Distribution transformer insulation state multi-dimensional perception and fault intelligent early warning method

Abstract

The invention relates to a distribution transformer insulation state multi-dimensional perception and fault intelligent early warning method, and belongs to the technical field of power equipment state monitoring and fault diagnosis. According to the technical scheme, the method comprises the following steps: constructing a temperature prediction model based on Bi-LSTM and a time sequence attention mechanism, decoupling and stripping temperature rise pressure fluctuation under a normal working condition in combination with a thermodynamic state equation, and extracting a dynamic residual vector representing the essence of a fault; constructing an object-oriented fault diagnosis knowledge graph, and mapping the residual features into a fuzzy evidence body; the improved D-S evidence theory is used for carrying out adaptive weighted fusion on multi-source heterogeneous data, knowledge base rules are corrected on line based on an operation and maintenance feedback mechanism, and closed-loop evolution and high-precision early warning of diagnostic logic are achieved. The method has the beneficial effects that false alarm and missing alarm caused by strong coupling of oil pressure and oil temperature are avoided, fault types are accurately distinguished, operation and inspection departments are helped to quantify equipment health risks, and all-weather automatic monitoring, fault classification and closed-loop evolution early warning of the distribution transformer are achieved.

Description

Technical Field

[0001] This invention relates to a method for multi-dimensional perception of insulation status and intelligent early warning of faults in distribution transformers, belonging to the field of power equipment condition monitoring and fault diagnosis technology. Background Technology

[0002] As a core hub connecting the distribution network and power users, distribution transformers play a crucial role in power distribution and conversion. Their operational reliability directly affects power supply quality and user safety. Due to their large number, wide distribution, and complex operating environment, distribution transformers are susceptible to faults such as insulation aging, winding deformation, or oil circuit blockage caused by long-term load fluctuations, seasonal temperature differences, and electromagnetic stress. Unplanned outages can lead to severe economic losses and social impacts. Condition monitoring of distribution transformers mainly falls into three categories: 1) Offline preventative testing: Regularly performing insulation resistance, DC resistance, and offline oil chromatography tests according to industry standards (such as DL / T 596). This method lacks real-time capability, struggles to detect sudden faults, and its maintenance costs increase exponentially with the number of distribution network nodes. 2) High-precision online monitoring (DGA): Using photoacoustic spectroscopy or gas chromatography to monitor dissolved gas components in the oil online. While highly accurate, this method is costly and complex to maintain, typically only applied to large main transformers in the main grid, making large-scale deployment at the distribution network level difficult. 3) Multi-parameter simplified condition monitoring: With the development of the ubiquitous power Internet of Things, the use of low-cost sensors to collect key physical quantities such as top oil temperature, tank static pressure, trace water and hydrogen has become a mainstream trend. This technical approach aims to achieve real-time perception of the operating status of distribution transformers through non-invasive or low-invasive methods.

[0003] While multi-parameter simplified monitoring offers cost advantages, it faces two major challenges in engineering applications: multi-physics coupling interference and sensor reliability. 1) Thermal-fluid-solid multi-field coupling characteristics: The distribution transformer tank can be considered a quasi-constant-volume thermodynamic system. According to the principle of thermal expansion and contraction and Charles's law, Joule heating caused by the load current leads to an increase in the temperature and volume of the insulating oil, resulting in a significant increase in the gas chamber pressure. This pressure change caused by normal operating conditions (load fluctuations, ambient temperature rise) often masks the small pressure increments generated by early minor faults (such as partial discharge), causing diagnostic models to fail. 2) Data source uncertainty: Low-cost MEMS sensors deployed outdoors are susceptible to electromagnetic interference and environmental aging, leading to data zero-point drift or increased noise. Existing technologies lack effective data fusion and conflict resolution mechanisms, failing to maintain the robustness of diagnostic results when single sensor data is abnormal.

[0004] Existing technologies typically employ an "edge-cloud" architecture, installing pressure sensors, temperature sensors, and single hydrogen sensors on the transformer itself, with the edge gateway uploading data to the main station system. Its core diagnostic logic primarily relies on the following two methods: 1) Static threshold method: The system presets fixed over-limit alarm values ​​(e.g., setting a pressure threshold P). th =50kPa, oil temperature threshold T th =85℃). When the real-time monitoring value P(t) > P th Or T(t)>T th 1) When the temperature is high, an alarm is triggered directly. 2) Linear regression method: A simple temperature-pressure linear regression model P=aT+b is established. The theoretical pressure is calculated by substituting the real-time oil temperature into the equation. If the deviation between the measured pressure and the calculated value exceeds the fixed tolerance range, it is judged as abnormal. This type of scheme mainly focuses on data collection and simple display, lacks in-depth modeling of the complex thermodynamic processes inside the transformer, and does not consider the reliability conflict problem between multiple data sources. The main disadvantages are as follows:

