A dry-type air-core reactor full-life-cycle data management method and system

Abstract

The application relates to the technical field of data management, and provides a dry-type air-core reactor whole-life-cycle data management method and system. The method comprises the following steps: acquiring original monitoring data, constructing a time series distillation network according to the original monitoring data, mapping the original monitoring data through the time series distillation network, acquiring a physical state coding vector, acquiring an original coding vector, acquiring a preservation value through the physical state coding vector and the original coding vector and constructing a group aging clock space, acquiring a mirror value according to a scarcity factor of the group aging clock space, a demand degree factor of the same type of equipment on a state area and a time depth factor of the original monitoring data, fusing the preservation value and the mirror value, acquiring a comprehensive data value, and differentiating management of the original monitoring data according to the comprehensive data value. Through the whole-life-cycle data management method which can accurately strip the working condition interference, the effectiveness of the whole-life-cycle data of the dry-type air-core reactor is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and more specifically, to a method and system for full lifecycle data management of dry-type air-core reactors. Background Technology

[0002] As a core piece of equipment in the power system, dry-type air-core reactors have a long life cycle and complex operating conditions. During their entire life cycle, they generate a massive amount of raw monitoring data, including equipment status data such as temperature, vibration, and partial discharge, as well as operating condition data such as grid load and ambient temperature and humidity. This data serves as the basis for equipment condition assessment and operation and maintenance decisions.

[0003] Currently, existing data management methods for dry-type air-core reactors suffer from incomplete stripping of operating condition interference. This often leads to insufficient data validity, failing to provide reliable support for subsequent data value assessment and equipment condition evaluation. Specifically, during the operation of dry-type air-core reactors, raw monitoring data is easily affected by operating condition factors such as grid load fluctuations and changes in ambient temperature and humidity. Existing methods cannot effectively distinguish between the equipment's own aging state and the differences caused by operating condition interference. Directly using the interfered raw data for aging index calculation and data value assessment will result in significant errors, leading to biased operation and maintenance decisions and an inability to accurately determine the actual degree of equipment aging.

[0004] Therefore, this solution designs a full lifecycle data management method that can accurately isolate operating condition interference and ensure data validity. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for full lifecycle data management of dry-type air-core reactors, the method comprising:

[0006] The raw monitoring data of the dry air-core reactor throughout its entire life cycle is obtained. A time-series distillation network is constructed based on the raw monitoring data. The raw monitoring data is then mapped based on the time-series distillation network to obtain the physical state encoding vector.

[0007] Obtain the original encoding vector, and based on the physical state encoding vector and the original encoding vector, obtain the preservation value and construct the population aging clock space;

[0008] The mirror value is obtained based on the scarcity factor of the population aging clock space, the demand factor of the same type of in-operation equipment for the state region, and the time depth factor of the original monitoring data.

[0009] By integrating preservation value and mirror value, a comprehensive data value is obtained, and the original monitoring data is managed in a differentiated manner based on the comprehensive data value.

[0010] Furthermore, embodiments of the present invention also provide a full lifecycle data management system for dry-type air-core reactors, comprising:

[0011] The acquisition module is used to acquire the original monitoring data and the original encoding vector throughout the entire life cycle of the dry-type air-core reactor.

[0012] The mapping module maps the raw monitoring data based on the time-series distillation network to obtain the physical state encoding vector;

[0013] The construction module constructs a time-series distillation network based on the original monitoring data to build a population aging clock space;

[0014] The calculation module obtains the preservation value based on the physical state encoding vector and the original encoding vector, and obtains the mirror value based on the scarcity factor of the population aging clock space, the demand factor of the same type of operating equipment for the state region, and the time depth factor of the original monitoring data.

[0015] The fusion module integrates the preservation value and the mirror value to obtain a comprehensive data value.

[0016] The management module performs differentiated management of the raw monitoring data based on the comprehensive data value.

[0017] Based on the above, this application embodiment constructs a time-series distillation network and trains it using a phased annealing strategy. This effectively extracts the physical state features from the original monitoring data and outputs a physical state encoding vector independent of the operating condition statistics. Simultaneously, it quantifies the preservation value and mirror value of the data and integrates them into a comprehensive data value, which reflects the actual value of the original monitoring data. Through this full lifecycle data management method that can accurately remove operating condition interference, the effectiveness of the full lifecycle data of the dry-type air-core reactor is ensured. Attached Figure Description

[0018] Figure 1 This is a flowchart of the steps of a method for managing the entire life cycle data of a dry-type air-core reactor according to the present invention;

[0019] Figure 2 This is a schematic diagram of a dry-type air-core reactor full life cycle data management system according to the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0021] like Figure 1As shown, a method for full lifecycle data management of dry-type air-core reactors includes the following steps:

[0022] Step S1: Obtain the original monitoring data throughout the entire life cycle of the dry-type air-core reactor, construct a time-series distillation network based on the original monitoring data, map the original monitoring data based on the time-series distillation network, and obtain the physical state encoding vector.

[0023] Specifically, to ensure data integrity and continuity, a combination of online real-time acquisition and offline supplementary acquisition can be used to collect raw monitoring data. For example, dedicated monitoring sensors can be deployed on and around the reactor to collect various status data during equipment operation in real time. The acquisition frequency can be set to 1 time / minute to 1 time / hour depending on the data type, and the collected data can be transmitted to the background storage system in real time. For special scenarios such as initial commissioning tests, regular maintenance, and fault diagnosis, portable monitoring devices can be used to collect specific data to supplement the deficiencies of online acquisition.

[0024] It should be noted that the original monitoring data includes equipment status data, operating condition data, and maintenance-related data. The equipment status data includes winding temperature, core temperature, vibration amplitude, vibration frequency, partial discharge, insulation resistance, dielectric loss, etc., reflecting the thermal, mechanical, and insulation characteristics of the equipment. The operating condition data includes grid load, operating voltage, current, ambient temperature, and ambient humidity. The maintenance-related data includes data acquisition timestamps, acquisition locations, sensor numbers, and equipment operating status markers.

[0025] In some possible embodiments, taking a certain type of dry-type air-core reactor as an example, temperature sensors are deployed in the reactor windings and core, vibration sensors are deployed in the outer casing, and partial discharge sensors and insulation monitoring sensors are deployed at the outgoing terminals. Simultaneously, ambient temperature and humidity sensors and power grid parameter acquisition modules are deployed around the equipment. The vibration and partial discharge data acquisition frequency is set to once per minute, while the temperature, power grid load, and ambient temperature and humidity data acquisition frequency is set to once per hour. All acquired data is transmitted in real-time to the backend data storage server via industrial Ethernet. For offline supplementary data acquisition, for example, during the initial commissioning test in January 2023, a portable dielectric loss tester is used to collect dielectric loss values. The insulation resistance tester collects specific data on insulation resistance; during the routine maintenance in July 2023, a portable vibration analyzer was used to supplement the collection of vibration frequency data under different loads; the final raw monitoring data obtained includes equipment status data, such as winding temperature 25℃-65℃, vibration amplitude 0.1-0.8mm / s, partial discharge 0-500pC, etc.; operating condition data, such as grid load 50%-100%, operating voltage 10kV±5%, ambient temperature -15℃-38℃, etc.; and maintenance-related data, such as temperature data collected at 14:00 on June 15, 2023, with the collection location being winding phase A, sensor number TC-01, and equipment status marked as normal, etc.

