in the degradation trajectory map, the anomaly detection network can focus more intently on detecting the most common fault types. During the decommissioning phase of intelligent ring main units (RMUs), this approach allows the model to make more targeted lifetime predictions. After structural pruning, the resulting decommissioning lifetime prediction model can more accurately predict the remaining lifetime and potential major fault types of the RMU based on its current operating status and performance degradation, providing more precise decision-making support for the decommissioning management and maintenance of RMUs.

[0185] Step S1000: The degradation stage is matched with the real-time collected multimodal operation data through the retirement stage life prediction model. If the similarity between the current real-time steady-state feature vector and the feature before a certain degradation inflection point in the degradation trajectory map exceeds the preset matching threshold, the corresponding fault occurrence timestamp is extracted from the equipment fault record table, and the remaining life estimate is calculated based on the time difference between the current timestamp and the fault occurrence timestamp.

[0186] The real-time multimodal operating data is collected in real time by operating data acquisition nodes deployed at each electrical node of the intelligent ring main unit, including voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra. The decommissioning life prediction model is a model specifically designed for predicting the life of intelligent ring main units during the decommissioning stage, after structural pruning.

[0187] Degradation stage matching involves comparing and matching real-time acquired multimodal operating data with different degradation stages in a labeled degradation trajectory map to determine the current degradation stage of the intelligent ring main unit. The current real-time steady-state feature vector is extracted from the real-time acquired multimodal operating data and reflects the steady-state operating characteristics of the intelligent ring main unit at the current moment.

[0188] The preset matching threshold is a pre-set threshold used to determine whether the feature similarity is high enough. If the similarity between the current real-time steady-state feature vector and the features before a certain degradation inflection point in the degradation trajectory map exceeds the preset matching threshold, it means that the current operating state of the intelligent ring network cabinet is very similar to the state before that degradation inflection point, and it may soon face the fault corresponding to that degradation inflection point.

[0189] The equipment fault log records fault information that occurs during the operation of the intelligent ring main unit, including the timestamp of each fault occurrence. The corresponding fault occurrence timestamp is extracted from the equipment fault log, i.e., the fault occurrence time corresponding to the current degradation inflection point is found. The remaining lifespan estimate is calculated based on the time difference between the current timestamp and the fault occurrence timestamp. This time difference can be approximated as the time remaining before the corresponding fault occurs in the intelligent ring main unit from its current state, thus providing an important reference for the decommissioning and maintenance planning of the intelligent ring main unit.

[0190] Step S1100: Based on the ratio between the remaining life estimate and the preset safe operating cycle length, generate suggested replacement time points and corresponding spare parts scheduling instructions.

[0191] The remaining life estimate is the time remaining before the corresponding fault occurs for the intelligent ring main unit, calculated using a decommissioning phase life prediction model. The preset safe operating cycle length is a pre-set safe operating time standard that takes into account various factors such as the performance, reliability, and maintenance costs of the intelligent ring main unit.

[0192] Based on the ratio of the estimated remaining lifespan to the preset safe operating period, it can be determined whether the intelligent ring main unit can continue to operate within a safe range. If the estimated remaining lifespan is less than the preset safe operating period, it indicates that the intelligent ring main unit may soon fail and needs to be replaced promptly. In this case, based on the estimated remaining lifespan and the current time, a suggested replacement time is generated. This time can guide maintenance personnel to reasonably arrange the replacement plan for the intelligent ring main unit.

[0193] Spare parts dispatch instructions are generated based on the recommended replacement time points to dispatch the spare parts needed for replacing the intelligent ring main unit. They include information such as the type, quantity, dispatch time, and location of the spare parts. By generating recommended replacement time points and corresponding spare parts dispatch instructions, it can be ensured that suitable spare parts are available in a timely manner when the intelligent ring main unit needs replacement, reducing equipment downtime and improving the operational reliability and maintenance efficiency of the intelligent ring main unit.