[0005] (1) The inability to decouple the normal temperature rise effect leads to a high false alarm rate under heavy overload conditions. When the transformer operates under heavy overload or high temperature conditions, the load current I load Increased copper and iron losses lead to a higher top oil temperature T. top A significant increase occurs. According to ΔV=αVΔT (the formula for thermal expansion of liquids) and PV=nRT (the equation of state for gases), the internal pressure P of the oil tank will rise substantially. Existing static threshold methods cannot distinguish between "thermally induced pressure rise" and "fault-induced gas pressure rise." At this point, although the equipment insulation is normal, the pressure P(t) exceeds the static threshold P. th The system will mistakenly report it as an internal fault, resulting in a waste of operational and maintenance resources.

[0006] (2) The high rate of missed detection of latent faults is caused by the masking effect of environmental baseline values. Under low temperature or light load conditions, the base oil temperature inside the transformer is low, resulting in a low base oil pressure (e.g., 10 kPa). If an early partial discharge fault occurs at this time, a small amount of gas will be generated, causing the pressure to increase by ΔP. fault =5kPa, total pressure is only 15kPa. Conclusion of defect: Since the total pressure is far below the alarm threshold (e.g., 50kPa), the existing technology cannot detect such abnormal increments superimposed within the safe range, causing potential faults to remain dormant for a long time until they evolve into destructive accidents.

[0007] (3) Lack of multi-source information fusion mechanism, resulting in poor resistance to sensor drift and interference. In field operation, pressure sensors may experience falsely high readings due to zero-point drift, while hydrogen sensors may experience falsely low readings due to poisoning failure. Existing technologies typically employ simple logical OR gates or directly accept single data. There is a lack of conflict resolution mechanisms based on evidence theory. When sensor data conflict (e.g., extremely high pressure but normal temperature and hydrogen levels), existing systems cannot identify sensor malfunctions, leading to continuous false alarms or loss of system reliability.

[0008] (4) The model parameters are fixed and cannot adapt to the characteristic evolution of the equipment throughout its entire life cycle. After years of operation, the heat sink accumulates dust, leading to increased thermal resistance, and the aging of the seals causes micro-leakage (a decrease in base pressure). The physical characteristic parameters of the equipment (such as thermal time constant and airtightness) change over time. Existing technologies are based on fixed models / thresholds set at the factory or in the initial stage of operation, and cannot be corrected online according to actual operating data. As the equipment ages, the prediction residuals of the fixed model will gradually increase, resulting in a monotonically decreasing trend in diagnostic accuracy over time.

[0009] In summary, existing simple monitoring technologies face numerous technical bottlenecks in practical applications: First, significant multi-physics coupling effects exist within transformers. Normal high-load operation leads to increased oil temperature, which in turn causes an increase in tank pressure. Existing static threshold methods struggle to effectively distinguish between this "normal thermal-induced pressurization" and "fault-induced gas pressurization," easily resulting in false alarms. Second, low-cost sensors are prone to zero-point drift or data anomalies under long-term harsh environments. A single data source often leads to conflicting diagnostic results, lacking an effective information fusion mechanism to distinguish between true and false alarms. Finally, existing diagnostic models are mostly based on fixed expert rules and cannot self-evolve according to equipment aging and historical operating data. As operating time increases, diagnostic accuracy gradually decreases. Therefore, existing maintenance methods relying on offline testing or single threshold judgments can no longer meet the demands of the rapid development of the distribution IoT and precise equipment status perception. A low-cost, highly reliable technology that can decouple multi-physics interference, fuse multi-source sensor information, and possess self-evolution capabilities is urgently needed to solve the insulation diagnosis and fault prevention problems of distribution transformers. Summary of the Invention

[0010] The purpose of this invention is to provide a multi-dimensional perception method for insulation status and intelligent fault early warning of distribution transformers. By combining thermodynamic physical mechanisms, time-series deep learning, and multi-source evidence fusion theory, it avoids false alarms and missed alarms caused by strong coupling between oil pressure and oil temperature, accurately distinguishes fault types, and obtains the precise insulation status and fault type confidence of distribution transformers by decoupling and feature extraction of multi-physics field time-series data, combined with expert knowledge base and dynamic evidence fusion algorithm. This helps operation and maintenance departments quantify equipment health risks, realize all-weather automatic monitoring, fault classification, and closed-loop evolutionary early warning of distribution transformers, and solve the above-mentioned technical problems existing in the prior art.

