Intelligent power distribution room sensor fault early warning method and system based on multi-modal data fusion

Abstract

The invention provides an intelligent power distribution room sensor fault early warning method and system based on multi-modal data fusion. The method comprises the following steps: step 1, realizing working condition adaptive acquisition of multi-modal data; step 2, eliminating an alignment scheme of physical contradictions; step 3, noise analysis and filtering parameter adaptive cooperation are realized; 4, performing primary decomposition on the denoised signal, and analyzing a non-stationary signal; step 5, realizing security sharing and weighted aggregation of cross-node features; step 6, constructing a multi-modal sensor fault diagnosis model based on the time consistency fracture field; and step 7, optimizing hyper-parameters of the multi-modal sensor fault diagnosis model based on the time consistency fracture field. According to the invention, early warning of faults is realized, fault traceability and propagation path analysis capability are realized, deep support is provided for operation and maintenance decision, and the overall operation reliability and intelligent operation and maintenance level of the power distribution room are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of fault early warning technology, and in particular relates to a method and system for early warning of sensor faults in intelligent power distribution rooms based on multimodal data fusion. Background Technology

[0002] With the continuous growth of electricity demand and the ongoing expansion of the power grid, the importance of smart substations in the power system is becoming increasingly prominent. As a key node in power distribution and conversion, the stable operation of sensors in substations directly affects the reliability of power supply. Sensor failures can lead to inaccurate power monitoring data, potentially triggering power outages and causing significant economic losses and social impact. Internationally, developed countries and regions such as the United States and the European Union have been involved in smart grid construction for some time and have achieved certain results in early warning of sensor failures in substations. For example, the smart monitoring system developed with the support of the Electric Power Research Institute (EPRI) utilizes IoT and big data technologies to monitor the status of substation equipment in real time and establishes fault prediction models to detect potential sensor failures in advance. Some smart grid demonstration projects in the European Union employ multi-source data fusion algorithms to analyze and process sensor data, effectively improving the accuracy of fault warnings.

[0003] Research and application of smart distribution rooms in China are also developing rapidly. State Grid Corporation of China, China Southern Power Grid, and other enterprises are actively promoting the construction of smart grids and widely applying various sensors in the intelligent transformation of distribution rooms. However, traditional sensor fault early warning methods mostly rely on single types of data, making it difficult to comprehensively and accurately reflect the sensor's operating status, resulting in problems such as high false alarm rates and missed alarms. With the development of multimodal data fusion technology, its application to sensor fault early warning in smart distribution rooms has become a research hotspot, and it is expected to overcome the limitations of traditional methods and improve the reliability and timeliness of fault early warning. Summary of the Invention

[0004] Purpose of the Invention: This invention aims to address the problem of low accuracy in fault diagnosis of traditional power distribution room sensors. By fusing multimodal data such as temperature, humidity, current, and voltage, and combining advanced algorithms such as adaptive sampling adjustment, attention-guided diffusion denoising, and trend-aware Kalman filtering, a full-process fault early warning system from data acquisition to intelligent decision-making is constructed. This system enables early and accurate identification and warning of sensor faults in power distribution rooms, significantly improving fault diagnosis efficiency, ensuring the safe and stable operation of power distribution rooms, and effectively reducing operation and maintenance costs and the risk of power accidents.

[0005] This invention specifically provides a method for early warning of sensor faults in intelligent power distribution rooms based on multimodal data fusion, comprising the following steps: Step 1: Construct a chaotic index C(t) based on chaotic dynamics, determine the chaotic feature weights by combining the working condition classification model, and generate a dynamic sampling rate using a nonlinear adaptive sampling function to achieve working condition-adaptive acquisition of multimodal data. Step 2: For each core device, a dedicated physical mapping model corresponding to the device's operating mechanism is constructed. The Physical-Constrained Dynamic Time Warping (PC-DTW) method is adopted, and physical deviations are weighted and incorporated into the DTW distance (Dynamic Time Warping distance). A threshold is set to exclude alignment schemes with physical inconsistencies, so as to ensure the consistency between the timing alignment result and the physical components. The core devices include transformers, switchgear, and instrument transformers. Step 3: Based on the Attention-Driven Diffusion Model (A-DDM), the noise manifold features in the sensing signal are learned and noise feature parameters are output. Based on the noise feature parameters, the covariance matrix of the Trend-Aware Kalman Filter (T-AKF) is dynamically generated. The signal denoising is completed through state prediction and state correction processes, thereby achieving adaptive coordination of noise analysis and filtering parameters. Step 4: First, the denoised signal is decomposed using the entropy-constrained variational mode decomposition (E-VMD) method. Then, invalid modes are removed using the spectral kurtosis thresholding mechanism. Finally, the adaptive noise intensity CEEMDAN (AN-CEEMDAN) is used to adaptively adjust the noise intensity according to the signal complexity and complete the secondary decomposition, thereby achieving accurate analysis of non-stationary signals. Step 5: Extract the time-domain, frequency-domain, and time-frequency-domain features of each signal component, and use a multi-head attention mechanism to weight and filter the features to obtain key features; construct a node trust model for federated feature aggregation, and introduce device type weights to modify the node trust model, forming a Trust-Weighted Federated Aggregation (TWFA) framework; combine AES-128 encryption and key isolation mechanisms to achieve secure sharing and weighted aggregation of cross-node features; Step 6: Construct a multimodal sensor fault diagnosis model based on time-consistency fracture field (Temporal ConsistencyBreakageFieldModel, TCBFM): First, model the cross-modal time-consistency fracture field, calculate the consistency gradient tensor to characterize the dynamic changes and stability of the fracture field, then quantitatively describe the degree of structural damage through the consistency fracture energy density, and finally extract the fracture morphology fingerprint vector as the fault feature representation. Step 7: Use the multi-objective adaptive Bayesian optimization algorithm (MO-ABO) to optimize the hyperparameters of the multimodal sensor fault diagnosis model (TCBFM) based on the time-consistent fracture field.

[0006] Beneficial effects: This invention constructs a complete multimodal data processing chain by combining multi-source sensing acquisition and cross-modal timing calibration with a two-stage denoising and joint decomposition algorithm, which significantly improves signal quality and feature representation capabilities, and provides a reliable data foundation for fault early warning.

[0007] By using the TCBFM model to characterize cross-modal temporal consistency structure changes, the model accurately captures consistency breakage features caused by faults without relying on single-modal temporal features or fixed time scales. Combined with the MO-ABO multi-objective optimization strategy, the model maintains stable and accurate diagnostic performance under complex and variable operating conditions, significantly improving the robustness and generalization ability of the system.

[0008] This invention is based on the TWFA trust-weighted federated aggregation framework to achieve trusted fusion of cross-node features and edge-center collaborative modeling; it employs AES-128 encryption and key isolation mechanisms to ensure data privacy and security. Simultaneously, through a multi-model fusion decision module with attention weighting and feature consistency constraints, it achieves lightweight and scalable integrated diagnostics, significantly improving model convergence efficiency and final recognition accuracy.

[0009] Through the causal reasoning and root cause analysis modules, the system not only enables early warning of faults, but also has the ability to trace fault sources and analyze propagation paths, providing in-depth support for operation and maintenance decisions and effectively improving the overall operational reliability and intelligent operation and maintenance level of the power distribution room. Attached Figure Description

[0010] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0011] Figure 1 This is a flowchart of the method of the present invention.

