Cableway full life cycle health management method and system based on multi-mode AI

Abstract

The invention relates to a cableway full-life-cycle health management method and system based on multi-mode AI, in particular to the field of cableway full-life-cycle health management, and realizes high-fidelity and high-efficiency monitoring of running states of cableway electrical components through synchronous sensing and self-adaptive acquisition of multi-physics field data. Based on multi-modal feature fusion and a deep self-encoding technology, a health index capable of sensitively reflecting the overall degradation state of a part is constructed, a hybrid prediction framework of a physical model and a data driving model is combined, residual life accurate prediction with probability significance is generated, and the residual life prediction accuracy is improved by means of a digital twinborn environment. Closed-loop online optimization of a prediction model, active perception of a new fault mode and automatic expansion of a knowledge base are realized, so that post-maintenance and regular overhaul of an operation and maintenance mode are improved to accurate predictive maintenance based on states, the safety guarantee level and the operation economic benefits are remarkably improved, and the method is suitable for large-scale popularization and application. And the expert experience is converted into digital assets which can be continuously accumulated and iterated.

Description

Technical Field

[0001] This invention relates to the field of cableway full life-cycle health management, and more specifically, to a cableway full life-cycle health management method and system based on multimodal AI. Background Technology

[0002] As a vital mode of transportation connecting areas with complex terrain, the safety and reliability of passenger ropeways are of paramount importance. The core drive unit, control system, and safety circuits of a ropeway system are composed of numerous sophisticated electrical components, such as frequency converters that power the entire system, programmable logic controller modules that serve as the control center, relays that control circuit switching, and various sensors that sense the system's status. During long-term continuous operation, these components inevitably age gradually due to factors such as electrical stress, thermal stress, and environmental stress, leading to a slow decline in performance. Even more serious challenges include fluctuations in grid voltage, changes in ambient temperature and humidity, and electromagnetic interference at the operating site. External factors can accelerate this aging process and may cause microscopic degradation of component parameters. In the early stages, this degradation often manifests as extremely subtle electrical characteristic anomalies, such as a slow decrease in the resistance of insulating materials, a slight increase in the contact resistance of electrical connection points, or a slight distortion in the control signal waveform. At this stage, the component function has not yet been lost, and the system can still maintain basic operation, but the hidden dangers have already been laid. However, most traditional monitoring systems can only monitor whether macroscopic operating parameters such as voltage and current are within the normal range. They are powerless to reveal this deep-seated, microscopic electrical characteristic degradation phenomenon, thus resulting in a lack of early warning for its latent faults.

[0003] Currently, the industry generally relies on periodic power outages for health status assessments of cableway electrical components. Under this model, maintenance personnel must manually measure and assess critical components using specialized offline instruments such as insulation resistance testers and LCR meters. This method is not only inefficient, severely impacting the continuous operation of the cableway, but more importantly, it is a post-event verification rather than a pre-event warning, failing to capture the dynamic performance changes of components under actual operating conditions. Although some advanced online monitoring systems have been implemented, their functions are mostly limited to setting simple voltage or current threshold alarms. This binary alarm mechanism only works when component failures have progressed to a considerable extent and parameters have become severely compromised. The technology only triggers after a certain time limit, failing to detect the gradual decline in component performance and thus being extremely insensitive to potential failure risks. This current state of technology leads to passive and outdated maintenance strategies, making it difficult to achieve a fundamental shift from periodic preventive maintenance to condition-based predictive maintenance. Therefore, there is an urgent need in the field for a non-invasive or low-invasive online electrical inspection technology that can be performed seamlessly during normal cableway operation without interrupting power supply. This technology should be able to continuously and automatically monitor the evolution of microscopic electrical parameters of key electrical components and, with the help of intelligent data analysis methods, accurately characterize the performance degradation trend of components and accurately predict early failures, thereby providing a solid basis for proactive and precise operation and maintenance decisions. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for cableway full lifecycle health management based on multimodal AI, thereby solving the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for full life-cycle health management of cableways based on multimodal AI, specifically including the following steps: Step S1: In response to the activation of online monitoring of the cableway electrical system, a multi-physics sensor network deployed at key electrical component nodes synchronously collects raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values. A unified time stamp is applied to the collected raw time-series data to form and output an initial monitoring data package. Based on preset operating condition change sensing rules, the sampling frequency of the multi-physics sensor network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is dynamically adjusted to obtain a synchronous multi-physics monitoring dataset composed of updated raw time-series data. Each synchronous multi-physics monitoring dataset is stored chronologically to construct a historical monitoring dataset. Step S2: For the historical monitoring dataset, extract predefined native feature parameters related to component failure mechanisms from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values, and generate associated feature parameters based on a cross-physics coupling model; fuse and reduce the dimensionality of the native feature parameters and associated feature parameters to construct a fused health index characterizing the overall health status of the electrical components; store each constructed fused health index in chronological order to form a historical fused health index sequence; associate and store the historical monitoring dataset and the historical fused health index sequence together to form a cableway electrical component health knowledge base; Step S3: Obtain the historical fused health index sequence and simultaneously input the historical fused health index sequence into the pre-trained physical degradation baseline model and the data-driven residual prediction model; the physical degradation baseline model outputs the baseline remaining useful life prediction value based on the failure physics equation based on the historical fused health index sequence; the data-driven residual prediction model outputs the prediction deviation correction amount for the prediction result of the physical degradation baseline model based on the historical fused health index sequence; by fusing the baseline remaining useful life prediction value and the prediction deviation correction amount, the remaining useful life probability distribution of the electrical component is calculated. Step S4: Continuously inject synchronous multiphysics monitoring datasets, historical fused health indicator sequences, and actual component maintenance records obtained from the cableway maintenance system into the cableway digital twin; based on the comparison results between the remaining useful life probability distribution and the component failures recorded in the actual component maintenance records, trigger incremental learning and updates of the data-driven residual prediction model; simultaneously, based on the historical monitoring dataset, perform data pattern distribution change detection on the synchronous multiphysics monitoring dataset; when it is identified that the data pattern in the synchronous multiphysics monitoring dataset exceeds the data pattern distribution range of the historical monitoring dataset, generate a new fault mode warning, thereby completing the online optimization of the data-driven residual prediction model and the expansion of the cableway electrical component health knowledge base; In a preferred embodiment, the specific process of synchronously acquiring the raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration in step S1 is as follows: At key electrical component nodes in the cableway's electrical system, such as the DC bus of the drive frequency converter, the insulated gate bipolar transistor module, and the control relay contacts, a multi-physics field sensing network consisting of high-frequency current and voltage sensors, patch thermocouple sensors, and high-frequency vibration accelerometer sensors is deployed. All sensors in this multi-physics field sensing network are connected to a synchronous acquisition controller, which has a built-in high-precision temperature-controlled crystal oscillator clock source. For every global hardware trigger pulse generated by the high-precision constant temperature crystal oscillator clock source, the synchronous acquisition controller simultaneously sends a start sampling command to all connected sensors. At the same moment that the high-frequency current and voltage sensor, the patch thermocouple sensor, and the high-frequency vibration accelerometer sensor receive the start sampling command, they respectively collect the instantaneous voltage, current, temperature, and vibration acceleration values ​​of their respective nodes. This ensures that the instantaneous voltage, current, temperature, and vibration acceleration values ​​obtained in each sampling are marked with the same precise timestamp corresponding to the global hardware trigger pulse. The instantaneous voltage, current, temperature, and vibration acceleration values, which are collected from the same global hardware trigger pulse and have the same timestamp, are collectively used to form the initial monitoring data packet.

[0006] In a preferred embodiment, the process of dynamically adjusting the sampling frequency of the multi-physics sensing network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​based on preset operating condition change sensing rules specifically involves: A base sampling frequency, a higher sampling frequency than the base sampling frequency, a complexity threshold, a complexity change rate threshold, and a stable period number are predefined. For the instantaneous voltage, current, temperature, and vibration acceleration sequences in the initial monitoring data packet, the following operations are performed respectively: The comprehensive complexity index of this type of instantaneous voltage, current, temperature, or vibration acceleration sequence within the current sliding time window is calculated. The specific calculation process is as follows: First, based on a predefined maximum scale factor, different scale factors are selected to coarse out the instantaneous voltage, instantaneous current, instantaneous temperature, or instantaneous vibration acceleration sequences to construct subsequences at different time scales. Then, the permutation entropy value of each subsequence is calculated. Finally, the arithmetic mean of the permutation entropy values ​​corresponding to all scale factors is calculated, and this arithmetic mean is used as a comprehensive complexity index. Simultaneously, the absolute value of the difference between the comprehensive complexity index of the current sliding time window and the comprehensive complexity index of the previous sliding time window is calculated as the complexity change rate. When any one of the voltage instantaneous value sequence, current instantaneous value sequence, temperature instantaneous value sequence, and vibration acceleration instantaneous value sequence has a comprehensive complexity index exceeding a predefined complexity threshold, or its complexity change rate exceeding a predefined complexity change rate threshold, the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics field sensing network in the next acquisition cycle from the basic sampling frequency to a high sampling frequency. After adjusting the sampling frequency to a high sampling frequency, the comprehensive complexity index and complexity change rate are calculated for each type of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequence for each subsequent acquisition cycle. If the number of consecutive calculations reaches a stable number of cycles, and within these consecutive stable acquisition cycles, the comprehensive complexity index of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity threshold, and the complexity change rate of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity change rate threshold, then the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics sensing network from the high sampling frequency back to the basic sampling frequency. Through this process of dynamically adjusting the sampling frequency, a synchronous multiphysics monitoring dataset is obtained in each acquisition cycle. This synchronous multiphysics monitoring dataset contains instantaneous voltage, current, temperature, and vibration acceleration values ​​with a unified timestamp, and is accompanied by an identifier of the actual sampling frequency used in that acquisition cycle. The value of the actual sampling frequency identifier is either the basic sampling frequency or the high sampling frequency. The synchronous multiphysics monitoring datasets obtained in each acquisition cycle are stored in the order of their timestamps, thereby constructing and continuously updating the historical monitoring dataset.

