Operation and maintenance equipment health situation assessment method combining edge perception and digital twinning

Abstract

The present application relates to the technical field of equipment health assessment, and specifically provides a kind of operation and maintenance equipment health situation assessment method fusing edge perception and digital twinning, comprising: real-time collection operating data using edge perception node, operating data includes multi-source state perception parameter and operating condition parameter;According to operating condition parameter, a plurality of operating condition partitions are obtained by dividing multi-source state perception parameter, each operating condition partition is associated and fused to obtain cross-modal associated features and multi-source fusion features;Obtain each operating condition partition dynamic baseline, calculate normalized deviation;Based on multi-source fusion features and normalized deviation, construct digital twinning model, output state prediction result and prediction residual;Normalized deviation, state prediction result and prediction residual are input into pre-trained health situation assessment model, and output health grade, performance degradation trend and fault risk prediction result, to realize the automatic quantitative evaluation of equipment health situation, improve the accuracy and timeliness of evaluation and prediction.

Description

Technical Field

[0001] This invention relates to the field of equipment health assessment technology, specifically to a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins. Background Technology

[0002] With the deepening of industrial digital and intelligent transformation, modern industrial systems are placing higher demands on the reliability, availability, and maintenance costs of equipment operation and maintenance. The health status of maintained equipment is directly related to the continuity and security of the production system. Therefore, real-time perception, dynamic assessment, and visualization of equipment health status have become core aspects of intelligent industrial operation and maintenance. However, in actual industrial operation and maintenance scenarios, due to the wide distribution of equipment and the complexity of the operating environment, existing operation and maintenance models still face the following prominent problems: data collection relies on centralized architecture, which is prone to delays, resulting in insufficient real-time performance; equipment health status assessment relies heavily on manual experience, which is highly subjective and inconsistent; key indicators and potential hazards are difficult to perceive in real time, resulting in low early fault identification rates; the health status lacks global visualization, making it impossible to intuitively grasp the overall operating status of the equipment; and fault prediction and maintenance decisions lack sufficient data support, resulting in insufficient scientific operation and maintenance. These pain points severely restrict the improvement of operation and maintenance efficiency and quality.

[0003] To address the aforementioned pain points, industry experts have proposed integrating edge sensing technology with digital twin technology to assess the health status of maintenance equipment. Specifically, this approach typically involves deploying edge sensing nodes at key locations on the equipment to collect operational data such as vibration, temperature, current, and pressure. The collected data is then transmitted to the cloud or server to build a digital twin model of the equipment, mapping the physical device to the virtual model. Finally, based on this twin model, the equipment's health status is assessed, and alarms or maintenance recommendations are generated. This technical approach aims to reduce transmission pressure by utilizing localized processing at the edge and to provide an intuitive display and assisted assessment of equipment status through digital twins.

[0004] However, the accuracy of health status assessment and fault prediction of existing integrated assessment schemes using edge sensing technology and digital twin technology still needs to be improved. The main reasons are as follows: (1) The multi-source data fusion stage does not consider the coupling alignment of operating conditions: Existing schemes often use direct splicing or simple weighted averaging for data with different time scales and physical semantics, such as vibration, temperature, current, and pressure. They do not consider the coupling relationship and unified representation alignment mechanism between data under different operating conditions, which makes the extracted features used for assessment and prediction unable to accurately reflect the actual operating status and degradation law of the equipment. Feature distortion directly affects the accuracy of subsequent modeling and assessment. (2) The health assessment stage does not consider the dynamic correlation between key data: When conducting health status assessment, existing schemes often rely on fixed thresholds or isolated judgments from a single data source. They do not consider the dynamic deviation correlation between real-time sensing data and the historical operating baseline of the equipment, nor do they consider the physical semantic correlation between multi-source features (such as vibration and temperature, current and pressure). This results in the assessment results failing to fully reflect the actual degradation trend and potential fault risk of the equipment. The assessment is one-sided and has a high false alarm rate.

[0005] To address these issues, this invention provides a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins, in order to solve the problems in existing technologies.

[0007] One embodiment of the present invention provides a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins, comprising the following steps: The operation data of the operation and maintenance equipment is collected in real time by using edge sensing nodes deployed on the operation and maintenance equipment. The operation data includes multi-source status sensing parameters and operating condition parameters. The multi-source state perception parameters are divided according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters. The multi-source state perception parameters of each operating condition partition are then correlated and fused to obtain the cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The dynamic baseline of each working condition partition is obtained, and the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline is calculated. The dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. A digital twin model is constructed based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. The normalized deviation, state prediction results, and prediction residuals are input into the pre-trained health status assessment model to obtain the health status assessment index of the maintenance equipment. Based on the health status assessment index, the health level, performance degradation trend, and fault risk prediction results of the maintenance equipment are output.

[0008] This invention also relates to a health status assessment system for operation and maintenance equipment that integrates edge sensing and digital twins, comprising: The data acquisition module is used to collect the operation data of the operation and maintenance equipment in real time using the edge sensing nodes deployed on the operation and maintenance equipment. The operation data includes multi-source status sensing parameters and operation condition parameters. The parameter processing module is used to divide the multi-source state perception parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters, and to perform correlation and fusion processing on the multi-source state perception parameters of each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The deviation calculation module is used to obtain the dynamic baseline of each working condition partition and calculate the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline. The dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. The status prediction module is used to construct a digital twin model based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. The health assessment module is used to input the normalized deviation, state prediction results and prediction residuals into the pre-trained health status assessment model to obtain the health status assessment index of the operation and maintenance equipment, and output the health level, performance degradation trend and fault risk prediction results of the operation and maintenance equipment based on the health status assessment index.

[0009] The above embodiments provide a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins, which has the following beneficial effects: This invention reduces data latency caused by centralized data collection by using edge sensing nodes to collect real-time operational data from maintenance equipment. It divides multi-source state sensing parameters into operational condition partitions based on operating condition parameters and performs correlation and fusion processing on each partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each partition. This achieves unified representation and alignment of data with different physical semantics and time scales under operational condition coupling conditions, avoiding feature distortion caused by direct splicing or simple weighted averaging. Furthermore, it calculates the normalized bias of the cross-modal correlation features relative to the corresponding dynamic baseline based on the dynamic baseline of each operational condition partition. The system establishes a dynamic deviation correlation between real-time sensing data and historical operational baselines. Based on multi-source fusion features and normalized deviation, it constructs a digital twin model to dynamically map the operating status and degradation trend of maintenance equipment, outputting status prediction results and prediction residuals. Finally, the normalized deviation, status prediction results, and prediction residuals are input into a pre-trained health status assessment model to output the health level, performance degradation trend, and fault risk prediction results of maintenance equipment. This achieves automation and quantification of health status assessment, improves the accuracy and timeliness of assessment and prediction, reduces false alarms and missed alarms, and provides scientific data support for maintenance decisions. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins, as provided in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0012] Reference Figure 1 One embodiment of the present invention provides a method for assessing the health status of maintenance equipment by integrating edge awareness and digital twins, comprising the following steps: S10. Real-time collection of operational data of the maintenance equipment using edge sensing nodes deployed on the maintenance equipment, the operational data including multi-source status sensing parameters and operational condition parameters; S20. Divide the multi-source state perception parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters, and perform correlation and fusion processing on the multi-source state perception parameters of each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. S30. Obtain the dynamic baseline of each working condition partition, and calculate the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline, wherein the dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. S40. Construct a digital twin model based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. S50. Input the normalized deviation, state prediction result and prediction residual into the pre-trained health status assessment model to obtain the health status assessment index of the operation and maintenance equipment, and output the health level, performance degradation trend and fault risk prediction result of the operation and maintenance equipment based on the health status assessment index.

[0013] In this embodiment, specifically: As described in steps S10-S20 above, this solution first utilizes edge sensing nodes deployed on the maintenance equipment to collect real-time operational data. Taking the main transformer of a substation under the jurisdiction of a power grid as an example, edge sensing nodes are deployed at key locations such as windings, bushings, and tap changers. The collected operational data includes two categories: multi-source status sensing parameters and operating condition parameters. Multi-source status sensing parameters cover physical quantities reflecting the equipment's operating status, such as vibration signals, winding temperature, load current, and partial discharge signals. Operating condition parameters cover physical quantities characterizing the equipment's operating conditions, such as load rate, operating voltage, ambient temperature, tap changer position, and grid-connected operation mode. By completing data collection and preliminary processing at the edge, this solution avoids the transmission delay caused by sending all raw data back to the cloud, ensuring the real-time nature of the health status assessment.

[0014] After acquiring operational data, this solution divides the multi-source state sensing parameters into operating condition zones based on operating condition parameters. Multi-source state sensing parameters collected under the same or similar operating conditions are grouped into the same operating condition zone, thus obtaining several operating condition zones with corresponding multi-source state sensing parameters. Taking a 220kV main transformer as an example, the coupling relationships between various modal parameters differ significantly under different operating conditions such as rated load, light load, overload, and high ambient temperature. Therefore, it is necessary to partition the multi-source state sensing parameters according to operating conditions to avoid data confusion between different operating conditions.

[0015] Subsequently, this scheme performs correlation and fusion processing on the multi-source state perception parameters within each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The cross-modal correlation features characterize the correlation relationships between different modal parameters, while the multi-source fusion features comprehensively reflect the overall operating status of the equipment under that operating condition partition. The principle is that the physical semantic coupling relationships between modal parameters of maintenance equipment differ fundamentally under different operating conditions. If the data from all time periods is directly spliced ​​or weighted without distinguishing operating conditions, the correlation patterns under different operating conditions will be confused, resulting in the extracted features failing to accurately reflect the true operating status and degradation patterns of the equipment under specific operating conditions. By partitioning first and then fusing, this scheme achieves unified representation alignment of data with different physical semantics and time scales under operating condition coupling conditions, effectively avoiding feature distortion caused by direct splicing or simple weighted averaging, and providing an accurate feature foundation for subsequent dynamic baseline establishment and digital twin modeling.

[0016] As described in step S30 above, this scheme obtains the dynamic baseline of each operating condition zone and calculates the normalized deviation of the cross-modal correlation characteristics of each operating condition zone relative to the corresponding dynamic baseline. Taking a 220kV main transformer in a substation under a power grid as an example, different operating condition zones such as rated load steady state, light load, overload, and ambient high temperature correspond to their respective dynamic baselines. This dynamic baseline represents the normal correlation mode between modal parameters when the equipment is operating normally under the corresponding operating condition zone. The dynamic baseline of each operating condition zone is dynamically adjusted based on the historical operating data of the maintenance equipment to adapt to the slow changes in the operating status of the equipment over time. After obtaining the dynamic baseline, this scheme compares the current cross-modal correlation characteristics with the dynamic baseline of the corresponding operating condition zone and calculates the normalized deviation corresponding to the operating condition zone. This deviation is used to quantify the degree of deviation of the current state from the normal correlation mode of the operating condition zone. For example, when the current vibration and temperature correlation under the rated load steady state zone deviates significantly from the normal correlation mode represented by the dynamic baseline of the zone, the corresponding normalized deviation increases, reflecting that the current state of the equipment has deviated from the normal operating range.