[0011] The technical solution of this invention is:

[0012] A multi-dimensional perception method for insulation status and intelligent fault early warning of distribution transformers is proposed. First, a temperature prediction model based on Bi-LSTM and temporal attention mechanism is constructed. This model, combined with the thermodynamic equation of state (PDMA), decouples and separates temperature rise and pressure fluctuations under normal operating conditions, extracting dynamic residual vectors characterizing the fault's essence. Second, an object-oriented fault diagnosis knowledge graph is constructed, mapping residual features to fuzzy evidence bodies. Finally, an improved DS evidence theory (MSE-AFM) is used to adaptively weight and fuse multi-source heterogeneous data, and the knowledge base rules are corrected online based on an operation and maintenance feedback mechanism, achieving closed-loop evolution of diagnostic logic and high-precision early warning. The method includes the following steps:

[0013] S1. Construct a multi-dimensional condition monitoring dataset for distribution transformers, acquire time-series monitoring data on hydrogen concentration in oil, static pressure in oil tank, top oil temperature, load current, and ambient temperature, and perform data cleaning and synchronization alignment.

[0014] S2. Construct a physical field decoupling model based on bidirectional long short-term memory network (Bi-LSTM) coupled with gas state equation. Use historical normal operation data to train the model to predict theoretical oil temperature and theoretical oil pressure. Calculate the dynamic residual vector between measured and predicted values. Remove pressure fluctuations caused by normal load temperature rise and extract pure residual sequences that characterize fault features.

[0015] S3. Establish an object-oriented insulation diagnosis expert knowledge graph, transform transformer entity attributes and fault mechanisms into fuzzy production rules, and map the dynamic residual sequence obtained from the decoupled model into basic probability assignments (BPA) for different fault types through fuzzy membership functions.

[0016] S4. An improved DS evidence theory fusion algorithm based on evidence credibility correction is proposed. The algorithm calculates the conflict coefficient between evidence sources from various sensors, dynamically adjusts the weight of each evidence body, performs weighted fusion reasoning on multiple basic probability assignments generated, solves the problem of misjudgment and conflict when the fault characteristics of a single sensor are not obvious, and outputs the final fault type and confidence level.

[0017] S5. Generate multi-level early warning signals based on the fusion diagnostic results, and introduce a feedback self-evolution mechanism. Use actual operation and inspection results as verification signals to reverse the diagnostic rule thresholds in the knowledge graph and the network parameters of the prediction model, thereby realizing the self-iteration and accuracy improvement of the diagnostic logic.

[0018] In step S1, time-series operation data of distribution transformers of different regions and models are acquired through the online monitoring platform of the distribution network. This includes time-series monitoring data of hydrogen concentration in oil, static pressure in the oil tank, top oil temperature, load current, and ambient temperature. A multi-physics field insulation diagnostic database (MPID) for distribution transformers is constructed by combining historical maintenance records. This database covers typical states of normal aging, partial discharge, overheating, and abnormal oil circuits. It considers the influence of seasonal temperature differences, load fluctuations, and electromagnetic interference in the actual operating environment. Through data cleaning and time-series alignment, the problems of inconsistent sampling frequencies of different sensors and uneven distribution of fault samples are solved, ensuring the integrity and generalizability of the data foundation.

[0019] In step S2, a distribution transformer state decoupling model (Phy-LSTM) based on physical-data dual-drive is constructed. Its main body includes a time series prediction architecture based on a bidirectional long short-term memory network (Bi-LSTM) and a pressure correction module based on thermodynamic mechanisms. The specific implementation of the distribution transformer state decoupling model is as follows:

[0020] S201. A Bi-LSTM prediction network is constructed to perform high-precision inversion of the top oil temperature of the transformer. A temporal attention mechanism based on a multilayer perceptron is introduced. By calculating the energy score of the hidden layer state of the Bi-LSTM, attention weights are adaptively assigned to each moment in the historical time window to generate a context vector containing global thermal history information, which accurately captures the hysteresis temperature rise characteristics caused by load mutation.

[0021] S202 proposes a pressure dynamic correction algorithm (PDMA) based on the ideal gas law. It uses the predicted theoretical oil temperature to derive the theoretical static pressure of the oil tank. The algorithm is based on the physical assumption that the transformer oil tank is a quasi-constant volume container. By quantifying the expansion effect of the insulating oil volume with temperature change, it quantifies the normal pressure fluctuation caused by temperature and separates it from the measured pressure.