[0194] As an optional derivative implementation, after triggering a warning signal or model update command that matches a preset threshold in step S500, the method may further include the following steps S5001~S5010: Step S5001: Determine the corresponding electrical node isolation strategy according to the type of the warning signal. The electrical node isolation strategy includes a circuit breaker operation command to disconnect the target electrical node and a power supply command to switch the backup circuit.

[0195] The types of warning signals reflect different degrees of anomalies in the intelligent ring main unit, and different types of warning signals correspond to different abnormal states of electrical nodes. The electrical node isolation strategy is a strategy designed to prevent the anomaly from escalating further and to ensure the safety of the intelligent ring main unit and related power systems.

[0196] The circuit breaker operation command to disconnect a target electrical node is, for example, used when a warning signal indicates a serious abnormality at the target electrical node. This involves operating the circuit breaker to isolate the electrical node from the power system, preventing the fault from affecting other parts. The command to switch to a backup power circuit is used when an abnormality occurs at the target electrical node. To ensure continuous power supply, the power line is switched to a backup circuit, ensuring that the relevant equipment can continue to operate normally.

[0197] Determining the appropriate electrical node isolation strategy based on the type of warning signal allows for the implementation of the most suitable isolation measures for different abnormal situations, ensuring the safety of the power system while minimizing the impact on normal power supply. For example, for a Level 1 warning signal, it may be necessary to immediately disconnect the circuit breaker of the target electrical node and quickly switch to backup power supply; while for a Level 2 warning signal, it may only require local inspection and handling first, followed by appropriate isolation operations if necessary.

[0198] Step S5002: Generate a maintenance work order based on the electrical node isolation strategy. The maintenance work order includes a fault location identifier, a list of required spare parts, and a repair operation sequence. The maintenance work order is then pushed to the preset operation and maintenance management platform.

[0199] Maintenance work orders are important documents used to guide maintenance personnel in performing equipment maintenance and repair work. Maintenance work orders are generated based on electrical node isolation strategies, meaning that detailed maintenance plans and operating procedures are developed according to the determined electrical node isolation strategies.

[0200] Fault location markers clearly indicate the specific electrical node location where a fault has occurred in the intelligent ring main unit. This helps maintenance personnel quickly locate the fault and improves maintenance efficiency. A spare parts list details all necessary spare parts for repairing the faulty electrical node, including the name, model, and quantity of each part, ensuring sufficient spare parts are available during maintenance. A maintenance operation sequence outlines the logical and sequential maintenance steps, guiding maintenance personnel through the troubleshooting and repair process step by step.

[0201] Step S5003: Monitor the residual current and temperature change rate of the target electrical node after isolation in real time. If the residual current exceeds the safety threshold or the temperature change rate shows an exponential upward trend, trigger the emergency power-off command and activate the fire-fighting linkage device.

[0202] Even after isolation, when a target electrical node is disconnected from the main part of the power system, some residual current may still remain. Residual current, for example, is the current that remains in the circuit after electrical equipment has been disconnected from the power supply, due to the effects of components such as capacitors and inductors. The safety threshold is a pre-set maximum allowable residual current value. When the residual current exceeds the safety threshold, it may pose a safety hazard to equipment and personnel.

[0203] The rate of temperature change is, for example, how quickly the temperature of a target electrical node changes over time. If the rate of temperature change increases exponentially, it indicates that the target electrical node may have serious faults such as short circuits or overloads, leading to a rapid increase in temperature and potentially causing safety accidents such as fires.

[0204] Step S5004: After the fire alarm linkage device is activated, collect multimodal operating data and environmental parameters of the target electrical node before and after the fault, generate a fault analysis report and associate it with the work order number of the maintenance work order.

[0205] The activation of the fire alarm linkage device indicates a serious fault in the target electrical node. Collecting multimodal operating data of the target electrical node before and after the fault, including voltage fluctuation sequences, current harmonic distribution, temperature change curves, and mechanical vibration spectra, can reflect the changes in the operating state of the electrical node before and after the fault, helping to analyze the cause of the fault.