[0017] The principle behind this solution lies in the fact that the normal correlation patterns of maintenance equipment differ fundamentally across different operating condition zones. If a uniform, fixed threshold is used to determine the status of all operating conditions, misjudgments may occur due to these differences. Furthermore, the normal operating status of equipment is not static but slowly drifts with aging, environmental factors, and other influences, making a fixed baseline unsuitable for adapting to these changes. By configuring dynamic baselines for each operating condition zone and dynamically adjusting them based on historical operating data, while normalizing the deviation of current cross-modal correlation features relative to the corresponding baseline into a uniform and comparable deviation, this solution establishes a dynamic deviation correlation between real-time sensing data and the equipment's historical operating baseline. This replaces the traditional approach of relying on fixed thresholds or isolated judgments from a single data source, providing accurate anomaly measurement signals for subsequent digital twin modeling and health status assessment.

[0018] As described in steps S40-S50 above, this scheme constructs a digital twin model based on the multi-source fusion characteristics and normalized deviation corresponding to each operating condition zone. Taking a 220kV main transformer in a substation under a power grid as an example, the digital twin model is based on the multi-source fusion characteristics under various operating conditions such as rated load steady state, light load, overload, and high ambient temperature. It learns the evolution law of the normal operating state of the equipment under the corresponding operating condition zone and perceives the difference between the current state and the normal mode by combining the normalized deviation. The constructed digital twin model is used to dynamically map the operating state and degradation trend of the maintenance equipment, and outputs the state prediction result and prediction residual at the current moment. The state prediction result reflects the state that the equipment should have at the current moment under the normal evolution law, and the prediction residual reflects the difference between the actual state and the normal predicted state. For example, when the main transformer windings experience early overheating degradation, the difference between the normal state predicted by the digital twin model and the actual observed state increases, the prediction residual increases accordingly, and the normalized deviation also shows an increasing trend. Both together characterize the degree of degradation of the equipment.

[0019] After obtaining the state prediction results and prediction residuals, this scheme inputs the normalized deviation, state prediction results, and prediction residuals into a pre-trained health status assessment model to obtain health status assessment indicators for the maintenance equipment. Based on these indicators, it outputs the health level, performance degradation trend, and fault risk prediction results of the maintenance equipment. The health status assessment model is pre-trained based on the historical operating data of the maintenance equipment and can comprehensively assess the equipment's health status from three dimensions: the current degree of abnormality reflected by the normalized deviation, the expected normal evolution reflected by the state prediction results, and the actual degradation magnitude reflected by the prediction residuals. For example, when the normalized deviation of a 220kV main transformer continues to increase and the prediction residuals show an upward trend, the health level output by the health status assessment model is downgraded from normal to warning, the performance degradation trend shows that the winding temperature-related indicators are slowly deteriorating, and the fault risk prediction results indicate the risk of overheating faults.

[0020] The principle behind this approach is that the digital twin model learns the evolutionary patterns of equipment's normal operating state, transforming the degree of degradation into a residual between the actual state and the predicted normal state. This avoids the limitations of traditional methods that rely on a large number of fault samples for supervised learning. The health status assessment model integrates three complementary types of information: normalized deviation, state prediction results, and prediction residuals. The deviation reflects the static difference between the current state and the normal baseline, the prediction results reflect the dynamic expectation of normal evolution, and the residuals reflect the real-time magnitude of actual degradation. The combination of these three elements comprehensively depicts the health status of the equipment, avoiding biased assessments and false alarms caused by isolated judgments from a single data source. Through the cascading collaboration of the digital twin model and the health status assessment model, this solution achieves a dynamic mapping between the operating status and degradation trends of maintenance equipment, as well as automated and quantitative output of health levels, performance degradation trends, and fault risks. This effectively improves the accuracy, consistency, and timeliness of health status assessment and fault prediction, providing sufficient and scientific data support for industrial equipment maintenance decisions.

[0021] In summary, this solution ensures the real-time nature of operational data through real-time acquisition by edge sensing nodes, achieves unified representation of multi-source data under coupled operational conditions through condition zoning and correlation fusion, establishes dynamic deviation correlation between real-time status and historical normal patterns through dynamic baseline and normalized deviation, realizes dynamic mapping between the operating status and degradation trend of maintenance equipment through digital twin model and outputs status prediction results and prediction residuals, and achieves multi-dimensional quantitative output of health level, performance degradation trend and fault risk through pre-trained health status assessment model. This effectively solves the problems of multi-source data feature distortion, health assessment relying on fixed thresholds or single data sources, one-sided assessment and high false alarm rate in existing solutions, and realizes full-process automation and quantitative closed loop of health status of maintenance equipment from data acquisition, feature construction, anomaly measurement to assessment output.

[0022] In one embodiment, step S20, which involves dividing the multi-source state sensing parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state sensing parameters, specifically includes: S211. Perform coupling analysis on the operating condition parameters to determine a multi-dimensional operating condition space, and pre-set several operating condition partitions and stable operating condition judgment conditions corresponding to each operating condition partition in the multi-dimensional operating condition space. S212. Monitor the changes in the operating condition parameters in the multi-dimensional operating condition space in real time; S213. When the operating condition parameters meet the stable operating condition determination conditions of any operating condition partition, the multi-source state sensing parameters collected during the duration period that meets the stable operating condition determination conditions are assigned to the operating condition partition to obtain several operating condition partitions with corresponding multi-source state sensing parameters. Among them, the duration period that does not meet any stable operating condition determination conditions is marked as the operating condition transition state, and the multi-source state sensing parameters collected during the duration of the operating condition transition state are marked as transition state data. The transition state data does not participate in the partitioning.

[0023] In this embodiment, specifically: As described in steps S211-S213 above, this solution first performs a coupled analysis of the operating condition parameters to determine a multi-dimensional operating condition space. Taking a 220kV main transformer in a substation under the jurisdiction of a power grid as an example, the operating condition parameters include load rate, operating voltage, ambient temperature, tap changer position, etc. These parameters are coupled with each other. For example, the load rate is positively correlated with the winding temperature rise, and the ambient temperature is positively correlated with the winding temperature. Moreover, the load rate and ambient temperature jointly affect the thermal balance state of the equipment. Therefore, the operating condition cannot be divided based on a single parameter, but a multi-dimensional operating condition space needs to be constructed based on each operating condition parameter. After determining the multi-dimensional operating condition space, this solution pre-sets several operating condition partitions and corresponding stable operating condition judgment conditions for each partition within this space. For example, preset rated load steady-state zones (load rate between 70% and 100%, ambient temperature between -10℃ and 40℃, and lasting for more than 5 minutes), light load zones (load rate below 30% and lasting for more than 5 minutes), overload zones (load rate above 100% and lasting for more than 3 minutes), and high-temperature zones (ambient temperature above 40℃ and load rate above 50% and lasting for more than 5 minutes) can be used. The stable operating condition judgment conditions for each zone include both the value range of the operating condition parameters and the duration requirement of that value range, so as to avoid frequent zone switching caused by short-term fluctuations in operating condition parameters.

[0024] Subsequently, this solution monitors the changes in operating condition parameters in a multi-dimensional operating condition space in real time, continuously tracking the positional migration of each parameter within the space. When an operating condition parameter continuously meets the stable operating condition judgment condition of a certain operating condition zone within a certain time period, the multi-source state sensing parameters collected during that time period are assigned to that operating condition zone. For example, when the load rate of a 220kV main transformer is stable at 85% and the ambient temperature is stable at 25℃ for 8 minutes, the stable operating condition judgment condition of the rated load steady-state zone is met, and the multi-source state sensing parameters such as vibration, winding temperature, and load current collected during those 8 minutes are assigned to the rated load steady-state zone. For durations that do not meet any stable operating condition judgment condition, such as the period during which the load rate rapidly increases from 10% to 80% during equipment startup, or the period during which parameters rapidly migrate between multiple zones during load switching, these are marked as transitional operating conditions. The multi-source state sensing parameters collected during these periods are marked as transitional data and do not participate in the operating condition zone division.

[0025] The principle behind this solution is that the operating status of maintenance equipment is determined by the coupling of multiple operating condition parameters, and a single parameter cannot accurately characterize the operating condition of the equipment. Furthermore, the equipment is in an unstable state during operating condition switching, and the coupling relationships between various modal parameters are not yet stable. Including transitional data in partitions would result in the same partition containing features from different stable operating conditions, contaminating subsequent correlation and fusion processing. By constructing a multi-dimensional operating condition space through coupling analysis, setting stable operating condition judgment conditions with duration requirements, and excluding transitional data from the partitioning, this solution achieves accurate identification and partitioning of stable operating conditions. This ensures that the multi-source state perception parameters within each operating condition partition have a consistent operating condition background, providing a clean and stable data foundation for subsequent correlation and fusion processing within each partition.

[0026] It is worth mentioning that in existing technologies, operating condition division is usually based on only a single operating condition parameter (such as division by load rate) or simple segmentation using a fixed time window, without considering the coupling relationship between multiple parameters or distinguishing between stable operating conditions and transient states. This results in the division results failing to accurately reflect the actual operating conditions of the equipment, and the inclusion of transient state data in the partition features affects the accuracy of subsequent evaluations. In contrast, this embodiment achieves accurate identification and pure partitioning of stable operating conditions through multi-dimensional operating condition spatial coupling analysis, stable operating condition determination conditions with duration requirements, and a transient state data exclusion mechanism, effectively avoiding the problems of coarse operating condition division and transient state interference in existing technologies.

[0027] In one embodiment, step S20, which involves performing correlation and fusion processing on the multi-source state perception parameters of each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition, specifically includes: S221. For each working condition zone, perform time synchronization and alignment processing on the multi-source state perception parameters in the working condition zone to unify them to the same time reference, and obtain the aligned multi-source state perception parameters. S222. Based on the aligned multi-source state perception parameters, calculate the dynamic correlation metric between different modal parameters, construct a cross-modal correlation matrix, and use the cross-modal correlation matrix as the cross-modal correlation feature corresponding to the working condition partition. S223. Based on the aligned multi-source state perception parameters, feature extraction is performed on the state perception parameters of each mode within the same working condition partition to obtain the independent features of each mode. The independent features of each mode are then fused to obtain the multi-source fused features corresponding to the working condition partition.

[0028] In this embodiment, specifically: As described in steps S221-S223 above, this scheme first performs time synchronization and alignment processing on the multi-source state sensing parameters within each operating condition zone. Taking a 220kV main transformer in a substation under the jurisdiction of a power grid as an example, the multi-source state sensing parameters collected in the rated load steady-state zone, such as vibration signals, winding temperatures, load currents, and partial discharge signals, have different sampling frequencies. For example, the vibration signal sampling frequency is 10kHz, the winding temperature sampling frequency is 1Hz, and the load current sampling frequency is 4kHz. The sampling points of each modal parameter on the time axis are not aligned. This scheme unifies each modal parameter to the same time reference through methods well known to those skilled in the art, such as resampling or interpolation, to obtain aligned multi-source state sensing parameters, ensuring that the values ​​of different modal parameters at the same time have a temporal correspondence.