[0012] Figure 2 The flowchart is for the MO-ABO algorithm. Detailed Implementation

[0013] This embodiment provides a method for early warning of sensor faults in intelligent power distribution rooms based on multimodal data fusion, including the following steps:

[0014] Step 1: Construct a chaotic index C(t) based on chaotic dynamics, determine the chaotic feature weights by combining the working condition classification model, and generate a dynamic sampling rate using a nonlinear adaptive sampling function to achieve working condition-adaptive acquisition of multimodal data.

[0015] Step 2: For each core device, a dedicated physical mapping model corresponding to the device's operating mechanism is constructed. The Physical-Constrained Dynamic Time Warping (PC-DTW) method is adopted, and physical deviations are weighted and incorporated into the DTW distance (Dynamic Time Warping distance). A threshold is set to exclude alignment schemes with physical inconsistencies, so as to ensure the consistency between the timing alignment result and the physical components. The core devices include transformers, switchgear, and instrument transformers.

[0016] Step 3: Based on the Attention-Driven Diffusion Model (A-DDM), the noise manifold features in the sensing signal are learned and noise feature parameters are output. Based on the noise feature parameters, the covariance matrix of the Trend-Aware Kalman Filter (T-AKF) is dynamically generated. The signal denoising is completed through state prediction and state correction processes, thereby achieving adaptive coordination of noise analysis and filtering parameters.

[0017] Step 4: First, the denoised signal is decomposed using the entropy-constrained variational mode decomposition (E-VMD) method. Then, invalid modes are removed using the spectral kurtosis thresholding mechanism. Finally, the adaptive noise intensity CEEMDAN (AN-CEEMDAN) is used to adaptively adjust the noise intensity according to the signal complexity and complete the secondary decomposition, thereby achieving accurate analysis of non-stationary signals.

[0018] Step 5: Extract the time-domain, frequency-domain, and time-frequency-domain features of each signal component, and use a multi-head attention mechanism to weight and filter the features to obtain key features; construct a node trust model for federated feature aggregation, and introduce device type weights to modify the node trust model, forming a Trust-Weighted Federated Aggregation (TWFA) framework; combine AES-128 encryption and key isolation mechanisms to achieve secure sharing and weighted aggregation of cross-node features;

[0019] Step 6: Construct a multimodal sensor fault diagnosis model based on time-consistency fracture field (Temporal ConsistencyBreakageFieldModel, TCBFM): First, model the cross-modal time-consistency fracture field, calculate the consistency gradient tensor to characterize the dynamic changes and stability of the fracture field, then quantitatively describe the degree of structural damage through the consistency fracture energy density, and finally extract the fracture morphology fingerprint vector as the fault feature representation.

[0020] Step 7: Use the multi-objective adaptive Bayesian optimization algorithm (MO-ABO) to optimize the hyperparameters of the multimodal sensor fault diagnosis model (TCBFM) based on the time-consistent fracture field.

[0021] In step 1, the multi-source sensors are deployed as follows: the vibration sensor is installed about 50cm outside the transformer terminal block; the acoustic sensor is placed at the center of the switch cabinet door panel; the temperature sensor is placed about 30cm above the busbar; and the current and voltage sensors are installed at the standard CT / VT measuring points.

[0022] The system collects current, voltage, temperature, vibration, and acoustic signals to construct a timestamped multimodal time-series dataset, and labels each data sample with the corresponding device status tag. It also adds an adaptive sampling rate adjustment mechanism and a cross-modal data time-series calibration algorithm to solve the problems of redundant data or loss of key information caused by traditional fixed sampling rates, as well as data misalignment caused by clock deviations of different sensors.

[0023] 1) To address the challenges of traditional linear volatility-based sampling strategies failing to accurately capture fault precursors and the inability of fixed thresholds to adapt to various operating conditions, this invention proposes a nonlinear adaptive sampling rate adjustment method based on chaotic dynamics theory. This method calculates the chaotic characteristic indices of multi-source sensor signals, constructs nonlinear sampling rate adjustment rules, and combines diagnostic feedback to achieve self-calibration of the threshold, thereby forming a dynamically evolving sampling strategy.

[0024] First, the phase space of each sensor signal is reconstructed, and three chaotic features are calculated within a sliding window: the maximum Lyapunov exponent, the correlation dimension, and the local trajectory divergence rate. To achieve fusion quantization, a unified chaotic index C(t) is constructed to describe the chaotic intensity of the signal within the window.

[0025] ,

[0026] in, The normalized value of the maximum Lyapunov exponent; This is the normalized value of the correlation dimension; This is the normalized value of the local trajectory divergence rate (finite-time Lyapunov exponent); , and As weight;

[0027] To ensure that the chaos index reflects the actual disturbance under different operating conditions, a chaos feature sensitivity modeling mechanism driven by operating condition classification is introduced. By modeling the response characteristics of chaos features under different operating conditions, three types of feature sensitivity vectors F are obtained for weight calculation:

[0028] F=[F1,F2,F3],

[0029] Where F1 is the sensitivity component of the maximum Lyapunov exponent; F2 is the sensitivity component of the correlation dimension; and F3 is the sensitivity component of the local trajectory divergence rate.

[0030] The weights are obtained by normalizing the feature sensitivity:

[0031] ,

[0032] This scheme ensures the physical consistency of weights, interpretability of operating conditions, and high stability, avoiding the "collapse" problem common in weight iteration methods.

[0033] Based on model analysis and simulation verification, this invention clarifies the differences in sensitivity of three types of chaotic features and their optimal weight configuration under typical power distribution room conditions, as shown in Table 1.

[0034] Table 1

[0035]

[0036] Based on the interval where the chaotic index C(t) is located, the sampling rate f is determined using a piecewise nonlinear function. s (t):

[0037] ,

[0038] in The minimum sampling rate (the lower limit of the basic sampling rate corresponding to the signal's stable region); It is the time constant (the decay rate control parameter of the exponential decay function). is the first threshold of the chaos index; a, b, and c are polynomial coefficients (quadratic polynomial coefficients used for calculating the sampling rate in the chaos budding region). The second threshold of the chaos index (the critical value that divides the "chaotic nascent region" and the "strong chaotic state"); θ1 represents the maximum sampling rate (the upper limit of the fixed sampling rate used in a strongly chaotic state); e represents the natural constant. When C(t) < θ1, the signal is in the stable region, and an exponentially decreasing sampling rate is used to reduce redundant data. When θ1 ≤ C(t) ≤ θ2, the signal enters the chaotic nascent region, and the sampling rate increases nonlinearly with the chaos index in a polynomial manner. When C(t) > θ2, the system enters a strongly chaotic state, the sampling rate is fixed at the maximum value, and the edge nodes are triggered to perform short-time denoising, mode decomposition, and feature pre-extraction to reduce the communication pressure caused by high-frequency sampling.

[0039] Traditional methods use fixed thresholds, which cannot adapt to different operating conditions. This invention constructs a closed-loop "acquisition-diagnosis-calibration" system, updating the threshold every 5 minutes based on the accuracy of the diagnostic model.

[0040] ,

[0041] Where, θ i (t) represents the threshold of the i-th chaotic interval at time t; η is the learning rate, used to control the step size of threshold updates; Acc is the fault diagnosis accuracy obtained based on the current sampling strategy; This indicates the sensitivity of accuracy to a threshold, reflecting the direction and magnitude of the impact of threshold changes on diagnostic performance.

[0042] This mechanism enables automatic adaptation to operating condition switching, automatic correction of threshold deviation, and continuous evolution of sampling strategies, making the sampling strategies no longer static but possessing self-learning capabilities.