[0206] Environmental parameters include the environmental conditions of the target electrical node, such as temperature, humidity, and air pressure. Environmental factors may affect the performance of electrical equipment, so collecting environmental parameters is also valuable for fault analysis.

[0207] Step S5005: Input the fault analysis report into the incremental training dataset of the operation monitoring model, update the weight parameters related to the target electrical node in the operation monitoring model, and synchronize the updated parameters to all online instances.

[0208] The fault analysis report contains detailed information about the target electrical node fault, such as the cause and characteristics of the fault. Inputting the fault analysis report into the incremental training dataset of the operational monitoring model—that is, adding this new fault information to the training dataset—allows the model to learn more about the characteristics and patterns of the fault situation.

[0209] In the operational monitoring model, the weight parameters related to the target electrical node determine the model's ability to judge and analyze the operating status of the target electrical node. By updating these weight parameters, the model can better adapt to new situations of the target electrical node, improving the accuracy of fault detection and prediction.

[0210] As one implementation method, after synchronizing the updated parameters to all online instances in step S5005, the method may further include: step S5006: obtaining the completion status of the maintenance work order from the operation and maintenance management platform, and when the completion status is "maintenance successful", extracting the calibration operation data of the target electrical node after maintenance.

[0211] The operations and maintenance management platform is responsible for the unified management and scheduling of maintenance work orders, and records the execution and completion status of these work orders. Obtaining the completion status of maintenance work orders from the operations and maintenance management platform allows for timely understanding of the repair progress and results of target electrical nodes.

[0212] When the completion status is "Repair Successful," it indicates that the target electrical node has been repaired and restored to normal operation. At this point, calibration operation data of the target electrical node after repair is extracted. This calibration operation data includes, for example, multi-modal operation data of the target electrical node under normal operating conditions after repair, including voltage fluctuation sequences, current harmonic distribution, temperature change curves, and mechanical vibration spectra. This data reflects the actual operating performance of the target electrical node after repair, providing a foundation for subsequent performance evaluation and model validation.

[0213] Step S5007: Compare the calibration operation data with the pre-fault operation data in the fault analysis report, and calculate the recovery rate of the steady-state characteristic vector and the fluctuation attenuation coefficient of the transient characteristic matrix.

[0214] The calibration operation data is the actual operating data of the target electrical node after maintenance, while the pre-fault operation data in the fault analysis report is the operating data of the target electrical node before the fault occurred. Comparing the two allows for a clear view of the impact of maintenance on the operating status of the target electrical node.

[0215] Steady-state eigenvectors are vectors extracted from operational data that reflect the characteristic information of electrical nodes under stable operating conditions. The recovery rate of a steady-state eigenvector is, for example, the similarity between the steady-state eigenvector after maintenance and the steady-state eigenvector before the fault, which can be measured by some similarity index, such as cosine similarity. A higher recovery rate indicates a better recovery of the steady-state operating performance of the target electrical node after maintenance. The transient characteristic matrix is ​​a matrix that reflects the characteristic information of an electrical node during transient processes. The fluctuation attenuation coefficient of the transient characteristic matrix is, for example, the degree of attenuation of the fluctuation amplitude in the transient characteristic matrix after maintenance relative to the pre-fault state, which can be calculated by comparing the fluctuation range or standard deviation of the two. A larger fluctuation attenuation coefficient indicates improved stability of the target electrical node during transient processes after maintenance. By calculating the recovery rate of the steady-state eigenvector and the fluctuation attenuation coefficient of the transient characteristic matrix, the maintenance effect can be objectively evaluated, providing a basis for judging whether the target electrical node has fully recovered to normal operation.

[0216] Step S5008: If the recovery rate is lower than the preset recovery threshold or the fluctuation attenuation coefficient exceeds the preset attenuation threshold, a secondary maintenance instruction is generated and the maintenance work order is re-triggered.