[0007] In a preferred embodiment, the specific process of extracting predefined native feature parameters related to the component failure mechanism from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration in step S2 is as follows: For the instantaneous voltage, current, temperature, and vibration acceleration values ​​synchronously acquired with a unified timestamp in each acquisition cycle of the historical monitoring dataset constructed in step S1, multiple instantaneous voltage values ​​arranged in chronological order within the same acquisition cycle are formed into a voltage instantaneous value sequence, multiple instantaneous current values ​​into a current instantaneous value sequence, multiple instantaneous temperature values ​​into a temperature instantaneous value sequence, and multiple instantaneous vibration acceleration values ​​into a vibration acceleration instantaneous value sequence. Multiple native feature parameters characterizing component degradation are extracted from these sequences. For the voltage and current instantaneous value sequences, their nonlinear dynamic characteristics are calculated, including multi-scale ranking. Entropy is calculated for the instantaneous temperature value sequence, including its rising slope and spatial distribution gradient over time. For the instantaneous vibration acceleration value sequence, its spectral kurtosis within a predefined frequency band related to the bearing fault characteristic frequency is extracted. All native feature parameters are selected based on predefined physical failure mechanisms. The multi-scale permutation entropy is obtained by calculating the average permutation entropy of the subsequences of the signal at different time scales, and the spectral kurtosis is calculated by analyzing the fourth-order cumulant of the vibration signal in the frequency domain. After extraction, the multiple native feature parameters extracted from the instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration value sequences are arranged according to a predefined feature order to form the native feature parameter set corresponding to the acquisition period.

[0008] In a preferred embodiment, the specific process of generating associated feature parameters based on the cross-physical field coupling relationship model, and fusing and reducing the dimensionality of the original feature parameters and associated feature parameters to construct the fused health index is as follows: A cross-physics coupling model of electro-thermal-vibration relationships is constructed. This model takes native feature parameter elements from the native feature parameter set used to calculate the effective value of the current as input, and calculates multiple associated feature parameters through predefined coupling equations. The associated feature parameters include two types: The first type is the electro-thermal coupling characteristic parameter, whose value is obtained by squaring the effective value of the current in the instantaneous current value sequence, and then multiplying the result of the squaring operation by a preset scaling factor and a preset thermal resistance parameter based on the component material and structure. The second category is thermal-vibration coupling characteristic parameters, whose values ​​are obtained by multiplying the gradient of the instantaneous temperature value sequence with the linear thermal expansion coefficient of the component material, the elastic modulus of the component material, and the characteristic area determined according to the component structure. The calculated associated characteristic parameters are arranged in a predefined order to form an associated characteristic parameter set. The original feature parameter set and the associated feature parameter set are concatenated sequentially to form a high-dimensional hybrid feature set containing all original and associated feature parameters. Each high-dimensional hybrid feature set and its corresponding component health status label are defined as a training sample. The training samples are input into a deep convolutional autoencoder for fusion and dimensionality reduction. This deep convolutional autoencoder processes the input high-dimensional hybrid feature set and outputs a low-dimensional fused health index. During training, this deep convolutional autoencoder introduces a health status metric loss function, the calculation process of which is as follows: In each batch of the training process, for each training sample in that batch, several of its nearest neighbor training samples are selected to form a training sample pair. For each training sample pair, the Euclidean distance between the low-dimensional fused health indicators obtained after compression by the autoencoder is calculated, and the absolute value of the difference between the pre-acquired component health status label values ​​corresponding to the two training samples in that training sample pair is calculated. Then, the difference between the result of multiplying the absolute value of the Euclidean distance and the label value difference by a preset scaling factor is calculated, and the square of this difference is used as the loss contribution of this training sample pair. Finally, the loss contributions of all training sample pairs of all training samples in a training batch are summed to obtain the total health status metric loss value of that batch. Finally, the fusion health indicators generated in each collection cycle are stored in chronological order to form a historical fusion health indicator sequence. Then, the historical monitoring dataset, the historical fusion health indicator sequence, and the actual sampling frequency identifier recorded in the synchronous multiphysics monitoring dataset from step S1 are associated and mapped through a unified timestamp shared by the three. The complete data set after this association and mapping is then persistently stored, thereby forming a health knowledge base for cableway electrical components.

[0009] In a preferred embodiment, the specific process of inputting the historical fusion health index sequence into the pre-trained physical degradation benchmark model in step S3 is as follows: First, a physical degradation benchmark model based on the key failure mechanism of components is constructed. This physical degradation benchmark model is based on the Arrennis-type failure physical equation that describes the electrothermal aging law of insulating materials. Subsequently, the historical fusion health index sequence from the cableway electrical component health knowledge base is used as the input to the physical degradation benchmark model; the physical degradation benchmark model maps the instantaneous absolute temperature and instantaneous voltage value sequences required by the model equations based on the input historical fusion health index sequence. Based on the instantaneous absolute temperature value sequence and the instantaneous voltage value sequence, the cumulative aging amount from the initial moment to the current moment is calculated by integration. When the cumulative aging amount reaches the failure threshold preset according to the component material characteristics, the component is determined to fail, and the predicted time length from the current moment to the cumulative aging amount reaching the failure threshold is used as the baseline remaining useful life prediction value. Meanwhile, based on the errors or preset empirical values ​​in the process of identifying parameters of the physical degradation benchmark model, the benchmark prediction uncertainty corresponding to the benchmark remaining useful life prediction value is estimated. The benchmark remaining useful life prediction value and the benchmark prediction uncertainty together constitute the output of the physical degradation benchmark model.

[0010] In a preferred embodiment, the specific process of inputting the historical fused health indicator sequence into the data-driven residual prediction model and fusing the baseline remaining useful life prediction value to calculate the remaining useful life probability distribution is as follows: First, a data-driven residual prediction model is constructed, which takes a historical fusion health index sequence as input and its training objective is to learn the prediction residuals of the physical degradation benchmark model. During the training phase, the predicted residual corresponding to each historical moment is calculated from the cableway electrical component health knowledge base using recorded data containing known actual failure times. The predicted residual is the difference between the actual remaining useful life recorded from that historical moment and the baseline remaining useful life predicted by the physical degradation baseline model at the same historical moment. The data-driven residual prediction model is trained using the historical fused health index sequence and its corresponding predicted residual as training samples. In the application phase, the current historical fused health index sequence is input into the pre-trained data-driven residual prediction model. The data-driven residual prediction model outputs a prediction bias correction amount for the current moment and the correction uncertainty corresponding to the prediction bias correction amount. Then, the baseline remaining useful life prediction value and its baseline prediction uncertainty output by the physical degradation baseline model, together with the prediction bias correction amount and its correction uncertainty output by the data-driven residual prediction model, are input into a Bayesian fusion framework. In this Bayesian fusion framework, a prior Gaussian probability distribution is constructed using the baseline remaining useful life prediction value as the mean and the square of the baseline prediction uncertainty as the variance; a likelihood Gaussian probability distribution is constructed using the sum of the baseline remaining useful life prediction value and the prediction bias correction amount as the mean and the square of the correction uncertainty as the variance; based on the prior Gaussian probability distribution and the likelihood Gaussian probability distribution... However, the Gaussian probability distribution is used. The parameters of the posterior Gaussian probability distribution are calculated through Bayesian inference. The mean and variance of the posterior Gaussian probability distribution are calculated as follows: the posterior mean is equal to the quotient obtained by dividing the baseline remaining useful life prediction value by the square of the baseline prediction uncertainty, and the quotient obtained by dividing the sum of the baseline remaining useful life prediction value and the prediction deviation correction amount by the square of the correction uncertainty. The sum of these two is then divided by the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The posterior variance is equal to the reciprocal of the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The mean of this posterior Gaussian probability distribution is the best estimate of the remaining useful life after fusion; finally, this posterior Gaussian probability distribution is characterized as the probability distribution of the remaining useful life of electrical components.

[0011] In a preferred embodiment, the specific process of triggering incremental learning and updating of the data-driven residual prediction model based on the comparison results between the remaining useful life probability distribution and the actual component maintenance records in step S4 is as follows: First, in the cableway digital twin, the synchronous multiphysics monitoring dataset from step S1, the historical fused health indicator sequence from step S2, and the remaining useful life probability distribution from step S3 are continuously received and stored. When the actual component maintenance record that records the actual failure or maintenance operation of a component is obtained from the cableway maintenance system, the actual remaining useful life of the component is extracted from the record. Based on the remaining useful life probability distribution, calculate the probability density of the actual remaining useful life of the component falling within the range of possible values ​​described by the remaining useful life probability distribution, and use it as the relative likelihood value. Simultaneously, based on the remaining useful life probability distribution and a set of predefined maintenance decision rules, the decision uncertainty measure for making a definite maintenance decision for the electrical component at the current moment is calculated. The decision uncertainty measure is specifically the decision entropy, which is calculated as follows: First, a predefined set of maintenance decision actions containing multiple specific maintenance actions is defined; then, for each maintenance action in the set, the probability of taking the action is calculated based on all historical data on which the remaining useful life probability distribution is based, according to the remaining useful life probability distribution; finally, the product of the probability of each action and its logarithmic value is summed, and the negative value is taken. The result is the decision entropy. If the relative likelihood value is lower than a preset likelihood threshold, or the decision entropy is higher than a preset decision entropy threshold, it is determined to be a prediction uncertainty event that needs to be learned, and this is used as a trigger signal. When the trigger signal is generated, the historical fusion health indicator sequence associated with the current prediction time, the synchronous multiphysics monitoring dataset, and the real remaining useful life of the component extracted from the actual component maintenance record will be used together to form a new training sample with a real label. Then, in the simulation environment provided by the cableway digital twin, the historical fused health index sequence and synchronous multiphysics monitoring dataset in the new training sample are used to drive the digital twin model to perform state inference and generate the simulation degradation trajectory corresponding to the sample in order to verify the consistency between data characteristics and physical processes. Finally, the new training samples and their corresponding simulated degradation trajectories are submitted to an online incremental learning process for safe and incremental updates and optimizations of the parameters of the data-driven residual prediction model.