[0026] Step S1 includes:

[0027] Step S1-1: The time-series distillation network includes a time-series context encoder, a working condition decoupling module, and an autoregressive prediction module.

[0028] It should be noted that the specific structural parameters of the temporal context encoder include: the encoder adopts a 3-layer stacked dilated causal convolutional network, with a uniform kernel size of 3×3, and the dilation coefficients of each layer are set to 1, 2, and 4 respectively, to expand the temporal receptive field layer by layer, covering temporal features over a period of one month; each convolutional output is followed by a batch normalization layer (BN), a ReLU activation function, and a dropout layer, with a dropout rate of 0.1; the number of output nodes for each layer is 128, 64, and 32 respectively, ultimately outputting a 32-dimensional multi-scale dynamic feature vector; this structure can accurately capture the short-term operating condition fluctuations and long-term aging drift characteristics of reactors.

[0029] The specific structural parameters of the working condition decoupling module include a feature mapping fully connected layer and a working condition discriminator. The feature mapping fully connected layer is a 3-layer structure with 32 input nodes, 16 intermediate layer nodes, and 3 output nodes, ultimately outputting a 3D physical state encoding vector. Each layer is configured with a ReLU activation function, and the weight parameters are randomly initialized and optimized through training iterations. The working condition discriminator is a 4-layer fully connected network with 3 input nodes, 8, 16, and 8 hidden layer nodes, and 2 output nodes, used to distinguish between high and low working conditions to achieve adversarial decoupling.

[0030] The specific structural parameters of the autoregressive prediction module include a 4-layer fully connected prediction network, with input node 3, hidden layer nodes 16, 32, and 16, and output node 4, which respectively predict the winding temperature, vibration amplitude, and ambient temperature and humidity correlation parameters at the next moment. Each layer is configured with a ReLU activation function to realize the timing prediction of equipment status.

[0031] Furthermore, the autoregressive prediction loss function is specifically expressed as follows:

[0032] ;

[0033] Among them, the Represented as the autoregressive prediction loss value, the The total number of samples is represented as the total number of samples. Represented as a sample index, the Represented as the original detection value, the This is represented as a state prediction value.

[0034] The decoupling loss function for adversarial working conditions is specifically expressed as follows:

[0035] ;

[0036] Among them, the Represented as the decoupling loss value against adverse operating conditions, the This is represented by a label indicating the actual operating condition, where 0 represents a low-interference, stable operating condition and 1 represents a high-interference, fluctuating operating condition. This represents the probability of a high-interference operating condition, with a value range of 0-1.

[0037] In the early stages of training, only the autoregressive prediction loss function is retained. In the middle and later stages, the autoregressive prediction loss function and the adversarial working condition decoupling loss function are weighted, with specific weights of 0.7 and 0.3, to achieve a balance between preserving state features and removing working condition interference.

[0038] Step S1-1 includes:

[0039] Step S1-1-1: The temporal context encoder receives the original monitoring data and the corresponding working condition data based on the dilated causal convolutional network, and outputs multi-scale dynamic features.

[0040] Specifically, the temporal context encoder adopts an extended causal convolutional network structure. By setting reasonable expansion coefficients and convolutional kernel sizes, its receptive field can cover a time span of several weeks to several months, thereby acquiring multi-timescale dynamic features from rapid fluctuations to slow drifts in the raw monitoring data, such as instantaneous vibration peaks to long-term insulation aging trends. The module also receives raw monitoring data and corresponding operating condition data of the dry-type air-core reactor, such as grid load, ambient temperature, and humidity, and extracts features related to the physical state of the equipment to initially filter out operating condition interference.

[0041] Furthermore, the specific process of outputting multi-scale dynamic features includes: firstly, normalizing the received raw monitoring data and corresponding operating condition data to unify the data dimensions, such as normalizing temperature to the [0,1] interval and vibration amplitude to the [0,1] interval; then concatenating them into a time-series input matrix according to the timestamp order, where each row corresponds to all data at a time node; and then inputting this matrix into the input layer of an extended causal convolutional network; the extended causal convolutional network consists of multiple stacked convolutional layers, each layer using different dilation coefficients and fixed convolutional kernel sizes. Low dilation coefficient layers are responsible for capturing short-term rapid fluctuation features, such as instantaneous vibration peaks, while high dilation coefficient layers capture long-term slow drift features, such as long-term insulation aging trends, by expanding the receptive field. At the same time, the network fuses the features extracted from each layer through skip connections to avoid feature loss; the fused multi-scale features pass through a fully connected transition layer to filter out features strongly correlated with the physical state of the equipment and filter out redundant features caused by operating condition interference such as power grid load fluctuations and environmental temperature and humidity changes, finally outputting a multi-scale dynamic feature vector.

[0042] In some possible embodiments, continuing the above case, for a certain type of reactor, the expansion coefficient of the timing context encoder is set to 2, the convolution kernel size is 3, and the sensing field covers a time span of one month. The raw monitoring data of the reactor from January 2023 to January 2024 are received, such as temperature range -10℃ to 65℃, vibration amplitude 0.1-0.8mm / s, and corresponding operating condition data, such as ambient temperature -15℃ to 38℃, grid load 50%-100%. Through the expanded causal convolution operation, a multi-scale dynamic feature vector containing instantaneous vibration characteristics, short-term temperature change characteristics, and long-term insulation drift characteristics is output.

[0043] Step S1-1-2, the working condition decoupling module includes a fully connected layer and a working condition discriminator. The fully connected layer maps multi-scale dynamic features into physical state encoding vectors, and the working condition discriminator is used to train the time-series distillation network.

[0044] Specifically, the operating condition decoupling module comprises two sub-modules: a fully connected layer and an operating condition discriminator. These are used to remove operating condition interference and obtain a pure physical state encoding vector. The fully connected layer linearly maps and compresses the multi-scale dynamic features output by the temporal context encoder, transforming them into a physical state encoding vector with uniform dimensions that reflects only the true physical state of the device. The operating condition discriminator is used for adversarial training of the temporal distillation network. By identifying the operating condition information in the physical state encoding vector, it ensures that the fully connected layer outputs an encoding vector that is statistically independent of the operating conditions, guaranteeing that the encoding vector contains only the device's own state information and is unaffected by changes in operating conditions.

[0045] Furthermore, the process of mapping multi-scale dynamic features to physical state encoding vectors includes: inputting the multi-scale dynamic feature vectors output by the temporal context encoder into a fully connected layer; firstly, standardizing the feature vectors to eliminate the dimensional differences between features of different dimensions; the fully connected layer consists of multiple layers of linear neurons, which perform linear transformations on the standardized multi-scale dynamic features through preset weight matrices and bias vectors, gradually compressing the feature dimensions and mapping the high-dimensional multi-scale dynamic features to low-dimensional vectors, while retaining information strongly correlated with the physical state of the equipment and removing redundant features related to the operating conditions; after each layer of linear mapping, a ReLU nonlinear activation function is introduced to perform nonlinear transformations on the mapped features, further strengthening the state features and suppressing noise interference; through the linear mapping of the last fully connected layer, the optimized feature vectors are mapped to physical state encoding vectors of preset dimensions, and then the vectors are normalized to unify their value range to a preset interval, such as [0,3], and finally the physical state encoding vector is output.

