Energy system operation state real-time monitoring method and system based on digital twinning

Abstract

The present application relates to a kind of based on digital twinning energy system operating state real-time monitoring method and system, belong to energy system monitoring technical field.It includes the following steps: constructing energy system training sample with operating state annotation, define entity measurement matrix and digital twinning simulation matrix;Through channel adaptive bias correction, structured graph filtering and robust normalization processing, construct fusion input tensor;Partition time sequence state classification model with embedded digital twinning reference is constructed, feature is extracted and is updated state according to physical partition, with digital twinning reference vector weighted partition contribution, output state prediction probability and complete model optimization;Finally, the optimized model is deployed in online monitoring platform, and the real-time identification of equipment operating state is realized.The present application can improve the abnormal state recognition accuracy of energy system, anti-interference ability and model explainability, and is suitable for energy system intelligent operation and maintenance scene.

Description

Technical Field

[0001] This invention belongs to the field of energy system monitoring technology, specifically relating to a method and system for real-time monitoring of the operating status of an energy system based on digital twins. Background Technology

[0002] As modern energy systems evolve towards higher power density, stronger multi-physics coupling, and intelligent operation and maintenance, the need for real-time monitoring of the operational status of key equipment such as power generation equipment, energy storage converters, and drive motors is becoming increasingly urgent. However, energy systems in actual operation simultaneously undergo complex interactions of electrical, thermal, and mechanical dynamics. While providing massive amounts of monitoring data, field sensor networks also face severe challenges in data quality.

[0003] Existing technologies suffer from the following problems: Traditional state monitoring methods based on data-driven approaches or simple residual analysis typically employ uniform smoothing filters or mean subtraction techniques for all channels when processing sensor data. This fails to distinguish between differences in drift characteristics caused by electrical transients and thermodynamic inertia, easily leading to insufficient correction of slowly changing thermal parameters or over-smoothing of electrical dynamic characteristics. Existing methods largely rely on single-channel threshold analysis or general time-series feature extraction, lacking quantitative utilization of the inherent physical coupling relationships within the energy system (such as the correlation between excitation fluctuations and active power oscillations). This makes it difficult for models to effectively distinguish between "chain reactions caused by real physical faults" and "spurious fluctuations caused by single-point sensor anomalies or noise." Conventional depth Learning classification models (such as pure CNN, LSTM, or Transformer) typically treat multi-channel data as a fully connected grid for black-box feature learning, ignoring the inherent strong and weak coupling structures between electrical, thermal, and mechanical sections. Under imbalanced samples or complex operating conditions, they are prone to learning spurious correlations in the data, leading to decreased generalization ability and insufficient interpretability across operating conditions. Existing digital twin applications mostly remain at the level of offline simulation verification or as independent reference curve comparison, failing to deeply embed twin simulation data into real-time data stream processing and classification decision weight allocation. As a result, digital twins only play the role of bystanders in online monitoring systems, failing to fully leverage their calibration role for physical benchmarks and reference pattern recognition. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a method for real-time monitoring of the operating status of an energy system based on digital twins, comprising the following steps: S1. Construct training samples for the energy system and label their states; for each sample, define the entity measurement matrix and the digital twin simulation matrix; S2. Channel adaptive bias correction is performed using the digital twin simulation matrix to obtain the bias correction matrix. The physical correlation matrix obtained from the offline analysis of the digital twin is used to perform structured graph filtering on the bias correction matrix to extract residuals that cannot be explained by normal coupling relationships, thus obtaining the graph filtering feature matrix. This verifies the consistency of cross-channel physical coupling and isolates abnormal residuals that do not conform to physical laws from complex time series. Channel-wise robust normalization is performed on the bias correction matrix, digital twin simulation matrix, and graph filtering feature matrix to construct a fused input tensor. This achieves unified feature dimensions and robust handling of outliers, and organizes various complementary information into a standardized input suitable for convolutional networks. S3. Construct a partitioned temporal state classification model with embedded digital twin references. Extract local temporal and cross-channel joint features from the fused input tensor. Update the state according to the physical partition structure. Generate digital twin reference vectors using the normalized digital twin simulation matrix. Weight the contributions of different physical partitions and output the predicted probability of the running state. Use a loss function to train and optimize the partitioned temporal state classification model. S4. Deploy the trained and optimized final partitioned time-series state classification model on the energy system online monitoring platform to identify the equipment operating status in real time.

[0005] Furthermore, in step S1, multi-channel synchronous measurement data of the physical device are collected within a continuous time window, and synchronous simulation data output by the digital twin is read within the same time window: The defined entity measurement matrix is ​​as follows: and digital twin simulation matrix as All dimensions are Where 12 represents 12 physical attribute channels and 256 represents 256 consecutive sampling times; the 12 physical attribute channels are defined in a fixed order as: voltage fundamental amplitude, total harmonic distortion of current, active power oscillation component, reactive power fluctuation, frequency offset, bearing temperature gradient, winding hot spot temperature, shell vibration intensity, axial displacement, cooling medium flow deviation, excitation current fluctuation rate, and oil pressure change rate. For each sample, a running status label is given, denoted as . , No. The true category of each sample, the number of categories of the operating status label is fixed at 7, and they correspond to normal state, overload state, cooling failure state, bearing wear state, inter-turn short circuit state, oscillation instability state and sensor drift state in a fixed order.

[0006] Furthermore, in step S2, different bias estimation methods are used for different channels to obtain the bias correction matrix. By using digital twins as physical benchmarks, static bias and dynamic drift in physical sensor data are calibrated separately. This process suppresses noise and drift while maximizing the preservation of dynamic characteristics that reflect the true changes in the device's state. The specific steps are as follows: Read the first Entity measurement matrix of sample and digital twin simulation matrix The deviation matrix is ​​obtained by subtracting the values ​​point by point according to the channel and the sampling time. The 12 physical attribute channels are divided into full-window bias channels and slow-variable drift channels. For full-window bias channels, the median of the deviation sequence is calculated within the entire 256-point time window, and this median is used as the full-window bias of the current channel in the current sample. For slow-variable drift channels, the deviation sequence is segmented and statistically analyzed using a sliding segmentation method with a length of 32 and a step size of 16. The bias value is obtained by subtracting the bias value corresponding to the current sampling time from the physical measurement value, thereby eliminating non-state errors introduced by the sensor's own characteristics, environmental temperature drift or inaccurate reference, so that the corrected data is closer to the actual physical response of the device, rather than the original electrical or thermal measurement noise. Perform the above operations on all 12 physical attribute channels, and rearrange them according to the original channel order and time order to obtain the bias correction matrix. The size is .

[0007] Furthermore, in step S2, the specific process of obtaining the graph filtering feature matrix is ​​as follows: Offline construction of physical correlation matrix The size is Matrix elements Indicates the first The physical attribute channel for the first The standardized influence strength of each physical attribute channel; physical correlation matrix This was obtained through small perturbation sensitivity analysis using a digital twin model; For the bias correction matrix Processed column by column according to sampling time, the first Each sampling time corresponds to a 12-dimensional column vector. This represents the set of bias correction values ​​for the 12 physical attribute channels at the same sampling time; represented by the physical correlation matrix. Left multiply column vector This yields the physical prediction vector; the column vectors are then... Subtracting the weighted result of the physical prediction vector, we get the first... The output column vector of the graph filter at each sampling time. Perform the above operation one by one at each of the 256 sampling times, and then concatenate all the graph filter output column vectors in chronological order to obtain the graph filter feature matrix. .