[0012] In a preferred embodiment, the specific process of detecting changes in data pattern distribution of the synchronous multiphysics monitoring dataset based on historical monitoring datasets and generating new fault mode warnings to expand the knowledge base is as follows: While the incremental learning process of the model is executed in parallel, a continuously running data pattern distribution change detection process is started. The process uses the historical monitoring dataset constructed in step S1 as the data source of the baseline distribution. The historical monitoring dataset contains all historical synchronous multiphysics monitoring datasets generated by electrical components under known health conditions and known degradation modes. For the newly injected cableway digital twin, the real-time synchronous multiphysics monitoring dataset from step S1 is used as the input data to be detected. A deep kernel single-class classification model is used to calculate its novelty score relative to the benchmark distribution represented by the historical monitoring dataset. The deep kernel uniclass classification model comprises a deep feature extraction network, a predefined kernel function, a set of support vector samples learned from historical monitoring datasets, a set of weight coefficients corresponding to the support vector samples, and a decision threshold. The deep feature extraction network maps the input synchronous multiphysics monitoring dataset samples to a high-dimensional feature space. The process of calculating the novelty score in the deep kernel uniclass classification model is as follows: First, the deep feature extraction network maps the current input real-time synchronous multiphysics monitoring dataset samples and each support vector sample to a high-dimensional feature space. Then, in the high-dimensional feature space, the predefined kernel function is used to calculate the kernel function similarity value between the current input sample and each support vector sample. Next, the kernel function similarity value corresponding to each support vector sample is multiplied by the weight coefficient corresponding to that support vector sample to obtain a set of weighted similarity values. Subsequently, the weighted similarity values ​​corresponding to all support vector samples are summed to obtain a comprehensive similarity value. Finally, the decision threshold is subtracted from the comprehensive similarity value, and the difference is the novelty score. If the novelty score is less than zero, the current data pattern is determined to be outside the distribution range of data patterns in the historical monitoring dataset and is identified as a potential new fault mode. During the detection process, if the novelty score of the current synchronous multiphysics monitoring dataset is less than zero, it is determined that the current data pattern exceeds the data pattern distribution range of the historical monitoring dataset and is identified as a potential new fault mode. When a potential new failure mode is identified, a new failure mode warning is immediately generated. The warning includes the synchronous multiphysics monitoring dataset that triggered the identification, the corresponding fused health index, the timestamp, and the calculated novelty score. Subsequently, this new fault mode warning information is treated as a new knowledge entry with the label "Pending Review - Suspected New Mode", associated with its corresponding timestamp, and added to the cableway electrical component health knowledge base.

[0013] This application also provides a cableway full life cycle health management system based on multimodal AI, specifically including: a multi-physics field synchronous acquisition and adaptive sampling module, a multi-physics field feature fusion and health knowledge base construction module, a physical-data hybrid driven remaining life probability prediction module, and a digital twin driven model optimization and knowledge expansion module, wherein; Multi-physics synchronous acquisition and adaptive sampling module: It is configured to respond to the online monitoring start command of the cableway electrical system, control the sensor network deployed at key component nodes to synchronously acquire voltage, current, temperature and vibration data, and add a unified time stamp to the data; it dynamically adjusts the acquisition frequency based on preset sensing rules, outputs synchronous multi-physics monitoring dataset, and archives it to form historical monitoring dataset; Multiphysics Feature Fusion and Health Knowledge Base Construction Module: It is configured to extract and fuse features from historical monitoring datasets, generate fused health indicators that characterize the health status of components, and associate historical monitoring datasets with the sequentially stored sequence of fused health indicators to construct a health knowledge base for cableway electrical components; The physics-data hybrid driven remaining useful life probability prediction module is configured to obtain historical fused health indicator sequences from a knowledge base and perform collaborative prediction using a pre-trained physical model and a data-driven model, and output the remaining useful life probability distribution of electrical components after fusion processing. The digital twin-driven model optimization and knowledge expansion module is configured to continuously receive monitoring data, health indicators, and actual maintenance records in the cableway digital twin; trigger incremental learning optimization of the data-driven model based on the comparison between the life prediction distribution and the actual records; and simultaneously perform abnormal pattern detection on real-time monitoring data based on historical datasets, and generate early warning information to expand the knowledge base when new fault modes are identified.

[0014] The beneficial effects of this invention are as follows: Through synchronous sensing and adaptive acquisition of multi-physics field data, high-fidelity and high-efficiency monitoring of the operating status of cableway electrical components is achieved. Based on multi-modal feature fusion and deep self-encoding technology, a health index that can sensitively reflect the overall degradation status of components is constructed. Combining a hybrid prediction framework of physical model and data-driven model, a probabilistic and accurate prediction of remaining life is generated. With the help of a digital twin environment, closed-loop online optimization of the prediction model and proactive perception of new fault modes and automatic expansion of the knowledge base are realized. Thus, the operation and maintenance mode is upgraded from post-maintenance and periodic inspection to state-based accurate predictive maintenance, which significantly improves the level of safety assurance and operational economic benefits, and transforms expert experience into digital assets that can be continuously accumulated and iterated. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 This paper presents a method for full life-cycle health management of cableways based on multimodal AI, which specifically includes the following steps: Step S1: In response to the activation of online monitoring of the cableway electrical system, a multi-physics sensor network deployed at key electrical component nodes synchronously collects raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values. A unified time stamp is applied to the collected raw time-series data to form and output an initial monitoring data package. Based on preset operating condition change sensing rules, the sampling frequency of the multi-physics sensor network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is dynamically adjusted to obtain a synchronous multi-physics monitoring dataset composed of updated raw time-series data. Each synchronous multi-physics monitoring dataset is stored chronologically to construct a historical monitoring dataset. Step S2: For the historical monitoring dataset, extract predefined native feature parameters related to component failure mechanisms from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values, and generate associated feature parameters based on a cross-physics coupling model; fuse and reduce the dimensionality of the native feature parameters and associated feature parameters to construct a fused health index characterizing the overall health status of the electrical components; store each constructed fused health index in chronological order to form a historical fused health index sequence; associate and store the historical monitoring dataset and the historical fused health index sequence together to form a cableway electrical component health knowledge base; Step S3: Obtain the historical fused health index sequence and simultaneously input the historical fused health index sequence into the pre-trained physical degradation baseline model and the data-driven residual prediction model; the physical degradation baseline model outputs the baseline remaining useful life prediction value based on the failure physics equation based on the historical fused health index sequence; the data-driven residual prediction model outputs the prediction deviation correction amount for the prediction result of the physical degradation baseline model based on the historical fused health index sequence; by fusing the baseline remaining useful life prediction value and the prediction deviation correction amount, the remaining useful life probability distribution of the electrical component is calculated. Step S4: Continuously inject the synchronous multiphysics monitoring dataset, historical fused health indicator sequences, and actual component maintenance records obtained from the cableway maintenance system into the cableway digital twin. Based on the comparison results between the remaining useful life probability distribution and the component failures recorded in the actual component maintenance records, trigger incremental learning and updates of the data-driven residual prediction model. Simultaneously, based on the historical monitoring dataset, perform data pattern distribution change detection on the synchronous multiphysics monitoring dataset. When it is identified that the data pattern in the synchronous multiphysics monitoring dataset exceeds the data pattern distribution range of the historical monitoring dataset, generate a new fault mode warning, thereby completing the online optimization of the data-driven residual prediction model and the expansion of the cableway electrical component health knowledge base.