[0046] In some possible embodiments, continuing the above case, for a certain type of reactor, the fully connected layer of the operating condition decoupling module maps the multi-scale dynamic features output in step S1-1-1 into a 3D physical state encoding vector, corresponding to thermal characteristics, mechanical characteristics, and insulation characteristics, respectively. The operating condition discriminator takes this encoding vector as input and attempts to predict the corresponding operating condition data, such as ambient temperature and grid load. During the training process, the time-series distillation network makes it difficult for the operating condition discriminator to identify the operating condition information by using gradient inversion of the encoding vector output by the fully connected layer, and finally obtains a pure encoding vector that is unrelated to the operating condition. For example, the encoding vector for June 15, 2023 is [1.8, 1.4, 1.0], which only reflects the real physical state of the reactor at that moment.

[0047] In step S1-1-3, the autoregressive prediction module predicts the raw monitoring data for the next moment based on the physical state encoding vector and in conjunction with the prediction network.

[0048] Specifically, the autoregressive prediction module is used to verify the effectiveness of the physical state encoding vector and ensure that it completely preserves the physical state information of the device. The module takes the physical state encoding vector as input and combines it with a preset prediction network, such as a fully connected neural network structure, to predict the raw monitoring data at the next moment. By minimizing the error between the predicted value and the actual collected value, the parameters of the time-series distillation network are optimized in reverse to ensure that the encoding vector can accurately reflect the device state.

[0049] Furthermore, the specific prediction process includes inputting the standardized encoded vector into a pre-defined fully connected prediction network. This network consists of 3-5 layers of linear neurons, each layer incorporating the ReLU nonlinear activation function. Through a pre-defined weight matrix and bias vector, the input encoded vector undergoes multiple rounds of linear transformation and nonlinear optimization, gradually mapping it to a prediction vector with the same dimension as the original monitoring data at the next time step. For example, predicting temperature outputs a 1D value, while predicting both temperature and vibration outputs a 2D vector. The predicted value output by the prediction network is compared with the actual original monitoring data collected at the next time step. The mean squared error is used to calculate the deviation between the two. The error is backpropagated using the gradient descent algorithm, and all parameters of the time-series distillation network are adjusted to continuously reduce the prediction error. These steps are repeated until the prediction error stabilizes within an acceptable range, completing the prediction process.

[0050] It should be noted that the allowable error range is set according to the original data type of the monitoring data. The allowable error range for temperature monitoring data prediction is ±0.5℃, the allowable error range for vibration amplitude prediction is ±0.05mm / s, the allowable error range for partial discharge prediction is ±50pC, and the allowable error range for insulation resistance prediction is ±50MΩ, etc. If the error exceeds this range, the network parameters need to be re-optimized until the requirements are met.

[0051] In some possible embodiments, continuing the above case, for a certain type of reactor, the autoregressive prediction module takes the physical state encoding vector output in step S1-1-2 as input, such as [1.8, 1.4, 1.0] on June 15, 2023, and combines it with a 3-layer fully connected prediction network to predict the raw temperature monitoring data one hour after June 15, 2023. The predicted value is 42.3℃, the actual collected raw temperature data is 42.1℃, the prediction error is 0.2℃, the error is controlled within the allowable range, indicating that the physical state encoding vector is effective.

[0052] Step S1-2: Train the time-series distillation network based on the phased annealing strategy.

[0053] Specifically, the phased annealing strategy optimizes the network loss function in three stages. In the first stage, only the autoregressive prediction loss is optimized to warm up the network and ensure that the encoding vector can retain the core state information of the device. In the second stage, adversarial working condition decoupling loss is gradually introduced to separate the working condition information from the device state information. In the third stage, the learning rate is reduced and all loss terms are jointly optimized until the network converges, resulting in a trained temporal distillation network.

[0054] Furthermore, the specific training process includes first dividing the original monitoring data and corresponding operating condition data throughout the entire life cycle of the dry-type air-core reactor into a training set, a validation set, and a test set in a ratio of 7:2:1; initializing all parameters of the time-series distillation network, such as the weight matrix and bias vector; and setting parameters such as the initial learning rate, the number of training rounds, and the loss function threshold. The initial learning rate is preset to 0.001, the total number of training rounds is preset to 300, and the convergence threshold is set to 0.0001.

[0055] The first training phase consists of 1-100 training epochs. In this phase, only the autoregressive prediction loss function is enabled, and adversarial training of the condition decoupling module is disabled. The original monitoring data of the training set is input into the temporal distillation network. Features are extracted by the temporal context encoder, mapped to physical state encoding vectors by the fully connected layer, and then the autoregressive prediction module predicts the original monitoring data at the next time step. The mean square error between the predicted value and the actual value is calculated, which is the autoregressive prediction loss. The parameters of the temporal context encoder and the autoregressive prediction module are adjusted in reverse using the gradient descent algorithm. The model performance is verified with a validation set every 10 training epochs. If the validation set loss does not decrease for 3 consecutive times and is not lower than the convergence threshold, the model can proceed to the next phase ahead of schedule.

[0056] In the second training phase, spanning 101-200 training epochs with a learning rate of 0.001, an adversarial working condition decoupling loss function is gradually introduced. Adversarial training of the working condition discriminator is then enabled. The working condition discriminator receives physical state encoding vectors and attempts to predict the corresponding working condition data. The working condition discrimination loss, i.e., the adversarial working condition decoupling loss, is calculated. The autoregressive prediction loss and the adversarial working condition decoupling loss are summed with a 1:1 weight to obtain the total loss. Gradient inversion is used to make the fully connected layer output an encoding vector independent of the working condition. The model is validated every 10 training epochs to ensure the working condition separation effect of the encoding vectors, while monitoring the changes in the validation set loss.

[0057] In the third training phase, spanning 201-300 training epochs, the learning rate is reduced to 0.0001. The two loss terms, autoregressive prediction loss and adversarial decoupling loss, are jointly optimized, with the loss weights adjusted to 2:1. Priority is given to ensuring the state preservation capability of the encoded vector. All parameters of the entire temporal distillation network are continuously adjusted through gradient descent. Validation is performed every 5 training epochs. Training is stopped when the total loss on the validation set is below the convergence threshold for 5 consecutive times and the prediction error on the test set is controlled within the allowable range, thus obtaining the temporal distillation network.

[0058] In some possible embodiments, for the time-series distillation network of the above-mentioned reactor type, a staged annealing strategy is adopted for training. In the first stage, only the autoregressive prediction loss is optimized, and the learning rate is set to 0.001. The network is warmed up by minimizing the error between the predicted temperature and the actual temperature. In the second stage, the adversarial operating condition decoupling loss is gradually introduced, and the learning rate is kept at 0.001 to remove the operating condition interference from the encoding vector. In the third stage, the learning rate is reduced to 0.0001, and the autoregressive prediction loss and the adversarial operating condition decoupling loss are jointly optimized until the network converges. After training, the network outputs the physical state encoding vector.

[0059] Steps S1-3: Input the raw monitoring data into the trained time-series distillation network and output the physical state encoding vector.

[0060] For example, all 8,760 sets of raw monitoring data for the above-mentioned reactor model from January 2023 to January 2024 are input one by one into a trained time-series distillation network, which outputs 8,760 corresponding 3D physical state encoding vectors. Among them, the encoding vector for January 1, 2023 is [1.2, 0.8, 0.5], which is the encoding vector at the initial stage of commissioning, i.e., the raw encoding vector; the encoding vector for June 15, 2023 is [1.8, 1.4, 1.0]; and the encoding vector for January 1, 2024 is [2.5, 2.1, 1.6], etc.