[0217] The preset recovery threshold is a pre-set threshold used to determine whether the recovery degree of the steady-state eigenvector is acceptable, and the preset decay threshold is a pre-set threshold used to determine whether the decay degree of the transient eigenma fluctuation is acceptable.

[0218] If the recovery rate is lower than the preset recovery threshold, it indicates that the steady-state performance of the target electrical node after maintenance has not reached the expected recovery level, and there may still be some potential problems. If the fluctuation attenuation coefficient exceeds the preset attenuation threshold, it indicates that the stability of the target electrical node during transient processes has not been effectively improved after maintenance, and there may also be potential faults.

[0219] In this scenario, a secondary maintenance instruction is generated, explicitly requiring a second maintenance of the target electrical node to thoroughly resolve any potential issues. The maintenance work order is then re-triggered, sending a new work order to the operations and maintenance management platform, where appropriate maintenance personnel are assigned to perform the secondary maintenance. By promptly identifying and addressing any remaining issues after the initial maintenance, the target electrical node can be fully restored to normal operation, improving the reliability and stability of the intelligent ring main unit.

[0220] Optionally, in step S5008, after generating the secondary maintenance instruction and re-triggering the push of the maintenance work order, the method may also include the following steps S50081~S50085: Step S50081: Collect the spare parts model and operation log replaced during the secondary maintenance process, match the spare parts model with the historical spare parts replacement records, and filter out the installation locations of other spare parts in the same batch.

[0221] During secondary maintenance, some spare parts may be replaced to resolve issues at the target electrical nodes. The model numbers of the replaced spare parts and operation logs are collected during the secondary maintenance. The spare part model numbers clearly identify the specific spare parts replaced, and the operation logs record the detailed process of the secondary maintenance, including the time of replacement and the operating steps.

[0222] Historical spare parts replacement records are information on spare parts replacements recorded during previous maintenance of the intelligent ring main unit, including spare parts model, replacement time, and installation location. Matching spare parts models with historical replacement records identifies instances where the same spare parts model was used in the past. This matching process allows for filtering out the installation locations of other spare parts from the same batch.

[0223] Spare parts from the same batch may have the same quality problems or potential faults. By identifying the installation locations of other spare parts from the same batch, these spare parts can be given special attention and preventive monitoring. This can help detect potential problems in advance, prevent similar faults from recurring at other electrical nodes, and improve the overall reliability of the intelligent ring main unit.

[0224] Step S50082: Based on the installation location of spare parts in the same batch, conduct preventive operation monitoring on the associated electrical nodes, focusing on analyzing the similarity of harmonic components in the steady-state characteristic vector and the consistency of the trend of temperature change curves.

[0225] Based on the installation locations of spare parts from the same batch, the associated electrical nodes requiring preventative operation monitoring were identified. Preventative operation monitoring involves monitoring the operational status of equipment before any obvious faults occur, thereby identifying potential malfunctions in advance.

[0226] The similarity of harmonic components in a steady-state eigenvector is, for example, the degree of similarity of harmonic components in the steady-state eigenvectors of associated electrical nodes. Anomalies in harmonic components can lead to problems such as equipment overheating and increased losses. By analyzing the similarity of harmonic components, it can be determined whether similar harmonic problems exist in associated electrical nodes.

[0227] The consistency of temperature change curves refers to the degree of similarity in the trends of temperature change curves at associated electrical nodes. Temperature is an important indicator reflecting the operating status of electrical equipment, and anomalies in temperature change curves may indicate equipment malfunctions. Focusing on analyzing the consistency of temperature change curve trends can help detect potential temperature anomalies at associated electrical nodes in a timely manner.

[0228] By conducting preventative operational monitoring of associated electrical nodes, focusing on analyzing the similarity of harmonic components in the steady-state characteristic vector and the consistency of the temperature change curve trend, potential fault hazards that may be caused by spare parts from the same batch can be identified in advance, and corresponding measures can be taken to handle them, ensuring the safe and stable operation of the intelligent ring main unit.