[0043] In step 2, to solve the problem of "time alignment but physical inconsistency" in traditional dynamic time warping (DTW), this invention proposes a device-specific physical constraint PC-DTW method for core equipment such as transformers, switchgear, and instrument transformers in power distribution rooms. Through differentiated physical mapping, innovative constraint formulas, and operating condition adaptation parameters, it achieves a dual improvement in timing alignment accuracy and physical consistency.

[0044] Based on the working principle of the equipment, a cross-modal signal mapping model that conforms to electrical and mechanical laws is constructed. The core mapping relationships are as follows: Transformer: Focusing on the electrical coupling of "voltage-current-power" and the thermal correlation of "oil temperature-load", a dual constraint is imposed by power deviation (reflecting electrical consistency) and oil temperature deviation (reflecting thermal rationality); Switchgear: In view of the characteristics of mechanical faults, a mapping is established between "vibration-contact resistance" (the core indicator of contact loosening) and "acoustic-closing and opening time" (the status indicator of the operating mechanism) to capture mechanical anomalies; Current transformer: Based on the transformation ratio characteristics, the deviation is calculated by "secondary side signal-primary side load" to ensure electrical matching of high and low voltage side signals.

[0045] Based on traditional DTW time similarity, a novel physical constraint is introduced, as shown in the following formula:

[0046] Weighted integration of equipment-specific physical deviations into the DTW local distance balances time alignment and physical consistency.

[0047] ,

[0048] in, Let be the physically constrained DTW local distance between the i-th time and the j-th time. The maximum deviation value of the k-th type of physical mapping; It is the locally defined dynamic time warping (DTW) distance with physical constraints between the signal to be matched at time i and the reference signal at time j, which is used to simultaneously measure the time alignment and physical parameter consistency of the signals at these two times. It is the deviation value of the k-th type of physical mapping between the signal to be matched at time i and the reference signal at time j; It is the weight of the physical mapping term. Adjust the priority of different physical constraints; K is the number of physical mapping items corresponding to the equipment (K=2 for transformers / switchgear, K=1 for instrument transformers).

[0049] Set physical threshold (Range 0.01~0.1, normalization bias), if If the local distance between the sampling point pairs is set to infinity, the alignment scheme that would otherwise be physically contradictory is forcibly eliminated, ensuring that the calibration results conform to engineering practice.

[0050] To balance timing calibration accuracy and fault diagnosis support capabilities, a multi-objective optimization strategy is adopted to determine the physical threshold. With weight The optimal parameters are: to minimize the post-calibration data reconstruction error. recon Ensure time alignment accuracy and maximize fault feature identification accuracy (Acc) feat The subsequent diagnosis is supported by two objectives. The Pareto optimal solution is searched within the parameter range using the NSGA-II algorithm to balance the diversity of solutions and engineering applicability. Finally, after verifying the generalization of parameters on different operating conditions and equipment models, the parameters are fixed according to the "equipment-operating condition" dimension. The parameters are shown in Table 2.

[0051] Table 2

[0052]

[0053] Step 3 includes:

[0054] To address the shortcomings of traditional two-stage denoising, which suffers from "no linkage between denoising and filtering and static covariance matrix," this invention innovatively proposes a noise manifold-driven trend-aware filtering mechanism. This breaks through the conventional path of diffusion models that "directly output denoised signals," allowing the attention-guided diffusion model (A-DDM) to focus on learning the high-dimensional manifold features of noise distribution. This dynamically generates the core parameters of the trend-aware Kalman filter (T-AKF), achieving deep synergy between "noise feature analysis and adaptive filtering parameters," significantly improving signal purity and fault feature retention under complex interference.

[0055] The core function of A-DDM has shifted from "signal denoising" to "noise manifold modeling." It simulates noise evolution through forward diffusion and analyzes the noise structure through backward diffusion, ultimately outputting the key features driving the filtering, rather than the denoised signal: Forward diffusion modeling: Gaussian noise is gradually added to the original signal x0 to generate a noisy sequence x at time t. t The mathematical expression for simulating the evolution of noise from weak to strong is:

[0056] ,

[0057] in, Let be the diffusion coefficient, which increases linearly from 0.001 to 0.02 at t = 1, 2, ..., T, where T is the total number of time points in the forward diffusion process. Standard Gaussian noise is used to allow the model to learn the noise distribution pattern.

[0058] Backdiffusion and noise manifold extraction: Training a noise prediction network , through reverse diffusion from x T During the recovery of x0, the focus is on learning the high-dimensional distribution pattern of the noise (noise manifold), ultimately outputting two core features: noise manifold coordinates z(t): describing the geometric structure of the noise in high-dimensional space (such as correlation and clustering), with the dimension consistent with the signal sampling length; and local noise energy map E(t): combining the signal's local attention weights. The statistical distribution of local noise intensity reflects the local aggregation characteristics of noise.

[0059] Based on the z(t) and E(t) outputs of the attention-guided diffusion model (A-DDM), T-AKF overcomes the limitation of "static Q and R settings" by constructing a dynamic covariance generation logic, enabling the filtering parameters to match the current noise and signal trends in real time. Specifically, this includes:

[0060] R(t) is the noise intensity of the measurement process at time t, which is composed of the "global interference intensity of the noise manifold coordinates z(t)" and the "signal trend factor". "A combined driving force ensures that trend-based fault signals are not over-filtered, as shown in the formula:"

[0061] ,

[0062] in, This is an intermediate parameter; the value used in this embodiment is... m is the signal sampling length; The larger the value, the stronger the global noise interference, and R(t) increases accordingly; exp is the natural exponential function; Let be the i-th component of the noise manifold coordinates at time t, and be the local noise feature at time t output by the attention-guided diffusion model (A-DDM).

[0063] Q(t) describes the noise variation during signal transmission, which is determined by the concentration intensity of the local noise energy map E(t) and the signal prediction trend. Driven by this, the ability to suppress strong local interference is enhanced, and the formula is:

[0064] ,

[0065] in, and This is an intermediate parameter; in this embodiment, it takes the following value: , ; The larger the value of (the maximum value of the local noise energy map at time t), the stronger the local impulse interference at time t, and (Q(t)) increases accordingly; The trend characteristics corresponding to the predicted value of the signal at time t.

[0066] Based on dynamically generated Q(t) and R(t), T-AKF completes filtering according to the following steps:

[0067] State prediction: using the results of the previous time step With state transition matrix To obtain the predicted state ;

[0068] Gain Calculation and State Correction: The Kalman gain is calculated by combining dynamic Q(t) and R(t), and then the sensor's measured value y is used. k Correct the prediction state and output the denoised clean signal. .

[0069] In step 4, an improved joint mode decomposition strategy is applied to the denoised signal from step 3, including:

[0070] Step 4-1, Entropy-Constrained Variational Mode Decomposition (E-VMD): Add modal entropy constraint terms to the variational functional of VMD. Construct a new objective function :

[0071] ,

[0072] in, For the k-th mode Shannon entropy, is the entropy constraint coefficient (set to 0.05); K is the number of modes. Here, j is the time partial derivative operator, and j is the imaginary unit. Let be the center frequency of the k-th mode. The penalty factor is x(t), and the original input time-domain signal is x(t). The functional is solved by the alternating direction multiplier method to complete the mode decomposition of E-VMD.