[0061] The first dimension is the thermal characteristic encoding value, which corresponds to the average temperature of the reactor winding. Excluding the interference of the iron core temperature, the winding temperature is an indicator reflecting the heating aging and heat loss accumulation of the reactor. The original input physical quantity is the winding temperature, which is mapped to the range of 20℃-65℃. After normalization and feature purification by the time-series distillation network, it is mapped to the encoding range [1.0, 3.0]. The higher the temperature, the more severe the thermal aging, and the larger the value. In the example, the 25℃ state corresponds to the encoding 1.2, which is in line with the thermal characteristic benchmark of a brand-new low-temperature device.

[0062] The second dimension is the mechanical characteristic encoding value, which corresponds to the vibration amplitude of the reactor body and reflects the degree of winding loosening, structural fatigue, and mechanical aging. The original input physical quantity is the steady-state vibration amplitude of the reactor body, with a mapping range of 0.1 mm / s to 0.8 mm / s. After network purification, it is mapped to the encoding range [0.8, 2.0]. The larger the vibration amplitude and the more serious the mechanical defect, the larger the value. In the example, the steady-state vibration of 0.1 mm / s corresponds to the encoding 0.8.

[0063] The third dimension is the insulation characteristic code value, which corresponds to the main insulation resistance value of the reactor and reflects the degree of insulation aging, moisture, and deterioration. The original input physical quantity is insulation resistance, with a mapping range of 1000MΩ-200MΩ. The lower the insulation resistance and the more severe the aging, the larger the code value after the network reverse mapping. The code range is [0.5, 2.0]. In the example, the device with a high insulation resistance of 1000MΩ corresponds to a code of 0.5.

[0064] Step S2: Obtain the original encoding vector, obtain the preservation value based on the physical state encoding vector and the original encoding vector, and construct the population aging clock space.

[0065] Step S2 includes:

[0066] Step S2-1: The original encoding vector is the physical state encoding vector of the dry air-core reactor in the initial stage of commissioning. The original encoding vector is obtained as a reference.

[0067] Specifically, the original encoding vector is the physical state encoding vector obtained by mapping the original monitoring data at the initial stage of operation of the dry air-core reactor, that is, after the factory acceptance test is qualified, the equipment is in a brand new state, without aging or defects, through a trained time-series distillation network; this vector reflects the initial physical state of the equipment.

[0068] In some possible embodiments, continuing the above case, for a certain type of reactor, its initial commissioning was January 1, 2023. After passing the factory acceptance test, the original monitoring data at that time was collected, in which the temperature was 25℃, the vibration amplitude was 0.1mm / s, and the insulation resistance was 1000MΩ. The data was input into the trained time-series distillation network, and the output physical state encoding vector was [1.2,0.8,0.5], which is the original encoding vector of the reactor.

[0069] Step S2-2: Obtain the Mahalanobis distance between the physical state encoding vector and the original encoding vector, and define the Mahalanobis distance as an aging index.

[0070] Specifically, the aging index is a quantitative indicator that measures the degree of physical aging of a dry-type air-core reactor. It is obtained by the Mahalanobis distance between the physical state encoding vector and the original encoding vector. The Mahalanobis distance can effectively measure the similarity between two vectors, taking into account the correlation of features in various dimensions, and is more in line with the needs of equipment condition assessment than the Euclidean distance. The Mahalanobis distance is defined as the aging index. The larger the aging index, the greater the deviation between the current equipment state and the initial state, and the more serious the aging of the equipment. Conversely, the smaller the aging index, the closer the equipment state is to the initial state, and the less severe the aging.

[0071] For example, for the aforementioned type of reactor, its original encoding vector is [1.2, 0.8, 0.5]. The physical state encoding vectors of June 15, 2023 [1.8, 1.4, 1.0] and January 1, 2024 [2.5, 2.1, 1.6] are selected. The Mahalanobis distance between the two vectors and the original encoding vector is calculated. The covariance matrix is ​​preset to be a diagonal matrix with all diagonal elements being 0.1. The calculated Mahalanobis distance is then divided by the normalization coefficient 3.89 to obtain the final aging index. The aging index of June 15, 2023 is 0.8, and the aging index of January 1, 2024 is 1.5.

[0072] It should be noted that, in order to normalize the original Mahalanobis distance to the aging index range of 0-5 that is consistent with the actual operation and maintenance of the project, this embodiment sets the normalization coefficient to 3.89. The specific setting process includes: obtaining the normalization coefficient based on the historical operation data of a group of equipment of the same model, selecting the original Mahalanobis distance data of 10 equipment of this model throughout their entire life cycle, calculating the statistical maximum value of all original Mahalanobis distances, and finding that the maximum value of the original Mahalanobis distance of this model of equipment throughout its entire life cycle is about 19.45. Combined with the preset reasonable aging index range of 0-5, the normalization coefficient, i.e., 3.89, is calculated by dividing the statistical maximum value by the upper limit of the target range.

[0073] Step S2-3: Obtain the change in aging indicators between the time of collecting the original monitoring data and the current time, and combine it with the preservation formula to obtain the preservation value of the original monitoring data at the current time.

[0074] The preservation formula is specifically expressed as follows: ;

[0075] Among them, the The value of the original monitoring data at the current moment is expressed as the preservation value of the data. Represented as the attenuation sensitivity coefficient, the This is expressed as the change in aging indicators.

[0076] Specifically, the preservation value is an indicator that measures the reference value of the original monitoring data to the current state of the equipment. It is calculated based on the change in aging index between the acquisition time and the current time (current time aging index - acquisition time aging index), combined with a preset preservation formula for quantification. The attenuation sensitivity coefficient is preset according to the aging characteristics of the same model of equipment, with a value range of 0.1-0.5. The larger the change in aging index, the greater the degree of aging of the equipment from the acquisition time to the current time, and the lower the preservation value of the data. Conversely, the smaller the change in aging index, the higher the preservation value. The preservation value ranges from 0 to 1, and the closer it is to 1, the higher the reference value of the data.

[0077] It should be noted that the attenuation sensitivity coefficient is set based on the aging characteristics of dry-type hollow reactors and the quantitative requirements of engineering operation and maintenance for preservation value. Dry-type hollow reactors are slow-aging devices with a stable aging process throughout their entire life cycle. The attenuation sensitivity coefficient needs to be set within a reasonable range, such as 0.1-0.5, to avoid the coefficient being too large, which would cause the preservation value to decay too quickly and fail to reflect the short-term reference value of the data, or the coefficient being too small, which would cause the preservation value to decay too slowly and fail to distinguish the differences in data value between different aging levels. Operation and maintenance personnel need to quickly determine the reference value of data through preservation value. The setting of the attenuation sensitivity coefficient should match the attenuation trend of preservation value with the actual aging trend of the equipment, so as to facilitate the differentiation of high-value, medium-value, and low-value data.

[0078] In this embodiment, for this type of reactor, the attenuation sensitivity coefficient can be set to 0.3. Based on the statistical data obtained from the group aging of this type of equipment, the aging rate of this type of reactor is stable from the initial to the middle stage of operation, and the change in aging index is usually between 0 and 2. An attenuation sensitivity coefficient of 0.3 can make the preservation value exhibit a reasonable exponential decay with the change in aging index. When the change in aging index is 0.7, the preservation value is about 0.81, which is in the high value range and meets the reference value of the data at this stage. When the change in aging index is 2, the preservation value is about 0.55, which is in the medium value range. This is consistent with the judgment of the operation and maintenance personnel on the value of the data at this stage, while taking into account the data value differentiation and calculation rationality, and avoiding the problem of excessively fast or slow attenuation.