[0229] Step S50083: If the similarity of harmonic components exceeds the preset similarity threshold or the correlation coefficient of trend consistency is lower than the preset correlation threshold, a batch quality warning is generated and the batch spare parts are marked as high-risk spare parts.

[0230] The preset similarity threshold is a pre-set threshold used to determine whether the similarity of harmonic components is too high, and the preset correlation threshold is a pre-set threshold used to determine whether the correlation coefficient of the temperature change curve trend consistency is too low.

[0231] If the similarity of harmonic components exceeds a preset similarity threshold, it indicates that the harmonic components of the associated electrical nodes are very similar, which may be due to quality problems in the same batch of spare parts causing harmonic anomalies. If the correlation coefficient of trend consistency is lower than a preset correlation threshold, it indicates that the temperature change curves of the associated electrical nodes have large differences in trend, which may also suggest potential faults in the same batch of spare parts.

[0232] Step S50084: Update the spare parts list of the maintenance work order according to the high-risk spare parts mark, force the replacement of spare parts of the same batch in all associated electrical nodes, and archive the replacement spare parts operation data to the incremental training dataset.

[0233] The spare parts list of the maintenance work order is updated based on the high-risk spare parts marking. That is, the maintenance work order clearly lists the information of the high-risk spare parts of the same batch that need to be replaced, so as to ensure that these spare parts can be accurately replaced during the maintenance process.

[0234] Forced replacement of spare parts from the same batch in all associated electrical nodes is to eliminate potential fault hazards caused by spare parts from the same batch, ensuring the safe and stable operation of the intelligent ring main unit. Timely replacement of high-risk spare parts can prevent similar faults from recurring in other electrical nodes, improving equipment reliability. The operational data of the replaced spare parts is archived into an incremental training dataset. This data includes multimodal operational data of the new spare parts during operation, such as voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra. Archived into the incremental training dataset, this allows the operational monitoring model to learn the operational characteristics of the new spare parts, improving the model's accuracy in monitoring and predicting equipment operating status, and better adapting to equipment changes.

[0235] Step S50085: Enhance the anomaly detection network in the operation monitoring model by using archived data, optimize its sensitivity to predicting failures of high-risk spare parts, and deploy the optimized model parameters to all online instances in real time.

[0236] Archived data, such as operational data of replaced spare parts, contains the operational characteristics and information of the new spare parts. The anomaly detection network in the operational monitoring model is reinforced using this archived data; that is, the anomaly detection network is additionally trained using this new data, enabling the model to better identify potential fault characteristics of high-risk spare parts.

[0237] Optimize its sensitivity to predicting failures of high-risk spare parts, that is, through enhanced training, enable the anomaly detection network to more accurately predict the failure status of high-risk spare parts, issue early warning signals in advance, and reduce the impact of failures on intelligent ring network cabinets.

[0238] Deploying optimized model parameters to all online instances in real time ensures that all running monitoring models can use the latest, optimized parameters for monitoring and evaluation. This guarantees the consistency and accuracy of the entire intelligent ring main unit monitoring system, enables timely detection and handling of potential faults in high-risk spare parts, and improves the operational reliability and safety of the intelligent ring main unit.

[0239] Step S5009: When the calibration operation data meets the preset recovery conditions, release the isolation state of the target electrical node and generate a validation set of the operation monitoring model based on the calibration operation data.

[0240] Calibration operation data refers to multimodal operation data collected after maintenance of the target electrical node, reflecting its actual operating performance. This data includes voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra. Preset recovery conditions are a series of pre-defined standards used to determine whether the target electrical node has fully recovered to normal operating status. These conditions may be determined based on indicators such as the recovery rate of the steady-state eigenvector and the fluctuation attenuation coefficient of the transient eigenma matrix.