[0029] After completing time synchronization alignment, this scheme calculates the dynamic correlation metric between different modal parameters based on the aligned multi-source state sensing parameters, constructs a cross-modal correlation matrix, and uses this cross-modal correlation matrix as the cross-modal correlation feature corresponding to the operating condition partition. Specifically, for any two modal parameters, an equal-length time series after time synchronization alignment is taken, and a sliding time window is used to calculate the correlation metric between the two modal parameters within the window. The dynamic correlation metric changes with time as the time window slides. For modal parameters with a linear correlation (such as vibration signal and winding temperature), the Pearson correlation coefficient can be used to calculate the correlation metric, and the calculation formula is:

[0030] in, , The two modal parameters are respectively in the first... The value at each sampling time, , These are the mean values ​​of the two modal parameters within the sliding time window, respectively. The value range is [-1, 1], and the larger the absolute value, the stronger the correlation between the two modal parameters; For modal parameters with nonlinear correlations (such as load current and partial discharge signal), mutual information can be used to calculate the correlation metric. The calculation formula is as follows:

[0031] in, The joint probability distribution of the two modal parameters, , These are the marginal probability distributions of the two modal parameters, respectively. The larger the mutual information value, the stronger the nonlinear correlation between the two modal parameters.

[0032] After obtaining the dynamic correlation metrics between each pair of modal parameters, they are arranged in matrix form according to modal order to construct a cross-modal correlation matrix. Taking the rated load steady-state zone as an example, vibration and temperature exhibit a stable positive correlation pattern in this zone, and the corresponding element in the cross-modal correlation matrix remains stable at a high level; however, when equipment degradation occurs, the correlation metrics will shift.

[0033] Meanwhile, based on the aligned multi-source state perception parameters, this scheme extracts features from the state perception parameters of each mode within the same operating condition zone, obtaining independent features for each mode. These independent features are then fused to obtain the multi-source fused features corresponding to that operating condition zone. For example, time-domain features such as root mean square and kurtosis are extracted from vibration signals; features such as mean and rate of change are extracted from winding temperatures; features such as harmonic content and imbalance are extracted from load currents; and features such as the number of discharges and average amplitude are extracted from partial discharge signals. The independent features of each mode are then spliced ​​or weighted and fused. The weights used for weighted fusion can be determined based on the preset importance weights of each mode, or adaptively determined based on the contribution of each mode's features to the equipment's operating state. Taking the splicing and fusion method as an example, the multi-source fusion feature can be represented as a feature vector composed of the root mean square of vibration, vibration kurtosis, average temperature, temperature change rate, current harmonic content, current imbalance, partial discharge number, and average amplitude of partial discharge. This feature vector comprehensively reflects the overall operating status of the equipment under this working condition zone.

[0034] The principle behind this approach is that different modal parameters have different sampling frequencies. Without time synchronization alignment, the correlation measurement between modes will be distorted due to temporal misalignment. The cross-modal correlation matrix can capture the physical semantic coupling relationship between modes such as vibration-temperature and current-partial discharge, reflecting the correlation pattern of the equipment under specific operating conditions. The independent feature extraction of each mode retains the specific information of each mode, and the fusion results in a comprehensive representation that takes into account both inter-modal correlation and intra-modal features. Through the collaborative processing of time synchronization alignment, cross-modal correlation matrix construction, and multi-modal feature fusion, this scheme achieves unified representation alignment of data with different physical semantics and different time scales under operating condition coupling, providing an accurate feature foundation for subsequent dynamic baseline establishment and digital twin modeling.

[0035] It is worth mentioning that in existing technologies, multi-source data fusion typically employs direct splicing or simple weighted averaging, failing to consider the time synchronization issues between different modal parameters or explicitly construct the correlation relationships between modalities. This results in the fused features failing to accurately reflect the actual operating status and degradation patterns of the equipment under specific working conditions. In contrast, this embodiment eliminates timing misalignment through time synchronization alignment, explicitly represents the coupling relationships between modalities through a cross-modal correlation matrix, and balances specificity and integrity through the fusion of independent features extracted from each modality. This effectively avoids feature distortion caused by direct splicing or simple weighted averaging.

[0036] In one embodiment, step S30 specifically includes: S311. For each operating condition partition, based on the cross-modal correlation features generated by the multi-source state perception parameters in the historical operating data that meets the normal operating conditions under that operating condition partition, a dynamic baseline corresponding to that operating condition partition is established, and the dynamic baseline is continuously updated. The dynamic baseline is updated by only including the cross-modal correlation features generated by the multi-source state perception parameters that meet the stable operating condition determination conditions. S312. Obtain the current cross-modal correlation features of each working condition partition; S313. For each working condition partition, the current cross-modal correlation features of the working condition partition are compared element by element with the dynamic baseline corresponding to the working condition partition to obtain the correlation difference between each modal parameter pair. The correlation difference between each modal parameter pair is weighted based on the preset modal importance weight to obtain the weighted correlation difference. S314. Normalize the weighted correlation differences of each modal parameter pair to obtain the normalized deviation corresponding to the working condition partition.

[0037] In this embodiment, specifically: As described in steps S311-S314 above, this scheme first establishes a corresponding dynamic baseline for each operating condition zone. Taking a 220kV main transformer in a substation under a power grid as an example, the dynamic baseline of the rated load steady-state zone is established based on historical operating data that meets the normal operating conditions under that zone: the cross-modal correlation characteristics generated by multi-source state sensing parameters over a past period (e.g., the past three months) are taken, and their statistical mean or median is calculated as the baseline value, forming a baseline matrix that characterizes the normal correlation mode between various modal parameters when the equipment is operating normally under that zone. For example, in the dynamic baseline of the rated load steady-state zone, the correlation coefficient baseline value between vibration signal and winding temperature is 0.85, and the mutual information baseline value between load current and partial discharge signal is 0.6. Once the dynamic baseline is established, it is continuously updated to adapt to the slow changes in equipment operating status over time. During the update, only cross-modal correlation features generated by multi-source state sensing parameters that meet the stable operating condition determination conditions are included. That is, only stable operating condition data with a load rate of 70% to 100% and an ambient temperature of -10℃ to 40℃ for more than 5 minutes are included. Transitional operating condition data are not included in the baseline update to avoid unstable data contaminating the baseline.

[0038] After establishing a dynamic baseline, this scheme acquires the current cross-modal correlation features of each operating condition zone. For each operating condition zone, the current cross-modal correlation features of that zone are compared element-by-element with the corresponding dynamic baseline to obtain the correlation differences between each modal parameter pair. Specifically, both the cross-modal correlation features and the dynamic baseline are in matrix form. Element-by-element comparison involves calculating the difference between corresponding elements in the two matrices. For example, if the correlation coefficient between the current vibration signal and the winding temperature is 0.72 and the baseline value is 0.85, then the correlation difference for that modal parameter pair is 0.13. After obtaining the correlation differences for each modal parameter pair, the correlation differences are weighted based on preset modal importance weights to obtain weighted correlation differences. The modal importance weights are preset according to the sensitivity of each modal parameter pair to the health status of the equipment. For example, the weight of the vibration-temperature pair is 0.3, the weight of the current-partial discharge pair is 0.4, and the weight of the temperature-current pair is 0.3. The weighted correlation difference is the sum of the products of the correlation differences of each modal parameter pair and its corresponding weight. Assume there are a total of [number missing] modal parameters within the operating condition zone. The modal parameter pair, the first The correlation differences of each modal parameter pair are The corresponding preset modal importance weights are Then the weighted correlation difference The calculation formula is:

[0039] in, For the index of the modal parameter pair, .

[0040] Finally, this scheme normalizes the weighted correlation differences of each modal parameter pair to obtain the normalized deviation corresponding to the operating condition partition. Normalization can be achieved by dividing by the norm of the dynamic baseline matrix or by using min-max normalization to map the weighted correlation differences to the [0,1] interval, thus providing a uniform and comparable scale for the deviations of different operating condition partitions and different modal parameter pairs. For example, the current weighted correlation difference of the rated load steady-state partition is 0.08, and after normalization, the normalized deviation is 0.15, indicating a slight deviation from the normal correlation pattern of this partition.

[0041] The principle behind this approach is that the dynamic baseline is established based on normal operating data, representing the normal correlation pattern of the equipment under specific operating conditions. The degree of deviation of the current feature from the baseline reflects the degree of abnormality of the equipment. Only stable operating condition data is included to update the baseline, avoiding baseline drift caused by unstable transitional data. Element-by-element comparison can pinpoint which specific modal parameter pair has changed its correlation, while modal importance weights highlight the contribution of modal pairs that are more sensitive to the health status of the equipment. Normalization eliminates the dimensional differences of different modal correlation measures, making the deviation comparable across different operating condition intervals. Through the collaborative processing of dynamic baseline establishment and updating, element-by-element comparison, weighting, and normalization, this scheme establishes a dynamic deviation correlation between real-time sensing data and the equipment's historical normal baseline, providing accurate anomaly measurement signals for subsequent digital twin modeling and health status assessment.

[0042] It is worth noting that in existing technologies, health assessments typically employ fixed thresholds or isolated judgments from a single data source. These assessments fail to consider the dynamic deviation between real-time sensing data and the equipment's historical operating baseline, nor do they account for the physical and semantic relationships between multiple feature sources. This results in assessments that cannot fully reflect the equipment's true degradation trends and potential failure risks, leading to biased assessments and high false alarm rates. In contrast, this embodiment establishes and dynamically updates baselines for each operating condition zone based on normal historical data, accurately quantifies correlation differences through element-by-element comparison and modal importance weighting, and achieves comparable deviation measurements across operating conditions through normalization processing. This effectively avoids the biased assessments and false alarms caused by fixed thresholds and single data source judgments.

[0043] In one embodiment, step S40, which involves constructing a digital twin model based on the multi-source fusion features and normalized deviations corresponding to each working condition partition, specifically includes: S411. Construct a hierarchical collaborative architecture including a bridging decision unit and several partition state mapping units that correspond one-to-one with each working condition partition, wherein each partition state mapping unit is used to output the state prediction result under the corresponding working condition partition, and the bridging decision unit is used to determine the output method of the state prediction result based on the association matching result. S412. Configure a shared feature extraction layer and an independent evolution mapping layer for each partition state mapping unit. The shared feature extraction layer of each partition state mapping unit is configured with the same preset parameter extraction rules, which are used to extract common basic features in multi-source fusion features. The independent evolution mapping layer of each partition state mapping unit is configured with mutually independent preset learning mapping rules, which are used to learn the normal operation state and evolution law under the corresponding working condition partition based on the common basic features, and output the state prediction result under the corresponding working condition partition. S413. Configure association matching and routing rules for the bridging decision unit, so that the bridging decision unit obtains the current cross-modal association features based on the current multi-source state perception parameters, performs association matching between the cross-modal association features and the normal association distribution pattern represented by the dynamic baseline corresponding to each working condition partition, outputs the current state prediction result according to the matching result, and determines and outputs the prediction residual according to the difference between the current actual state observation value and the current state prediction result. S414. Based on historical operating data of maintenance equipment meeting normal operating conditions, extract multi-source fusion features corresponding to each operating condition partition, train the state mapping unit of each partition, so that each partition state mapping unit learns the normal state evolution law under the corresponding operating condition partition, and obtains the trained digital twin model.