[0079] In some possible embodiments, continuing the above case, for this type of reactor, the preset attenuation sensitivity coefficient is 0.3. Assuming the current time is January 1, 2024, the current aging index is 1.5. The original monitoring data collected on June 15, 2023 is selected. The aging index at the time of collection is 0.8. The change in aging index is calculated to be 0.7. Substituting into the preservation formula, the preservation value is obtained as 0.81.

[0080] Steps S2-4: Using the trajectory of aging indicators over time as individual aging clocks, collect all individual aging clocks to construct a group aging clock space.

[0081] Specifically, the individual aging clock reflects the aging process of a single dry-type air-core reactor. It connects the aging indicators at all moments throughout the entire life cycle of the equipment with time as the horizontal axis and aging indicators as the vertical axis, forming a trajectory of the aging indicators changing over time. This trajectory is the individual aging clock of the equipment, which can reflect the aging rate and degree of the equipment from the initial operation to the present moment. The group aging clock space gathers the individual aging clocks of all equipment of the same model, including retired equipment. It forms the physical state code vectors of all equipment in the same dimensional space, which is used to reflect the group aging pattern of equipment of the same model and to provide a spatial carrier for the calculation of mirror value.

[0082] In some possible embodiments, continuing the above case, for a certain type of reactor, with the operating month as the horizontal axis and the aging index as the vertical axis, all aging indices from January 2023 to January 2024 are connected. For example, the aging index for January is 0.1, the aging index for June is 0.8, and the aging index for December is 1.5, etc., forming the individual aging clock of the reactor. The trajectory shows a slow upward trend, indicating that the equipment is in the early stable aging stage. At the same time, the individual aging clocks of the other 9 equipment of the same model are collected, with 7 in operation and 3 retired. The physical state encoding vectors of all equipment are plotted in the same 3D space to construct the group aging clock space. The area near the origin in the space is the design reference point. The encoding vector trajectory of each equipment starts from the area near the origin and spreads outward to form the group aging trajectory.

[0083] For example, the specific data expression form of the individual aging clock is a two-dimensional time-series numerical sequence of a single reactor, consisting of a timestamp and a one-dimensional aging index. Taking a single reactor as an example in this embodiment, the specific real data examples of its individual aging clock include {(0h,0.00), (8760h,1.22), (17520h,2.15), (26280h,3.48)}. Among them, 0h corresponds to the moment when the equipment is newly put into operation, and the aging index is 0. 8760h, 17520h, and 26280h correspond to the specific time nodes of 1 year, 2 years, and 3 years of operation, respectively, and the corresponding aging index is the specific aging quantification score calculated in real time. Each aging index value is calculated from the three-dimensional encoding vector of the corresponding moment. For example, the three-dimensional encoding vector corresponding to moment 8760h is [1.4,0.9,0.7], and the aging index 1.22 is obtained after normalization by Mahalanobis distance.

[0084] The population aging clock space is a pure three-dimensional real-valued matrix set composed of the three-dimensional physical encoding vectors of all devices of the same model at all times. Taking 10 reactors in this embodiment as an example, the specific real data samples of the population aging clock space include the state points of device 1 at various times: [1.2, 0.8, 0.5], [1.4, 0.9, 0.7], [1.9, 1.5, 1.1], etc.; the state points of device 2 at various times: [1.2, 0.8, 0.5], [1.5, 1.0, 0.8], [2.0, 1.6, 1.2], etc.; the same applies to the other devices. All three-dimensional numerical points are uniformly collected into the same three-dimensional coordinate system. All data are specific quantified values, and time is only a data acquisition label and does not participate in the spatial coordinate composition. Among them, individual sampling is discrete time-series data, while spatial modeling is a continuous feature space. Therefore, the continuous population aging pattern can be fitted using discrete historical state points of devices.

[0085] Step S3: Obtain mirror value based on the scarcity factor of the population aging clock space, the demand factor of the same type of in-operation equipment for the state region, and the time depth factor of the original monitoring data.

[0086] Step S3 includes:

[0087] Step S3-1: Obtain the sample density of the physical state encoding vector in the population aging clock space based on kernel density estimation, i.e., the kernel density estimate. Perform negative log normalization on the kernel density estimate to obtain the scarcity factor.

[0088] Specifically, the scarcity factor is an indicator that measures the scarcity of the region where the physical state encoding vector is located. It is obtained by kernel density estimation method to obtain the sample density of the encoding vector in the population aging clock space, that is, the kernel density estimate. The smaller the sample density, the scarcer the encoding vector in the region, the scarcer the corresponding data, and the larger the scarcity factor; conversely, the larger the sample density, the smaller the scarcity factor. In order to unify the value range of the scarcity factor to 0-1, the kernel density estimate needs to be negative log normalized to ensure the standardization and comparability of the calculation results.

[0089] Furthermore, the process of obtaining the kernel density estimate includes determining the kernel density estimation parameters. In this embodiment, a Gaussian kernel function is used, with a preset bandwidth h=0.2. The bandwidth is set according to the state fluctuation range of the encoding vectors of the same model of equipment, with a value range of 0.1-0.3, to ensure that it can capture the local distribution characteristics of the sample while avoiding overfitting. Then, the sample set is determined. The population aging clock space contains 87,600 sets of physical state encoding vectors of 10 devices of the same model. These 87,600 sets of vectors are used as the sample set for kernel density estimation, denoted as x1, x2, ..., x i , i=n=87600; Substitute into the kernel density estimation formula to calculate, the kernel density estimation formula is expressed as:

[0090] ;

[0091] Among them, the Represented as the target physical state encoding vector The kernel density estimate in the population aging clock space, i.e., the sample density, is... Represented as the target physical state encoding vector, the The total number of samples for kernel density estimation is represented by the following. Represented as bandwidth, the Represented as the dimension of the population aging clock space. Represented as a Gaussian kernel function, the Let be the th in the sample set Group physical state encoding vector.

[0092] The Gaussian kernel function is specifically represented as follows:

[0093] ;

[0094] Among them, the Represented as a Gaussian kernel function, the Represented as the normalized vector difference, i.e. .

[0095] Next, the target vector, sample set, bandwidth, and Gaussian kernel function are substituted into the formula, and the Gaussian kernel function value is calculated by substituting each vector in the sample set one by one. After summing, the result is divided by nh^d to obtain the kernel density estimate of the target physical state encoding vector, which is 0.08 in this embodiment.

[0096] Furthermore, the negative logarithmic normalization process includes: determining a normalization benchmark; statistically estimating the kernel density of all samples in the population aging clock space; taking the maximum value as the normalization benchmark (in this embodiment, the maximum kernel density estimate is 0.25); and substituting this value into the negative logarithmic normalization formula, which is:

[0097] ;

[0098] Among them, the Represented as a scarcity factor, the This is represented as the kernel density estimate of the target physical state encoding vector in the population aging clock space. This is expressed as the maximum kernel density estimate.

[0099] Next, the kernel density estimate of the target vector (0.08) and the maximum kernel density estimate (0.25) are substituted into the formula to obtain the scarcity factor S = 0.62.