[0008] Furthermore, in step S2, robust normalization and fusion of the input tensor are constructed: On the training set, robust centers and robust scales for each of the 12 physical attribute channels are statistically analyzed; the robust center used for the bias correction matrix is ​​denoted as... , Indicates the first The median of the bias correction values ​​for each physical attribute channel in the training set; the corresponding robustness metric is denoted as . , Indicates the first The interquartile range of the bias correction values ​​for each physical attribute channel in the training set; the robust metric used for the graph filtering feature matrix is ​​denoted as... , Indicates the first The interquartile range of the graph-filtered feature values ​​of each physical attribute channel in the training set; For the bias correction matrix Robust normalization is performed on each channel to obtain the normalized bias correction matrix. ; for digital twin simulation matrix Use the same as the bias correction matrix and Perform normalization to obtain the normalized digital twin simulation matrix. ;Graph filtering feature matrix Divide by the corresponding channel The normalized graph filtering feature matrix is ​​obtained. ; Normalized bias correction matrix Subtract the normalized digital twin simulation matrix Obtain the normalized difference matrix between the physical entity and the digital twin. ; Normalized bias correction matrix Normalized graph filtering feature matrix and normalized difference matrix Stacking them according to the feature plane direction yields the fused input tensor. The size is .

[0009] Furthermore, in step S3, a two-dimensional convolutional encoder is used to perform temporal compression and local joint feature extraction on the fused input tensor to obtain the convolutional feature tensor. The size is Where 32 represents the number of convolutional feature channels, 12 represents the number of physical attribute channels, and 64 represents the number of time steps after compression, realizing the local joint feature compression and extraction of multi-dimensional time-series data along both time and channel dimensions; the two-dimensional convolutional encoder adopts a two-layer stacked two-dimensional convolutional layer structure, including a first-layer two-dimensional convolution and a second-layer two-dimensional convolution; the first-layer two-dimensional convolution has 16 convolution kernels with a kernel size of 3×7, where 3 covers the entire information plane and 7 corresponds to local continuous segments on the time axis; the time step is 2, and the channel direction uses a size-preserving padding method; batch normalization and ReLU activation function are applied after convolution; the second-layer two-dimensional convolution has 32 convolution kernels with a kernel size of The time step is 2, and the channel direction continues to use the size-preserving padding method; after convolution, batch normalization and ReLU activation function are also applied.

[0010] Furthermore, in step S3, the state is updated according to the physical partition structure: Three physical zones are defined based on a fixed set of channels: electrical zone, thermal flow zone, and mechanical zone. From the physical correlation matrix Constructing a partition association matrix The size is : Physical correlation matrix The average value of the elements belonging to the corresponding partition channel sub-blocks is used to obtain the original association strength between the three partitions; each row is normalized so that the sum of the elements in each row is 1, resulting in the partition association matrix. Partitioned Association Matrix It can be used to describe the restricted information flow relationships between electrical zones, thermal flow zones, and mechanical zones.

[0011] Traverse the convolutional feature tensor at the compressed 64 time steps For the first Each time step is compressed to extract the convolutional feature tensor. The slice at this time step has a size of [size missing]. The channels within the three physical partitions are averaged and aggregated to obtain the input vectors for the electrical partition, the thermal flux partition, and the mechanical partition, with each partition input vector having a dimension of 32. A gated loop unit is established for each of the three physical zones to perform state updates, resulting in the final state vector of the electrical zone. The final state vector of the heat flow partition and the final state vector of the mechanical partition .

[0012] Furthermore, in step S3, the normalized digital twin simulation matrix is... The digital twin reference vector is extracted, and the final state vectors of the three partitions are weighted according to the digital twin reference vector to obtain the final fused feature vector. And complete the classification of 7 operating states: Normalized digital twin simulation matrix The input is a reference encoder, and the output is a reference feature map. The reference encoder extracts local patterns of the digital twin reference trajectory along the time axis using two layers of one-dimensional convolution: the first layer has a kernel length of 5, 16 output channels, and a stride of 2; the second layer has a kernel length of 5, 32 output channels, and a stride of 2. Each convolution is followed by a ReLU activation function. Global average pooling is performed on the reference feature map along the physical attribute channel dimension and the time dimension to obtain the digital twin reference vector. ; The final state vector of the electrical partition The final state vector of the heat flow partition and the final state vector of the mechanical partition Respectively with digital twin reference vectors The scores are concatenated and input into three independent scoring layers to obtain three partition scores. Softmax normalization is applied to these three partition scores to obtain electrical partition weights, thermal flux partition weights, and mechanical partition weights. The three independent scoring layers are fully connected layers. The final state vectors of the electrical, thermal flux, and mechanical partitions are weighted and summed using the obtained three partition weights to obtain the final fused feature vector. ; The final fused feature vector Input a fully connected classification layer, and output classification logistic values ​​for 7 categories; perform a Softmax transformation on the classification logistic values ​​to obtain the predicted probability vector. .

[0013] Furthermore, in step S3, the training phase employs class-weighted cross-entropy loss to perform end-to-end optimization of the partitioned temporal state classification model: The number of samples in the class's runtime state is denoted as The maximum number of samples in all categories is denoted as , No. The loss weights for the class's runtime state are To reduce the impact of minority class samples on parameter updates, the optimizer used is the Adam optimizer, with an initial learning rate of 10. 3. The batch size is 32, and the total number of training rounds is 120. During the validation phase, the macro-average F1 score of the 7 running states is calculated, and the model parameters corresponding to the highest macro-average F1 score are saved as the final partitioned time-series state classification model.

[0014] This invention also provides a real-time monitoring system for the operating status of an energy system based on digital twins, which executes the above-described real-time monitoring method for the operating status of an energy system based on digital twins, including: Data acquisition module: used to collect real-time monitoring data on the operating status of the energy system and construct physical measurement matrix and digital twin simulation matrix; The data preprocessing module is used to perform channel adaptive bias correction using the digital twin simulation matrix to obtain the bias correction matrix; to perform structured graph filtering on the bias correction matrix using the physical correlation matrix obtained from the offline analysis of the digital twin to obtain the graph filtering feature matrix; and to perform channel-wise robust normalization on the bias correction matrix, the digital twin simulation matrix, and the graph filtering feature matrix to construct the fused input tensor. Model building module: Constructs a partitioned temporal state classification model with embedded digital twin references, extracts local temporal and cross-channel joint features from the fused input tensor, updates the state according to the physical partition structure, generates digital twin reference vectors using the normalized digital twin simulation matrix, weights the contributions of different physical partitions, and outputs the predicted probability of the running state; uses a loss function to train and optimize the partitioned temporal state classification model. Real-time monitoring module: This module is used to deploy the trained and optimized final partitioned time-series state classification model on the energy system online monitoring platform to identify the operating status of equipment in real time.