[0044] In this embodiment, specifically: As described in steps S411-S414 above, this scheme first constructs a hierarchical collaborative architecture for the digital twin model. This architecture includes a bridging decision unit and several partition state mapping units that correspond one-to-one with each operating condition partition. Specifically, the hierarchical collaborative architecture adopts a hierarchical organization method of upper-level decision-making and lower-level mapping: the lower layer consists of several partition state mapping units, each corresponding one-to-one with an operating condition partition. Each partition state mapping unit includes a feature extraction part and a state evolution prediction part. Its input is multi-source fusion features. After the feature extraction part extracts common basic features, the state evolution prediction part learns the normal state evolution law under the corresponding operating condition partition and outputs the state prediction result under that partition. The upper layer consists of bridging decision units, which are connected to each partition state mapping unit. The bridging decision unit includes an association matching submodule and a routing output submodule. The association matching submodule is used to calculate the matching degree between the current cross-modal association features and the dynamic baseline of each operating condition partition. The routing output submodule is used to determine the output method of the state prediction result based on the matching result. The bridging decision unit finally outputs the state prediction result and the prediction residual. Both the partition state mapping unit and the bridging decision unit are software functional modules that can be implemented based on machine learning models such as neural networks, support vector regression, and Gaussian process regression. This embodiment uses neural networks as an example. Each partition state mapping unit calculates the state prediction result for its corresponding operating condition partition based on multi-source fusion features. The bridging decision unit determines the operating condition partition corresponding to the current state and the output method of the state prediction result based on the correlation matching result between the current cross-modal correlation features and the dynamic baseline of each operating condition partition. It also determines the prediction residual based on the difference between the actual state observation value and the state prediction result at the current moment. Taking a 220kV main transformer in a substation under a power grid as an example, four partition state mapping units are set up for four operating condition partitions: rated load steady state, light load, overload, and ambient high temperature. Each partition state mapping unit is used to output the state prediction result for its corresponding operating condition partition. Simultaneously, a bridging decision unit is set up to determine the output method of the state prediction result based on the correlation matching result.

[0045] After constructing a hierarchical collaborative architecture, this solution configures a shared feature extraction layer and an independent evolutionary mapping layer for each partition state mapping unit. The shared feature extraction layer of each partition state mapping unit is configured with the same preset parameter extraction rules; that is, the shared feature extraction layer of each partition state mapping unit uses the same network structure and parameters (weight sharing), which can be implemented using fully connected layers or convolutional layers. The preset parameter extraction rules refer to the pre-configuration of the network structure of the shared feature extraction layer, whose weight parameters are learned during subsequent training. The input of the shared feature extraction layer is multi-source fusion features, and the output is common basic features, used to extract basic features common to each operating condition partition that reflect the overall operating status of the equipment from the multi-source fusion features. Taking a 220kV main transformer as an example, the multi-source fusion features are an 8-dimensional feature vector containing vibration root mean square, vibration kurtosis, average temperature, temperature change rate, current harmonic content, current imbalance, partial discharge frequency, and average amplitude of partial discharge. A shared feature extraction layer maps this vector to a 4-dimensional common basic feature vector through nonlinear transformation, extracting common information related to the overall operating status of the equipment. Sharing the same feature extraction rules reduces the number of model parameters, improves the generalization ability of feature extraction, and ensures that the common basic features extracted under different operating conditions have a unified feature space. The independent evolution mapping layers of each partition state mapping unit are configured with mutually independent preset learning mapping rules. That is, the independent evolution mapping layers of each partition state mapping unit adopt mutually independent network structures and parameters, which can be implemented using fully connected regression networks or time-series prediction networks such as LSTM and GRU. The preset learning mapping rules refer to the pre-configuration of the network structure of the independent evolution mapping layers, whose weight parameters are independently learned during subsequent training. The input to the independent evolutionary mapping layer is the common basic features output from the shared feature extraction layer, and the output is the state prediction result for the corresponding operating condition partition. This state prediction result is the feature vector (with the same dimension as the common basic features or multi-source fusion features) corresponding to the predicted normal operating state of the equipment at the current or next time moment, used to characterize the state that the equipment should have under normal evolutionary patterns. Using temporal prediction networks such as LSTM or GRU can capture the temporal dependencies of equipment state evolution. The parameters of each partition's independent evolutionary mapping layer are independent of each other, and can capture the specific patterns of equipment state evolution under different operating conditions such as rated load steady state, light load, overload, and high ambient temperature. The forward propagation process of each partition's state mapping unit is as follows: multi-source fusion features are input into the shared feature extraction layer to obtain common basic features; these common basic features are then input into the independent evolutionary mapping layer of the corresponding partition to obtain the state prediction result for that partition.

[0046] Subsequently, this scheme configures association matching and routing rules for the bridging decision unit. Based on the current multi-source state perception parameters, the bridging decision unit obtains the current cross-modal association features according to the association fusion processing method of steps S221-S223 (time synchronization alignment, calculation of dynamic association metrics, and construction of cross-modal association matrices). These cross-modal association features are then matched with the normal association distribution patterns represented by the dynamic baselines corresponding to each operating condition partition. The dynamic baselines of each operating condition partition are the statistical baselines of the cross-modal association matrices when the equipment is operating normally in the corresponding partition. The normal association distribution pattern is the distribution pattern of the association metrics between each pair of modal parameters in the baseline matrix. Specifically, association matching can be achieved by calculating the similarity between the current cross-modal association features and the dynamic baselines of each operating condition partition: first, the current cross-modal association feature matrix and the dynamic baseline matrices of each partition are expanded into vectors by row, and then the similarity between the current feature vector and the baseline vectors of each partition is calculated using metrics such as cosine similarity, Euclidean distance, or Mahalanobis distance. For example, the cosine similarity calculation formula is:

[0047] in, This is the vectorized form of the current cross-modal correlation features. This is the vectorized form of the dynamic baseline of a certain working condition partition; the working condition partition with the highest similarity is taken as the working condition partition corresponding to the current state, and the output method of the state prediction result is determined according to the matching result, that is, whether to use the output of a single partition state mapping unit or combine the outputs of multiple partition state mapping units, and output the state prediction result at the current time.

[0048] Simultaneously, the bridging decision unit determines and outputs the prediction residual based on the difference between the current actual state observation and the current state prediction result. The current actual state observation is either a multi-source fusion feature obtained from the aforementioned correlation and fusion processing of the multi-source state perception parameters collected at the current time, or a common basic feature further extracted by the shared feature extraction layer; it is a feature vector of the same dimension as the state prediction result. The prediction residual is the difference vector obtained by subtracting the actual state observation from the state prediction result element by element.

[0049] Or it could be the scalar residual in the L2 norm form of the difference vector:

[0050] in, For residual vectors The One element, This is the feature dimension. The prediction residual reflects the difference between the actual state and the normal predicted state.

[0051] Finally, this solution extracts multi-source fusion features corresponding to each operating condition partition based on historical operating data of the maintenance equipment meeting normal operating conditions, and trains the state mapping unit of each partition. Specifically, firstly, following the operating condition partitioning method in steps S211-S213, the historical normal operating data is divided into operating condition partitions, that is, only historical data that meets the stable operating condition judgment conditions of each operating condition partition is assigned to the corresponding partition, and the operating condition transition data is not included in the training; then, following the association fusion processing method in steps S221-S223, the multi-source fusion features corresponding to each partition are extracted as training input. The training label is the actual state at the corresponding time under normal operating conditions, and this actual state is the multi-source fusion feature or common basic feature of the same dimension as the training input; when the training objective is to predict the state at the next time moment, the training label is... The multi-source fusion features at each time point are used as input, and with The actual state at each time step is the label; when the training objective is to reconstruct the state at the current time step, the label is used. The multi-source fusion features at each time point are used as input, and with The actual state at each time step is represented by a label. During training, the mean squared error is used as the loss function, calculated using the following formula:

[0052] in, For training labels, For the model prediction results, The number of training samples. This serves as the sample index. The parameters of each partition state mapping unit are iteratively updated using the backpropagation algorithm, employing optimizers such as stochastic gradient descent (SGD) or Adam, until the loss function meets a preset convergence condition, such as the loss value falling below a preset threshold or failing to decrease after multiple iterations. The parameters of the shared feature extraction layer are shared and updated uniformly across all partition state mapping units, while the parameters of each partition's independent evolution mapping layer are updated independently, enabling each partition state mapping unit to learn the normal state evolution pattern under its corresponding operating condition partition, resulting in a trained digital twin model. During training, only historical data under normal operating conditions is used; fault or degradation data is not introduced. Normal operating conditions can be determined based on equipment operation records, maintenance records, or dynamic baseline deviation thresholds, ensuring that each partition state mapping unit learns the normal evolution trajectory of the equipment under its corresponding operating condition. During runtime, the model predicts the state it should have at the current moment under the normal evolution trajectory; the difference between this prediction and the actual observation is the prediction residual, the magnitude of which reflects the degree to which the equipment deviates from its normal state.

[0053] The principle behind this approach lies in the fundamental differences in the state evolution patterns of maintenance equipment under different operating conditions. Therefore, it is necessary to set up independent evolutionary mapping layers for each region to capture specificities. Simultaneously, the basic state characteristics of equipment under different operating conditions share commonalities, and sharing a feature extraction layer can reduce redundancy and improve generalization ability. The bridging decision unit is responsible for determining the output method based on the matching results between the current cross-modal correlation features and the dynamic baselines of each region, adapting to possible cross-operating condition states that may occur during equipment operation. The digital twin model is trained using only normal data, reflecting the degree of degradation through a "predicting normal state - calculating residuals" approach, avoiding the limitations of traditional methods that rely on a large number of fault samples for supervised learning. Through a hierarchical collaborative architecture, a mapping mechanism combining sharing and independence, and the collaboration between normal behavior modeling and residual anomaly detection, this scheme achieves dynamic mapping of the operating state of maintenance equipment and reflects degradation trends through residual prediction, providing state prediction results and prediction residuals for subsequent health status assessment.

[0054] It is worth noting that in existing technologies, digital twin modeling typically uses a single model to train uniformly on data from all operating conditions, failing to distinguish the differences in evolutionary patterns under different operating conditions, nor considering the cross-operating condition states of equipment during operation. This leads to a decrease in prediction accuracy during operating condition switching or transitional states. Furthermore, existing solutions often rely on fault samples for supervised learning, resulting in insufficient model generalization ability in scenarios where fault samples are scarce. In contrast, this embodiment effectively improves the prediction accuracy and generalization ability of the digital twin model in multi-operating condition scenarios by setting independent evolutionary mapping layers for different operating conditions, using shared feature extraction layers to ensure generalization, employing bridging decision units to handle cross-operating condition states, and achieving residual anomaly detection through training only with normal data.