[0022] In step S3, an object-oriented insulation diagnosis expert knowledge graph is constructed, transforming transformer entity attributes and fault mechanisms into fuzzy production rules. This step is specifically implemented as follows:

[0023] S301, establish the mapping relationship from the dynamic residual vector R(t) to the specific fault type, and use the fuzzy membership function to discretize the continuous residual values ​​into semantic variables. For each feature component, construct the membership function μ(x) of the combination of trapezoidal and Cauchy distributions. For example, the membership function of pressure residual is defined as follows:

[0024]

[0025] The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows:

[0026] In the formula, r P λ1 and λ2 are the starting threshold and saturation threshold of the pressure residual, respectively, which are determined by historical statistical data in the knowledge base; c and k are the shape parameters of the control function; through this transformation, the numerical characteristics at the physical level are converted into fuzzy trust values ​​at the logical level.

[0027] S302 proposes a Basic Probability Assignment (BPA) method based on fuzzy correlation matrices. This method constructs a feature-fault correlation matrix describing the degree of support each feature quantity provides for different fault types, and calculates the basic probability assignment function m(·) according to the following formula:

[0028]

[0029] The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows:

[0030] In the formula, m j (F k ) represents the relationship between the j-th sensor characteristic (such as oil temperature residual) and the fault type F. k Basic level of trust; w jk For predefined association weights in the knowledge graph; w jn Ω is the association weight of the j-th feature with the n-th fault in the identification framework, which is obtained by accumulating the weighted support of this feature for all potential fault types; Ω is the uncertainty coefficient, used to characterize the cognitive uncertainty when the feature value is in the fuzzy edge range.

[0031] S303, design a hierarchical rule inference engine to handle combinational fault logic. The inference engine includes an atomic rule base for handling single-parameter out-of-bounds behavior and a composite rule base for handling multi-parameter temporal cooperative behavior. The rule storage format adopts the following optimized production rules:

[0032]

[0033] The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows:

[0034] The above formula describes the i-th production rule in the inference engine.i Specifically, it is used for logical discrimination of high-energy arc faults. In the formula, the logical operator ∧ represents an "AND" relationship, meaning that all conditions of the antecedent must be met simultaneously to trigger the inference; r P is High and v H2 "is High" indicates that both the pressure residual characteristic and the hydrogen production rate characteristic after fuzzification are in a significantly high semantic state; the CF=0.95 on the arrow represents the certainty factor of this rule as 0.95, meaning that under the concurrent occurrence of the aforementioned pressure mutation and hydrogen production surge characteristics, the system determines the current discharge fault type (F) with a 95% confidence level. D The reasoning logic accurately corresponds to the physical mechanism by which high-energy discharge causes the insulating oil to decompose violently and generate shock waves.

[0035] In step S4, a multi-source evidence adaptive fusion model (MSE-AFM) is constructed, which includes a conflict coefficient calculation module and a weighted inference engine; the specific implementation is as follows:

[0036] S401 proposes an evidence reliability assessment method based on Jousselme distance (RA-ED) to calculate the distance d between the i-th evidence source and the j-th evidence source. ij and the credibility weight w of the i-th piece of evidence i

[0037]

[0038]

[0039] The meanings of the letters in the formula are as follows:

[0040] In the formula, m i and m j , respectively, are evidence vectors from different sensors; D is the Jaccard similarity matrix, used to describe the intersection relationship of different fault sets in the identification framework; n represents the total number of sensor evidence sources participating in the fusion; (1-d ij The value represents the similarity (i.e., the degree of mutual support) between the i-th and j-th evidence sources, and is determined by the Jousselme distance d between them. ij The smaller the distance, the higher the similarity; 1-d kj This represents the similarity between the k-th and j-th pieces of evidence when traversing all evidence sources to calculate the total support of the system;

[0041] In S402, a weighted probabilistic orthogonal fusion strategy (WPOF) is proposed, utilizing the weights w obtained in S401. iGenerate the corrected average body of evidence The Dempster orthogonal sum rule is applied to calculate the final failure confidence level m. final (A):

[0042]

[0043]

[0044] The meanings of the letters in the formula are as follows:

[0045] In the formula, A, B, and C are subsets of the fault hypothesis set (such as discharge and overheating); K is the normalization factor, reflecting the degree of conflict after fusion. The preprocessing is equivalent to smoothing out the conflicts in the evidence space, ensuring that K<1, thereby guaranteeing the convergence and rationality of the fusion result.