[0073] Step 4-2, Mode Filtering Mechanism Based on Spectral Kujicity Threshold: To reduce invalid modes entering the secondary decomposition stage, a spectral kujicity (SK) filtering mechanism is introduced after E-VMD. For each primary mode u k (t), calculate the envelope kurtosis. :

[0074] ,

[0075] if Less than the threshold If the threshold value is between 2.5 and 3.8, it is determined to be a weakly correlated or noisy mode, and the primary mode u is... k (t) are removed and do not enter the adaptive noise intensity CEEMDAN reconstruction stage; through this screening mechanism, the amount of secondary decomposition can be reduced by 40%~60%, which greatly improves the processing efficiency. It is the mathematical expectation operator, used to calculate the mean of signal statistics, and is the core operator in the spectral kurtosis formula for statistically averaging signal features; u k It is the primary mode u k (t) The envelope spectrum signal after envelope spectrum transformation; since the transformation from the time domain to the envelope spectrum domain has been completed, the signal no longer changes with time t, so the time variable t is omitted;

[0076] Step 4-3, Adaptive Noise Intensity CEEMDAN (AN-CEEMDAN): Define signal complexity The approximate entropy of the signal and the mean of the sample entropy are:

[0077] ,

[0078] in, The approximate entropy of the signal. The entropy of the signal sample; according to Adjust the intensity of the added Gaussian white noise :

[0079] ,

[0080] in, This represents the standard deviation of the original signal.

[0081] Step 5 includes: constructing a dynamic feature selection and feature interaction term mining module based on an attention mechanism. The dynamic feature selection and feature interaction term mining module first calculates the time-frequency entropy, kurtosis and envelope spectrum features of each component obtained in step 4, then generates a feature weight vector through a multi-head attention mechanism, and filters features with weight values ​​greater than the threshold δ to complete the dynamic selection. Finally, it combines the Trust Weighted Federated Aggregation (TWFA) framework to realize feature sharing across distribution room nodes, effectively solving the aggregation deviation problem caused by the difference in data quality between different nodes.

[0082] Because the operating characteristics and state sensitivities of different types of equipment in a power distribution room vary significantly, traditional TWFA does not differentiate the importance of nodes, which can easily lead to insufficient weighting of key node features during the aggregation process. Therefore, this invention constructs a node equipment type weight set. :

[0083] ,

[0084] in, It is the transformer node weight; It is the node weight of the switchgear; These are the capacitor node weights; These are the weights of other auxiliary nodes (corresponding to the node weights of auxiliary equipment such as voltage transformers, current transformers, surge arresters, and busbar monitoring).

[0085] The node device type weights are determined based on the device's importance in the power distribution system, historical fault impact factors, and data reliability assessment results. By introducing node type weights, key nodes contribute more to global aggregation, thereby enhancing the overall feature representation capability.

[0086] To address the issue of insufficient node diversity caused by traditional methods relying solely on data quality assessment for trust scores, this invention couples node device type weights with the trust score calculation process, proposing a modified trust score model:

[0087] ,

[0088] in It is the corrected trust level of the i-th node; The device type weight for the i-th node; These are node feature quality metrics, such as feature stability, reconstruction error, variational consistency, etc. This represents the node distribution deviation, used to reflect the correlation between nodes and the global feature distribution; This is a correlation penalty term used to suppress the impact of abnormal nodes on aggregation.

[0089] This correction method can simultaneously reflect the importance of nodes and feature confidence, allowing key nodes to be expressed first during the aggregation process, while reducing the impact of drifting or noisy nodes.

[0090] To address the potential data leakage risks during the aggregation of multi-node features, this invention introduces a lightweight secure transmission module into the TWFA architecture to achieve end-to-end encryption of cross-node feature vectors. The process includes:

[0091] Before uploading the feature vector, the node processes the original vector f. i Perturbation and encryption processes are performed to obtain the encrypted features. :

[0092] ,

[0093] in, AES-128 encryption function; It is designed to reduce noise disturbances and enhance resistance to reverse propulsion.

[0094] Encrypted features cannot be directly parsed during transmission, preventing data leakage at the source.

[0095] To avoid excessive access to node data by the aggregator, this invention employs a key isolation mechanism: keys are partitioned and stored in different secure areas; the aggregator only decrypts necessary feature statistics in segments; it does not decrypt complete original features, thus achieving the principle of minimum exposure across the entire link.

[0096] This method ensures that the aggregation center only obtains the low-dimensional statistical features required for global aggregation, further reducing the risk of data leakage.

[0097] In step 6, to overcome the problem that existing fault diagnosis methods mainly rely on time series feature learning and are difficult to effectively characterize the implicit temporal consistency structure changes between multimodal sensing signals, a multimodal sensor fault diagnosis model based on the temporal consistency breakage field (TCBFM) is established.

[0098] This model does not model the time variation characteristics of a single mode or a fixed time scale, but rather models the cross-modal time consistency relationship formed by multimodal sensing signals under different time lag conditions as a continuous consistency fracture field structure. When the equipment or sensor is in normal operation, this consistency fracture field remains continuous in the time-lag space; when a fault occurs, the consistency fracture field will exhibit local fractures or structural damage. This invention achieves the identification of fault types and warning levels by extracting and discriminating consistency fracture features. Specifically, it includes the following steps:

[0099] Step 6-1: Construct a time-consistent fracture field;

[0100] Let the sampled value of the m-th sensing mode at time t be... To address the physical coupling relationships between different modes, a cross-modal time-consistent fracture field is constructed:

[0101] ,

[0102] in It is a cross-modal time-consistent fracture field, measuring the difference between the m-th mode (time t) and the n-th mode (lag). The degree of physical coupling consistency; It is the sampled value of the m-th sensing mode at time t; The nth sensing mode lags behind at time t. The sampled value at time; This refers to the time lag. (⋅) is an asymmetric physical consistency kernel function built based on the device operation mechanism, used to describe the degree of consistency between the responses of mode m and mode n under a given time delay; For modal coupling weights;

[0103] By accumulating the consistency measures under different modal combinations and lag conditions, a time consistency fracture field in the time-lag two-dimensional space is formed to describe the overall consistency structure state of the system at the current moment.

[0104] Step 6-2: Calculate the consistency gradient tensor;

[0105] To characterize the variation features of the time-consistent fracture field in both the time and lag dimensions, a consistency gradient tensor is defined on the constructed consistency fracture field. :

[0106] ,

[0107] in, It is the component of the gradient in the time direction, used to reflect the dynamic rate of change of the uniform structure; It is the component of the gradient in the hysteresis direction, used to reflect the stability of the cross-modal time alignment relationship;

[0108] When the power equipment operation monitoring system in the power distribution room is in normal working condition, the above gradient changes gradually; when a fault occurs in the equipment in the system, the gradient will show a significant abrupt change in the corresponding fault area.

[0109] Step 6-3: Calculate the uniform fracture energy density;

[0110] To quantitatively describe the degree of damage to the uniform fracture field structure, a uniform fracture energy density function is defined. :

[0111] ,

[0112] in, Represents the vector norm; θ is the fracture detection threshold; The indicator function is d; d is the differential symbol.

[0113] This energy density is used to characterize the fracture strength that cannot maintain continuity in a time-consistent fracture field, and its magnitude is closely related to the degree of failure.

[0114] Step 6-4, Fault identification driven by fracture morphology;

[0115] Based on the temporal distribution characteristics of the uniform fracture energy density, a fracture morphology fingerprint vector is constructed. :

[0116] ,

[0117] in, These represent the mean, skewness, and kurtosis of the fracture energy distribution, respectively.

[0118] Based on the aforementioned feature analysis of time-consistent fracture field and consistent gradient tensor, the currently constructed fracture morphology fingerprint vector is matched and calculated with the preset sensor fault feature library. By comparing the similarity of vector features, the corresponding sensor fault type and warning level are output, and finally the diagnosis and warning of sensor operating status are realized.