[0055] In one embodiment, the step of outputting the current state prediction result based on the matching result specifically includes: When the matching result indicates that the current state is consistent with the associated distribution pattern of a single working condition partition, the output of the partition state mapping unit corresponding to that working condition partition is directly called as the state prediction result at the current moment. When the matching result indicates that the current state exhibits a correlation distribution characteristic across multiple operating condition zones, the operating condition bridging mode is activated. The normalized deviation of the current cross-modal correlation characteristic relative to the dynamic baseline of each operating condition zone is used as the bridging signal. The activation weight of each zone state mapping unit is calculated, and the outputs of each zone state mapping unit are weighted and fused to obtain the current state prediction result.

[0056] In this embodiment, specifically: This scheme divides the output of the state prediction result into two cases based on the matching result. When the matching result indicates that the current state is consistent with the correlation distribution pattern of a single operating condition zone, the output of the zone state mapping unit corresponding to that operating condition zone is directly called as the state prediction result at the current moment. Specifically, when the similarity between the current cross-modal correlation feature and the dynamic baseline of a certain operating condition zone is higher than the preset similarity threshold and significantly higher than the similarity with the dynamic baselines of other operating condition zones, it is determined that the current state is consistent with the correlation distribution pattern of that single operating condition zone; at this time, the bridging decision unit directly selects the output of the zone state mapping unit corresponding to that operating condition zone as the final state prediction result, without performing weighted fusion of multi-zone outputs. Taking a 220kV main transformer as an example, when the cosine similarity between the current cross-modal correlation feature and the dynamic baseline of the rated load steady-state zone is 0.92, while the similarity with the light load, overload, and ambient high temperature zones is all lower than 0.6, it is determined that the current state is consistent with the rated load steady-state zone, and the output of the zone state mapping unit corresponding to the rated load steady-state zone is directly called as the state prediction result at the current moment.

[0057] When the matching result indicates that the current state exhibits a cross-modal correlation distribution characteristic across multiple operating condition zones, the operating condition bridging mode is activated. Specifically, when the similarity between the current cross-modal correlation feature and the dynamic baselines of two or more operating condition zones is higher than a preset similarity threshold, or when the similarity difference between the zones is not significant, it is determined that the current state exhibits a cross-modal correlation distribution characteristic across multiple operating condition zones. Typical scenarios include operating condition transition processes such as equipment startup and load switching. At this time, the normalized deviation of the current cross-modal correlation feature relative to the dynamic baseline of each operating condition zone is used as the bridging signal. The normalized deviation is calculated according to the aforementioned steps S311-S314 and reflects the degree of deviation between the current state and the normal correlation mode of each operating condition zone. The smaller the deviation, the closer the current state is to the normal mode of that operating condition zone, and the higher the reliability of the output of the zone state mapping unit.

[0058] The activation weight of the state mapping unit for each operating condition partition is calculated based on the normalized deviation corresponding to each operating condition partition. Let the total number of operating condition partitions participating in the bridging be... The current cross-modal association features are relative to the first The normalized deviation of the dynamic baseline for each operating condition zone is Then the first Activation weights of each partition state mapping unit The deviation is normalized using the inverse of the deviation method, and the formula is as follows:

[0059] in, For partitioned indexes, And satisfy That is, the sum of the activation weights of each partition is 1.

[0060] After obtaining the activation weights of each partition state mapping unit, the outputs of each partition state mapping unit are weighted and fused to obtain the current state prediction result, as shown in the formula:

[0061] in, For the first The output of each partition state mapping unit.

[0062] Taking a 220kV main transformer as an example, when the equipment is in the process of load switching, the similarity between the current cross-modal correlation feature and the rated load steady-state partition is 0.75, and the similarity with the light load partition is 0.72, both of which are higher than the preset threshold of 0.7, and it is judged as a cross-partition state. The normalized deviation relative to the rated load steady-state partition is calculated to be 0.2, and the normalized deviation relative to the light load partition is 0.25. Based on the reciprocal normalization of the deviation, the activation weight of the rated load steady-state partition is calculated to be 0.55, and the activation weight of the light load partition is 0.45. The outputs of the two partition state mapping units are weighted and fused according to the weights of 0.55 and 0.45 to obtain the state prediction result at the current moment.

[0063] The principle is that under stable operating conditions, the current state of the equipment clearly belongs to a certain operating condition zone, and directly calling the output of the corresponding zone unit can ensure prediction accuracy. However, in the transitional operating condition state, the current state simultaneously exhibits the associated distribution characteristics of multiple operating condition zones. The output of a single zone unit cannot accurately reflect the true state in the transitional state, and a weighted fusion of the outputs of multiple zone units is needed to achieve a smooth transition. The normalized deviation is used as the bridging signal because the deviation reflects the degree of closeness between the current state and the normal mode of each zone. The smaller the deviation, the higher the reliability of the output of the zone. The activation weight calculated accordingly can reasonably allocate the contribution of the output of each zone. Through the synergy of two output methods—direct calling of a single zone and cross-zone weighted fusion—this scheme ensures both prediction accuracy under stable operating conditions and the continuity of prediction results under the transitional operating condition state, avoiding prediction residual anomalies caused by output jumps during zone switching.

[0064] It is worth mentioning that in existing technologies, digital twin models typically employ a hard handover approach, directly switching to the model of the corresponding partition based on the operating condition determination result. This approach fails to consider the scenario where the current state simultaneously possesses characteristics of multiple partitions during the operating condition transition, leading to abrupt changes in prediction results and abnormally increased prediction residuals at the time of operating condition switching. This embodiment, however, achieves a smooth transition of prediction results during the operating condition transition by activating the operating condition bridging mode under cross-partition conditions, calculating activation weights using the normalized deviation as the bridging signal, and weighted fusing multiple partition outputs. This effectively avoids the output abrupt change problem caused by hard handover.

[0065] In one embodiment, the digital twin model dynamically maps the degradation trend of the maintenance equipment, including: S421. Based on the prediction residual, determine the temporal cumulative characteristics of the prediction residual within a preset time window; S422. Based on the time-series cumulative characteristics and the changing trend of the normalized deviation corresponding to each operating condition partition, determine and output the degradation trend of the maintenance equipment.

[0066] In this embodiment, specifically: This scheme first determines the temporal cumulative characteristics of the prediction residual within a preset time window based on the prediction residual. The prediction residual is the difference between the actual state observation value at the current time and the state prediction result at the current time, as output by the bridging decision unit in step S413 (i.e., the scalar residual). The preset time window is a continuous period of time preceding the current moment, such as the past 1 hour, the past 24 hours, or the past 7 days, and can be preset according to the equipment's operating characteristics and degradation rate. The time-series cumulative feature is obtained by time-series aggregation of the prediction residuals at each moment within the preset time window. It can be calculated using methods such as moving average, cumulative sum, or exponentially weighted moving average. For example, a simple moving average can be used to calculate the time-series cumulative feature. The formula is:

[0067] in, This represents the total number of sampling moments within the preset time window. For the first in the window The prediction residual (scalar) at each time step. This serves as the index for the sampling time within the window. The time-series cumulative feature characterizes the overall level of the prediction residuals of the equipment within the preset time window. Random fluctuations in the residuals at a single moment are smoothed out, while persistent residual biases are accumulated and amplified. Taking a 220kV main transformer as an example, if the moving average of the prediction residuals within this window gradually increases from 0.05 to 0.15, the time-series cumulative feature shows an increasing trend, indicating that the prediction bias of the equipment is continuously accumulating.

[0068] After obtaining the time-series cumulative characteristics, this scheme determines and outputs the degradation trend of the maintenance equipment based on the changing trends of the time-series cumulative characteristics and the normalized deviation corresponding to each operating condition partition. The normalized deviation corresponding to each operating condition partition is calculated according to the aforementioned steps S311-S314. Its changing trend can be obtained by performing linear regression or monotonicity tests on the normalized deviation at each moment within a preset time window. For example, the slope of the deviation changing with time can be calculated; a positive slope significantly greater than zero indicates an upward trend in the deviation. The determination of the degradation trend comprehensively considers both the time-series cumulative characteristics and the trend of deviation changes: when the time-series cumulative characteristics continuously increase and the normalized deviation shows an upward trend, it is determined that the equipment has a real degradation trend, and the degradation rate can be characterized by the changing slope of the time-series cumulative characteristics or the deviation; when the time-series cumulative characteristics increase but the normalized deviation remains stable, it may be a local anomaly or model error, and is not determined as an overall degradation trend; when the normalized deviation increases but the time-series cumulative characteristics remain stable, it may be a deviation caused by changes in operating conditions, and is not determined as a degradation trend. The output of degradation trends can include degradation rate, degradation level (e.g., no degradation, mild degradation, moderate degradation, severe degradation), or degradation curve. Taking a 220kV main transformer as an example, if the time-series cumulative characteristic of the predicted residual increases from 0.05 to 0.15 in the past 24 hours, and the normalized deviation of the rated load steady-state partition increases monotonically from 0.1 to 0.3, then the equipment is determined to have a degradation trend related to winding overheating, with a degradation rate of approximately 0.02 per day and a degradation level of mild degradation.

[0069] The principle behind this approach is that the predicted residual at a single moment is easily affected by factors such as measurement noise and instantaneous fluctuations in operating conditions, making it unreliable in reflecting degradation trends. By accumulating the residual over a preset time window, random noise can be smoothed out, and persistent degradation signals can be amplified. Furthermore, relying solely on the predicted residual cannot distinguish between "true degradation" and "operating condition changes or model errors," while the normalized deviation reflects the degree of deviation between the current state and the normal baseline. Its trend can serve as a cross-validation signal for degradation judgment: when residual accumulation and a rise in deviation occur simultaneously, the reliability of the degradation judgment is significantly improved. Through the synergy of these two signals—the time-series accumulation characteristic and the normalized deviation trend—this solution achieves an accurate and robust mapping of the degradation trend of maintenance equipment, avoiding false alarms and missed alarms caused by single-signal judgments.

[0070] It is worth mentioning that in existing technologies, degradation trend judgment usually relies on only a single indicator (such as only based on the residual threshold or only based on the deviation threshold), without considering the time-series cumulative effect or performing multi-signal cross-validation. This results in instantaneous noise easily triggering false alarms and slow degradation being easily missed. In contrast, this embodiment smooths noise by predicting the time-series cumulative characteristics of the residuals and cross-validates the normalized deviation change trend, achieving accurate and robust judgment of degradation trends and effectively reducing the false alarm rate and the missed alarm rate.