[0020] In this embodiment, the specific process of synchronously acquiring the raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration in step S1 is as follows: At key electrical component nodes in the cableway's electrical system, such as the DC bus of the drive inverter, the insulated gate bipolar transistor module, and the control relay contacts, a multi-physics sensing network is deployed, consisting of high-frequency current and voltage sensors, patch thermocouple sensors or infrared thermal imaging units, and high-frequency vibration accelerometer sensors. All sensors in this multi-physics sensing network are connected to a synchronous acquisition controller, which has a built-in high-precision temperature-controlled crystal oscillator clock source. This high-precision temperature-controlled crystal oscillator clock source can provide a frequency reference with an accuracy better than ±0.1ppm, ensuring that the timing accumulation error of long-term sampling is minimal. The period of the global hardware trigger pulse, i.e., the basic sampling period, can be set according to monitoring requirements. For example, for steady-state electrical monitoring, it can be set to 10 milliseconds, corresponding to a basic sampling frequency of 100 Hz. For every global hardware trigger pulse generated by the high-precision constant temperature crystal oscillator clock source, the synchronous acquisition controller simultaneously sends a start sampling command to all connected sensors. At the same moment that a high-frequency current and voltage sensor, a patch thermocouple sensor or an infrared thermal imaging unit, and a high-frequency vibration accelerometer sensor receive a start sampling command, they respectively collect the instantaneous voltage, current, temperature and vibration acceleration values ​​of their respective nodes. This ensures that the instantaneous voltage, current, temperature and vibration acceleration values ​​obtained in each sampling are marked with the same precise timestamp corresponding to the global hardware trigger pulse. The instantaneous voltage, current, temperature, and vibration acceleration values, which are collected by the same global hardware trigger pulse and have the same timestamp, are collectively used to form the initial monitoring data packet. To facilitate transmission and processing, the initial monitoring data packet can adopt a predefined structured format, such as a data frame format that includes a header, timestamp field, data fields of each sensor, and a check field. The synchronous acquisition controller can temporarily cache this data packet for subsequent processing. Based on preset operating condition change sensing rules, the process of dynamically adjusting the sampling frequency of the multi-physics sensing network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is as follows: A base sampling frequency, a higher sampling frequency than the base sampling frequency, a complexity threshold, a complexity change rate threshold, and a stable period number are predefined. The complexity threshold and the complexity change rate threshold are not fixed values; their initial values ​​can be determined by calculating the above comprehensive complexity index from the historical instantaneous voltage, current, temperature, and vibration acceleration value sequences of the cableway electrical system under known healthy conditions, and then statistically analyzing their normal distribution. For example, the complexity threshold can be set to twice the standard deviation of the historical index average, and the complexity change rate threshold can be set to 1.5 times the standard deviation of the historical change rate average. In practical applications, these thresholds can be optimized and adjusted through a learning process. The stable period number is used to prevent the sampling frequency from frequently switching near the threshold, acting as a "hysteresis" effect. Its value can be set according to actual engineering needs. For example, setting it to 10 acquisition cycles means that the system needs to meet the stability condition for 10 consecutive cycles before the operating condition is considered to have returned to stability. For the voltage, current, temperature, and vibration acceleration value sequences in the initial monitoring data packet, the following operations are performed respectively: The comprehensive complexity index of this type of instantaneous voltage, current, temperature, or vibration acceleration sequence within the current sliding time window is calculated. The specific calculation process is as follows: First, based on a predefined maximum scale factor, different scale factors are selected to coarse-grain the instantaneous voltage, current, temperature, or vibration acceleration sequences to construct subsequences at different time scales. Then, the permutation entropy value of each subsequence is calculated. Finally, the arithmetic mean of the permutation entropy values ​​corresponding to all scale factors is calculated, and this arithmetic mean is used as a comprehensive complexity index. Here, the permutation entropy value is an index that measures the complexity and randomness of a time series and is sensitive to dynamic changes. Through multi-scale analysis, the irregularities of the signal at multiple time resolutions can be captured simultaneously. For example, the maximum scale factor can be set to 5, and the permutation entropy value of the subsequence at each scale can be calculated (the embedding dimension can be set to 3, and the time delay can be set to 1). Finally, the permutation entropy values ​​of the 5 scales are averaged. The closer the index value is to 1, the closer the sequence is to random noise; the lower the value, the more regular the sequence. For stable cableway electrical signals, this index is usually at a low level. Simultaneously, the absolute value of the difference between the comprehensive complexity index of the current sliding time window and the comprehensive complexity index of the previous sliding time window is calculated as the complexity change rate. When any one of the voltage instantaneous value sequence, current instantaneous value sequence, temperature instantaneous value sequence, and vibration acceleration instantaneous value sequence has a comprehensive complexity index exceeding a predefined complexity threshold, or its complexity change rate exceeding a predefined complexity change rate threshold, the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics field sensing network in the next acquisition cycle from the basic sampling frequency to a high sampling frequency. After adjusting the sampling frequency to a high sampling frequency, the comprehensive complexity index and complexity change rate are calculated for each type of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequence for each subsequent acquisition cycle. If the number of consecutive calculations reaches a stable number of cycles, and within these consecutive stable acquisition cycles, the comprehensive complexity index of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity threshold, and the complexity change rate of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity change rate threshold, then the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics sensing network from the high sampling frequency back to the basic sampling frequency. Through this process of dynamically adjusting the sampling frequency, a synchronous multiphysics monitoring dataset is obtained in each acquisition cycle. This synchronous multiphysics monitoring dataset contains instantaneous voltage, current, temperature, and vibration acceleration values ​​with a unified timestamp, and also includes an identifier of the actual sampling frequency used in that acquisition cycle. Each synchronous multiphysics monitoring dataset can be considered as a record in the cableway electrical component health knowledge base when stored. This record contains the following fields: unique record ID, timestamp, instantaneous voltage, instantaneous current, instantaneous temperature, instantaneous vibration acceleration, and actual sampling frequency identifier. The actual sampling frequency identifier is stored in the form of an enumeration value. For example, '0' represents the basic sampling frequency (e.g., 100Hz), and '1' represents the high sampling frequency (e.g., 5kHz). The value of the actual sampling frequency identifier is either the basic sampling frequency or the high sampling frequency. The synchronous multiphysics monitoring datasets obtained in each acquisition cycle are stored in the order of their included timestamps, thereby constructing and continuously updating a historical monitoring dataset that serves as the input data source for step S2.

[0021] In this embodiment, it is specifically necessary to explain the process in step S2, which involves extracting predefined native feature parameters related to the component failure mechanism from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values, as follows: For the instantaneous voltage, current, temperature, and vibration acceleration values ​​synchronously acquired with a unified timestamp in each acquisition cycle of the historical monitoring dataset constructed in step S1, multiple instantaneous voltage values ​​arranged in chronological order within the same acquisition cycle are formed into a voltage instantaneous value sequence, multiple instantaneous current values ​​into a current instantaneous value sequence, multiple instantaneous temperature values ​​into a temperature instantaneous value sequence, and multiple instantaneous vibration acceleration values ​​into a vibration acceleration instantaneous value sequence. Multiple native feature parameters characterizing component degradation are extracted from these sequences. For the voltage and current instantaneous value sequences, their nonlinear dynamic characteristics are calculated, including multi-scale... Multi-scale permutation entropy is used to capture the complexity and randomness of signal waveform changes. The calculation process for multi-scale permutation entropy can be as follows: For example, a maximum scale factor of 5, an embedding dimension of 3, and a time delay of 1 can be set. Permutation entropy is calculated for coarse-grained subsequences of voltage or current sequences at different scales. Finally, the arithmetic mean of the entropy values ​​at the five scales is calculated to obtain a comprehensive complexity index. The closer this index is to 1, the closer the signal is to random noise; the lower it is, the more regular the signal. For stable cableway electrical signals, this index is usually at a low level, for example, below 0.4. For instantaneous temperature value sequences, the rising slope and spatial distribution gradient over time are calculated to reflect heat accumulation and diffusion characteristics. For instantaneous vibration acceleration value sequences, the frequency related to bearing fault characteristics is extracted. The spectral kurtosis within a predefined frequency band is used to highlight the impact vibration component. This predefined frequency band can be determined based on the fault characteristic frequencies (such as inner ring, outer ring, and rolling element fault frequencies) calculated from the geometric parameters and rotational speed of the cableway drive motor bearing. The spectral kurtosis can be calculated using a fast spectral kurtosis analysis method, performed in the time-frequency domain through short-time Fourier transform and window functions. For example, the center frequency can be set as the bearing outer ring fault characteristic frequency, and the bandwidth as its sideband range. The spectral kurtosis value within this frequency band can be calculated, effectively amplifying the periodic impact component in the signal. All native characteristic parameters are selected based on predefined physical failure mechanisms. The multi-scale permutation entropy is obtained by calculating the average permutation entropy of the subsequences of the signal at different time scales. The spectral kurtosis is determined by analyzing the vibration signal... The number is calculated from the fourth-order cumulant in the frequency domain. After extraction, multiple original feature parameters extracted from the instantaneous voltage value sequence, instantaneous current value sequence, instantaneous temperature value sequence, and instantaneous vibration acceleration value sequence are arranged according to a predefined feature order to form the original feature parameter set corresponding to the acquisition period. The original feature parameter set is used to characterize the original health characteristics of electrical components in the multi-physics field within the acquisition period. For example, the original feature parameter set may include, in order: voltage multi-scale arrangement entropy, current multi-scale arrangement entropy, current effective value, temperature rise slope, temperature spatial gradient, and vibration spectrum kurtosis. Among them, the current effective value is calculated from the current instantaneous value sequence. All feature values ​​can be normalized before splicing to eliminate the influence of dimensions. The specific process of generating associated feature parameters based on a cross-physics coupling relationship model, and then fusing and reducing the dimensionality of the original feature parameters and associated feature parameters to construct a fused health index is as follows: A cross-physics coupling model of electro-thermal-vibration relationships is constructed. This model takes native feature parameter elements from the native feature parameter set used to calculate the effective value of the current as input, and calculates multiple associated feature parameters through predefined coupling equations. The associated feature parameters include two types: The first type is the electro-thermal coupling characteristic parameter, whose value is obtained by squaring the effective value of the current in the instantaneous current value sequence, and then multiplying the result of the squaring operation by a preset scaling factor and a preset thermal resistance parameter based on the component material and structure. The second category is thermal-vibration coupling characteristic parameters, whose values ​​are obtained by multiplying the gradient of the instantaneous temperature value sequence with the linear thermal expansion coefficient of the component material, the elastic modulus of the component material, and the characteristic area determined according to the component structure. The calculated associated characteristic parameters are arranged in a predefined order to form an associated characteristic parameter set. The original feature parameter set and the associated feature parameter set are concatenated sequentially to form a high-dimensional hybrid feature set containing all original and associated feature parameters. Each high-dimensional hybrid feature set and its corresponding component health status label are defined as a training sample. The component health status label is a numerical value used for supervised training. It is obtained in the following ways: for historical data, the label can be calculated as a percentage of the remaining useful life based on the actual time interval from the data collection to the next planned maintenance or failure of the component; for new components or data lacking failure records, it can be labeled by domain experts based on experience or simulation results based on physical models. The label value can be between 0 and 1, where 1 represents a brand new health status and 0 represents complete failure. The training samples are input into a deep convolutional autoencoder for fusion and dimensionality reduction. The deep convolutional autoencoder can contain two parts: an encoder and a decoder. The encoder can be composed of several one-dimensional convolutional layers, activation function layers, and pooling layers stacked together to compress the high-dimensional hybrid feature set into a low-dimensional fused health index. The decoder structure is symmetrical to it and is used to reconstruct the input features from the low-dimensional representation. The goal of model training is to minimize the sum of reconstruction error and health status metric loss. When the model is deployed (inference), only the encoder part is used to generate the fused health metric. This deep convolutional autoencoder processes the high-dimensional mixed feature set of the input and outputs a low-dimensional fused health metric. The deep convolutional autoencoder introduces a health status metric loss function during training, and the calculation process of the health status metric loss function is as follows: In the initial training phase of the model, a historical high-dimensional hybrid feature set and its corresponding historical health status labels are extracted from the existing cableway electrical component health knowledge base to form an offline training dataset. This dataset is divided into a training set, a validation set, and a test set. Each batch during training is a set of several records (e.g., 32 or 64 records) randomly selected from the training set. Within each batch, for each training sample, several nearest neighbor training samples are selected to form a training sample pair. For each training sample pair, the Euclidean distance between the low-dimensional fused health indicators obtained after autoencoder compression of the two training samples in the pair is calculated, and the absolute value of the difference between the pre-acquired component health status label values ​​corresponding to the two training samples in the pair is also calculated. Then, the Euclidean distance and the label are calculated. The difference between the absolute value of the difference and the result of multiplying by a preset scaling factor (the preset scaling factor is used to balance the dimensions between the distance metric and the label difference, and its value can be determined through cross-validation, for example, it can be initially set to 1.0 and adjusted on the validation set according to the model performance) is used as the loss contribution of this training sample pair. Finally, the loss contributions of all training sample pairs in a training batch are summed to obtain the total health status metric loss value of the batch. The health status metric loss function is used to constrain the autoencoder so that in the latent space, the distance between any two low-dimensional fused health indicators is proportional to the absolute difference between their corresponding component health status labels, so that the output fused health indicators maintain a monotonic correlation with the component health status in the latent space. Finally, the fused health indicators generated in each collection cycle are stored in chronological order to form a historical fused health indicator sequence. Then, the historical monitoring dataset, the historical fused health indicator sequence, and the actual sampling frequency identifier recorded in the synchronous multiphysics monitoring dataset from step S1 are associated and mapped through a unified timestamp shared by the three. The association mapping can be achieved at the database level through primary and foreign key relationships in relational data tables. Specifically, the timestamp can be used as the primary key to create an overview table. Each record in the table contains the timestamp and is associated with three sub-tables: a sub-table storing the original monitoring data, a sub-table storing the fused health indicators, and a sub-table storing the collection condition information (including the actual sampling frequency identifier). By performing a joint query through the timestamp, complete multimodal health status information at any point in time can be obtained. Persistent storage can be achieved using a time-series database or a relational database, and an optimized storage structure indexed by the timestamp can be established. The complete data set after this association mapping is then persistently stored, thus forming a health knowledge base for cableway electrical components.