[0119] Step 7 includes: MO-ABO, based on Bayes' theorem, approximates the black-box function by constructing a joint probability model of multiple objective functions. Let there be m objective functions, denoted as... ,in To optimize variables (hyperparameter vector). Let m be the m-th objective function. Bayes' theorem is used to update the posterior distribution of the objective function:

[0120] ,

[0121] in, It is the joint posterior probability of the objective function given the observed dataset D; It is the likelihood function of the objective function, representing the probability of the observed data given the objective function value; It is the joint prior probability of the objective function, usually assumed to be the Gaussian Process (GP) prior; It is the marginal probability, used for normalization;

[0122] The probabilistic transitive model employs a Gaussian process, with each objective function... Modeled as a Gaussian process GP prior:

[0123] ,

[0124] in It is a mean function, usually set to a constant or zero; It is the covariance function (kernel function), such as the squared exponential kernel, used to capture the smoothness of the objective function; the prior distribution is initialized based on historical data and the posterior distribution is obtained through Bayesian update;

[0125] Acquisition functions are used to balance exploration and exploitation. In multi-objective optimization problems, Expected Hypervolume Improvement (EHVI) is commonly used as the core acquisition function. Its core function is to quantify the expected improvement effect of new sampling points on the Pareto front hypervolume.

[0126] ,

[0127] in It is the expected hypervolume improvement acquisition function value, used to quantify the expected improvement of the Pareto front hypervolume by the new sampling point x; This is the hypervolume improvement, calculated as the volume of the region dominated by the newly added points. The adaptive mechanism adapts to the complexity of the objective function by dynamically adjusting the parameters of the acquisition function or the model hyperparameters, thereby improving optimization efficiency.

[0128] Establishing optimization variables: To achieve automated tuning of the Time-Consistent Fracture Field-Based Multimodal Sensor Fault Diagnosis Model (TCBFM), this invention uses the key structural hyperparameters of the TCBFM model as optimization variables, assuming the hyperparameter vector x is:

[0129] ,

[0130] Where x i This represents the i-th structural hyperparameter to be optimized;

[0131] Structural hyperparameters include: the range of time lag values ​​τ. range This is used to define the effective lag interval of cross-modal time consistency relationships, affecting the integrity of the fracture field construction; the width σ of the physical consistency kernel function. k The sensitivity of the metric used to adjust the consistency of responses between modalities; the feature dimension D of the fracture morphology fingerprint vector. f This is used to set the dimension size for representing fault features; the hidden layer dimension H of the fault discrimination module dThis is used to adjust the fitting ability and generalization performance of fault classification;

[0132] Establish the objective function: The objective function is used to evaluate the performance of the hyperparameter vector x. In the fault diagnosis model, considering two or more competing objectives, such as diagnostic accuracy, computational efficiency, and model robustness, the following four objective functions are defined:

[0133] The fault diagnosis accuracy is represented by maximizing the F1 score on the test set.

[0134] Minimize the model inference time (in seconds);

[0135] This represents the number of model parameters (in the millions), which should be minimized to control complexity.

[0136] This indicates the maximum accuracy of anti-interference diagnosis.

[0137] objective function vector for:

[0138] ,

[0139] Since the objective function is a black-box function, its value is obtained by training the TCBFM model and evaluating it on the validation set. The multi-objective optimization problem is expressed as:

[0140] ,

[0141] in X is the vector of structural hyperparameters to be optimized, i.e., a vector composed of multiple structural hyperparameters to be optimized; The feasible region, i.e., the set of values ​​for all structural hyperparameter vectors that conform to the actual physical constraints and algorithmic constraints, is used to limit the... The valid value boundaries.

[0142] Note: Accuracy and Transforming into a minimization problem is equivalent to minimizing The optimization objective is to find the Pareto optimal solution set. , of which for any There are no other Make Not inferior to in any goal And it is better at least in one objective;

[0143] MO-ABO optimizes hyperparameters through an iterative process, the specific steps of which are as follows:

[0144] Step a1, Initialization: Generate an initial observation set based on prior information (such as empirical values ​​or random sampling). ,in This represents the initial number of sampling points;

[0145] Step a2, Probabilistic Model Update: For each objective function f k (x), use Gaussian process regression to fit the observed data, update the posterior distribution, and the posterior mean. and variance for:

[0146] ,

[0147] ,

[0148] in It is the kernel matrix corresponding to the k-th target, which is composed of the kernel function values ​​of all sample pairs in the sampled point matrix X, and is used to characterize the similarity between sampled points; This is the matrix of sampled points. For the corresponding target value vector, Let I be the noise variance; I is the identity matrix.

[0149] The adaptive mechanism dynamically adjusts the kernel function hyperparameters by maximizing the marginal likelihood. :

[0150] ,

[0151] in These are the hyperparameters of the kernel function; This indicates that the sampled point matrix X and the kernel function hyperparameters... Under the condition, the target value vector The corresponding marginal likelihood probability density is used to measure the reasonableness of the target value distribution under the current hyperparameters.

[0152] Step a3, acquisition function optimization: Use the EHVI acquisition function to select the next sampling point. :

[0153] ,

[0154] in The desired hypervolume improvement is the acquisition function value; the EHVI calculation is based on the current Pareto front P and the hypervolume reference point r; in terms of adaptation, the acquisition function may incorporate uncertainty weights, such as adding an entropy term to enhance exploration.

[0155] Step a4, Evaluation and Update: At point Evaluation of objective function value (Through model training and validation), update the observation set:

[0156] ,

[0157] Where n represents the current iteration round number, D n It is the observation set after the nth iteration;

[0158] Step a5, Convergence check: Repeat steps a2 to a4 until the convergence condition is met (such as the maximum number of iterations, the hypervolume change being less than a threshold). (Or calculate budget exhaustion). Output the Pareto front after convergence. .

[0159] This embodiment also provides an intelligent power distribution room sensor fault early warning system based on multimodal data fusion implemented using the aforementioned method, comprising:

[0160] Multi-source sensor acquisition module: A multi-source sensor array deployed at key nodes in the power distribution room, integrating adaptive sampling rate adjustment and cross-modal timing calibration functions;

[0161] Dual-stage denoising module: realizes the cascaded processing of attention-guided diffusion denoising (A-DDM) and trend-aware Kalman filtering (T-AKF);

[0162] Signal parsing module: Executes the VMD-CEEMDAN joint decomposition algorithm based on entropy constraints;

[0163] Federated Feature Selection Module: Integrates attention mechanism dynamic feature selection and interaction term mining unit, supports Trust Weighted Federated Aggregation (TWFA).

[0164] Hybrid diagnostic module: Based on the TCBFM model, extract time-consistent fracture features (including consistency gradient tensor, fracture energy density, and fracture morphology fingerprint vector). Under MO-ABO optimization, input the fracture morphology fingerprint vector into the decision module to generate the final diagnostic output.

[0165] Early warning output module: generates fault category labels, anomaly scores and root cause location reports, and provides fault propagation path analysis results to enable early warning and intelligent operation and maintenance decisions.

[0166] This embodiment also provides an electronic device, including a processor and a memory, wherein the memory stores program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0167] This embodiment also provides a storage medium storing a computer program or instructions that, when run on a computer, execute the steps of the method described.

[0168] In this embodiment, a smart distribution room in a 110kV substation in a certain region is used as the operating environment to provide a specific example to verify and illustrate the sensor fault early warning method for smart distribution rooms based on multimodal data fusion proposed in this invention. This distribution room contains one main transformer, six high-voltage switchgear cabinets, and several instrument transformers. Its operating conditions are complex, and it has long suffered from problems such as sensor signal drift, frequent instantaneous interference, and a high false alarm rate.