[0015] The advantages of this invention are: This invention proposes a digital twin-guided channel adaptive bias correction method. Using digital twin simulation data as a dynamic benchmark, it employs median and sliding piecewise linear interpolation for differentiated bias subtraction of channels with different physical characteristics (full-window bias and slow-varying drift) in physical measurements, rather than a simple one-size-fits-all filtering. This method eliminates sensor drift noise while maximizing the preservation of true dynamic disturbance characteristics. Furthermore, it proposes a structured graph filtering enhancement mechanism based on the physical correlation matrix. Utilizing the cross-channel coupling relationship matrix established through offline sensitivity analysis of the digital twin, it explicitly extracts predictive components explainable by normal physical laws from the time-series signal and extracts anomalous residuals that violate physical consistency as independent feature planes, enabling the model to perceive the differences in cross-channel physical coupling. Consistency; A time-series state classification model with embedded physical partition structure constraints is proposed. By forcibly dividing the system channel into three physical partitions—electrical, thermal-fluid, and mechanical—and introducing partition association weights aggregated from the physical association matrix in the gated loop unit update, the information flow between different physical domains can be restricted, avoiding the weakening of physical constraints and parameter redundancy caused by fully connected hybrid modeling. A partition weighted decision mechanism guided by digital twin reference vectors is established. In the classification stage, reference vectors are generated using normalized twin simulation data, and decision weights for different physical partitions are dynamically evaluated and allocated. This enables the model to adaptively focus on the most relevant physical domain according to the twin reference mode of the current window, realizing a collaborative closed loop of twin calibration at the data layer and decision layer. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 A comparison chart of physical measurement and digital twin simulation of voltage fundamental amplitude; Figure 3 A comparison chart showing the physical measurement and digital twin simulation of the winding hot spot temperature; Figure 4 A comparison diagram of axial displacement physical measurement and digital twin simulation; Figure 5 A comparison chart showing the physical measurement and digital twin simulation of excitation current fluctuation rate; Figure 6 This is a schematic diagram of the deviation correction for the entire window offset channel; Figure 7 This is a schematic diagram of deviation correction for a slow-drift channel. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a method for real-time monitoring of the operating status of an energy system based on digital twins, the specific steps of which include: S1. Energy System Training Sample Construction and State Labeling In actual operation, energy systems simultaneously exhibit electrical, thermal, and mechanical dynamics. Relying solely on field sensor data makes them susceptible to electromagnetic interference, transmission noise, sensor zero drift, and asynchronous responses across channels. Relying solely on digital twin outputs can also lead to the omission of local anomalies in physical equipment that are not yet fully represented in the model.

[0020] This invention pairs physical measurement data with digital twin synchronous simulation data, first completing physical consistency recalibration, and then classifying operating states, thereby improving the ability to distinguish between cooling failure states, bearing wear states, inter-turn short circuit states, oscillation instability states, and sensor drift states. The specific steps are as follows: Collect multi-channel synchronous measurement data of the physical equipment within a continuous time window, and read the synchronous simulation data output by the digital twin within the same time window.

[0021] Specifically, regarding the first Sample, define the entity measurement matrix as The digital twin simulation matrix is Entity measurement matrix and digital twin simulation matrix All dimensions are Where 12 represents 12 physical attribute channels and 256 represents 256 consecutive sampling times; when the sampling rate is 100Hz, 256 sampling times correspond to a time window of 2.56s.

[0022] In specific implementation scenarios, the 12 physical attribute channels are defined in a fixed order as follows: voltage fundamental amplitude (V); total harmonic distortion rate (%); active power oscillation component (kW); reactive power fluctuation (kVar); frequency offset (Hz); bearing temperature gradient (°C / s); winding hot spot temperature (°C); housing vibration intensity (mm / s); axial displacement (μm); cooling medium flow rate deviation (%); excitation current fluctuation rate (%); and oil pressure change rate (kPa / s). All subsequent steps follow this fixed order, without altering the channel arrangement.

[0023] Furthermore, a running status label is given for each sample. The running status label is denoted as... , Used to indicate the first The sample data represents the true category; the number of categories for the operating status tags is fixed at 7, corresponding in a fixed order to normal state, overload state, cooling failure state, bearing wear state, inter-turn short circuit state, oscillation instability state, and sensor drift state. Tags can be determined jointly based on maintenance records, protection action records, expert confirmation results, and digital twin offline playback results.

[0024] Furthermore, the total number of training samples is denoted as... Save during training phase The entity measurement matrix, digital twin simulation matrix, and running status labels form the input set for subsequent preprocessing and model training.

[0025] In one embodiment, a comparison is made between physical measurements and digital twin simulations (partial channels). The original physical measurement curves (solid red line) and the synchronous digital twin simulation curves (dashed blue line) for four representative channels (voltage fundamental amplitude, winding hotspot temperature, axial displacement, and excitation current fluctuation rate) selected from the 12 physical channels of the energy system are compared within a 2.56-second time window. Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown, the physical measurement data does indeed contain full-window offsets, slow drifts, and localized abnormal fluctuations that deviate from the digital twin reference trajectory, demonstrating the objective fact that relying solely on field sensor data is susceptible to sensor zero drift and dynamic disturbances. The horizontal axis represents time in seconds (s); the vertical axis represents the physical quantities corresponding to each channel, in volts (V), degrees Celsius (°C), micrometers (μm), and percentages (%), respectively.

[0026] S2. Digital twin-guided multidimensional temporal-physical consistency recalibration and structured graph filtering preprocessing The physical measurement matrix contains not only real operational disturbances, but also slow-varying biases, intra-segment drift, and cross-channel physical coupling distortions. If directly input into the classification model, the model is prone to mistaking slow-varying biases for fault symptoms and misidentifying asynchronous sampling between channels as state changes.

[0027] This step first uses a digital twin simulation matrix for channel adaptive bias correction, then extracts graph filtering features that violate physical coupling relationships based on the offline constructed physical correlation matrix, and finally completes robust normalization and input tensor construction to obtain a more suitable structured input for subsequent classification model learning. The specific steps are as follows: S201, Twin-guided channel adaptive bias correction Different channels in the physical measurement matrix exhibit different drift characteristics. Channels such as voltage fundamental amplitude, frequency offset, housing vibration intensity, and axial displacement are more prone to full-window offset; while channels such as bearing temperature gradient, winding hot spot temperature, cooling medium flow rate deviation, and oil pressure change rate are more prone to intra-segment drift that changes slowly over time. If a fixed offset is uniformly subtracted from all channels, it can easily lead to under-correction or over-correction of thermodynamically related channels.

[0028] This step uses different bias estimation methods for different channels to obtain the bias correction matrix. Furthermore, by using a digital twin as a physical benchmark, static deviations and dynamic drifts in physical sensor data are calibrated separately. This process suppresses noise and drift while maximizing the preservation of dynamic characteristics that reflect the true changes in the device's state. The specific steps are as follows: 1) Read the first Entity measurement matrix of sample and digital twin simulation matrix The deviation matrix is ​​obtained by subtracting the values ​​point by point according to the channel and the sampling time. The elements in the deviation matrix represent the offset of the physical measurement value relative to the digital twin simulation value at the same channel and the same sampling time.

[0029] 2) The 12 physical attribute channels are divided into full-window bias channels and slow-variable drift channels. Among them, the full-window bias channels are fixed as follows: voltage fundamental amplitude, current total harmonic distortion rate, active power oscillation component, reactive power fluctuation, frequency offset, shell vibration intensity, axial displacement, and excitation current fluctuation rate; the slow-variable drift channels are fixed as follows: bearing temperature gradient, winding hot spot temperature, cooling medium flow deviation, and oil pressure change rate.

[0030] The reason for dividing the channels in this way is that the first group of channels usually exhibits overall bias within a single 2.56s window, while the second group of channels is more likely to show slow baseline drift within a single 2.56s window.

[0031] 3) For the full-window bias channel, calculate the median of the bias sequence within the entire 256-point time window, and use this median as the full-window bias of the current channel in the current sample.

[0032] It should be noted that using the median instead of the mean can reduce the impact of instantaneous spikes, local jumps, and occasional communication noise on bias estimation.

[0033] 4) For the slowly drifting channel, the deviation sequence is segmented and statistically analyzed using a sliding segmentation method with a length of 32 and a step size of 16. The median of the deviation sequence is calculated in each segment to obtain a set of segmented bias values; then the segmented bias values ​​are linearly interpolated according to the time position to expand into a time-varying bias sequence with a length of 256.

[0034] Based on this, the slow drift channel can have different bias correction values ​​at the beginning and end of the window, making it more suitable for describing slow baseline changes caused by thermal inertia and fluid inertia.