[0022] In this embodiment, the specific process of inputting the historical fused health index sequence into the pre-trained physical degradation benchmark model in step S3 is as follows: First, a physical degradation benchmark model based on the key failure mechanism of the component is constructed. This physical degradation benchmark model is based on the Arrennis-type failure physical equation that describes the electrothermal aging law of insulating materials. The Arrennis-type failure physical equation defines the exponential relationship between the component aging rate and the reciprocal of the absolute temperature of the material and the power of the voltage it withstands. Subsequently, the historical fusion health index sequence from the cableway electrical component health knowledge base is used as the input to the physical degradation benchmark model; based on the input historical fusion health index sequence, the physical degradation benchmark model maps or derives the absolute temperature instantaneous value and voltage instantaneous value sequence required for the model equation; Based on the instantaneous absolute temperature and voltage value sequences, the cumulative aging amount from the initial moment to the current moment is calculated through integration. The integration can be implemented using numerical integration methods, such as the trapezoidal integration method, by discretely integrating the instantaneous aging rate sequence calculated from the instantaneous absolute temperature and voltage value sequences according to the Arrhenius equation. When the cumulative aging amount reaches the failure threshold preset according to the component material characteristics, the component is determined to have failed. The predicted time length from the current moment to the cumulative aging amount reaching the failure threshold is used as the baseline remaining useful life prediction value. The failure threshold can be determined based on the accelerated aging test data of the component's insulation material, and is characterized as the maximum cumulative aging amount that the material performance can withstand when it degrades to a safe critical point. For example, for epoxy insulation materials, the cumulative aging amount corresponding to the breakdown voltage dropping to a certain percentage (such as 50%) of the initial value can be determined through high temperature and high voltage accelerated aging tests as the failure threshold. Meanwhile, based on the errors or preset empirical values ​​in the process of identifying parameters of the physical degradation benchmark model, the benchmark prediction uncertainty corresponding to the benchmark remaining useful life prediction value is estimated. The benchmark remaining useful life prediction value and the benchmark prediction uncertainty together constitute the output of the physical degradation benchmark model. The specific process of inputting historical fused health indicator sequences into a data-driven residual prediction model and fusing baseline remaining useful life predictions to calculate the probability distribution of remaining useful life is as follows: First, a data-driven residual prediction model is constructed, which takes a historical fusion health index sequence as input and its training objective is to learn the prediction residuals of the physical degradation benchmark model. During the training phase, the predicted residuals for each historical moment are calculated from the cableway electrical component health knowledge base using recorded data containing known actual failure times. The predicted residuals are the difference between the actual remaining useful life recorded from that historical moment and the baseline remaining useful life predicted by the physical degradation baseline model at the same historical moment. The data-driven residual prediction model is trained using the historical fused health index sequence and its corresponding predicted residuals as training samples. In the application phase, the current historical fused health index sequence is input into the pre-trained data-driven residual prediction model. The data-driven residual prediction model outputs a prediction bias correction amount for the current moment and the correction uncertainty corresponding to the prediction bias correction amount. Then, the baseline remaining useful life prediction value and its baseline prediction uncertainty output from the physical degradation baseline model, along with the prediction bias correction and its correction uncertainty output from the data-driven residual prediction model, are input into a Bayesian fusion framework. Within this framework, a prior Gaussian probability distribution is constructed using the baseline remaining useful life prediction value as the mean and the square of the baseline prediction uncertainty as the variance. A likelihood Gaussian probability distribution is constructed using the sum of the baseline remaining useful life prediction value and the prediction bias correction as the mean and the square of the correction uncertainty as the variance. Based on the prior Gaussian probability distribution... The parameters of the posterior Gaussian probability distribution are calculated using Bayesian inference, based on the likelihood Gaussian probability distribution. The mean and variance of the posterior Gaussian probability distribution are calculated as follows: the posterior mean is equal to the quotient obtained by dividing the baseline remaining useful life prediction value by the square of the baseline prediction uncertainty, and the quotient obtained by dividing the sum of the baseline remaining useful life prediction value and the prediction bias correction by the square of the correction uncertainty. The sum of these two is then divided by the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The posterior variance is equal to the reciprocal of the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The mean of the posterior Gaussian probability distribution is the best estimate of the remaining useful life after fusion, and its variance represents the overall prediction uncertainty after fusion. Finally, the posterior Gaussian probability distribution is represented as the probability distribution of the remaining useful life of electrical components.