[0169] (1) Multimodal data acquisition and adaptive sampling results

[0170] like Figure 1 As shown, the system first acquires current, voltage, temperature, vibration, and acoustic signals from the transformer and switchgear through a multi-source sensor acquisition module. The initial base sampling rate for current and voltage is set to 2kHz, for temperature signals to 1Hz, and for vibration and acoustic signals to 8kHz.

[0171] The system performs sliding window processing (window length 2s, step size 0.5s) on vibration and acoustic signals at edge nodes, calculates the maximum Lyapunov exponent, correlation dimension, and local trajectory divergence rate, and obtains a unified chaos index C(t). During normal equipment operation, C(t) is mainly distributed in the range of 0.15 to 0.25; when slight signs of loosening appear in the switchgear contacts, the chaos index of vibration and acoustic signals rises to above 0.42.

[0172] Based on the changes in the chaotic index, the system dynamically increases the sampling rate of vibration and acoustic signals from 8kHz to 16kHz, while keeping the sampling rate of current and voltage signals unchanged. This allows for the focused capture of potential fault-related information without significantly increasing the overall data volume. Compared to a fixed sampling rate scheme, this method effectively reduces redundant data by approximately 30% and avoids missing early fault characteristics.

[0173] (2) Cross-modal timing calibration results

[0174] Due to differences in sampling frequency and response characteristics among different sensors, there is a significant time misalignment in the raw data. For example... Figure 1 As shown in the uncalibrated state, the peak value of the vibration signal is offset by about 12ms relative to the current change.

[0175] The system employs a Physically Constrained Dynamic Time Warping (PC-DTW) algorithm for switchgear equipment, introducing physical mapping constraints for "vibration-contact resistance" and "acoustic-closing / opening time," with a physical threshold set to 0.06. After PC-DTW processing, the key features of each modal signal achieve physical consistency alignment on the time axis, eliminating the "time alignment but physical inconsistency" problem that may occur with traditional DTW. The alignment result is as follows: Figure 1 The calibrated curve is shown in the figure.

[0176] (3) Results of two-stage denoising and signal analysis

[0177] Under the influence of strong electromagnetic interference and environmental noise, the original vibration signal contains significant random pulse interference. The system first models the noise distribution using an attention-guided diffusion model (A-DDM), outputting the noise manifold coordinates z(t) and the local noise energy map E(t). The results show that before the switchgear malfunction, the local noise energy is significantly concentrated in the interval between 3.2 s and 3.8 s.

[0178] Based on the aforementioned noise characteristics, the Trend-Aware Kalman Filter (T-AKF) dynamically adjusts the covariance matrix, enabling the filtering process to suppress random interference while preserving the slowly evolving trend characteristics. Compared to fixed-parameter Kalman filtering, the fault-related weak vibration components in the denoised signal are more clearly visible.

[0179] Subsequently, the system performed joint E-VMD and AN-CEEMDAN decomposition on the denoised signal, ultimately obtaining four effective modal components. Among them, the second mode showed a significant energy enhancement in the 200–350 Hz frequency band, which is highly consistent with the characteristic frequency band of switchgear contact loosening, providing a reliable basis for subsequent feature extraction.

[0180] (4) Feature fusion and time-consistent fracture field construction results

[0181] The system extracts time-frequency entropy, kurtosis, and envelope spectrum features from the aforementioned modal components and dynamically weights them using a multi-head attention mechanism. The feature weight distribution results show that the weights of vibration and acoustic modes significantly increase during abnormal phases, while the weights of temperature and voltage features remain stable.

[0182] Based on this, the system constructs a cross-modal time-consistent fracture field. For example... Figure 2 As shown, during normal equipment operation, the uniform fracture field remains continuous and smooth in the time-hysteresis space; however, about 15 minutes before contact loosening occurs, local high-energy regions appear in the fracture field.

[0183] By calculating the consistency gradient tensor and fracture energy density, the system extracts fracture morphology fingerprint vectors. The mean energy and kurtosis of the vectors are significantly higher than the threshold of the normal operating condition feature library, indicating the presence of potential sensor anomalies.

[0184] (5) Comparison of fault early warning results and technical effects

[0185] The system's final output of the warning results shows that the model had accurately triggered a medium-level warning about 12 minutes before the switch cabinet contacts became loose and caused abnormal sensor signals. When the fault further developed and the contact resistance of the contacts increased to 1.2Ω (exceeding the normal threshold of 0.5Ω), the warning level was dynamically upgraded to a high level. At the same time, the abnormal sensors were accurately located as the switch cabinet vibration sensor and acoustic sensor, with a positioning accuracy of 100%.

[0186] To verify the technical superiority of the method of the present invention, existing early warning methods based on a single vibration threshold and diagnostic methods based on traditional statistical characteristics were selected as comparison objects, and performance tests were conducted in the same experimental environment (110kV intelligent power distribution room, covering 3 typical working conditions and 5 common sensor faults). Test results show that the average early warning time of the method of this invention is 12.3 min, which is much higher than the 4.1 min of the single vibration threshold method and the 5.7 min of the traditional statistical feature method; the fault diagnosis accuracy reaches 98.6%, which is 16.3 percentage points and 11.9 percentage points higher than the two comparative methods, respectively; the false alarm rate is only 1.2%, which is 7.7 percentage points and 5.3 percentage points lower than the comparative methods, respectively; the multi-condition adaptation pass rate exceeds 99.2%, which is more than 15 percentage points higher than the traditional method; the anti-interference diagnosis accuracy (with 15dB Gaussian noise added) reaches 95.8%, which is 23.2 percentage points and 16.7 percentage points higher than the two comparative methods, respectively; the fault location accuracy is 100%, which is 10-15 percentage points higher than the comparative methods; the model inference time is 0.32s, and the number of parameters is controlled within a reasonable range.

[0187] In summary, the data in this embodiment fully demonstrates that the multimodal sensor fault early warning method proposed in this invention is significantly superior to existing technologies in terms of early warning timeliness, diagnostic accuracy, adaptability to operating conditions, and anti-interference capability. It can effectively solve problems such as delayed fault detection, high false alarm rate, and insufficient collaborative utilization of multimodal data, and has extremely high engineering application value and promotion prospects.

[0188] This invention provides a method and system for early warning of sensor faults in intelligent power distribution rooms based on multimodal data fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for early warning of sensor faults in intelligent power distribution rooms based on multimodal data fusion, characterized in that, Includes the following steps: Step 1: Construct a chaotic index C(t) based on chaotic dynamics, determine the chaotic feature weights by combining the working condition classification model, and generate a dynamic sampling rate using a nonlinear adaptive sampling function to achieve working condition-adaptive acquisition of multimodal data. Step 2: For each core device, a dedicated physical mapping model corresponding to the device's operating mechanism is constructed. The Physical Constraint Dynamic Time Warping (PC-DTW) method is used to incorporate physical deviations into the DTW distance with weights, set a threshold, and eliminate alignment schemes that contradict physical principles. The core devices include transformers, switchgear, and instrument transformers. Step 3: Based on the attention-guided diffusion model (A-DDM), the noise manifold features in the sensing signal are learned and noise feature parameters are output. The covariance matrix of the trend-aware Kalman filter (T-AKF) is dynamically generated according to the noise feature parameters. The signal denoising is completed through state prediction and state correction processes, thereby achieving adaptive coordination of noise analysis and filtering parameters. Step 4: First, the denoised signal is decomposed using the entropy-constrained variational mode decomposition method E-VMD. Then, invalid modes are removed using the spectral kurtosis threshold screening mechanism. Finally, the noise intensity is adaptively adjusted by the adaptive noise intensity CEEMDAN according to the signal complexity to complete the secondary decomposition, thereby achieving accurate analysis of non-stationary signals. Step 5: Extract the time-domain, frequency-domain, and time-frequency-domain features of each signal component, and use a multi-head attention mechanism to weightedly filter the features to obtain key features; construct a node trust model for federated feature aggregation, and introduce device type weights to modify the node trust model, forming a trust-weighted federated aggregation framework TWFA; combine AES-128 encryption and key isolation mechanisms to achieve secure sharing and weighted aggregation of cross-node features; Step 6: Construct a multimodal sensor fault diagnosis model TCBFM based on time-consistent fracture field: First, model the cross-modal time-consistent fracture field, calculate the consistency gradient tensor to characterize the dynamic changes and stability of the fracture field, then quantitatively describe the degree of structural damage through the consistency fracture energy density, and finally extract the fracture morphology fingerprint vector as the fault feature representation. Step 7: Use the multi-objective adaptive Bayesian optimization algorithm MO-ABO to optimize the hyperparameters of the TCBFM multimodal sensor fault diagnosis model based on time-consistent fracture field.