[0035] 5) Subtract the bias value corresponding to the current sampling time from the actual measurement value to obtain the bias correction value. This eliminates non-state errors introduced by sensor characteristics, environmental temperature drift, or inaccurate references, making the corrected data closer to the actual physical response of the equipment, rather than the original electrical or thermal measurement noise. This is expressed as: ; in, Indicates the first The sample in the first The first physical attribute channel, the first The bias correction value at each sampling time; Indicates the first The sample in the first The first physical attribute channel, the first The entity measurement value at each sampling time; Indicates the first The sample in the first The physical property channel in the first The bias amount used at each sampling time.

[0036] When the When a physical attribute channel belongs to the whole window offset channel Take the whole window offset; when the first When a physical property channel belongs to a slowly varying drift channel Take the time-varying bias sequence at the 1st The value taken at each sampling time.

[0037] In one embodiment, for example, in the first When the channel is a slowly drifting channel (such as a bearing temperature gradient), assuming and The bias obtained at time intervals through piecewise statistics and linear interpolation. , So in The time will be subtracted from the physical measurement value. ;exist Time, subtract Therefore, this reflects that the baseline temperature drift of the sensor due to thermal inertia changes over time within a 2.56-second window.

[0038] 6) Perform the above operations on all 12 physical attribute channels, and rearrange them according to the original channel order and time order to obtain the bias correction matrix. Bias correction matrix The size is still This characterizes multidimensional time-series samples that have completed twin-guided baseline alignment.

[0039] It should be noted that this step transforms the digital twin simulation matrix from "additional reference data" into an "offset correction scale" by using digital twin simulation data as a dynamic benchmark to calculate and subtract full-window or time-varying offsets for different physical channels. Based on this, full-window offsets and slow-varying drifts in the physical measurement matrix are weakened, while real anomalies such as local overheating, insufficient cooling, mechanical shock, and electrical disturbances are more easily preserved.

[0040] In one embodiment, the deviation correction of the full window offset and the slowly drifting channel is illustrated as follows: Figure 6 , Figure 7 As shown, the key quantities in the bias correction process are illustrated using a typical channel as an example. Figure 6 The diagram shows the deviation sequence (blue line) and the median offset (red dashed line) of the full-window bias channel (taking the active power oscillation component as an example). The horizontal axis represents the sampling time (without unit time steps), and the vertical axis represents the deviation value (units consistent with the corresponding physical quantity, here in kW). The median line does not change with time, indicating that the channel mainly exhibits a fixed offset within the 2.56-second window. Subtracting from the median can effectively suppress occasional spike interference. Figure 7 The diagram shows the original bias sequence (blue line) and the time-varying bias sequence (red line) obtained by piecewise median linear interpolation for a slowly drifting channel (using a bearing temperature gradient as an example). The horizontal axis represents the sampling time, and the vertical axis represents the bias value (unit: °C / s). The slow change of the red line over time indicates that the baseline drift of this channel is dynamic, demonstrating the necessity of using sliding piecewise estimation of the time-varying bias. This allows for more accurate correction of slowly varying errors caused by thermal or fluid inertia.

[0041] S202, Enhanced Structured Graph Filtering Based on Physical Correlation Matrix Even after bias correction is completed, local inconsistencies that do not conform to normal coupling relationships may still exist between channels of different physical properties. For example, under normal operation, there is usually a stable correlation between excitation current fluctuation rate, active power oscillation component, and frequency offset; when entering an oscillating instability state, this correlation will be broken.

[0042] This step utilizes the physical correlation matrix obtained from offline digital twin analysis to perform structured graph filtering on the bias correction matrix, extracting residuals that cannot be explained by normal coupling relationships, and obtaining the graph filtering feature matrix. This method verifies the consistency of physical coupling across channels and isolates abnormal residuals that do not conform to physical laws from complex timing sequences. The specific steps are as follows: 1) Offline construction of physical association matrix Physical correlation matrix The size is Matrix elements Indicates the first The physical attribute channel for the first The standardized influence strength of each physical attribute channel, ranging from 0 to 1. Physical correlation matrix. This can be obtained through small perturbation sensitivity analysis using a digital twin model. In practical implementation, a small unit perturbation can be applied to each physical attribute channel, the response amplitudes of the remaining channels can be recorded, and then normalization can be performed on each row to obtain the relative weight distribution of the influence of other channels on the current channel. To reduce the influence of weak coupling noise, the influence intensity less than 0.05 can be directly set to zero.

[0043] In one embodiment, as an example, in a digital twin model, after applying a small unit perturbation to the excitation current volatility channel (channel 11), it was found that the response amplitude of the active power oscillation component channel (channel 3) was three times that of the frequency offset channel (channel 5). After normalization, the physical correlation matrix... elements in It may be for ,and for Based on this, it can be concluded that under normal physical coupling, active power oscillation is mainly explained by excitation fluctuations, with frequency shift being secondary.

[0044] 2) For the bias correction matrix Processed column by column according to sampling time. Specifically, the first... Each sampling time point corresponds to a 12-dimensional column vector, denoted as . Column vector This represents the set of bias correction values ​​for 12 physical attribute channels at the same sampling time.

[0045] Furthermore, using the physical correlation matrix Left multiply column vector This yields the physical prediction vector. The physical prediction vector represents the predicted response that should be obtained after mutual interpretation among the current 12 physical attribute channels under the normal coupling relationship given by the digital twin.

[0046] Furthermore, the column vector Subtracting the weighted result of the physical prediction vector, we get the first... The output column vector of the graph filter at each sampling time. , is represented as: ; in, Indicates the first The sample in the first The graph filter output column vector at each sampling time has a dimension of 12; This represents the graph filter intensity coefficient, used to control the degree of physical prediction component subtraction. Preferably, A value of 0.8 to 0.9 is acceptable, for example, 0.85. The larger the value, the more fully the stationary components that conform to normal coupling relationships are subtracted, and the more obviously abnormal residuals are highlighted.

[0047] 3) Perform the above operation one by one for each of the 256 sampling times, and concatenate all the graph filter output column vectors in chronological order to obtain the graph filter feature matrix. Graph filtering feature matrix The size is This characterizes the abnormal features of cross-channel physical inconsistencies that are more prominent after structured graph filtering.

[0048] In one embodiment, for example, if the physical correlation matrix In the middle, the normalized influence strength of the excitation current fluctuation rate channel (channel 11) on the active power oscillation component channel (channel 3) is... If the frequency offset channel has an influence strength of 0.7 on the active power oscillation component channel, and 0.2, then under normal operating conditions, most of the offset correction value of the active power oscillation component channel can be explained by these two related channels together, and the graph filter output residual is small. When the system enters an oscillating instability state, the active power oscillation component channel will deviate significantly from the normal coupling relationship, and the graph filter output residual will increase significantly. The subsequent classification model can more easily use this difference to complete the identification.

[0049] It should be noted that this step is not a conventional smoothing of a single channel, but rather, through the cross-channel prediction relationship defined by the physical correlation matrix, it subtracts the portion of the original signal that can be explained by the physical coupling of other channels, extracting the "cross-channel physical inconsistency" as an anomaly clue. Based on this, the graph filtering feature matrix... It retains the real disturbances while minimizing the stable components that can be explained by physical coupling under normal operating conditions.

[0050] S203, Robust Normalization and Construction of Fusion Input Tensors bias correction matrix Sum graph filtering feature matrix The numerical range may still have differences in scale between channels; if directly input into the network, the channel with larger dimensions may dominate the training process.