[0023] In this embodiment, the specific process of triggering incremental learning and updating of the data-driven residual prediction model based on the comparison results between the remaining useful life probability distribution and the actual component maintenance records in step S4 is as follows: First, in the cableway digital twin, the synchronous multiphysics monitoring dataset from step S1, the historical fused health indicator sequence from step S2, and the remaining useful life probability distribution from step S3 are continuously received and stored. When the actual component maintenance record that records the actual failure or maintenance operation of a component is obtained from the cableway maintenance system, the actual remaining useful life of the component is extracted from the record. Based on the remaining useful life probability distribution, calculate the probability density of the actual remaining useful life of the component falling within the range of possible values ​​described by the remaining useful life probability distribution, and use it as the relative likelihood value. Simultaneously, based on the remaining useful life probability distribution and a set of predefined maintenance decision rules, the decision uncertainty measure for making a definite maintenance decision for the electrical component at the current moment is calculated. Specifically, the decision uncertainty measure is the decision entropy, which is calculated as follows: First, a predefined set of maintenance decision actions containing multiple specific maintenance actions is defined. Then, for each maintenance action in this set, based on the remaining useful life probability distribution, the probability of taking that action under all historical data upon which the remaining useful life probability distribution is based is calculated. The probability calculation can be achieved by integrating the remaining useful life probability distribution over the time decision interval corresponding to each maintenance action. For example, if the predefined maintenance actions are "immediate repair," "planned repair within one month," and "continued monitoring," their corresponding time decision intervals can be [0, 24) hours, [24, 720) hours, and [720, +∞) hours, respectively. When calculating the probability, the remaining useful life probability density function is integrated over the corresponding interval. Finally, the product of the probability of each action and its logarithmic value is summed, and the negative value is taken. The result is the decision entropy. The higher the decision entropy value, the greater the difficulty in making a definite maintenance decision based on the current model. If the relative likelihood value is lower than a preset likelihood threshold, or the decision entropy is higher than a preset decision entropy threshold, it is determined to be a prediction uncertainty event that needs to be learned, and this is used as a trigger signal. The likelihood threshold and decision entropy threshold are adjustable parameters, and their initial values ​​can be determined based on the historical prediction accuracy statistics. For example, the likelihood threshold can be initially set to 0.05, which means that when the actual lifetime value falls in the probability density range of the model's predicted probability distribution below 5%, the prediction deviation is considered significant. The decision entropy threshold can be initially set to 1.0 (natural logarithmic unit). When the decision uncertainty is higher than this value, the model's judgment of the current state is considered too vague and needs to be optimized. In actual deployment, these thresholds can be fine-tuned according to the warning frequency and actual needs over a period of time. When the trigger signal is generated, the historical fusion health indicator sequence associated with the current prediction time, the synchronous multiphysics monitoring dataset, and the real remaining useful life of the component extracted from the actual component maintenance record will be used together to form a new training sample with a real label. Then, in the simulation environment provided by the cableway digital twin, the historical fused health index sequence and synchronous multiphysics monitoring dataset in the new training sample are used to drive the digital twin model to perform state inference and generate the simulation degradation trajectory corresponding to the sample in order to verify the consistency between data characteristics and physical processes. Finally, the new training samples and their corresponding simulated degradation trajectories are submitted to an online incremental learning process for safe and incremental updates and optimizations of the parameters of the data-driven residual prediction model. The specific steps of the online incremental learning process are as follows: The parameters to be optimized in the data-driven residual prediction model are treated as random variables, and a set of variational parameters is maintained to describe their probability distribution. During each incremental learning iteration, the new training samples that trigger learning and their corresponding simulation degradation trajectories constitute a micro-batch of new data. An online variational learning objective is constructed, consisting of two terms: the first is the expected log-likelihood term, which represents the expected probability that the new data will be correctly predicted by the model based on the parameter distribution described by the current variational parameters; this term drives the model to learn new knowledge. The second term is the KL divergence regularization term, which measures the difference between the new parameter distribution described by the current variational parameters and the old parameter distribution described by the old variational parameters saved before this learning iteration; this term penalizes drastic changes in the parameter distribution to retain old knowledge. The online variational learning objective is the first term minus a preset tradeoff coefficient multiplied by the second term. The preset tradeoff coefficient is used to balance the model's ability to "learn new knowledge" and "retain old knowledge." Its value is greater than zero. A smaller coefficient (such as 0.1) makes the model more inclined to adapt to new data quickly, but may forget old patterns; a larger coefficient (such as 10.0) makes the model update more conservative, with high stability, but a slower rate of learning new patterns. This coefficient can be determined by simulating the incremental learning process on historical data and fine-tuning it based on the model's comprehensive performance on the validation set. This mechanism effectively alleviates the "catastrophic forgetting" problem in neural networks during continuous learning. By adjusting the variational parameters to maximize the online variational learning objective, the parameters of the data-driven residual prediction model are updated. This process balances the learning of new knowledge and the retention of old knowledge. The specific process of detecting changes in data pattern distribution on the synchronous multiphysics monitoring dataset based on historical monitoring datasets and generating new fault mode early warnings to expand the knowledge base is as follows: While the incremental learning process of the model is executed in parallel, a continuously running data pattern distribution change detection process is started. The process uses the historical monitoring dataset constructed in step S1 as the data source of the baseline distribution. The historical monitoring dataset contains all historical synchronous multiphysics monitoring datasets generated by electrical components under known health conditions and known degradation modes. For the newly injected cableway digital twin, the real-time synchronous multiphysics monitoring dataset from step S1 is used as the input data to be detected. A deep kernel single-class classification model is used to calculate its novelty score relative to the benchmark distribution represented by the historical monitoring dataset. The deep kernel uniclass classification model includes a deep feature extraction network, a predefined kernel function, a set of support vector samples learned from historical monitoring datasets, a set of weight coefficients corresponding to the support vector samples, and a decision threshold. The predefined kernel function can be a Gaussian kernel function, which can map samples to an infinite-dimensional feature space, effectively handling non-linearly separable patterns. Its bandwidth parameter is one of the key hyperparameters that needs to be determined during model training, affecting the model's sensitivity to distribution changes. The deep feature extraction network is used to map the input synchronous multiphysics monitoring dataset samples to a high-dimensional feature space. The process of calculating the novelty score in the deep kernel uniclass classification model is as follows: First, a deep feature extraction network maps the current input real-time synchronous multiphysics monitoring dataset sample and each support vector sample to a high-dimensional feature space. Then, in the high-dimensional feature space, a predefined kernel function is used to calculate the kernel function similarity value between the current input sample and each support vector sample. Next, the kernel function similarity value corresponding to each support vector sample is multiplied by its corresponding weight coefficient to obtain a set of weighted similarity values. Subsequently, the weighted similarity values ​​corresponding to all support vector samples are summed to obtain a comprehensive similarity value. Finally, a decision threshold is subtracted from the comprehensive similarity value, and the resulting difference is the novelty score. If the novelty score is less than zero, the current data pattern is determined to be outside the distribution range of the data patterns in the historical monitoring dataset and is identified as a potential new fault mode. This joint training process is usually achieved by optimizing a single-class classification objective function, such as minimizing the range described by the support vectors, while ensuring that most training samples are correctly described as "normal". The support vector samples are representative samples that are automatically selected from the training set and located near the "boundary" of the known distribution during this optimization process. The weight coefficients and decision thresholds are the direct outputs of the optimization process. The training process of the deep kernel uniclass classification model is as follows: using samples from the synchronous multiphysics monitoring dataset in the historical monitoring dataset, the parameters of the deep feature extraction network, the composition of the support vector samples, the weight coefficients, and the decision threshold are jointly learned and determined. This joint training process is usually achieved by optimizing a uniclass classification objective function, such as minimizing the range described by the support vectors, while ensuring that most training samples are correctly described as "normal". The support vector samples are representative samples that are automatically selected from the training set and located near the "boundary" of the known distribution during this optimization process. The weight coefficients and decision thresholds are the direct outputs of the optimization process, enabling the model to calculate a high novelty score for the samples in the historical monitoring dataset, representing that they belong to the known distribution. During the detection process, if the novelty score of the current synchronous multiphysics monitoring dataset is less than zero, it is determined that the current data pattern exceeds the data pattern distribution range of the historical monitoring dataset and is identified as a potential new fault mode. When a potential new failure mode is identified, a new failure mode warning is immediately generated. The warning includes the synchronous multiphysics monitoring dataset that triggered the identification, the corresponding fused health index, the timestamp, and the calculated novelty score. Subsequently, this new fault mode warning information is treated as a new knowledge entry with the label "Pending Review - Suspected New Mode", associated with its corresponding timestamp, and added to the cableway electrical component health knowledge base; In this way, the health knowledge base of cableway electrical components can be expanded in real time without interrupting the online monitoring and prediction process. This provides a data foundation for subsequent expert diagnosis, pattern confirmation, and model retraining. The knowledge base can be designed with a dedicated "early warning and pending review mode" data table. In addition to raw data, features, scores, and timestamps, the newly added entries can also reserve fields such as "expert diagnosis conclusion," "fault mode classification," and "processing status." The platform can provide a human-computer interaction interface for domain experts to review, annotate, confirm, or reject early warning entries. Data confirmed as real new fault modes can be further used as high-quality samples to trigger the centralized retraining of data-driven residual prediction models and even deep kernel single-class classification models, thereby completing a complete knowledge loop from detection, early warning, expert confirmation to model enhancement.

[0024] Example 2 This embodiment provides, for example Figure 2 The diagram illustrates a cableway full lifecycle health management system based on multimodal AI, specifically comprising: a multi-physics synchronous acquisition and adaptive sampling module, a multi-physics feature fusion and health knowledge base construction module, a physics-data hybrid-driven remaining lifespan probability prediction module, and a digital twin-driven model optimization and knowledge expansion module, wherein; Multi-physics synchronous acquisition and adaptive sampling module: It is configured to respond to the online monitoring start command of the cableway electrical system, control the sensor network deployed at key component nodes to synchronously acquire voltage, current, temperature and vibration data, and add a unified time stamp to the data; it dynamically adjusts the acquisition frequency based on preset sensing rules, outputs synchronous multi-physics monitoring dataset, and archives it to form historical monitoring dataset; Multiphysics Feature Fusion and Health Knowledge Base Construction Module: It is configured to extract and fuse features from historical monitoring datasets, generate fused health indicators that characterize the health status of components, and associate historical monitoring datasets with the sequentially stored sequence of fused health indicators to construct a health knowledge base for cableway electrical components; The physics-data hybrid driven remaining useful life probability prediction module is configured to obtain historical fused health indicator sequences from a knowledge base and perform collaborative prediction using a pre-trained physical model and a data-driven model, and output the remaining useful life probability distribution of electrical components after fusion processing. The digital twin-driven model optimization and knowledge expansion module is configured to continuously receive monitoring data, health indicators, and actual maintenance records in the cableway digital twin; trigger incremental learning optimization of the data-driven model based on the comparison between the life prediction distribution and the actual records; and simultaneously perform abnormal pattern detection on real-time monitoring data based on historical datasets, and generate early warning information to expand the knowledge base when new fault modes are identified.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations.Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for full life-cycle health management of cableways based on multimodal AI, characterized in that, Specifically, the following steps are included: Step S1: In response to the activation of online monitoring of the cableway electrical system, a multi-physics sensor network deployed at key electrical component nodes synchronously collects raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values. A unified time stamp is applied to the collected raw time-series data to form and output an initial monitoring data package. Based on preset operating condition change sensing rules, the sampling frequency of the multi-physics sensor network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is dynamically adjusted to obtain a synchronous multi-physics monitoring dataset composed of updated raw time-series data. Each synchronous multi-physics monitoring dataset is stored chronologically to construct a historical monitoring dataset. Step S2: For the historical monitoring dataset, extract predefined native feature parameters related to component failure mechanisms from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values, and generate associated feature parameters based on a cross-physics coupling model; fuse and reduce the dimensionality of the native feature parameters and associated feature parameters to construct a fused health index characterizing the overall health status of the electrical components; store each constructed fused health index in chronological order to form a historical fused health index sequence; associate and store the historical monitoring dataset and the historical fused health index sequence together to form a cableway electrical component health knowledge base; Step S3: Obtain the historical fused health index sequence and simultaneously input the historical fused health index sequence into the pre-trained physical degradation baseline model and the data-driven residual prediction model; the physical degradation baseline model outputs the baseline remaining useful life prediction value based on the failure physics equation based on the historical fused health index sequence; the data-driven residual prediction model outputs the prediction deviation correction amount for the prediction result of the physical degradation baseline model based on the historical fused health index sequence; by fusing the baseline remaining useful life prediction value and the prediction deviation correction amount, the remaining useful life probability distribution of the electrical component is calculated. Step S4: Continuously inject the synchronous multiphysics monitoring dataset, historical fused health indicator sequences, and actual component maintenance records obtained from the cableway maintenance system into the cableway digital twin. Based on the comparison results between the remaining useful life probability distribution and the component failures recorded in the actual component maintenance records, trigger incremental learning and updates of the data-driven residual prediction model. Simultaneously, based on the historical monitoring dataset, perform data pattern distribution change detection on the synchronous multiphysics monitoring dataset. When it is identified that the data pattern in the synchronous multiphysics monitoring dataset exceeds the data pattern distribution range of the historical monitoring dataset, generate a new fault mode warning, thereby completing the online optimization of the data-driven residual prediction model and the expansion of the cableway electrical component health knowledge base.