[0241] When the calibration operation data meets the preset recovery conditions, it indicates that the target electrical node has recovered to a state where it can be normally connected to the power system. At this point, the isolation state of the target electrical node is lifted, allowing it to re-participate in the operation of the power system. A validation set for the operation monitoring model is generated based on the calibration operation data. This validation set is a dataset used to verify the performance of the operation monitoring model. By organizing and filtering the calibration operation data to form a dataset suitable for model validation, the operation monitoring model can be tested using this validation set. This evaluates the model's ability to monitor and predict the operating status of the target electrical node after maintenance, providing a basis for subsequent model optimization.

[0242] Step S5010: Evaluate the performance of the synchronized running monitoring model using the validation set. If the health score deviation in the evaluation result exceeds the preset deviation threshold, roll back to the model parameters of the previous version and re-trigger incremental training.

[0243] The synchronized operational monitoring model is, for example, an operational monitoring model that has already synchronized the updated parameters to all online instances. Its performance is evaluated using a validation set, which involves testing the model with a validation set generated based on calibration operational data. The model then analyzes and predicts data from the validation set to obtain corresponding health scores and other results.

[0244] Health score deviation refers to the degree of discrepancy between the health score predicted by the model and the actual situation in the validation set. The preset deviation threshold is a pre-defined allowable deviation range used to determine whether the model's prediction accuracy meets the requirements. If the health score deviation in the evaluation results exceeds the preset deviation threshold, it indicates that the synchronized operation monitoring model has significant errors in monitoring and predicting the operating status of the target electrical nodes after maintenance, and the model's performance may be affected.

[0245] Roll back to the previous version of the model parameters, restoring the model's parameters to their state before the update, to ensure the model can return to its previous level of accurate prediction. Re-trigger incremental training, using an incremental training dataset containing new data to train the model again, allowing the model to learn and adjust based on the new data, thereby improving the model's adaptability and prediction accuracy to the operating status of the target electrical nodes after maintenance, ensuring the model can continuously and accurately monitor the operating status of the intelligent ring main unit.

[0246] This invention provides an operational data monitoring system, such as... Figure 2 As shown, the operational data monitoring system 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the operational data monitoring system 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one, and the structure of this operational data monitoring system 100 does not constitute a limitation on the embodiments of the present invention.

[0247] This invention provides a runtime data monitoring system, which includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processors, they implement the method provided in the above embodiments.

Claims

1. A method for monitoring the full-cycle operation data of intelligent ring network cabinets, characterized in that, The method includes: real-time acquisition of multimodal operating data throughout the entire cycle via operating data acquisition nodes deployed at each electrical node of the intelligent ring main unit; the multimodal operating data includes voltage fluctuation sequences, current harmonic distributions, temperature change curves, and mechanical vibration spectra; segmenting and aligning the multimodal operating data based on preset time-series feature extraction rules to generate full-cycle state parameters corresponding to each electrical node; the full-cycle state parameters include steady-state feature vectors, transient feature matrices, and cumulative degradation factors; training an operating monitoring model adapted to the intelligent ring main unit based on the full-cycle state parameters; the operating monitoring model includes a state assessment network and an anomaly detection network, and the output of the state assessment network and the input of the anomaly detection network are connected through a dynamic weight allocation layer; the operating monitoring model performs layer-by-layer analysis on the real-time acquired multimodal operating data to generate a health score and anomaly confidence interval corresponding to the current operating stage; triggering an early warning signal or model update command matching a preset threshold based on the health score and anomaly confidence interval, and dynamically adjusting the network parameters of the operating monitoring model based on the model update command.