[0051] This step performs channel-wise robust normalization on the bias correction matrix, digital twin simulation matrix, and graph filtering feature matrix, and constructs a fused input tensor. This achieves unified feature units and robust outlier handling, and organizes various complementary information into a standardized input suitable for convolutional networks. The specific steps are as follows: 1) On the training set, determine the robust centers and robust scales for each of the 12 physical attribute channels. Specifically, the robust centers used for the bias correction matrix are denoted as... , Indicates the first The median of the bias correction values ​​for each physical attribute channel in the training set; the corresponding robustness metric is denoted as . , Indicates the first The interquartile range of the bias correction values ​​for each physical attribute channel in the training set.

[0052] Furthermore, the robust metric used for the graph filtering feature matrix is ​​denoted as . , Indicates the first The interquartile range of the filtered feature values ​​for each physical attribute channel in the training set is plotted. The interquartile range is used to characterize the degree of dispersion and is insensitive to extreme values.

[0053] 2) For the bias correction matrix Robust normalization is performed on each channel to obtain the normalized bias correction matrix. ; Digital twin simulation matrix Use the same as the bias correction matrix and Perform normalization to obtain the normalized digital twin simulation matrix. ;Graph filtering feature matrix Divide by the corresponding channel The normalized graph filtering feature matrix is ​​obtained. .

[0054] In practice, to suppress extreme outliers, the normalized values ​​can be truncated to the range of [-5, 5].

[0055] Based on this, the dimensional unification and noise suppression of three types of key information (corrected entity values, physical inconsistency residuals, and entity-twin differences) were completed, enabling subsequent models to learn in a fair and noise-resistant numerical space.

[0056] It should be noted that robust normalization is a data preprocessing technique designed to eliminate the influence of dimensionality without being affected by extreme outliers (such as transient spikes or noise) in the data. Instead of using the "mean" and "standard deviation," which are susceptible to extreme values, it uses the robust "median." (representing data centers) and "interquartile range" The data is scaled using the formula (representing the normal distribution range of the data). This can be understood as aligning the data to a scaled size. Centered on, with In a coordinate system with unit width.

[0057] 3) Compute the normalized difference matrix between the physical entity and the digital twin. Specifically, the normalized difference matrix From the normalized bias correction matrix Subtract the normalized digital twin simulation matrix The obtained size is This characterizes the degree of deviation of the current sample from the digital twin reference trajectory in each channel and at each time.

[0058] Furthermore, the normalized bias correction matrix is... Normalized graph filtering feature matrix and normalized difference matrix Stacking them according to the feature plane direction yields the fused input tensor. Fusion input tensor The size is The three planes in the first dimension correspond to the normalized bias correction matrix, the normalized graph filter feature matrix, and the normalized difference matrix, respectively.

[0059] Fusion Input Tensor It also preserves the corrected entity dynamics, cross-channel physical inconsistency residuals, and deviation information of the entity relative to the digital twin.

[0060] It should be noted that this step does not simply feed the original values ​​of the entity into the model, but rather organizes the three types of information, namely "corrected entity values", "physical coupling residuals" and "entity relative digital twin deviations", into a unified input, so that the subsequent model can simultaneously observe amplitude information, structural residual information and reference deviation information in the same sample.

[0061] S3. Construction and Training of a Partitioned Temporal State Classification Model Embedded with Digital Twin References After processing in step S2, the... The sample is composed of the fused input tensor and normalized digital twin simulation matrix Common Representation. This step constructs a partitioned temporal state classification model. First, it extracts local temporal and cross-channel joint features from the fused input tensor. Then, it updates the state based on the physical partition structure. Finally, it generates a digital twin reference vector using a normalized digital twin simulation matrix, weights the contributions of different physical partitions, and outputs the predicted probabilities of seven types of operating states. The specific steps are as follows: S301, Convolutional Coding with Fusion Input Tensors Fusion Input Tensor Containing three information planes and twelve physical attribute channels, this approach requires both extracting local dynamic patterns along the time axis and preserving local correlations between adjacent channels. This step employs a shallow two-dimensional convolutional encoder to perform temporal compression and local joint feature extraction on the fused input tensor, yielding a convolutional feature tensor. This method enables the compression and extraction of local joint features along both time and channel dimensions of multidimensional time-series data. The specific steps are as follows: 1) Merge input tensors Input a 2D convolutional encoder. The input size of the 2D convolutional encoder is... Where 3 represents 3 information planes, 12 represents the number of physical attribute channels, and 256 represents the time length. The two-dimensional convolutional encoder employs a structure of two stacked two-dimensional convolutional layers because its input tensor... It includes two spatial dimensions: "information plane" and "physical channel," as well as a "time" dimension. It uses a kernel size of... Two-dimensional convolution can slide simultaneously across three information planes and seven consecutive time points, directly learning the "local combination pattern among bias correction values, graph filter residuals, and twin deviation values ​​at a certain moment," which is better at capturing instantaneous correlations across information planes than one-dimensional temporal convolution or fully connected layers.

[0062] 2) Perform the first layer of two-dimensional convolution.

[0063] Specifically, the number of convolution kernels in the first layer of 2D convolution can be 16, and the kernel size can be... In this layer, 3 covers the entire information plane, and 7 corresponds to a local continuous segment on the time axis; the time step can be 2, and the channel direction uses a size-preserving padding method; after convolution, batch normalization and ReLU activation function are applied. After this layer, the time length of the output feature map is reduced from 256 to 128.

[0064] 3) Perform the second layer of two-dimensional convolution.

[0065] Specifically, the number of convolution kernels in the second layer of 2D convolution can be 32, and the kernel size can be... The time step can be set to 2, and the padding method in the channel direction continues to maintain the size. After convolution, batch normalization and ReLU activation are applied. After this layer, the convolutional feature tensor is obtained. Convolutional feature tensor The size is , where 32 represents the number of convolutional feature channels, 12 represents the number of physical attribute channels, and 64 represents the number of time steps after compression.

[0066] Among them, the convolutional feature tensor Each position in the kernel corresponds to a local representation of "a certain convolutional feature channel, a certain physical property channel, and a certain compressed time step". For example, when a convolutional kernel covers 5 consecutive sampling points on the time axis, the kernel can simultaneously perceive short-term combined patterns of local heating trends, local vibration peaks, and local frequency shifts, rather than just seeing a single sampling point.

[0067] It should be noted that this step uses shallow convolutional encoding instead of a deep network. The purpose is to control the parameter scale while preserving local dynamic patterns, and to avoid overfitting when the number of abnormal samples is limited.

[0068] S302, Time-series state update based on physical partition association The energy system has 12 physical attribute channels with a clearly defined physical partition structure. Voltage fundamental amplitude, total harmonic distortion of current, active power oscillation component, reactive power fluctuation, frequency offset, and excitation current fluctuation rate mainly belong to the electrical partition; bearing temperature gradient, winding hot spot temperature, cooling medium flow deviation, and oil pressure change rate mainly belong to the thermal flow partition; and shell vibration intensity and axial displacement mainly belong to the mechanical partition. If all channels are modeled in a completely mixed manner, the structural characteristics of "strong coupling within partitions and restricted coupling between partitions" are easily weakened. This step utilizes the physical partitions and partition association matrices to perform time-series state updates, obtaining the final state vectors of the three partitions. This achieves the evolution and aggregation of time-series features that follow physical structural constraints, resulting in partition-specific state vectors. The specific steps are as follows: 1) Define three physical partitions based on a fixed set of channels. Specifically, the electrical partition corresponds to channel set {1, 2, 3, 4, 5, 11}; the heat flux partition corresponds to channel set {6, 7, 10, 12}; and the mechanical partition corresponds to channel set {8, 9}. All subsequent partitioning operations will use this fixed partitioning.