2. The method according to claim 1, characterized in that, In step 1, the multi-source sensors are deployed as follows: vibration sensors are installed on the outside of the transformer terminals; acoustic sensors are arranged in the center of the switchgear door panel; temperature sensors are arranged above the busbars; and current and voltage sensors are installed at the standard CT / VT measuring points. Current, voltage, temperature, vibration, and acoustic signals are collected to construct a timestamped multimodal time-series dataset, and each data sample is labeled with a corresponding device status tag. Simultaneously, an adaptive sampling rate adjustment mechanism and a cross-modal data time-series calibration algorithm are added. First, the phase space of each sensor signal is reconstructed, and three chaotic features are calculated within a sliding window: the maximum Lyapunov exponent, the correlation dimension, and the local trajectory divergence rate. A unified chaos index C(t) is constructed to describe the chaotic intensity of the signal within the window. ， in, The normalized value of the maximum Lyapunov exponent; This is the normalized value of the correlation dimension; This is the normalized value of the local trajectory divergence rate; , and As weight; A chaotic feature sensitivity modeling mechanism driven by operating condition classification is introduced. By modeling the response characteristics of chaotic features under different operating conditions, three types of feature sensitivity vectors F are obtained for weight calculation: F=[F1,F2,F3], Where F1 is the sensitivity component of the maximum Lyapunov exponent; F2 is the sensitivity component of the correlation dimension; and F3 is the sensitivity component of the local trajectory divergence rate. The weights are obtained by normalizing the feature sensitivity: ， Based on the interval where the chaotic index C(t) is located, the sampling rate f is determined using a piecewise nonlinear function. s (t): ， in Minimum sampling rate; It is a time constant; is the first threshold of the chaos index; a, b, and c are polynomial coefficients; The second threshold for the chaos index; The maximum sampling rate is represented by e; e represents the natural constant. Update the threshold based on the accuracy of the diagnostic model: ， Where, θ i (t) represents the threshold of the i-th chaotic interval at time t; η is the learning rate; Acc is the fault diagnosis accuracy obtained based on the current sampling strategy; This indicates the sensitivity of accuracy to a threshold.

3. The method according to claim 2, characterized in that, In step 2, the equipment-specific physical deviations are incorporated into the DTW local distance according to their weights: ， in, Let be the physically constrained DTW local distance between the i-th time and the j-th time. The maximum deviation value of the k-th type of physical mapping; It is the physically constrained dynamic time warping (DTW) local distance between the signal to be matched at time i and the reference signal at time j. It is the deviation value of the k-th type of physical mapping between the signal to be matched at time i and the reference signal at time j; It is the weight of the physical mapping term. K represents the number of physical mapping items corresponding to the device. Set physical threshold ,if If the local distance between the sampling point pairs is set to infinity, then the local distance between them is set to infinity. Determining the physical threshold using a multi-objective optimization strategy With weight The optimal parameters are: to minimize the error in reconstructing the calibrated data. recon Ensure time alignment accuracy and maximize fault feature recognition accuracy. feat Supporting subsequent diagnosis has two objectives: using the NSGA-II algorithm to search for Pareto optimal solutions within the parameter range to balance solution diversity and engineering applicability.

4. The method according to claim 3, characterized in that, Step 3 includes: Gaussian noise is gradually added to the original signal x0 to generate a noisy sequence x at time t. t The mathematical expression for simulating the evolution of noise from weak to strong is: ， in, Let be the diffusion coefficient, which increases linearly from 0.001 to 0.02 at t = 1, 2, ..., T, where T is the total number of time points in the forward diffusion process. Standard Gaussian noise; Backdiffusion and noise manifold extraction: Training a noise prediction network , through reverse diffusion from x T During the recovery of x0, the focus is on learning the high-dimensional distribution of noise, ultimately outputting two types of core features: Noise manifold coordinates z(t): describe the geometric structure of noise in high-dimensional space, with the dimension consistent with the signal sampling length; Local noise energy map E(t): combined with signal local attention weights The statistical distribution of local noise intensity reflects the local aggregation characteristics of noise; Based on the outputs z(t) and E(t) of the attention-guided diffusion model (A-DDM), a dynamic covariance generation logic is constructed to ensure that the filtering parameters are matched with the current noise and signal trends in real time. Specifically, this includes: R(t) is the noise intensity of the measurement process at time t, which is determined by the global interference intensity of the noise manifold coordinates z(t) and the signal trend factor. Driven by both forces, the formula is: ， in, Here, m is the intermediate parameter, m is the signal sampling length, and exp is the natural exponential function. Let i be the i-th component of the noise manifold coordinates at time t; Q(t) describes the noise variation during signal transmission, which is determined by the concentration intensity of the local noise energy map E(t) and the signal prediction trend. The driver, the formula is: ， in, and These are intermediate parameters. The trend characteristics corresponding to the predicted signal value at time t; Based on dynamically generated Q(t) and R(t), T-AKF completes filtering according to the following steps: State prediction: using the results of the previous time step With the state transition matrix To obtain the predicted state ; Gain Calculation and State Correction: The Kalman gain is calculated by combining dynamic Q(t) and R(t), and then the sensor's measured value y is used. k Correct the prediction state and output the denoised clean signal. .

5. The method according to claim 4, characterized in that, In step 4, an improved joint mode decomposition strategy is applied to the denoised signal from step 3, including: Step 4-1, Entropy-Constrained Variational Mode Decomposition (E-VMD): Add modal entropy constraint terms to the variational functional of VMD. Construct a new objective function : ， in, For the k-th mode Shannon entropy, K is the entropy constraint coefficient; K is the number of modes. Here, j is the time partial derivative operator, and j is the imaginary unit. Let be the center frequency of the k-th mode. The penalty factor is x(t), and the original input time-domain signal is x(t). Step 4-2, Mode Selection Mechanism Based on Spectral Kujicity Threshold: A spectral kujicity screening mechanism is introduced after E-VMD. For each primary mode u k (t), calculate the envelope kurtosis. : ， if Less than the threshold If it is determined to be a weakly correlated or noisy mode, the primary mode u will be... k (t) are removed and do not enter the adaptive noise intensity CEEMDAN reconstruction stage; It is the mathematical expectation operator; u k It is the primary mode u k (t) The envelope spectrum signal after envelope spectrum transformation; Step 4-3, Adaptive Noise Intensity CEEMDAN: Define signal complexity The approximate entropy of the signal and the mean of the sample entropy are: ， in, The approximate entropy of the signal. The entropy of the signal sample; according to Adjust the intensity of the added Gaussian white noise : ， in, This represents the standard deviation of the original signal.