[0071] In one embodiment, step S50, the construction and training of the pre-trained health status assessment model, specifically includes: S501. Construct a health status assessment model architecture, which includes an input encoding layer, a feature fusion layer, and three output branches that correspond one-to-one with the health level, performance degradation trend, and fault risk prediction results. S502. Obtain historical operating data of pre-collected maintenance equipment, perform operating condition zoning and cross-modal correlation fusion processing to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition zoning, and calculate the normalized deviation corresponding to each operating condition zoning based on the dynamic baseline corresponding to each operating condition zoning; input the historical operating data into the digital twin model to obtain the state prediction results and prediction residuals corresponding to each operating condition zoning; construct a model training set using the normalized deviation, state prediction results and prediction residuals as training samples, and label each training sample in the model training set with a health level label based on the temporal distribution characteristics of the normalized deviation and the normal deviation range represented by the dynamic baseline corresponding to each operating condition zoning. S503. Configure a multi-task joint training objective for the health status assessment model architecture, determine the classification loss of the health level output branch, the regression loss of the performance degradation trend output branch, and the classification loss of the fault risk output branch respectively, and sum the weighted multi-task loss according to the prediction residuals of each training sample; and introduce a correlation consistency constraint term between the prediction residual and the normalized deviation to constrain the health status assessment model architecture to learn the physical correlation between the prediction residual and the normalized deviation; combine the weighted multi-task loss and the correlation consistency constraint term to obtain the joint loss. S504. Iteratively train the health status assessment model architecture based on the model training set and joint loss until the joint loss meets the preset convergence condition, and obtain the trained health status assessment model.

[0072] In this embodiment, specifically: As described in steps S501-S504 above, this scheme first constructs a health status assessment model architecture, which includes an input encoding layer, a feature fusion layer, and three output branches corresponding one-to-one with the health level, performance degradation trend, and fault risk prediction results. Specifically, the input encoding layer is used to encode the normalized deviation of the input, the state prediction result, and the prediction residual, respectively. It can be implemented using a fully connected layer, mapping three types of inputs with different dimensions and scales to encoded features of a unified dimension. The feature fusion layer is used to fuse the three types of encoded features output by the input encoding layer. It can be implemented using concatenation or attention mechanisms to obtain the fused health status assessment features. The three output branches make predictions based on the fused features. The health level output branch and the fault risk output branch are classification branches, which can be implemented using a fully connected layer with a softmax activation function, outputting the probability distribution of each category. The performance degradation trend output branch is a regression branch, which can be implemented using a fully connected layer, outputting the continuous predicted value of the degradation trend. The input encoding layer encodes the normalized deviation (scalar), the state prediction result (feature vector), and the prediction residual (feature vector or scalar) into feature vectors of the same dimension. The feature fusion layer concatenates the three into a fused feature. The three output branches output the health level (normal / attention / abnormal / severe), the performance degradation trend (degradation rate), and the fault risk (no risk / low risk / medium risk / high risk), respectively.

[0073] After constructing the model architecture, this solution acquires pre-collected historical operating data of the maintenance equipment to build a model training set. Specifically, the pre-collected historical operating data is processed by operating condition partitioning according to steps S211-S213, and cross-modal correlation fusion processing according to steps S221-S223 to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. Following steps S311-S314, the normalized deviation corresponding to each operating condition partition is calculated based on the dynamic baseline. The historical operating data is input into the trained digital twin model (trained according to steps S411-S414) to obtain the state prediction results and prediction residuals corresponding to each operating condition partition. The normalized deviation, state prediction results, and prediction residuals are used as inputs to the training samples to construct the model training set. Simultaneously, each training sample in the model training set is labeled with a health level tag. The health level tag is automatically labeled based on the temporal distribution characteristics of the normalized deviation and the normal deviation range represented by the dynamic baseline corresponding to each operating condition zone: when the normalized deviation of the training sample is within the normal deviation range and the temporal distribution is stable, it is labeled as normal; when the deviation exceeds the normal deviation range but the increase is small, it is labeled as caution; when the deviation significantly exceeds the normal deviation range and shows a continuous upward trend, it is labeled as abnormal; when the deviation is far beyond the normal deviation range, it is labeled as severe. Taking a 220kV main transformer as an example, the normal deviation range of the rated load steady-state zone is a normalized deviation of less than 0.2. Based on this, the training samples can be automatically labeled as four levels: normal (<0.2), caution (0.2-0.5), abnormal (0.5-0.8), and severe (>0.8).

[0074] Subsequently, this scheme configures a multi-task joint training objective for the health status assessment model architecture. The loss functions for the three output branches are determined as follows: the health level output branch uses cross-entropy loss as the classification loss, the performance degradation trend output branch uses mean squared error loss as the regression loss, and the fault risk output branch uses cross-entropy loss as the classification loss.

[0075] Specifically, suppose there are a total of training batches For the nth sample, for the nth One sample ( : The cross-entropy loss of the health level output branch is:

[0076] in, For the number of health level categories, For the first One-hot encoding of the health level label for each sample. The model predicts the first The probability of each health level.

[0077] The mean squared error loss of the output branch with performance degradation trend is:

[0078] in, The dimension of the degradation trend vector. For the first The true degradation trend label value of each sample These are the model's predicted values.

[0079] The cross-entropy loss of the fault risk output branch is:

[0080] in, For the number of fault risk categories, For the first One-hot encoding of each sample's fault risk label. The model predicts the first The probability of a failure risk.

[0081] The weighted multi-task loss is obtained by summing the prediction residuals of each training sample using weighted averages. Residual weights of each sample The predicted residuals of the sample Positive correlation, one calculation method is based on softmax normalization of the in-batch prediction residuals:

[0082] in, The temperature coefficient controls the smoothness of the weight distribution; samples with larger prediction residuals (more likely to be in a degraded state) receive higher weights during training, allowing the model to prioritize learning the features of degraded samples. The weighted multi-task loss is:

[0083] Simultaneously, a correlation consistency constraint term is introduced between the predicted residuals and the normalized deviation to constrain the health status assessment model to learn the physical correlation between the predicted residuals and the normalized deviation, namely:

[0084] in, For the first The prediction residuals for each sample This is the corresponding normalized deviation. The preset scaling factor is used to map the normalized deviation to a numerical range similar to the predicted residual, thus constraining the physical consistency between the two.

[0085] Combining the weighted multi-task loss with the association consistency constraint term yields the joint loss:

[0086] in, These are the weighting coefficients for the constraint terms, used to balance the contributions between multi-task classification / regression accuracy and physical consistency constraints; The value can be adjusted based on the performance of the validation set, with typical values ​​ranging from 0.1 to 1.0.

[0087] Finally, this scheme iteratively trains the health status assessment model architecture based on the model training set and joint loss. Specifically, the model training set is divided into training batches, and an optimizer such as stochastic gradient descent (SGD) or Adam is used to iteratively update the model parameters through backpropagation. In each iteration, the joint loss is calculated and the parameters are updated until the joint loss meets the preset convergence condition, such as when the loss value is lower than a preset threshold or no longer decreases significantly after several consecutive iterations. The trained health status assessment model is then obtained. The trained health status assessment model can be used to assess the health status of maintenance equipment, taking normalized deviation, state prediction results, and prediction residuals as inputs, and outputting health level, performance degradation trend, and fault risk prediction results.

[0088] The underlying principle is that health status assessment involves three different tasks: health level classification, degradation trend regression, and fault risk classification. Multi-task joint training allows the model to share feature representations, improving training efficiency and generalization ability. Weighted training based on predicted residuals ensures the model focuses on degraded samples, preventing normal samples from dominating training. Association consistency constraints leverage the physical correlation between predicted residuals and normalized deviations, enhancing the model's physical interpretability and prediction consistency. Through the synergy of a three-branch multi-task architecture, residual-weighted training, and association consistency constraints, this scheme achieves multi-dimensional and high-precision health status assessment.

[0089] It is worth noting that in existing technologies, health assessment models typically employ a single-task architecture, training health level classification, degradation trend prediction, and failure risk prediction models separately. This fails to utilize shared features across multiple tasks, resulting in low training efficiency and inconsistent prediction results across tasks. Furthermore, existing solutions often rely on manual labeling of health levels, leading to high labeling costs and strong subjectivity. In contrast, this embodiment effectively improves model training efficiency, prediction consistency, and assessment accuracy through a three-branch multi-task joint architecture, automatic labeling based on normalized deviation, residual weighted training, and association consistency constraints.

[0090] In one embodiment, step S50 specifically includes: S511. Input the current normalized deviation, state prediction result and prediction residual into the trained health status assessment model. The input encoding layer encodes the data and the feature fusion layer fuses the encoded data to obtain the health status assessment index. S512. Match the current cross-modal correlation features with the dynamic baselines corresponding to each working condition partition, and determine the working condition partition corresponding to the current state based on the correlation matching results; use the normalized deviation of the current cross-modal correlation features relative to the dynamic baselines of each working condition partition as bridging signals, and calculate the bridging weights corresponding to each working condition partition. S513. Based on the normal deviation range represented by the dynamic baseline of the working condition zone corresponding to the current state, determine the evaluation judgment parameter corresponding to the working condition zone, and use the evaluation judgment parameter to perform zone calibration on the health status evaluation index to obtain the calibrated health status evaluation index. S514. Output the corresponding results through the corresponding branch of the health status assessment model, specifically including: The health level output branch outputs the health level of the maintenance equipment based on the calibrated health status assessment indicators. The performance degradation trend output branch is based on the change trend of the calibrated health status assessment index within a preset assessment time window, and combined with the temporal change characteristics of the bridging weight within the preset assessment time window, to output the performance degradation trend. The fault risk output branch outputs fault risk prediction results based on the time-series cumulative characteristics of the calibrated health status assessment indicators and the prediction residuals.

[0091] In this embodiment, specifically: As described in steps S511-S514 above, this scheme first inputs the current normalized deviation, state prediction result, and prediction residual into the trained health status assessment model. The trained health status assessment model is constructed and trained according to the aforementioned steps S501-S504, including an input encoding layer, a feature fusion layer, and three output branches.

[0092] The input encoding layer encodes the normalized deviation, the state prediction result, and the prediction residual separately, which can be implemented using a fully connected layer. Let the first... The input path is ( =1, 2, 3, corresponding to the normalized deviation, state prediction result, and prediction residual respectively, then the encoding is calculated as follows:

[0093] in, This is the weight matrix. For bias vectors, The activation function (such as ReLU) maps three types of inputs with different dimensions and scales to encoded feature vectors of a unified dimension. .

[0094] The feature fusion layer performs feature fusion on the encoding results, which can be done in one of the following two ways: (1) Concatenation method: The three types of encoded features are concatenated end to end into a high-dimensional fusion vector:

[0095] (2) Attention Mechanism: The weights of each encoded feature are calculated through an attention network, and then the fused features are obtained by weighted summation. Specifically, the attention weights... The calculation formula is:

[0096] The fusion features are:

[0097] in, , These are the learnable parameters of the attention layer. For learnable query vectors, satisfy Reflecting the first The degree to which the input contributes to the health status assessment under the current state.