[0069] 2) From the physical correlation matrix Constructing a partition association matrix Partition association matrix The size is The matrix elements represent the average influence strength of one physical partition on another physical partition.

[0070] In practical implementation, the physical correlation matrix The average value of elements belonging to the corresponding sub-block of the partition channel is taken to obtain the original association strength between the three partitions; then, each row is normalized so that the sum of the elements in each row is 1. Based on this, the partition association matrix is ​​obtained. It can be used to describe the restricted information flow relationships between electrical zones, thermal flow zones, and mechanical zones.

[0071] It should be noted that the partition association matrix The elements are derived from the physical correlation matrix. This is aggregated, quantifying the physical impact of changes in different physical zones (such as the electrical zone) on another zone (such as the heat flow zone). During state updates, the gated loop unit (GRU) needs to refer to the states of other zones from the previous time step, and... The weights in the equation determine the "intensity limit" of this cross-regional information flow. If the influence weight of the electrical region on the heat flow region is 0.1, then when the heat flow region is updated, only 10% of its attention will be focused on the historical state of the electrical region. This avoids the model learning spurious correlations in the data (such as electrical fluctuations always coincidentally with temperature fluctuations in time), and ensures that the state transition follows the true physical laws.

[0072] 3) Traverse the convolutional feature tensor according to the compressed 64 time steps. Specifically, regarding the first... Each time step is compressed to extract the convolutional feature tensor. The slice at this time step, the size of the slice is Then, the channels within the three physical partitions are averaged and aggregated to obtain the input vectors for the electrical partition, the thermal flow partition, and the mechanical partition, with each partition input vector having a dimension of 32.

[0073] In one embodiment, for example, in the first Each time step is compressed to extract the convolutional feature tensor. Slice at this time step (size) The electrical partition contains channel indices 1, 2, 3, 4, 5, and 11. The 32-dimensional feature vectors of these six channels are averaged bitwise to obtain a 32-dimensional vector, which is the "electrical partition input vector," representing the comprehensive local feature representation of the electrical-related channels at that compression moment.

[0074] 4) Establish a gated loop unit for each of the three physical partitions, with the hidden state dimension of each gated loop unit set to 32. Specifically, the initial hidden states of the electrical partition, thermal flow partition, and mechanical partition are all set to a vector of all zeros. For the current time step, first, based on the partition correlation matrix... The hidden states of the three partitions in the previous time step are weighted and summed to obtain the cross-partition context vector of the current partition. The "current partition input vector" and the "current partition cross-partition context vector" are then concatenated and sent to the corresponding gated loop unit to complete the state update of the current time step.

[0075] In the specific implementation, the gated loop unit adopts the conventional update gate, reset gate and candidate state calculation method. The update gate and reset gate adopt the Sigmoid activation function, and the candidate state adopts the hyperbolic tangent activation function.

[0076] In one embodiment, heat flow partitions (including channels for temperature, flow rate, etc.) are used in... Taking the time-based update as an example, the specific operation is as follows: a) Obtain cross-region context: Query the partition association matrix Find the influence weight of the "electrical region on the heat flow region" (e.g., 0.6) and the influence weight of the "mechanical region on the heat flow region" (e.g., 0.2). Multiply the hidden state of the electrical region at the previous time step by 0.6, add the hidden state of the mechanical region at the previous time step by 0.2, and add its own previous time step state multiplied by its own influence weight (e.g., 0.2) to obtain a "cross-partition context vector" that mixes historical information from other partitions.

[0077] b) Current information fusion: The "heat flow partition input vector" calculated at the current moment is concatenated with the "cross-partition context vector" obtained in the previous step to form a longer vector.

[0078] c) State update: Input this concatenated vector into the gated loop unit dedicated to the heat flow partition. The gated loop unit determines how much old state to remember and how much new information to update through the gating mechanism, and finally outputs the hidden state of the heat flow partition at the current moment.

[0079] 5) By iterating along the 64 compressed time steps, the final state vector of the electrical partition is finally obtained. The final state vector of the heat flow partition and the final state vector of the mechanical partition The final state vectors for the three partitions all have a dimension of 32, representing the combined state evolution results of the current sample on the electrical, thermal, and mechanical sides, respectively.

[0080] In one embodiment, for example, if the physical correlation matrix The fluctuation rate channel of the excitation current has a strong influence on the temperature channel of the winding hot spot, therefore... The resulting partition association matrix In this context, the electrical partition has a higher association weight with the thermal flow partition. Therefore, when the thermal flow partition gating loop unit is updated, it will refer more to the hidden state of the electrical partition at the previous moment; conversely, if there is no obvious physical influence between two partitions, the corresponding partition association weight is lower, thereby avoiding unreasonable information mixing.

[0081] It should be noted that sequential recursion is the standard operating procedure for recurrent neural networks. That is, starting from the first compression time step, the data from the first step is input to obtain the hidden state of the first step; then this hidden state is compared with the partition association matrix. Together, they are used to calculate the cross-regional context in step 2, and then combined with the input data from step 2 to calculate the hidden state in step 2; this process is repeated, like a chain, transferring information from... Keep passing on and updating The hidden state output in step 64 is the "final state vector".

[0082] It should also be noted that this step directly writes the physical partition structure into the time-series state update process, instead of letting the network rely entirely on the data to learn all the channel coupling relationships on its own. This can reduce parameter redundancy, improve training stability on small sample anomaly categories, and enhance the interpretability of the model.

[0083] S303, Partition-weighted classification guided by digital twin reference vectors Using only the final state vectors of three partitions for classification still doesn't fully utilize the overall reference trajectory provided by the digital twin within the current window. This step involves using the normalized digital twin simulation matrix... The digital twin reference vector is extracted, and then the final state vectors of the three partitions are weighted according to the digital twin reference vector to obtain the final fused feature vector. And complete the classification of 7 operating states, the specific steps are as follows: 1) Normalize the digital twin simulation matrix Input the reference encoder.

[0084] In practical implementation, the reference encoder extracts local patterns from the digital twin reference trajectory along the time axis. The structure can employ two one-dimensional convolutional layers: the first layer has a kernel length of 5, 16 output channels, and a stride of 2; the second layer has a kernel length of 5, 32 output channels, and a stride of 2. Each convolutional layer is followed by a ReLU activation function. The reference encoder outputs a reference feature map, the size of which can be selected from the provided text. .

[0085] Furthermore, global average pooling is performed on the reference feature map along the physical attribute channel dimension and the time dimension to obtain the digital twin reference vector. Digital twin reference vector The dimension is 32, which represents the overall operational characteristics of the digital twin reference trajectory within the current time window.

[0086] 2) Convert the final state vector of the electrical partition. The final state vector of the heat flow partition and the final state vector of the mechanical partition Respectively with digital twin reference vectors The scores are concatenated and input into three independent scoring layers to obtain three partition scores. Then, Softmax normalization is applied to the three partition scores to obtain the weights for the electrical, thermal, and mechanical partitions. The sum of the three partition weights is 1, representing the relative importance of the three physical partitions in the current sample classification.

[0087] In the actual implementation, the three independent scoring layers have the same structure, each being a simple fully connected layer (linear transformation). Specifically, the partitioned state vector is compared with the digital twin reference vector. After concatenation, we get a dimension of The scoring layer maps this 64-dimensional vector to a single real number (i.e., the score value), which is then normalized by Softmax and converted into the weight coefficients of each partition.

[0088] Furthermore, the final state vectors of the electrical partition, the thermal flow partition, and the mechanical partition are weighted and summed using the obtained three partition weights to obtain the final fused feature vector. Finally, the feature vectors are fused. The dimension is 32, representing the global discriminative features after calibration by the digital twin reference vector.