6. The method according to claim 5, characterized in that, Step 5 includes: constructing a dynamic feature selection and feature interaction term mining module based on an attention mechanism. The dynamic feature selection and feature interaction term mining module first calculates the time-frequency entropy, kurtosis and envelope spectrum features of each component obtained in step 4, then generates a feature weight vector through a multi-head attention mechanism, and filters features with weight values ​​greater than the threshold δ to complete the dynamic selection. Finally, it combines the trust-weighted federated aggregation (TWFA) framework to realize feature sharing across distribution room nodes. Construct a set of node device type weights : ， in, It is the transformer node weight; It is the node weight of the switchgear; These are the capacitor node weights; These are the weights of other auxiliary nodes; By coupling the node device type weight with the trust calculation process, a revised trust model is proposed: ， in It is the corrected trust level of the i-th node; The device type weight for the i-th node; For node characteristic quality indicators; The degree of deviation in node distribution; This is a relevance penalty item; A lightweight secure transmission module is introduced into the TWFA architecture to achieve end-to-end encryption of cross-node feature vectors, including: before uploading the feature vector, the node encrypts the original vector f. i Perturbation and encryption processes are performed to obtain the encrypted features. : ， in, AES-128 encryption function; For small disturbance noise; A key isolation mechanism is adopted: the key is stored in different secure areas.

7. The method according to claim 6, characterized in that, Step 6 involves establishing a multimodal sensor fault diagnosis model (TCBFM) based on a time-consistent fracture field, which includes the following steps: Step 6-1: Construct a time-consistent fracture field; Let the sampled value of the m-th sensing mode at time t be... To address the physical coupling relationships between different modes, a cross-modal time-consistent fracture field is constructed: ， in It is a cross-modal time-consistent fracture field; It is the sampled value of the m-th sensing mode at time t; The nth sensing mode lags behind at time t. The sampled value at time; This refers to the time lag. (⋅) represents the asymmetric physical consistency kernel function constructed based on the device operation mechanism; For modal coupling weights; Step 6-2: Calculate the consistency gradient tensor; Define a consistent gradient tensor on the constructed consistent fracture field. : ， in, It is the component of the gradient in the time direction; It is the component of the gradient in the hysteresis direction; Step 6-3: Calculate the uniform fracture energy density; Define the uniform fracture energy density function : ， in, The vector norm is represented; θ is the fracture detection threshold. The function is an indicator function; d is the differential symbol. Step 6-4, Fault identification driven by fracture morphology; Based on the temporal distribution characteristics of the uniform fracture energy density, a fracture morphology fingerprint vector is constructed. : ， in, These represent the mean, skewness, and kurtosis of the fracture energy distribution, respectively. The currently constructed fracture morphology fingerprint vector is matched and calculated with the preset sensor fault feature library. By comparing the similarity of the vector features, the corresponding sensor fault type and warning level are output, and finally the diagnosis and warning of the sensor operating status are realized. Step 7 includes: MO-ABO, based on Bayes' theorem, approximates the black-box function by constructing a joint probability model of multiple objective functions: Let there be m objective functions, denoted as... ,in To optimize variables, Let m be the m-th objective function. Bayes' theorem is used to update the posterior distribution of the objective function: ， in, It is the joint posterior probability of the objective function given the observed dataset D; It is the likelihood function of the objective function, representing the probability of the observed data given the objective function value; It is the joint prior probability of the objective function; It is the marginal probability; The probabilistic transitive model employs a Gaussian process, with each objective function... Modeled as a Gaussian process GP prior: ， in It is the mean function; It is the covariance function; the prior distribution is initialized based on historical data, and the posterior distribution is obtained through Bayesian update; The desired hypervolume improved EHVI is used as the core acquisition function: ， in The desired improvement is in the acquisition function value through super-volume. It is a volume improvement measure; Establish optimization variables: The key structural hyperparameters of the TCBFM model are used as optimization variables. Let the hyperparameter vector x be: ， Where x i This represents the i-th structural hyperparameter to be optimized; Structural hyperparameters include: the range of time lag values ​​τ. range Used to define the effective lag interval of cross-modal time consistency relationships; physical consistency kernel function width σ k The sensitivity of the metric used to adjust the consistency of responses between modalities; the feature dimension D of the fracture morphology fingerprint vector. f This is used to set the dimension size for representing fault features; the hidden layer dimension H of the fault discrimination module d This is used to adjust the fitting ability and generalization performance of fault classification; Establish the objective function: The objective function is used to evaluate the performance of the hyperparameter vector x. In the fault diagnosis model, considering two or more competing objectives, the following four objective functions are defined: Indicates the accuracy of fault diagnosis; Indicates the model inference time; Indicates the number of model parameters; Indicates the accuracy of anti-interference diagnosis; objective function vector for: ， The multi-objective optimization problem is represented as: ， in X is the structure hyperparameter vector to be optimized; X is the structure hyperparameter vector. The feasible domain; MO-ABO optimizes hyperparameters through an iterative process, the specific steps of which are as follows: Step a1, Initialization: Generate an initial observation set based on prior information. ,in This represents the initial number of sampling points; Step a2, Probabilistic Model Update: For each objective function f k (x), use Gaussian process regression to fit the observed data, update the posterior distribution, and the posterior mean. and variance for: ， ， in It is the kernel matrix corresponding to the k-th target; This is the matrix of sampled points. For the corresponding target value vector, Let I be the noise variance; I is the identity matrix. The adaptive mechanism dynamically adjusts the kernel function hyperparameters by maximizing the marginal likelihood. : ， in These are the hyperparameters of the kernel function; This indicates that the sampled point matrix X and the kernel function hyperparameters... Under the condition, the target value vector The corresponding marginal likelihood probability density; Step a3, acquisition function optimization: Use the EHVI acquisition function to select the next sampling point. : ， in The desired hypervolume improvement acquisition function value; EHVI calculation is based on the current Pareto front P and the hypervolume reference point r; Step a4, Evaluation and Update: At point Evaluation of objective function value Update the observation set: ， Where n represents the current iteration round number, D n It is the observation set after the nth iteration; Step a5, Convergence check: Repeat steps a2 to a4 until the convergence condition is met. After convergence, output the Pareto front. .

8. A smart power distribution room sensor fault early warning system based on multimodal data fusion, implemented using the method described in any one of claims 1 to 7, characterized in that, include: Multi-source sensor acquisition module: A multi-source sensor array deployed at key nodes in the power distribution room, integrating adaptive sampling rate adjustment and cross-modal timing calibration functions; Dual-stage denoising module: realizes the cascaded processing of attention-guided diffusion denoising (A-DDM) and trend-aware Kalman filtering (T-AKF); Signal parsing module: Executes the VMD-CEEMDAN joint decomposition algorithm based on entropy constraints; Federated Feature Selection Module: Integrates attention mechanism dynamic feature selection and interaction item mining unit, supports Trust-Weighted Federated Aggregation (TWFA); Hybrid diagnostic module: Based on the TCBFM model, time-consistent fracture features are extracted. Under MO-ABO optimization, the fracture morphology fingerprint vector is input into the decision module to generate the final diagnostic output. Early warning output module: generates fault category labels, anomaly scores and root cause location reports, and provides fault propagation path analysis results to enable early warning and intelligent operation and maintenance decisions.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 7.

Images