[0100] Step S3-2: In the population aging clock space, define a reference correlation radius with the physical state encoding vector as the center, obtain a hypersphere, count the number of operating devices within the hypersphere range, and obtain the demand factor by combining the total number of operating devices.

[0101] Specifically, the demand factor is an indicator that measures the reference demand of the same type of in-operation equipment for the physical state region. In the group aging clock space, with the target physical state encoding vector as the center, a reference correlation radius r is defined. A hypersphere with the vector as the center and r as the radius is delineated in the group space. The hypersphere has the same dimension as the group space. The number of the same type of in-operation equipment falling into the hypersphere is counted, and the ratio of this number to the total number of in-operation equipment is the demand factor. The larger the ratio, the more equipment currently needs this data as a reference, and the higher the demand.

[0102] Furthermore, the specific process of counting the number of the same type of in-operation equipment falling within the hypersphere includes: firstly, obtaining the center point and radius r of the hypersphere; then, comparing the state encoding vector of the same type of in-operation equipment with the state encoding vector of the center point of the hypersphere in each dimension; only if the difference in each dimension is less than or equal to the radius r is it determined that the equipment falls within the hypersphere; if there is a dimension greater than the radius r, the equipment is not included in the statistics.

[0103] It should be noted that the demand factor is used to quantify the degree of reference demand of the current target historical data for a group of operating equipment of the same model. Its physical essence is that in the current power system, what proportion of operating reactors of the same model are in a physical aging state that is highly similar to the target historical data? The larger the demand factor value, the more the current operating equipment replicates the historical aging conditions and equipment state, and the higher the practical reference value and benchmarking reuse value of this set of historical data. The smaller the value, the less the current operating equipment is in this state, and the lower the practical reference demand of this set of data.

[0104] It should be noted that the reference correlation radius is specifically set based on the dimensional characteristics of the group aging clock space, the state fluctuation patterns of equipment of the same model, and the distinguishability of the demand factor. In this embodiment, the group space is 3-dimensional, corresponding to the thermal, mechanical, and insulation characteristics of the equipment. The value of the reference correlation radius must match the numerical range of the 3-dimensional physical state encoding vector to avoid the radius being too large, resulting in an excessively wide coverage of the hypersphere and inability to accurately locate similar state areas, or the radius being too small, resulting in an excessively narrow coverage of the hypersphere and inability to capture operating equipment in similar states. At the same time, it must also fit the state fluctuation range of equipment of the same model. Although equipment of the same model has the same aging pattern, the physical state encoding vector of the same aging stage may have slight fluctuations due to the influence of the installation environment and operation and maintenance level. The value of the reference correlation radius must cover this fluctuation range to ensure that equipment in similar aging states can be included in the same hypersphere. It is also necessary to ensure the distinguishability of the demand factor. The value range of the reference correlation radius is set to 0.4-0.6, which can make the demand factor of different state areas show obvious differences, avoiding the need for all areas to have the same demand factor due to an unreasonable radius, making it impossible to distinguish the reference demand differences of different data.

[0105] In this embodiment, the reference radius can be set to 0.5. In the early stable aging stage of this type of reactor, the fluctuation range of each dimension of the physical state encoding vector is within ±0.2. The maximum deviation of the encoding vector of similar state devices in 3D space is about 0.45. The value of 0.5 can not only completely cover this deviation range and ensure that the operating devices in similar states can fall within the hypersphere, but also avoid the coverage range being too wide. For example, 0.6 would cause the hypersphere to cover multiple devices in different aging stages.

[0106] In some possible embodiments, continuing the above case, for a certain type of reactor, the preset reference correlation radius r is 0.5. In the population aging clock space, a 3D hypersphere is defined with the encoding vector [1.8, 1.4, 1.0] as the center. The physical state encoding vectors of the current operating equipment of the same type are counted. The encoding vectors of 2 equipment fall into the hypersphere. Therefore, the demand factor R = 2 / 7 ≈ 0.29, indicating that a small number of operating equipment need this historical data as a state reference, and the demand is moderate.

[0107] Step S3-3: Obtain the time difference between the time when the original monitoring data was collected and the current time, and combine it with the time depth coefficient to obtain the time depth factor.

[0108] Specifically, the time depth factor is an indicator that measures the time value of raw monitoring data. It is reflected by the time difference between the time when the raw monitoring data was collected and the current time. The larger the time difference, the higher the historical value of the data and the closer the time depth factor is to 1; conversely, the smaller the time difference, the closer the time depth factor is to 0.

[0109] Furthermore, the time depth factor is obtained through an exponential function, which is specifically expressed as:

[0110] ;

[0111] in Represented as a time depth coefficient, the Representing the current moment, the... This represents the original monitoring data collection time, and the time difference between the original monitoring data collection time and the current time is in years.

[0112] It should be noted that the time depth factor is set based on the lifecycle characteristics of this equipment model and the quantitative requirements of the time depth factor. First, it must be adapted to the entire lifecycle of the equipment. In this embodiment, the dry-type air-core reactor has a design life of 20 years, which is a long-term operating device. The value of the time depth factor must ensure that the change of the time depth factor with time difference closely matches the time scale of the equipment's lifecycle. If the time depth factor is too small, such as 0.05, it will cause the time depth factor to grow too slowly, and even if the data is stored for many years, it will not be able to reflect the value of long-term historical data. If the time depth factor is too large, such as 0.2, it will cause the time depth factor to grow too quickly. Short-term data, such as factors for 1-2 years, tend to be close to 1, making it impossible to distinguish the value differences of historical data of different durations. Secondly, in line with the quantitative range requirements of the time depth factor, the preset value range of the time depth factor is 0-1, and the time depth coefficient of 0.1 can make a reasonable correspondence between the time difference and the factor. When the time difference is 10 years, the time depth factor A=0.63, which is in the medium to high range and is consistent with the value positioning of historical data at this stage. When the time difference reaches 20 years, the factor A=0.86, which is close to 1, fully reflecting the high historical value of long-term historical data, which is consistent with the value perception of historical data of decommissioned equipment in operation and maintenance.

[0113] In some possible embodiments, continuing the above case, for a certain type of reactor, the preset time depth factor is 0.1. Assuming the current time is January 1, 2024, and the data collection time is June 15, 2023, the calculation time difference is 6.5 months, which is equivalent to 0.54 years. Substituting into the formula A=0.053, it shows that the time depth of the data is short and the time depth factor is low.

[0114] Steps S3-4: Obtain mirror value based on the product of scarcity factor, demand factor, and time depth factor.

[0115] It should be noted that the mirror value is a comprehensive quantitative indicator of the cross-temporal and spatial reference value of data. It is the product of the scarcity factor, the demand factor, and the time depth factor. These three factors reflect the mirror value of data from three dimensions: state scarcity, reference demand, and historical depth, respectively. The three factors complement and synergistically affect each other, and the product result is the mirror value, which ranges from 0 to 1. The larger the mirror value, the higher the reference value of the data for the same type of equipment, showing an evolutionary trend opposite to that of preservation value.

[0116] For example, continuing with the above case, combining the calculation results of steps S3-1, S3-2, and S3-3, the scarcity factor is 0.62, the demand factor is 0.29, and the time depth factor is 0.053. Multiplying these three together, we get a mirror value M≈0.0096, indicating that the current mirror value of this set of historical data is low, mainly because the time depth is short. As time goes on, the time depth factor increases, and the mirror value will gradually increase.