2. The method according to claim 1, characterized in that, The process involves segmenting and aligning the multimodal operating data based on preset time-series feature extraction rules to generate full-cycle state parameters corresponding to each electrical node. This includes: extracting periodic amplitude offsets and phase abrupt change points from the voltage fluctuation sequence, and dividing the voltage segment intervals according to the phase abrupt change points; stripping the fundamental component from the current harmonic distribution to obtain the energy proportion and cumulative distortion of each harmonic, and filtering key harmonic components based on preset harmonic energy thresholds; calculating the gradient rise rate and thermal equilibrium stabilization time based on the temperature change curve, and determining the mechanical loosening index and contact resistance degradation trend by combining the frequency domain energy distribution of the mechanical vibration spectrum; fusing the periodic amplitude offset, key harmonic components, mechanical loosening index, and contact resistance degradation trend into the steady-state feature vector; constructing the transient feature matrix based on the maximum instantaneous deviation within the voltage segment intervals and the thermal equilibrium stabilization time, and generating the cumulative degradation factor by weighted summation of the cumulative distortion.

3. The method according to claim 2, characterized in that, The step of fusing the periodic amplitude offset, key harmonic components, mechanical loosening index, and contact resistance degradation trend into the steady-state feature vector includes: normalizing the periodic amplitude offset to obtain a standardized voltage offset coefficient; calculating the harmonic influence weight based on the energy proportion and cumulative distortion of the key harmonic components, and weighting and superimposing the key harmonic components based on the harmonic influence weight to obtain a weighted key harmonic component; converting the mechanical loosening index into a first health score through a preset mechanical health mapping table, and mapping the contact resistance degradation trend into a second health score using linear interpolation; and concatenating the standardized voltage offset coefficient, the weighted key harmonic component, the first health score, and the second health score into the steady-state feature vector according to a preset dimension.

4. The method according to claim 1, characterized in that, The step of training an operation monitoring model adapted to the intelligent ring network cabinet based on the full-cycle state parameters includes: dividing the full-cycle state parameters into a training set and a validation set, and injecting noise to enhance the steady-state feature vectors and transient feature matrices in the training set; constructing a multi-layer convolutional channel of the state assessment network, wherein the first convolutional layer is used to extract the spatial correlation of the steady-state feature vectors, and the second convolutional layer is used to capture the temporal dependence of the transient feature matrix; setting a bidirectional long short-term memory unit in the anomaly detection network to generate a time-series prediction result of the health score based on the fusion features output by the dynamic weight allocation layer; calculating the mean squared error loss function by comparing the time-series prediction result with the cumulative degradation factor in the validation set, and updating the weight parameters of the multi-layer convolutional channel and the bidirectional long short-term memory unit using an adaptive momentum optimization algorithm; freezing the parameters of the state assessment network and the anomaly detection network when the mean squared error loss function is lower than a preset convergence threshold, thus completing the training of the operation monitoring model.

5. The method according to claim 4, characterized in that, The step of calculating the mean squared error loss function by comparing the time-series prediction results with the cumulative degradation factors in the validation set includes: extracting the ground truth sequence of the cumulative degradation factors corresponding to the current training batch from the validation set; obtaining the health score prediction sequence output by the anomaly detection network and trend-aligning the health score prediction sequence using exponential smoothing; calculating the sum of squared point-by-point differences between the trend-aligned health score prediction sequence and the ground truth sequence of the cumulative degradation factors; determining the mean squared error loss value of the current training batch based on the sum of squared point-by-point differences and the total number of samples in the validation set; and accumulating the mean squared error loss values ​​of all training batches to obtain the mean squared error loss function.

6. The method according to claim 1, characterized in that, The step of analyzing the real-time multimodal operation data layer by layer through the operation monitoring model to generate a health score and anomaly confidence interval corresponding to the current operation stage includes: inputting the real-time collected voltage fluctuation sequence and current harmonic distribution into the first convolutional layer of the state assessment network to extract the real-time steady-state feature vector; inputting the real-time temperature change curve and mechanical vibration spectrum into the second convolutional layer of the state assessment network to generate a real-time transient feature matrix; performing feature fusion on the real-time steady-state feature vector and the real-time transient feature matrix through the dynamic weight allocation layer to obtain an enhanced spatiotemporal feature tensor; inputting the enhanced spatiotemporal feature tensor into the bidirectional long short-term memory unit of the anomaly detection network to output a health score sequence for the next N operation cycles, where N > 2; and calculating the anomaly confidence interval based on the variance and maximum-minimum range of the health score sequence.