2. The method for cableway full lifecycle health management based on multimodal AI according to claim 1, characterized in that: In step S1, the specific process of synchronously acquiring the raw time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration is as follows: At key electrical component nodes in the cableway's electrical system, such as the DC bus of the drive frequency converter, the insulated gate bipolar transistor module, and the control relay contacts, a multi-physics field sensing network consisting of high-frequency current and voltage sensors, patch thermocouple sensors, and high-frequency vibration accelerometer sensors is deployed. All sensors in this multi-physics field sensing network are connected to a synchronous acquisition controller, which has a built-in high-precision temperature-controlled crystal oscillator clock source. For every global hardware trigger pulse generated by the high-precision constant temperature crystal oscillator clock source, the synchronous acquisition controller simultaneously sends a start sampling command to all connected sensors. At the same moment that the high-frequency current and voltage sensor, the patch thermocouple sensor, and the high-frequency vibration accelerometer sensor receive the start sampling command, they respectively collect the instantaneous voltage, current, temperature, and vibration acceleration values ​​of their respective nodes. This ensures that the instantaneous voltage, current, temperature, and vibration acceleration values ​​obtained in each sampling are marked with the same precise timestamp corresponding to the global hardware trigger pulse. The instantaneous voltage, current, temperature, and vibration acceleration values, which are collected from the same global hardware trigger pulse and have the same timestamp, are collectively used to form the initial monitoring data packet.

3. The method for cableway full life-cycle health management based on multimodal AI according to claim 2, characterized in that: Based on preset operating condition change sensing rules, the process of dynamically adjusting the sampling frequency of the multi-physics sensing network for instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is as follows: A base sampling frequency, a higher sampling frequency than the base sampling frequency, a complexity threshold, a complexity change rate threshold, and a stable period number are predefined. For the instantaneous voltage, current, temperature, and vibration acceleration sequences in the initial monitoring data packet, the following operations are performed respectively: The comprehensive complexity index of this type of instantaneous voltage, current, temperature, or vibration acceleration sequence within the current sliding time window is calculated. The specific calculation process is as follows: First, based on a predefined maximum scale factor, different scale factors are selected to coarse out the instantaneous voltage, instantaneous current, instantaneous temperature, or instantaneous vibration acceleration sequences to construct subsequences at different time scales. Then, the permutation entropy value of each subsequence is calculated. Finally, the arithmetic mean of the permutation entropy values ​​corresponding to all scale factors is calculated, and this arithmetic mean is used as a comprehensive complexity index. Simultaneously, the absolute value of the difference between the comprehensive complexity index of the current sliding time window and the comprehensive complexity index of the previous sliding time window is calculated as the complexity change rate. When any one of the voltage instantaneous value sequence, current instantaneous value sequence, temperature instantaneous value sequence, and vibration acceleration instantaneous value sequence has a comprehensive complexity index exceeding a predefined complexity threshold, or its complexity change rate exceeding a predefined complexity change rate threshold, the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics field sensing network in the next acquisition cycle from the basic sampling frequency to a high sampling frequency. After adjusting the sampling frequency to a high sampling frequency, the comprehensive complexity index and complexity change rate are calculated for each type of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequence for each subsequent acquisition cycle. If the number of consecutive calculations reaches a stable number of cycles, and within these consecutive stable acquisition cycles, the comprehensive complexity index of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity threshold, and the complexity change rate of all instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration sequences is lower than the complexity change rate threshold, then the synchronous acquisition controller will adjust the sampling frequency of all sensors in the multi-physics sensing network from the high sampling frequency back to the basic sampling frequency. Through this process of dynamically adjusting the sampling frequency, a synchronous multiphysics monitoring dataset is obtained in each acquisition cycle. This synchronous multiphysics monitoring dataset contains instantaneous voltage, current, temperature, and vibration acceleration values ​​with a unified timestamp, and is accompanied by an identifier of the actual sampling frequency used in that acquisition cycle. The value of the actual sampling frequency identifier is either the basic sampling frequency or the high sampling frequency. The synchronous multiphysics monitoring datasets obtained in each acquisition cycle are stored in the order of their timestamps, thereby constructing and continuously updating the historical monitoring dataset.

4. The method for cableway full life-cycle health management based on multimodal AI according to claim 3, characterized in that: In step S2, the specific process of extracting predefined native feature parameters related to the component failure mechanism from the time-series data of instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration values ​​is as follows: For the instantaneous voltage, current, temperature, and vibration acceleration values ​​synchronously acquired with a unified timestamp in each acquisition cycle of the historical monitoring dataset constructed in step S1, multiple instantaneous voltage values ​​arranged in chronological order within the same acquisition cycle are formed into a voltage instantaneous value sequence, multiple instantaneous current values ​​into a current instantaneous value sequence, multiple instantaneous temperature values ​​into a temperature instantaneous value sequence, and multiple instantaneous vibration acceleration values ​​into a vibration acceleration instantaneous value sequence. Multiple native feature parameters characterizing component degradation are extracted from these sequences. For the voltage and current instantaneous value sequences, their nonlinear dynamic characteristics are calculated. This includes multi-scale permutation entropy; for the instantaneous temperature value sequence, its rising slope and spatial distribution gradient over time are calculated; for the instantaneous vibration acceleration value sequence, its spectral kurtosis within a predefined frequency band related to the bearing fault characteristic frequency is extracted. All native feature parameters are selected based on predefined physical failure mechanisms. The multi-scale permutation entropy is obtained by calculating the average permutation entropy of the subsequences of the signal at different time scales, and the spectral kurtosis is calculated by analyzing the fourth-order cumulant of the vibration signal in the frequency domain. After extraction, multiple native feature parameters extracted from the instantaneous voltage, instantaneous current, instantaneous temperature, and instantaneous vibration acceleration value sequences are arranged according to a predefined feature order to form the native feature parameter set corresponding to the acquisition period.

5. The method for cableway full life-cycle health management based on multimodal AI according to claim 4, characterized in that: The specific process of generating associated feature parameters based on the cross-physics coupling relationship model, and fusing and dimensionality-reducing the original feature parameters and associated feature parameters to construct the fused health index is as follows: A cross-physics coupling model of electro-thermal-vibration relationships is constructed. This model takes native feature parameter elements from the native feature parameter set used to calculate the effective value of the current as input, and calculates multiple associated feature parameters through predefined coupling equations. The associated feature parameters include two types: The first type is the electro-thermal coupling characteristic parameter, whose value is obtained by squaring the effective value of the current in the instantaneous current value sequence, and then multiplying the result of the squaring operation by a preset scaling factor and a preset thermal resistance parameter based on the component material and structure. The second category is thermal-vibration coupling characteristic parameters, whose values ​​are obtained by multiplying the gradient of the instantaneous temperature value sequence with the linear thermal expansion coefficient of the component material, the elastic modulus of the component material, and the characteristic area determined according to the component structure. The calculated associated characteristic parameters are arranged in a predefined order to form an associated characteristic parameter set. The original feature parameter set and the associated feature parameter set are concatenated sequentially to form a high-dimensional hybrid feature set containing all original and associated feature parameters. Each high-dimensional hybrid feature set and its corresponding component health status label are defined as a training sample. The training samples are input into a deep convolutional autoencoder for fusion and dimensionality reduction. This deep convolutional autoencoder processes the input high-dimensional hybrid feature set and outputs a low-dimensional fused health index. During training, this deep convolutional autoencoder introduces a health status metric loss function, the calculation process of which is as follows: In each batch of the training process, for each training sample in that batch, several of its nearest neighbor training samples are selected to form a training sample pair. For each training sample pair, the Euclidean distance between the low-dimensional fused health indicators obtained after compression by the autoencoder is calculated, and the absolute value of the difference between the pre-acquired component health status label values ​​corresponding to the two training samples in that training sample pair is calculated. Then, the difference between the result of multiplying the absolute value of the Euclidean distance and the label value difference by a preset scaling factor is calculated, and the square of this difference is used as the loss contribution of this training sample pair. Finally, the loss contributions of all training sample pairs of all training samples in a training batch are summed to obtain the total health status metric loss value of that batch. Finally, the fusion health indicators generated in each collection cycle are stored in chronological order to form a historical fusion health indicator sequence. Then, the historical monitoring dataset, the historical fusion health indicator sequence, and the actual sampling frequency identifier recorded in the synchronous multiphysics monitoring dataset from step S1 are associated and mapped through a unified timestamp shared by the three. The complete data set after this association and mapping is then persistently stored, thereby forming a health knowledge base for cableway electrical components.