[0117] Step S4: Integrate the preservation value and the mirror value to obtain the comprehensive data value, and perform differentiated management of the original monitoring data based on the comprehensive data value.

[0118] Step S4 includes:

[0119] Step S4-1: Obtain the fusion weights based on the device's aging state, specifically as follows:

[0120] ;

[0121] Among them, the Represented as fusion weights, the Represented as the transition coefficient, the Represented as The aging indicators at any given time, the This represents the typical median aging index value for the same model of equipment.

[0122] Specifically, the fusion weight is a parameter that achieves a dynamic balance between preservation value and mirror value. Its value ranges from 0 to 1 and is determined by the current aging state of the equipment, rather than being a fixed constant. The transition coefficient is used to adjust the steepness of the transition of the fusion weight. When the equipment is in the early stable period, the fusion weight approaches 1, and the preservation value dominates. When the equipment is in the accelerated degradation period, the fusion weight approaches 0, and the mirror value dominates.

[0123] It should be noted that the entire life cycle of this type of reactor is divided into an early stable period, an aging index of 0-3.5, and an accelerated degradation period. When the aging index is greater than 3.5, the transition between the two stages needs to be smooth and controllable to avoid abrupt changes in the fusion weight, which could lead to abnormal calculation of the comprehensive data value. If the transition coefficient is too small, such as 0.5, the transition is too smooth, resulting in an unclear distinction between the fusion weights of the two aging stages. If the transition coefficient is too large, such as 1.5, the transition is too steep, causing a sudden change in the fusion weight when the aging index approaches the typical median aging index value, making it difficult to adapt to the smooth transition of the equipment's aging state. Therefore, in this embodiment, the transition coefficient can be set to 0.8.

[0124] Understandably, based on the aging pattern of 10 devices of the same model, the aging index data of all devices throughout their entire life cycle were statistically analyzed. After sorting the data from smallest to largest, the aging index value corresponding to the middle position was taken, that is, the typical median aging index value is 3.5.

[0125] In some possible embodiments, continuing the above case, for a certain type of reactor, the preset transition coefficient is 0.8, the typical median aging index value of the same type of equipment is 3.5, the aging index of the current time January 1, 2024 is 1.5, and the fusion weight calculated by substituting into the formula is 0.83, indicating that the current equipment is in the early stable period, the fusion weight is biased towards the preservation value, and the preservation value plays a dominant role in the comprehensive data value.

[0126] Step S4-2: Based on the fusion weight, the preservation value and mirror value are fused to obtain the comprehensive data value.

[0127] Specifically, the preservation value and mirror value are integrated through a weighted summation method, and the formula is expressed as follows:

[0128] ;

[0129] Among them, the Represented as comprehensive data value, the aforementioned Expressed as freshness value, the stated Represented as mirror value, the This is represented as the fusion weight.

[0130] In some possible embodiments, continuing the above case, for a certain type of reactor, combining the fusion weight of 0.83, the preservation value of 0.81, and the mirror value of 0.0096, the comprehensive data value V is calculated as 0.673 by substituting into the formula.

[0131] Step S4-3: Set the first threshold and the second threshold, and combine the comprehensive data value to carry out differentiated management of the original monitoring data.

[0132] It should be noted that, considering the distribution pattern of comprehensive data value and equipment operation and maintenance needs, this embodiment sets the first threshold to 0.7 and the second threshold to 0.3. The specific reasons for these settings include: based on the comprehensive value statistics of 87,600 sets of data from 10 devices of the same model, the comprehensive data value exhibits a distribution pattern of being concentrated in the middle and dispersed at both ends. Data with a value greater than or equal to 0.7 accounts for approximately 15%, and this type of data is mostly high-preservation-value or high-mirror-value data, playing a core role in equipment status diagnosis, model training, and reference for the same model; setting this as the first threshold can filter high-value data. Data between 0.3 and 0.7 accounts for approximately 65%, which is medium-value data and does not require high-speed storage resources; setting 0.3 as the second threshold can distinguish between medium-value and low-value data. Data less than or equal to 0.3 accounts for approximately 20%, which is mostly low-value redundant data, and targeted cleaning or archiving strategies can be implemented.

[0133] Furthermore, in operation and maintenance, priority should be given to ensuring high-speed access to high-value data. A threshold of 0.7 can ensure that only core data is stored on high-speed media, such as real-time status data with high freshness and historical typical data with high mirror value. A threshold of 0.3 can filter out redundant data without actual reference value, avoid waste of storage resources, and at the same time retain the key information of medium-value data.

[0134] Step S4-3 includes:

[0135] Step S4-3-1: If the comprehensive data value is greater than or equal to the first threshold, then execute the normal management strategy.

[0136] Specifically, when the overall value of the original monitoring data is greater than or equal to the first threshold, it indicates that the overall value of the data is high and normal management strategies should be implemented. The normal management strategy is to retain all the original monitoring data and store the data in high-speed storage media, such as SSDs.

[0137] For example, if the original monitoring data from December 1, 2023 is selected, its comprehensive data value V=0.72, which is greater than the first threshold, then the normal management strategy is implemented for this set of data, and the original monitoring data such as temperature, vibration, and insulation resistance are fully retained and stored in a high-speed SSD with a response time ≤10ms. It is given priority for the condition diagnosis of the reactor and the comparative analysis of the same type of equipment to ensure that the data can be quickly retrieved.

[0138] Step S4-3-2: If the overall data value is less than the first threshold but greater than or equal to the second threshold, then perform a storage degradation operation.

[0139] Specifically, when the overall value of the original monitoring data is less than the first threshold but greater than or equal to the second threshold, it indicates that the overall value of the data is moderate. It is not necessary to retain the entire data in high-speed storage media, and a storage degradation operation needs to be performed. The storage degradation operation specifically involves retaining the statistical characteristics of the data, such as the maximum, minimum, and average values, as well as key snapshots, such as data at the moment of an abnormal state. The original data is compressed and archived and stored in ordinary storage media, such as mechanical hard drives. This does not occupy high-speed storage resources and can be queried and accessed normally, balancing data value and storage cost.

[0140] For example, if the comprehensive data value of the raw monitoring data on June 15, 2023 is 0.673, and 0.3 < 0.673 < 0.7, then a storage degradation operation is performed on this set of data. The statistical characteristics such as the maximum temperature of 42.1℃ and the average vibration amplitude of 0.4mm / s at that moment are retained, as well as the key snapshot of the instantaneous temperature peak. The raw monitoring data is compressed and archived and stored on a mechanical hard drive, which can be queried and accessed normally, but does not occupy high-speed storage resources, thus reducing storage costs.

[0141] In step S4-3-3, if the comprehensive data value is less than the second threshold, it is marked as cleanable. If the corresponding original monitoring data has not had any serious warnings or intentional tripping records, it can be automatically cleaned. If the corresponding original monitoring data has had any serious warnings or intentional tripping records, the corresponding original monitoring data is retained and managed and stored normally.

[0142] Specifically, when the overall value of the original monitoring data is less than or equal to 0.3, it indicates that the overall value of the data is low and it is marked as cleanable. The cleanup decision needs to distinguish whether the data is protected data. If the data has no records of serious alarms or fault trips, it means that it has no retention value and can be automatically cleaned up. If the data has records of serious alarms or fault trips, it means that the data is protected data. Even if the overall value is low, it must be fully retained and properly managed and stored for fault tracing and defect analysis to ensure the traceability of the data.