[0089] 3) The final fused feature vector The input is a fully connected classification layer, which outputs classification logistic values ​​for 7 categories. A Softmax transformation is then performed on these classification logistic values ​​to obtain the predicted probability vector. Prediction probability vector The dimension is 7, and its 7 components correspond to the normal state, overload state, cooling failure state, bearing wear state, inter-turn short circuit state, oscillation instability state and sensor drift state in a fixed order.

[0090] In its implementation, the fully connected classification layer is a linear transformation layer that fuses the final feature vector of dimension 32. As input, it passes through a weight matrix (of size ). ) and a bias vector (size is Perform a linear mapping and output a vector with a dimension of 7. Each element of the vector is the classification logical value of the corresponding category.

[0091] 4) During the training phase, class-weighted cross-entropy loss is used to perform end-to-end optimization of the model. The number of samples in the class's runtime state is denoted as The maximum number of samples in all categories is denoted as Then the first The loss weights for class running states can be chosen. This aims to increase the impact of minority class samples on parameter updates. The optimizer can be the Adam optimizer, with an initial learning rate of [value missing]. The batch size can be 32, and the total number of training rounds can be 120. During the validation phase, the macro-average F1 score of the 7 running states is calculated, and the model parameters corresponding to the highest macro-average F1 score are saved as the final classification model.

[0092] In one embodiment, for example, if a sample simultaneously exhibits both elevated temperature and abnormal flow rate, then the digital twin reference vector... Typically, the sample is closer to the heat flux anomaly reference mode. In this case, the weight of the heat flux partition increases, and the model is more inclined to classify the sample as a cooling failure state. If a sample mainly shows a long-term offset of a single sensing channel relative to the digital twin, but the other channels maintain normal coupling, then the graph filter feature matrix... The cross-channel residuals in the model will not be enhanced synchronously, and the model is more inclined to classify the sample as a sensor drift state rather than a cooling failure state or an overload state.

[0093] It should be noted that the digital twin in this step not only participates in front-end bias correction and difference construction, but also in back-end partition weight allocation. The front end uses the digital twin as a physical benchmark to complete data layer alignment, and the back end uses the digital twin reference vector to complete decision layer calibration, thus forming a collaborative recognition mechanism of "preprocessing alignment + classification calibration".

[0094] S4, Online Recognition Process After completing model training, the final classification model will be deployed on the energy system online monitoring platform for real-time identification of equipment operating status. The specific steps are as follows: During the online phase, the same sampling method as in the training phase was used to continuously collect on-site measurement data for 12 physical attribute channels, and to synchronously read the simulation data for the corresponding time windows from the digital twin. Each new 256-point time window constitutes a sample to be identified.

[0095] Furthermore, for each sample to be identified, the following steps are performed sequentially: S201 adaptive bias correction of the twin guide channel, S202 structured graph filtering enhancement, and S203 robust normalization and fusion input tensor construction, to obtain the fusion input tensor. and normalized digital twin simulation matrix .

[0096] Furthermore, the input tensor will be fused. and normalized digital twin simulation matrix Input the trained partitioned temporal state classification model to obtain the predicted probability vector. .

[0097] Furthermore, the prediction probability vector The category with the highest probability is used as the identification result for the current time window; when the highest probability exceeds the preset confidence threshold, the corresponding status alarm is output. The confidence threshold can be set according to site requirements, for example, 0.75. If the identification result is a sensor drift state, the sensor calibration process can be triggered; if the identification result is a cooling failure state, bearing wear state, or inter-turn short circuit state, the equipment operation and maintenance early warning or protection linkage process can be triggered.

[0098] It should be noted that the preprocessing methods, normalization parameters, and classification model parameters performed for each time window in the online phase are consistent with those in the training phase. This ensures the consistency of data distribution between offline training and online deployment, and reduces recognition bias caused by inconsistent processing procedures.

[0099] Example 2 This embodiment provides a real-time monitoring system for the operating status of an energy system based on digital twins, which executes the real-time monitoring method for the operating status of an energy system based on digital twins described in Embodiment 1, including: Data acquisition module: used to collect real-time monitoring data on the operating status of the energy system and construct physical measurement matrix and digital twin simulation matrix; The data preprocessing module is used to perform channel adaptive bias correction using the digital twin simulation matrix to obtain the bias correction matrix; to perform structured graph filtering on the bias correction matrix using the physical correlation matrix obtained from the offline analysis of the digital twin to obtain the graph filtering feature matrix; and to perform channel-wise robust normalization on the bias correction matrix, the digital twin simulation matrix, and the graph filtering feature matrix to construct the fused input tensor. Model building module: Constructs a partitioned temporal state classification model with embedded digital twin references, extracts local temporal and cross-channel joint features from the fused input tensor, updates the state according to the physical partition structure, generates digital twin reference vectors using the normalized digital twin simulation matrix, weights the contributions of different physical partitions, and outputs the predicted probability of the running state; uses a loss function to train and optimize the partitioned temporal state classification model. Real-time monitoring module: This module is used to deploy the trained and optimized final partitioned time-series state classification model on the energy system online monitoring platform to identify the operating status of equipment in real time.

[0100] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of the operating status of an energy system based on digital twins, characterized in that, Includes the following steps: S1. Construct training samples for the energy system and label their states; For each sample, define the entity measurement matrix and the digital twin simulation matrix; S2. Use the digital twin simulation matrix to perform channel adaptive bias correction and obtain the bias correction matrix; Using the physical correlation matrix obtained from offline analysis of digital twins, the bias correction matrix is ​​subjected to structured graph filtering to obtain the graph filtering feature matrix; The bias correction matrix, digital twin simulation matrix, and graph filtering feature matrix are robustly normalized channel by channel to construct the fused input tensor. The specific process for obtaining the graph filtering feature matrix is ​​as follows: Offline construction of physical correlation matrix The size is Matrix elements Indicates the first The physical attribute channel for the first The standardized influence strength of each physical attribute channel; physical correlation matrix This was obtained through small perturbation sensitivity analysis using a digital twin model; For the bias correction matrix Processed column by column according to sampling time, the first Each sampling time corresponds to a 12-dimensional column vector. This represents the set of bias correction values ​​for the 12 physical attribute channels at the same sampling time; represented by the physical correlation matrix. Left multiply column vector This yields the physical prediction vector; the column vectors are then... Subtracting the weighted result of the physical prediction vector, we get the first... The output column vector of the graph filter at each sampling time. Perform the above operation one by one at each of the 256 sampling times, and then concatenate all the graph filter output column vectors in chronological order to obtain the graph filter feature matrix. ; S3. Construct a partitioned temporal state classification model with embedded digital twin references. Extract local temporal and cross-channel joint features from the fused input tensor. Update the state according to the physical partition structure. Generate digital twin reference vectors using the normalized digital twin simulation matrix. Weight the contributions of different physical partitions and output the predicted probability of the running state. Use a loss function to train and optimize the partitioned temporal state classification model. S4. Deploy the trained and optimized final partitioned time-series state classification model on the energy system online monitoring platform to identify the equipment operating status in real time.

2. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S1, multi-channel synchronous measurement data of the physical device are collected within a continuous time window, and synchronous simulation data output by the digital twin is read within the same time window: The defined entity measurement matrix is ​​as follows: and digital twin simulation matrix as All dimensions are Where 12 represents 12 physical attribute channels and 256 represents 256 consecutive sampling times; the 12 physical attribute channels are defined in a fixed order as: voltage fundamental amplitude, total harmonic distortion of current, active power oscillation component, reactive power fluctuation, frequency offset, bearing temperature gradient, winding hot spot temperature, shell vibration intensity, axial displacement, cooling medium flow deviation, excitation current fluctuation rate, and oil pressure change rate. For each sample, a running status label is given, denoted as . , No. The true category of each sample, the number of categories of the operating status label is fixed at 7, and they correspond to normal state, overload state, cooling failure state, bearing wear state, inter-turn short circuit state, oscillation instability state and sensor drift state in a fixed order.

3. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S2, different bias estimation methods are used for different channels to obtain the bias correction matrix. : Read the first Entity measurement matrix of sample and digital twin simulation matrix The deviation matrix is ​​obtained by subtracting the values ​​point by point according to the channel and the sampling time. The 12 physical property channels are divided into full-window offset channels and slow-change drift channels; For the full-window bias channel, the median of the bias sequence is calculated over the entire 256-point time window, and this median is used as the full-window bias of the current channel in the current sample. For the slowly drifting channel, the deviation sequence is segmented and statistically analyzed using a sliding segmentation method with a length of 32 and a step size of 16; the bias correction value is obtained by subtracting the bias value corresponding to the current sampling time from the entity measurement value. Perform the above operations on all 12 physical attribute channels, and rearrange them according to the original channel order and time order to obtain the bias correction matrix. The size is .

4. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S2, robust normalization and fusion of input tensors are constructed: On the training set, robust centers and robust scales for each of the 12 physical attribute channels are statistically analyzed; the robust center used for the bias correction matrix is ​​denoted as... , Indicates the first The median of the bias correction values ​​for each physical attribute channel in the training set; the corresponding robustness metric is denoted as . , Indicates the first The interquartile range of the bias correction values ​​for each physical attribute channel in the training set; The robust metric used for the graph filtering feature matrix is ​​denoted as . , Indicates the first The interquartile range of the graph-filtered feature values ​​of each physical attribute channel in the training set; For the bias correction matrix Robust normalization is performed on each channel to obtain the normalized bias correction matrix. ; Digital twin simulation matrix Use the same as the bias correction matrix and Perform normalization to obtain the normalized digital twin simulation matrix. ; Graph filtering feature matrix Divide by the corresponding channel The normalized graph filtering feature matrix is ​​obtained. ; Normalized bias correction matrix Subtract the normalized digital twin simulation matrix Obtain the normalized difference matrix between the physical entity and the digital twin. ; normalize the bias correction matrix Normalized graph filtering feature matrix and normalized difference matrix Stacking them according to the feature plane direction yields the fused input tensor. The size is .

5. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S3, a two-dimensional convolutional encoder is used to perform temporal compression and local joint feature extraction on the fused input tensor to obtain the convolutional feature tensor. The size is Where 32 represents the number of convolutional feature channels, 12 represents the number of physical attribute channels, and 64 represents the number of time steps after compression; the two-dimensional convolutional encoder adopts a two-layer stacked two-dimensional convolutional structure, including a first-layer two-dimensional convolution and a second-layer two-dimensional convolution; the first-layer two-dimensional convolution has 16 convolutional kernels with a kernel size of 3×7, where 3 covers the entire information plane and 7 corresponds to local continuous segments on the time axis; the time step is 2, and the channel direction uses a size-preserving padding method; batch normalization and ReLU activation function are applied after convolution; the second-layer two-dimensional convolution has 32 convolutional kernels with a kernel size of... The time step is 2, and the channel direction continues to use the size-preserving padding method; after convolution, batch normalization and ReLU activation function are also applied.

6. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S3, the state is updated according to the physical partition structure: Three physical zones are defined based on a fixed set of channels: electrical zone, thermal flow zone, and mechanical zone. From the physical correlation matrix Constructing a partition association matrix The size is : Physical correlation matrix The average value of the elements belonging to the corresponding partition channel sub-blocks is used to obtain the original association strength between the three partitions; each row is normalized so that the sum of the elements in each row is 1, resulting in the partition association matrix. ; Traverse the convolutional feature tensor at the compressed 64 time steps For the first Each time step is compressed to extract the convolutional feature tensor. The slice at this time step has a size of [size missing]. The channels within the three physical partitions are averaged and aggregated to obtain the input vectors for the electrical partition, the thermal flux partition, and the mechanical partition, with each partition input vector having a dimension of 32. A gated loop unit is established for each of the three physical zones to perform state updates, resulting in the final state vector of the electrical zone. The final state vector of the heat flow partition and the final state vector of the mechanical partition .

7. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 6, characterized in that, In step S3, the normalized digital twin simulation matrix is ​​used... The digital twin reference vector is extracted, and the final state vectors of the three partitions are weighted according to the digital twin reference vector to obtain the final fused feature vector. And complete the classification of 7 operating states: Normalized digital twin simulation matrix Input a reference encoder, output a reference feature map; The reference encoder extracts local patterns of the digital twin reference trajectory along the time axis using a two-layer one-dimensional convolution: the first layer has a kernel length of 5, 16 output channels, and a stride of 2; the second layer has a kernel length of 5, 32 output channels, and a stride of 2. Each convolution is followed by a ReLU activation function. Global average pooling is performed on the reference feature map along the physical attribute channel dimension and the time dimension to obtain the digital twin reference vector. ; The final state vector of the electrical partition The final state vector of the heat flow partition and the final state vector of the mechanical partition Respectively with digital twin reference vectors The scores are concatenated and input into three independent scoring layers to obtain three partition scores. Softmax normalization is applied to these three partition scores to obtain electrical partition weights, thermal flux partition weights, and mechanical partition weights. The three independent scoring layers are fully connected layers. The final state vectors of the electrical, thermal flux, and mechanical partitions are weighted and summed using the obtained three partition weights to obtain the final fused feature vector. ; The final fused feature vector Input a fully connected classification layer, and output classification logistic values ​​for 7 categories; perform a Softmax transformation on the classification logistic values ​​to obtain the predicted probability vector. .

8. The method for real-time monitoring of the operating status of an energy system based on digital twins according to claim 1, characterized in that, In step S3, the training phase employs class-weighted cross-entropy loss to perform end-to-end optimization of the partitioned temporal state classification model: The number of samples in the class's runtime state is denoted as The maximum number of samples in all categories is denoted as , No. The loss weights for the class's runtime state are The optimizer used is the Adam optimizer, with an initial learning rate of [value missing]. The batch size is 32, and the total number of training rounds is 120. During the validation phase, the macro-average F1 score of the 7 running states is calculated, and the model parameters corresponding to the highest macro-average F1 score are saved as the final partitioned time-series state classification model.

9. A real-time monitoring system for the operating status of an energy system based on digital twins, executing the real-time monitoring method for the operating status of an energy system based on digital twins as described in claim 1, characterized in that, include: Data acquisition module: used to collect real-time monitoring data on the operating status of the energy system and construct physical measurement matrix and digital twin simulation matrix; Data preprocessing module: used to perform channel adaptive bias correction using digital twin simulation matrix to obtain bias correction matrix; Using the physical correlation matrix obtained from offline analysis of digital twins, the bias correction matrix is ​​subjected to structured graph filtering to obtain the graph filtering feature matrix; The bias correction matrix, digital twin simulation matrix, and graph filtering feature matrix are robustly normalized channel by channel to construct the fused input tensor. Model building module: Constructs a partitioned temporal state classification model with embedded digital twin references, extracts local temporal and cross-channel joint features from the fused input tensor, updates the state according to the physical partition structure, generates digital twin reference vectors using the normalized digital twin simulation matrix, weights the contributions of different physical partitions, and outputs the predicted probability of the running state; uses a loss function to train and optimize the partitioned temporal state classification model. Real-time monitoring module: This module is used to deploy the trained and optimized final partitioned time-series state classification model on the energy system online monitoring platform to identify the operating status of equipment in real time.

Images