[0098] The feature fusion layer can further map the fused features into health status assessment indicators through a fully connected layer:

[0099] in, For fusion features (in the splicing method) Attention mode , , Here are the weight matrix and bias vector of the mapping layer. As a health status assessment index, it is a feature vector that comprehensively represents the equipment health status reflected by the current normalized deviation, status prediction results, and prediction residuals.

[0100] Taking the aforementioned 220kV main transformer as an example, the normalized deviation (1-dimensional, value 0.25), the state prediction result (4-dimensional feature vector), and the prediction residual (1-dimensional, value 0.1) at the current moment are respectively input into the input coding layer, and each is processed... The mapping is to 16-dimensional encoded features; the feature fusion layer concatenates the three into a 48-dimensional fusion vector, and then... It is mapped to a 4-dimensional health status assessment index.

[0101] After obtaining the health status assessment indicators, this scheme determines the operating condition partition and bridging weights corresponding to the current state. Specifically, the current cross-modal correlation features (obtained according to the aforementioned steps S221-S223, in matrix form) are matched with the dynamic baselines (also in matrix form) corresponding to each operating condition partition. During matching, the current cross-modal correlation feature matrix and the dynamic baseline matrix of each partition are first expanded into vector form by rows, and then the similarity between the current feature vector and the baseline vector of each partition is calculated using metrics such as cosine similarity or Euclidean distance.

[0102] The formula for calculating cosine similarity is:

[0103] in, This is the vectorized form of the current cross-modal correlation features. This is the vectorized form of the dynamic baseline of a certain working condition zone; the cosine similarity value ranges from [-1, 1], with a larger value indicating greater similarity.

[0104] The Euclidean distance calculation formula is:

[0105] in, For feature dimension, , They are vectors , The Each element has a distance; the smaller the distance, the more similar they are. Distance can be converted to similarity, for example:

[0106] After calculating the similarity between the current cross-modal correlation features and the dynamic baselines of each operating condition partition, the operating condition partition with the highest similarity is taken as the operating condition partition corresponding to the current state. A similarity threshold can be preset. When the highest similarity is lower than the threshold, it is determined that the current state does not belong to any stable operating condition partition and is in the operating condition transition state.

[0107] Simultaneously, using the normalized deviation of the current cross-modal correlation features relative to the dynamic baseline of each operating condition partition (calculated according to steps S311-S314 above) as bridging signals, the bridging weight corresponding to each operating condition partition is calculated. Let the total number of operating condition partitions be... The current cross-modal association features are relative to the first The normalized deviation of the dynamic baseline for each operating condition zone is Then the first Bridge weight of each partition The deviation is normalized using the inverse of the deviation method, and the formula is as follows:

[0108] in, For partitioned indexes, And satisfy That is, the sum of the bridging weights of each partition is 1. The bridging weight represents the degree of correlation between the current state and each working condition partition; the smaller the deviation, the greater the bridging weight of the partition.

[0109] Taking the aforementioned 220kV main transformer as an example, the current cross-modal correlation feature has the highest similarity to the rated load steady-state partition (0.85), and the operating condition partition corresponding to the current state is determined to be the rated load steady-state partition. The deviations relative to the rated load steady-state partition are calculated to be 0.2, relative to the light load partition is 0.5, and relative to the overload partition is 0.8. Based on the reciprocal normalization of the deviations, the bridging weights of each partition are 0.65, 0.26, and 0.09, respectively.

[0110] Subsequently, this scheme performs zonal calibration of the health status assessment indicators. Based on the normal deviation range represented by the dynamic baseline of the operating condition zone corresponding to the current state, the assessment judgment parameters corresponding to that operating condition zone are determined. The assessment judgment parameters may include the normal deviation range threshold, calibration coefficient, or calibration mapping function of that zone. Zonal calibration can be performed using linear calibration or mapping function calibration: In linear calibration, the calibrated health status assessment indicator = calibration coefficient × health status assessment indicator + calibration bias, where the calibration coefficient and calibration bias are determined based on the normal deviation range of that operating condition zone, so that the calibrated indicator has a unified dimension and health meaning in different operating condition zones; In mapping function calibration, min-max normalization can be used to map the health status assessment indicator to a unified interval (such as [0,1]), and the parameters of the mapping function are determined based on the normal deviation range and historical indicator distribution of that operating condition zone. The reason for zonal calibration is that the normal deviation ranges of different operating condition zones are different, and the health status represented by the same health status assessment index varies under different operating condition zones. For example, a deviation of 0.2 under the rated load steady-state zone may be within the normal range, while a deviation of 0.2 under the overload zone may have exceeded the normal range. Therefore, it is necessary to calibrate the index according to the assessment judgment parameters of the current zone so that the calibrated index has a consistent health meaning across operating conditions. For example, after calibration, the index is uniformly mapped to the [0,1] interval, where 0 represents complete health and 1 represents complete failure. Taking a 220kV main transformer as an example, the current state corresponds to the rated load steady-state zone, and the normal deviation range of this zone is less than 0.2. Based on this, the assessment judgment parameters are determined (such as a calibration coefficient of 1.2 and a calibration bias of 0). The 4-dimensional health status assessment index is linearly calibrated dimension by dimension to obtain the calibrated health status assessment index.

[0111] Finally, this solution outputs results through the corresponding branch of the health status assessment model.

[0112] The health level output branch, based on the calibrated health status assessment index, outputs the probability distribution of each health level through a classification network with a fully connected layer and a softmax activation function. Specifically, let the calibrated health status assessment index be... The fully connected layer outputs logits as After softmax normalization, the th The probability of each health level is:

[0113] in, Given the number of health level categories (e.g., normal, caution, abnormal, severe), the level with the highest probability is taken as the health level of the equipment being maintained.

[0114] The performance degradation trend output branch is based on the changing trend of the calibrated health status assessment indicators within a preset assessment time window, and combines the temporal variation characteristics of the bridging weights within the preset assessment time window to output the performance degradation trend. Specifically, the preset assessment time window can be the past 7 days or the past 30 days, and the window contains... At each sampling time, the calibrated index is scalarized as follows: Then, the slope of the index changing over time is calculated using linear regression. :

[0115] Simultaneously, the changes in bridging weights at each time point within the window are calculated, and the migration direction of the bridging weights (e.g., whether the partition corresponding to the maximum weight has changed) and migration rate (e.g., the weight difference between the start and end times) are extracted. ) as a time-series variation characteristic; slope After concatenating with the bridging weight change features, the input is a fully connected regression network, which outputs the performance degradation trend (e.g., degradation rate). Combining these two factors can distinguish between true degradation and index fluctuations caused by changes in operating conditions: when an upward trend in the index occurs simultaneously with the bridging weight migrating to more severe operating condition zones, it may be a fluctuation caused by changes in operating conditions; when the index continues to rise while the bridging weight remains stable, it is more likely to be true degradation.

[0116] The fault risk output branch outputs the fault risk prediction result based on the calibrated health status assessment index and the time-series cumulative characteristics of the predicted residuals. Specifically, the time-series cumulative characteristics of the predicted residuals are calculated according to the aforementioned step S421 (the moving average or cumulative sum of the residuals within a preset time window), and set as a scalar. ; calibrated health status assessment indicators With time-series cumulative features The features are concatenated into joint features, and a classification network with a fully connected layer and a softmax activation function outputs the probability distribution of each fault risk level:

[0117] in, Output for fully connected layers The One element, The number of fault risk categories (e.g., no risk, low risk, medium risk, high risk) is used; the category with the highest probability is taken as the fault risk prediction result. By combining the calibrated health status assessment indicators (reflecting the current health status) with the time-series cumulative characteristics of the prediction residuals (reflecting the degree of persistent abnormality), the probability and severity of the fault occurrence can be comprehensively judged.

[0118] Taking the aforementioned 220kV main transformer as an example, the calibrated 4D health status assessment index is input into the health level branch, and after passing through a fully connected layer and softmax, the probabilities of each level are output (normal 0.15, caution 0.65, abnormal 0.15, severe 0.05). The highest probability is taken as the output "caution" level; the linear regression slope of the index after calibration over the past 7 days... If the value is >0 (showing an upward trend) and the bridging weight remains stable in the rated load steady-state partition (without significant migration), it is judged as a true degradation. The performance degradation trend branch outputs "winding temperature-related indicators are slowly deteriorating, with a degradation rate of approximately 0.02 / day". Combining the time-series cumulative characteristics of the predicted residuals (the moving average of the residuals over the past 7 days has increased from 0.08 to 0.12), the fault risk branch outputs "medium risk, with the possibility of overheating faults".

[0119] The underlying principle is that the three-branch structure of the health status assessment model can simultaneously output three types of results: health level, degradation trend, and failure risk, avoiding inconsistencies in predictions from multiple models. Partition calibration ensures that assessment indicators have consistent health meanings under different operating conditions, avoiding misjudgments across operating conditions. The temporal variation characteristics of the bridging weights can help distinguish between actual degradation and operating condition migration, improving the accuracy of degradation trend judgment. The temporal cumulative characteristics of the prediction residuals can reflect persistent anomalies, improving the reliability of failure risk prediction. Through input encoding fusion, partition determination and bridging weight calculation, partition calibration, and the synergy of the three-branch output, this scheme achieves multi-dimensional, cross-operating condition consistent assessment of health status.

[0120] It is worth noting that in existing technologies, health assessments typically use a uniform threshold to determine indicators under all operating conditions, failing to consider the differences in normal ranges under different operating conditions, leading to misjudgments across operating conditions. Simultaneously, degradation trend judgment relies solely on indicator changes, neglecting the impact of operating condition migration, easily misjudging changes in operating conditions as degradation. This embodiment, however, effectively avoids misjudgments across operating conditions and misjudgments due to operating condition migration by using zonal calibration based on the dynamic baseline of the current operating condition zone and by assisting in degradation trend judgment through bridging weight time-series change characteristics, thus improving the accuracy and robustness of the assessment.

[0121] In one embodiment, a health status assessment system for maintenance equipment integrating edge sensing and digital twins is provided. This system corresponds to the health status assessment method for maintenance equipment integrating edge sensing and digital twins described in the previous embodiment. The system includes: The data acquisition module is used to collect the operation data of the operation and maintenance equipment in real time using the edge sensing nodes deployed on the operation and maintenance equipment. The operation data includes multi-source status sensing parameters and operation condition parameters. The parameter processing module is used to divide the multi-source state perception parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters, and to perform correlation and fusion processing on the multi-source state perception parameters of each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The deviation calculation module is used to obtain the dynamic baseline of each working condition partition and calculate the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline. The dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. The status prediction module is used to construct a digital twin model based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. The health assessment module is used to input the normalized deviation, state prediction results and prediction residuals into the pre-trained health status assessment model to obtain the health status assessment index of the operation and maintenance equipment, and output the health level, performance degradation trend and fault risk prediction results of the operation and maintenance equipment based on the health status assessment index.