[0143] For example, the raw monitoring data from February 1, 2023, has a comprehensive data value of 0.28, which is less than the second threshold and is marked as cleanable. Upon verification, it was found that no serious alarms or fault trips occurred in the reactor during the data collection period, so an automatic cleanup operation was performed on this data set. If another set of data from May 10, 2023, has a comprehensive data value of 0.25 and a serious partial discharge alarm occurred during the collection period, it is considered protected data and will not be cleaned. It will be fully retained and stored normally in the high-speed medium for subsequent fault tracing analysis.

[0144] Figure 2The diagram illustrates a full lifecycle data management system for dry-type air-core reactors, which can realize the ideas of this application, according to some embodiments of this application.

[0145] Specifically, a dry-type air-core reactor full lifecycle data management system includes:

[0146] The acquisition module is used to acquire the original monitoring data and the original encoding vector throughout the entire life cycle of the dry-type air-core reactor.

[0147] The mapping module maps the raw monitoring data based on the time-series distillation network to obtain the physical state encoding vector;

[0148] The construction module constructs a time-series distillation network based on the original monitoring data to build a population aging clock space;

[0149] The calculation module obtains the preservation value based on the physical state encoding vector and the original encoding vector, and obtains the mirror value based on the scarcity factor of the population aging clock space, the demand factor of the same type of operating equipment for the state region, and the time depth factor of the original monitoring data.

[0150] The fusion module integrates the preservation value and the mirror value to obtain a comprehensive data value.

[0151] The management module performs differentiated management of the raw monitoring data based on the comprehensive data value.

[0152] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0153] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for full lifecycle data management of dry-type air-core reactors, characterized in that, It includes the following steps: The raw monitoring data of the dry air-core reactor throughout its entire life cycle is obtained. A time-series distillation network is constructed based on the raw monitoring data. The raw monitoring data is then mapped based on the time-series distillation network to obtain the physical state encoding vector. Obtain the original encoding vector, and based on the physical state encoding vector and the original encoding vector, obtain the preservation value and construct the population aging clock space; The mirror value is obtained based on the scarcity factor of the population aging clock space, the demand factor of the same type of in-operation equipment for the state region, and the time depth factor of the original monitoring data. By integrating preservation value and mirror value, a comprehensive data value is obtained, and the original monitoring data is managed in a differentiated manner based on the comprehensive data value.

2. The method for full lifecycle data management of a dry-type air-core reactor according to claim 1, characterized in that, The raw monitoring data of the dry-type air-core reactor throughout its entire life cycle is obtained. A time-series distillation network is constructed based on the raw monitoring data. The raw monitoring data is then mapped using the time-series distillation network to obtain the physical state encoding vector, including: The temporal distillation network includes a temporal context encoder, a condition decoupling module, and an autoregressive prediction module; The temporal context encoder receives raw monitoring data and corresponding working condition data based on an extended causal convolutional network, and outputs multi-scale dynamic features. The working condition decoupling module includes a fully connected layer and a working condition discriminator. The fully connected layer maps multi-scale dynamic features into physical state encoding vectors, and the working condition discriminator is used to train the time-series distillation network. The autoregressive prediction module is based on the physical state encoding vector and combined with the prediction network to predict the raw monitoring data at the next moment; The time-series distillation network is trained based on a phased annealing strategy; The raw monitoring data is input into the trained time-series distillation network, which outputs a physical state encoding vector.

3. The method for full lifecycle data management of a dry-type air-core reactor according to claim 1, characterized in that, Obtain the original encoded vector, extract the preservation value based on the physical state encoded vector and the original encoded vector, and construct the population aging clock space, including: The original encoding vector is the physical state encoding vector of the dry air-core reactor in the initial stage of commissioning, and the original encoding vector is obtained as a reference. Obtain the Mahalanobis distance between the physical state encoding vector and the original encoding vector, and define the Mahalanobis distance as an aging index; Obtain the changes in aging indicators between the time of data collection and the current time, and combine them with the preservation formula to obtain the preservation value of the original monitoring data at the current time. Using the trajectory of aging indicators over time as an individual aging clock, we collect all individual aging clocks to construct a group aging clock space.

4. The method for full lifecycle data management of a dry-type air-core reactor according to claim 3, characterized in that, Preservation formulas include: The preservation formula is specifically expressed as follows: ; Among them, the The value of the original monitoring data at the current moment is expressed as the preservation value of the data. Represented as the attenuation sensitivity coefficient, the This is expressed as the change in aging indicators.

5. The method for full lifecycle data management of a dry-type air-core reactor according to claim 1, characterized in that, The mirror value is obtained based on the scarcity factor of the population aging clock space, the demand factor of the same type of in-operation equipment for the state region, and the time depth factor of the original monitoring data, including: The sample density of the physical state encoding vector in the population aging clock space is obtained based on kernel density estimation, i.e., kernel density estimate. The kernel density estimate is then normalized by negative logarithm to obtain the scarcity factor. In the population aging clock space, a reference correlation radius is defined with the physical state encoding vector as the center to obtain a hypersphere. The number of operating devices within the hypersphere is counted, and the demand factor is obtained by combining the total number of operating devices. Obtain the time difference between the time when the original monitoring data was collected and the current time, and combine it with the time depth coefficient to obtain the time depth factor; The mirror value is obtained by multiplying the scarcity factor, the demand factor, and the time depth factor.

6. The method for full lifecycle data management of a dry-type air-core reactor according to claim 1, characterized in that, By integrating preservation value and mirror value, a comprehensive data value is obtained, including: The fusion weight is obtained based on the device's aging state, specifically as follows: ； Among them, the Represented as fusion weights, the Represented as the transition coefficient, the Represented as The aging indicators at any given time, the This represents the typical median aging index value for the same model of equipment; The preservation value and mirror value are merged based on the fusion weight to obtain comprehensive data value.

7. The method for full lifecycle data management of a dry-type air-core reactor according to claim 1, characterized in that, Differentiated management of raw monitoring data based on comprehensive data value includes: Set a first threshold and a second threshold; If the total data value is greater than or equal to the first threshold, then the normal management strategy will be implemented. If the total data value is less than the first threshold but greater than or equal to the second threshold, then a storage degradation operation will be performed. If the total data value is less than the second threshold, it is marked as cleanable. If the corresponding original monitoring data has not had any serious warnings or intentional tripping records, it can be automatically cleaned. If the corresponding original monitoring data has had any serious warnings or intentional tripping records, the corresponding original monitoring data is retained and managed and stored normally.

8. A dry-type air-core reactor full life cycle data management system, used to implement the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire the original monitoring data and the original encoding vector throughout the entire life cycle of the dry-type air-core reactor. The mapping module maps the raw monitoring data based on the time-series distillation network to obtain the physical state encoding vector; The construction module constructs a time-series distillation network based on the original monitoring data to build a population aging clock space; The calculation module obtains the preservation value based on the physical state encoding vector and the original encoding vector, and obtains the mirror value based on the scarcity factor of the population aging clock space, the demand factor of the same type of operating equipment for the state region, and the time depth factor of the original monitoring data. The fusion module integrates the preservation value and the mirror value to obtain a comprehensive data value. The management module performs differentiated management of the raw monitoring data based on the comprehensive data value.

Images