7. The method according to claim 6, characterized in that, The step of fusing the real-time steady-state feature vector and the real-time transient feature matrix through the dynamic weight allocation layer to obtain the enhanced spatiotemporal feature tensor includes: expanding the dimension of the real-time steady-state feature vector to generate a first extended feature that matches the dimension of the real-time transient feature matrix; calculating the feature weights of each time step in the real-time transient feature matrix through an attention mechanism, and resampling the real-time transient feature matrix according to the feature weights; concatenating the first extended feature and the resampled real-time transient feature matrix according to the channel dimension to obtain an initial fused feature; reducing the dimension of the initial fused feature through a fully connected layer, and using a gated linear unit to perform nonlinear activation on the dimension-reduced feature to generate the enhanced spatiotemporal feature tensor.

8. The method according to claim 1, characterized in that, The step of triggering a warning signal or model update command matching a preset threshold based on the health score and abnormal confidence interval includes: generating a first-level warning signal and triggering an emergency shutdown command when the health score is lower than a first threshold and the upper limit of the abnormal confidence interval exceeds a second threshold; generating a second-level warning signal and triggering a local maintenance command when the health score is between the first and third thresholds and the width of the abnormal confidence interval exceeds a fourth threshold; and generating a model update command and starting an incremental learning mode when the rate of decline of the health score exceeds a fifth threshold for N consecutive operating cycles. The incremental learning mode includes: adding the multimodal operating data collected in the current cycle to the end of the training set, and... The weight parameters of the operation monitoring model are updated using a sliding window method. The step of initiating incremental learning includes: retaining the parameters of the first K convolutional channels of the state evaluation network in the operation monitoring model unchanged, and unlocking the parameters of the last L layers of the bidirectional long short-term memory units of the anomaly detection network; fine-tuning the parameters of the unlocked last L layers using a local training set, while fixing the parameters of the remaining layers; calculating the mean absolute percentage error (MASE) of the fine-tuned anomaly detection network on the validation set; if the MASE exceeds a preset fine-tuning threshold, rolling back to the parameter state before unlocking; and synchronizing the fine-tuned parameters to all online operation monitoring model instances when the MASE does not exceed the preset fine-tuning threshold.

9. The method according to claim 1, characterized in that, The method further includes: during the decommissioning phase of the intelligent ring main unit, collecting historical full-cycle status parameters and corresponding equipment fault record tables, and dividing the historical full-cycle status parameters into initial operating parameters, stable operating parameters, and degradation critical parameters according to the operating phase; generating a baseline degradation trajectory based on the difference between the initial operating parameters and the stable operating parameters, and marking key degradation inflection points on the baseline degradation trajectory by superimposing abnormal fluctuation events in the degradation critical parameters; associating and mapping the key degradation inflection points with the fault types in the equipment fault record table to generate a labeled degradation trajectory map, wherein each degradation inflection point is associated with at least one fault type and a corresponding fault occurrence timestamp; and inputting the labeled degradation trajectory map into a pre-training system. In the transfer learning framework, the anomaly detection network of the operation monitoring model is structurally pruned, retaining only the network branches related to the most frequent fault types in the degradation trajectory map, forming a decommissioning stage life prediction model. The decommissioning stage life prediction model is used to match the degradation stage of real-time collected multimodal operation data. If the similarity between the current real-time steady-state feature vector and the features before a certain degradation inflection point in the degradation trajectory map exceeds a preset matching threshold, the corresponding fault occurrence timestamp is extracted from the equipment fault record table, and the remaining life estimate is calculated based on the time difference between the current timestamp and the fault occurrence timestamp. Based on the ratio between the remaining life estimate and the preset safe operating cycle length, suggested replacement time points and corresponding spare parts scheduling instructions are generated.

10. A system for monitoring operational data, characterized in that, include: One or more processors; Memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, implement the method as described in any one of claims 1 to 9.

Images