[0122] Specific limitations regarding the health status assessment system for maintenance equipment integrating edge sensing and digital twins can be found in the limitations of the health status assessment method for maintenance equipment integrating edge sensing and digital twins described above, and will not be repeated here. Each module in the aforementioned health status assessment system for maintenance equipment integrating edge sensing and digital twins can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0123] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for assessing the health status of maintenance equipment that integrates edge sensing and digital twins, characterized in that, Includes the following steps: The operation data of the operation and maintenance equipment is collected in real time by using edge sensing nodes deployed on the operation and maintenance equipment. The operation data includes multi-source status sensing parameters and operating condition parameters. The multi-source state perception parameters are divided according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters. The multi-source state perception parameters of each operating condition partition are then correlated and fused to obtain the cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The dynamic baseline of each working condition partition is obtained, and the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline is calculated. The dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. A digital twin model is constructed based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. The normalized deviation, state prediction results, and prediction residuals are input into the pre-trained health status assessment model to obtain the health status assessment index of the maintenance equipment. Based on the health status assessment index, the health level, performance degradation trend, and fault risk prediction results of the maintenance equipment are output.

2. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 1, characterized in that, The step of dividing the multi-source state sensing parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state sensing parameters specifically includes: The operating condition parameters are coupled and analyzed to determine a multi-dimensional operating condition space, and several operating condition partitions and stable operating condition judgment conditions corresponding to each operating condition partition are preset in the multi-dimensional operating condition space. Real-time monitoring of the changes in the operating parameters within the multi-dimensional operating space; When the operating condition parameters meet the stable operating condition determination conditions of any operating condition partition, the multi-source state sensing parameters collected during the duration period that meets the stable operating condition determination conditions are assigned to that operating condition partition, so as to obtain several operating condition partitions with corresponding multi-source state sensing parameters. Among them, the duration period that does not meet any stable operating condition determination conditions is marked as the operating condition transition state, and the multi-source state sensing parameters collected during the duration of the operating condition transition state are marked as transition state data. The transition state data does not participate in the partitioning.

3. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twins as described in claim 2, characterized in that, The step of performing correlation and fusion processing on the multi-source state perception parameters of each working condition partition to obtain the cross-modal correlation features and multi-source fusion features corresponding to each working condition partition specifically includes: For each working condition zone, the multi-source state perception parameters within that working condition zone are time-synchronized and aligned to the same time reference to obtain aligned multi-source state perception parameters. Based on the aligned multi-source state perception parameters, the dynamic correlation metric between different modal parameters is calculated, a cross-modal correlation matrix is ​​constructed, and the cross-modal correlation matrix is ​​used as the cross-modal correlation feature corresponding to the working condition partition. Based on the aligned multi-source state perception parameters, feature extraction is performed on the state perception parameters of each mode within the same working condition partition to obtain the independent features of each mode. The independent features of each mode are then fused to obtain the multi-source fused features corresponding to the working condition partition.

4. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twins as described in claim 2, characterized in that, The steps of obtaining the dynamic baseline of each working condition partition and calculating the normalized deviation of the cross-modal correlation features of each working condition partition relative to the corresponding dynamic baseline specifically include: For each operating condition partition, a dynamic baseline is established based on the cross-modal correlation features generated from the multi-source state perception parameters in the historical operating data that meet the normal operating conditions under that operating condition partition, and the dynamic baseline is continuously updated. The dynamic baseline only includes the cross-modal correlation features generated from the multi-source state perception parameters that meet the stable operating condition determination conditions when it is updated. Obtain the current cross-modal correlation features for each operating condition partition; For each working condition partition, the current cross-modal correlation features of the working condition partition are compared element by element with the dynamic baseline corresponding to the working condition partition to obtain the correlation differences between each modal parameter pair. The correlation differences between each modal parameter pair are weighted based on the preset modal importance weight to obtain the weighted correlation difference. The weighted correlation differences of each modal parameter pair are normalized to obtain the normalized deviation corresponding to the working condition partition.

5. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 1, characterized in that, The steps for constructing a digital twin model based on the multi-source fusion features and normalized deviations corresponding to each working condition partition specifically include: A hierarchical collaborative architecture is constructed, comprising a bridging decision unit and several partition state mapping units that correspond one-to-one with each working condition partition. Each partition state mapping unit is used to output the state prediction result under the corresponding working condition partition, and the bridging decision unit is used to determine the output method of the state prediction result based on the association matching result. Each partition state mapping unit is configured with a shared feature extraction layer and an independent evolution mapping layer. The shared feature extraction layer of each partition state mapping unit is configured with the same preset parameter extraction rules, which are used to extract common basic features in multi-source fusion features. The independent evolution mapping layer of each partition state mapping unit is configured with mutually independent preset learning mapping rules, which are used to learn the normal operation state and evolution law under the corresponding working condition partition based on the common basic features, and output the state prediction result under the corresponding working condition partition. Configure association matching and routing rules for the bridging decision unit, so that the bridging decision unit can obtain the current cross-modal association features based on the current multi-source state perception parameters, perform association matching between the cross-modal association features and the normal association distribution patterns represented by the dynamic baselines corresponding to each working condition partition, output the current state prediction result based on the matching result, and determine and output the prediction residual based on the difference between the current actual state observation value and the current state prediction result. Based on historical operating data of maintenance equipment meeting normal operating conditions, multi-source fusion features corresponding to each operating condition partition are extracted, and the state mapping units of each partition are trained so that each partition state mapping unit learns the normal state evolution law under the corresponding operating condition partition, thus obtaining a trained digital twin model.

6. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 5, characterized in that, The step of outputting the current state prediction result based on the matching result specifically includes: When the matching result indicates that the current state is consistent with the associated distribution pattern of a single working condition partition, the output of the partition state mapping unit corresponding to that working condition partition is directly called as the state prediction result at the current moment. When the matching result indicates that the current state exhibits a correlation distribution characteristic across multiple operating condition zones, the operating condition bridging mode is activated. The normalized deviation of the current cross-modal correlation characteristic relative to the dynamic baseline of each operating condition zone is used as the bridging signal. The activation weight of each zone state mapping unit is calculated, and the outputs of each zone state mapping unit are weighted and fused to obtain the current state prediction result.

7. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 5, characterized in that, The digital twin model dynamically maps the degradation trend of the maintenance equipment, including: Based on the prediction residuals, determine the temporal cumulative characteristics of the prediction residuals within a preset time window; Based on the changing trends of the time-series cumulative characteristics and the normalized deviation corresponding to each operating condition partition, the degradation trend of the maintenance equipment is determined and output.

8. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 1, characterized in that, The construction and training of the pre-trained health status assessment model specifically includes: A health status assessment model architecture is constructed, which includes an input encoding layer, a feature fusion layer, and three output branches that correspond one-to-one with the health level, performance degradation trend, and fault risk prediction results. Historical operating data of pre-collected maintenance equipment is acquired and processed for operating condition partitioning and cross-modal correlation fusion to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. Based on the dynamic baseline corresponding to each operating condition partition, the normalized deviation is calculated. The historical operating data is input into the digital twin model to obtain the state prediction results and prediction residuals corresponding to each operating condition partition. The normalized deviation, state prediction results and prediction residuals are used as training samples to construct a model training set. Based on the temporal distribution characteristics of the normalized deviation and the normal deviation range represented by the dynamic baseline corresponding to each operating condition partition, a health level label is labeled for each training sample in the model training set. To configure a multi-task joint training objective for the health status assessment model architecture, the classification loss of the health level output branch, the regression loss of the performance degradation trend output branch, and the classification loss of the fault risk output branch are determined respectively. The weighted multi-task loss is obtained by weighting and summing the prediction residuals of each training sample. Furthermore, a correlation consistency constraint term is introduced between the prediction residual and the normalized deviation to constrain the health status assessment model architecture to learn the physical correlation between the prediction residual and the normalized deviation. The weighted multi-task loss and the correlation consistency constraint term are combined to obtain the joint loss. The health status assessment model architecture is iteratively trained based on the model training set and joint loss until the joint loss meets the preset convergence condition, thus obtaining the trained health status assessment model.

9. The method for assessing the health status of maintenance equipment by integrating edge perception and digital twin as described in claim 8, characterized in that, The steps of inputting the normalized deviation, state prediction results, and prediction residuals into a pre-trained health status assessment model to obtain health status assessment indicators for the maintenance equipment, and outputting the health level, performance degradation trend, and fault risk prediction results of the maintenance equipment based on the health status assessment indicators, specifically include: The current normalized deviation, state prediction results, and prediction residuals are input into the trained health status assessment model, encoded by the input encoding layer, and the encoding results are fused by the feature fusion layer to obtain the health status assessment index. The current cross-modal correlation features are matched with the dynamic baselines corresponding to each working condition partition, and the working condition partition corresponding to the current state is determined based on the correlation matching results. The normalized deviation of the current cross-modal correlation features relative to the dynamic baselines of each working condition partition is used as the bridging signal to calculate the bridging weights corresponding to each working condition partition. Based on the normal deviation range represented by the dynamic baseline of the current working condition zone, the evaluation and judgment parameters corresponding to the working condition zone are determined. The health status evaluation index is then calibrated by the evaluation and judgment parameters to obtain the calibrated health status evaluation index. The corresponding branch of the health status assessment model outputs the corresponding results, specifically including: The health level output branch outputs the health level of the maintenance equipment based on the calibrated health status assessment indicators. The performance degradation trend output branch is based on the change trend of the calibrated health status assessment index within a preset assessment time window, and combined with the temporal change characteristics of the bridging weight within the preset assessment time window, to output the performance degradation trend. The fault risk output branch outputs fault risk prediction results based on the time-series cumulative characteristics of the calibrated health status assessment indicators and the prediction residuals.

10. A health status assessment system for operation and maintenance equipment integrating edge sensing and digital twin, used to implement the steps of the health status assessment method for operation and maintenance equipment integrating edge sensing and digital twin as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect the operation data of the operation and maintenance equipment in real time using the edge sensing nodes deployed on the operation and maintenance equipment. The operation data includes multi-source status sensing parameters and operation condition parameters. The parameter processing module is used to divide the multi-source state perception parameters according to the operating condition parameters to obtain several operating condition partitions with corresponding multi-source state perception parameters, and to perform correlation and fusion processing on the multi-source state perception parameters of each operating condition partition to obtain cross-modal correlation features and multi-source fusion features corresponding to each operating condition partition. The deviation calculation module is used to obtain the dynamic baseline of each working condition partition and calculate the normalized deviation of the cross-modal correlation feature of each working condition partition relative to the corresponding dynamic baseline. The dynamic baseline of each working condition partition is dynamically adjusted based on the historical operating data of the maintenance equipment. The status prediction module is used to construct a digital twin model based on the multi-source fusion features and normalized deviation corresponding to each working condition partition. The digital twin model is used to dynamically map the operating status and degradation trend of the maintenance equipment and output the status prediction result and prediction residual. The health assessment module is used to input the normalized deviation, state prediction results and prediction residuals into the pre-trained health status assessment model to obtain the health status assessment index of the operation and maintenance equipment, and output the health level, performance degradation trend and fault risk prediction results of the operation and maintenance equipment based on the health status assessment index.

Images