6. The method for cableway full life-cycle health management based on multimodal AI according to claim 5, characterized in that: In step S3, the specific process of inputting the historical fusion health index sequence into the pre-trained physical degradation benchmark model is as follows: First, a physical degradation benchmark model based on the key failure mechanism of components is constructed. This physical degradation benchmark model is based on the Arrennis-type failure physical equation that describes the electrothermal aging law of insulating materials. Subsequently, the historical fusion health index sequence from the cableway electrical component health knowledge base is used as the input to the physical degradation benchmark model; the physical degradation benchmark model maps the instantaneous absolute temperature and instantaneous voltage value sequences required by the model equations based on the input historical fusion health index sequence. Based on the instantaneous absolute temperature value sequence and the instantaneous voltage value sequence, the cumulative aging amount from the initial moment to the current moment is calculated by integration. When the cumulative aging amount reaches the failure threshold preset according to the component material characteristics, the component is determined to fail, and the predicted time length from the current moment to the cumulative aging amount reaching the failure threshold is used as the baseline remaining useful life prediction value. Meanwhile, based on the errors or preset empirical values ​​in the process of identifying parameters of the physical degradation benchmark model, the benchmark prediction uncertainty corresponding to the benchmark remaining useful life prediction value is estimated. The benchmark remaining useful life prediction value and the benchmark prediction uncertainty together constitute the output of the physical degradation benchmark model.

7. The method for cableway full life-cycle health management based on multimodal AI according to claim 6, characterized in that: The specific process of inputting the historical fused health indicator sequence into the data-driven residual prediction model and fusing the baseline remaining useful life prediction value to calculate the remaining useful life probability distribution is as follows: First, a data-driven residual prediction model is constructed, which takes a historical fusion health index sequence as input and its training objective is to learn the prediction residuals of the physical degradation benchmark model. During the training phase, the predicted residual corresponding to each historical moment is calculated from the cableway electrical component health knowledge base using recorded data containing known actual failure times. The predicted residual is the difference between the actual remaining useful life recorded from that historical moment and the baseline remaining useful life predicted by the physical degradation baseline model at the same historical moment. The data-driven residual prediction model is trained using the historical fused health index sequence and its corresponding predicted residual as training samples. In the application phase, the current historical fused health index sequence is input into the pre-trained data-driven residual prediction model. The data-driven residual prediction model outputs a prediction bias correction amount for the current moment and the correction uncertainty corresponding to the prediction bias correction amount. Then, the baseline remaining useful life prediction and its baseline prediction uncertainty output by the physical degradation baseline model, together with the prediction bias correction and its correction uncertainty output by the data-driven residual prediction model, are input into a Bayesian fusion framework. In this Bayesian fusion framework, a prior Gaussian probability distribution is constructed using the baseline remaining useful life prediction value as the mean and the square of the baseline prediction uncertainty as the variance. A likelihood Gaussian probability distribution is constructed using the sum of the baseline remaining useful life prediction value and the prediction deviation correction amount as the mean and the square of the correction uncertainty as the variance. Based on the prior Gaussian probability distribution and the likelihood Gaussian probability distribution, the parameters of the posterior Gaussian probability distribution are calculated through Bayesian inference. The mean and variance of the posterior Gaussian probability distribution are calculated as follows: the posterior mean equals the quotient obtained by dividing the baseline remaining useful life prediction value by the square of the baseline prediction uncertainty, and the quotient obtained by dividing the sum of the baseline remaining useful life prediction value and the prediction deviation correction amount by the square of the correction uncertainty. These two are then added together and divided by the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The posterior variance equals the reciprocal of the sum of the reciprocal of the square of the baseline prediction uncertainty and the reciprocal of the square of the correction uncertainty. The mean of this posterior Gaussian probability distribution is the best estimate of the remaining useful life after fusion; finally, this posterior Gaussian probability distribution is characterized as the probability distribution of the remaining useful life of electrical components.

8. The method for cableway full life-cycle health management based on multimodal AI according to claim 7, characterized in that: In step S4, the specific process of triggering incremental learning and updating of the data-driven residual prediction model based on the comparison between the remaining useful life probability distribution and the actual component maintenance records is as follows: First, in the cableway digital twin, the synchronous multiphysics monitoring dataset from step S1, the historical fused health indicator sequence from step S2, and the remaining useful life probability distribution from step S3 are continuously received and stored. When the actual component maintenance record that records the actual failure or maintenance operation of a component is obtained from the cableway maintenance system, the actual remaining useful life of the component is extracted from the record. Based on the remaining useful life probability distribution, calculate the probability density of the actual remaining useful life of the component falling within the range of possible values ​​described by the remaining useful life probability distribution, and use it as the relative likelihood value. Simultaneously, based on the remaining useful life probability distribution and a set of predefined maintenance decision rules, the decision uncertainty measure for making a definite maintenance decision for the electrical component at the current moment is calculated. The decision uncertainty measure is specifically the decision entropy, which is calculated as follows: First, a predefined set of maintenance decision actions containing multiple specific maintenance actions is defined; then, for each maintenance action in the set, the probability of taking the action is calculated based on all historical data on which the remaining useful life probability distribution is based, according to the remaining useful life probability distribution; finally, the product of the probability of each action and its logarithmic value is summed, and the negative value is taken. The result is the decision entropy. If the relative likelihood value is lower than a preset likelihood threshold, or the decision entropy is higher than a preset decision entropy threshold, it is determined to be a prediction uncertainty event that needs to be learned, and this is used as a trigger signal. When the trigger signal is generated, the historical fusion health indicator sequence associated with the current prediction time, the synchronous multiphysics monitoring dataset, and the real remaining useful life of the component extracted from the actual component maintenance record will be used together to form a new training sample with a real label. Then, in the simulation environment provided by the cableway digital twin, the historical fused health index sequence and synchronous multiphysics monitoring dataset in the new training sample are used to drive the digital twin model to perform state inference and generate the simulation degradation trajectory corresponding to the sample in order to verify the consistency between data characteristics and physical processes. Finally, the new training samples and their corresponding simulated degradation trajectories are submitted to an online incremental learning process for safe and incremental updates and optimizations of the parameters of the data-driven residual prediction model.

9. The method for cableway full life-cycle health management based on multimodal AI according to claim 8, characterized in that: The specific process of detecting changes in data pattern distribution on the synchronous multiphysics monitoring dataset based on historical monitoring datasets and generating new fault mode early warnings to expand the knowledge base is as follows: While the incremental learning process of the model is executed in parallel, a continuously running data pattern distribution change detection process is started. The process uses the historical monitoring dataset constructed in step S1 as the data source of the baseline distribution. The historical monitoring dataset contains all historical synchronous multiphysics monitoring datasets generated by electrical components under known health conditions and known degradation modes. For the newly injected cableway digital twin, the real-time synchronous multiphysics monitoring dataset from step S1 is used as the input data to be detected. A deep kernel single-class classification model is used to calculate its novelty score relative to the benchmark distribution represented by the historical monitoring dataset. The deep kernel uniclass classification model comprises a deep feature extraction network, a predefined kernel function, a set of support vector samples learned from historical monitoring datasets, a set of weight coefficients corresponding to the support vector samples, and a decision threshold. The deep feature extraction network maps the input synchronous multiphysics monitoring dataset samples to a high-dimensional feature space. The process of calculating the novelty score in the deep kernel uniclass classification model is as follows: First, the deep feature extraction network maps the current input real-time synchronous multiphysics monitoring dataset samples and each support vector sample to a high-dimensional feature space. Then, in the high-dimensional feature space, the predefined kernel function is used to calculate the kernel function similarity value between the current input sample and each support vector sample. Next, the kernel function similarity value corresponding to each support vector sample is multiplied by the weight coefficient corresponding to that support vector sample to obtain a set of weighted similarity values. Subsequently, the weighted similarity values ​​corresponding to all support vector samples are summed to obtain a comprehensive similarity value. Finally, the decision threshold is subtracted from the comprehensive similarity value, and the difference is the novelty score. If the novelty score is less than zero, the current data pattern is determined to be outside the distribution range of data patterns in the historical monitoring dataset and is identified as a potential new fault mode. During the detection process, if the novelty score of the current synchronous multiphysics monitoring dataset is less than zero, it is determined that the current data pattern exceeds the data pattern distribution range of the historical monitoring dataset and is identified as a potential new fault mode. When a potential new failure mode is identified, a new failure mode warning is immediately generated. The warning includes the synchronous multiphysics monitoring dataset that triggered the identification, the corresponding fused health index, the timestamp, and the calculated novelty score. Subsequently, this new fault mode warning information is treated as a new knowledge entry with the tag "Pending Review - Suspected New Mode", associated with its corresponding timestamp, and added to the cableway electrical component health knowledge base.

10. A cableway full lifecycle health management system based on multimodal AI, applied to the cableway full lifecycle health management method based on multimodal AI as described in any one of claims 1-9, characterized in that: Specifically, it includes: The system includes a multi-physics synchronous acquisition and adaptive sampling module, a multi-physics feature fusion and health knowledge base construction module, a physics-data hybrid-driven remaining lifespan probability prediction module, and a digital twin-driven model optimization and knowledge expansion module. Multi-physics synchronous acquisition and adaptive sampling module: It is configured to respond to the online monitoring start command of the cableway electrical system, control the sensor network deployed at key component nodes to synchronously acquire voltage, current, temperature and vibration data, and add a unified time stamp to the data; it dynamically adjusts the acquisition frequency based on preset sensing rules, outputs synchronous multi-physics monitoring dataset, and archives it to form historical monitoring dataset; Multiphysics Feature Fusion and Health Knowledge Base Construction Module: It is configured to extract and fuse features from historical monitoring datasets, generate fused health indicators that characterize the health status of components, and associate historical monitoring datasets with the sequentially stored sequence of fused health indicators to construct a health knowledge base for cableway electrical components; The physics-data hybrid driven remaining useful life probability prediction module is configured to obtain historical fused health indicator sequences from a knowledge base and perform collaborative prediction using a pre-trained physical model and a data-driven model, and output the remaining useful life probability distribution of electrical components after fusion processing. The digital twin-driven model optimization and knowledge expansion module is configured to continuously receive monitoring data, health indicators, and actual maintenance records in the cableway digital twin; trigger incremental learning optimization of the data-driven model based on the comparison between the life prediction distribution and the actual records; and simultaneously perform abnormal pattern detection on real-time monitoring data based on historical datasets, and generate early warning information to expand the knowledge base when new fault modes are identified.

Images