[0046] In step S5, the fault warning release and knowledge base self-evolution are specifically implemented as follows:

[0047] S501 maps the final fault confidence vector output by S4 to an operation and maintenance decision signal, based on the confidence value m. final The distribution sets a tiered early warning threshold λ. When the fault confidence level is greater than λ... high A level one red alert is triggered at this time, due to λ low With λ high A level 2 yellow alert is triggered when the temperature is below λ. low Maintain a green, normal state at all times;

[0048] S502 proposes an online knowledge base correction strategy based on feedback error. It utilizes actual feedback results from on-site operations and maintenance as truth labels, and updates the weights of the expert system's inference rules in reverse according to the following formula:

[0049]

[0050] The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows:

[0051] In the formula, w is the predefined feature-fault association weight in the knowledge graph. jk η is the learning rate, which is usually taken as a minimum to ensure the stability of the update; Label is the true state confirmed by human verification (1 for confirmed, 0 for excluded); Pred is the prediction confidence of the model.

[0052] Key points of the invention concept:

[0053] (1) Physical-data dual-drive decoupling: Bi-LSTM is used to predict oil temperature and combined with thermodynamic equation (PDMA) to derive theoretical pressure. The normal temperature rise interference is eliminated by calculating the "residual". This is the core algorithm innovation of the present invention.

[0054] (2) Application of timing attention mechanism: The attention mechanism is introduced into oil temperature prediction to specifically solve the time lag problem caused by transformer thermal inertia and improve the accuracy of baseline prediction.

[0055] (3) Knowledge graph and fuzzy mapping: The numerical values ​​(residuals) of the physical layer are mapped to the probabilities (BPA) of the semantic layer through fuzzy functions, breaking down the barriers between numerical computation and logical reasoning;

[0056] (4) Conflict-based evidence fusion (MSE-AFM): Jousselme distance is used to quantify sensor conflicts and the weights of DS evidence theory are dynamically adjusted to solve the problem of unreliable sensor data;

[0057] (5) Closed-loop feedback self-evolution: The rule base weights are updated in reverse using the operation and maintenance results, realizing the online learning and self-iteration of the model.

[0058] The beneficial effects of this invention are as follows: The multi-dimensional perception and intelligent fault early warning method for distribution transformer insulation status provided by this invention constructs an insulation diagnosis system that integrates physical mechanisms and data-driven approaches. It builds a multi-physics field insulation diagnosis database (MPID) for distribution transformers, containing samples from multiple scenarios such as heavy overload and environmental interference, using cross-regional data from various equipment models. Data quality is ensured through data cleaning and alignment techniques. Furthermore, an innovative Phy-LSTM state decoupling model is designed, and its Bi-LSTM prediction network incorporates a temporal attention mechanism based on a multilayer perceptron. The mechanism accurately captures thermal inertia and hysteresis effects in long-sequence data. Combined with the PDMA pressure dynamic correction algorithm, it successfully decouples temperature rise and pressure fluctuations under normal operating conditions from fault-generated gas pressure increments, significantly improving the ability to extract weak fault features. An MSE-AFM multi-source evidence adaptive fusion model is constructed, utilizing RA-ED evidence reliability assessment and WPOF weighted orthogonal fusion strategy to dynamically adjust weights when sensor drift or fault feature conflicts occur. This effectively solves the "Zadeh paradox" problem of traditional DS theory, improving the anti-interference and robustness of diagnosis. The scheme introduces a feedback self-evolution mechanism, using actual operation and maintenance results to reverse-correct knowledge base rules, achieving self-iteration of diagnostic logic. Compared to expensive full-spectrum oil chromatography online monitoring schemes, this invention only relies on three core parameters—hydrogen, pressure, and temperature—to achieve high-precision diagnosis, significantly reducing the hardware cost and maintenance difficulty of large-scale distribution network deployment. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the Bi-LSTM temperature prediction model structure according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the physical field decoupling process according to an embodiment of the present invention;

[0062] Figure 4 This is the insulation diagnostic knowledge graph architecture of an embodiment of the present invention;

[0063] Figure 5 This is a fuzzy membership function diagram of an embodiment of the present invention. Detailed Implementation

[0064] The invention will be further illustrated below with reference to the accompanying drawings and examples.

[0065] like Figure 1 A method for multi-dimensional sensing of insulation status and intelligent fault early warning of distribution transformers is disclosed, comprising the following steps:

[0066] S1. Construct a multi-dimensional condition monitoring dataset for distribution transformers, acquire time-series monitoring data on hydrogen concentration in oil, static pressure in oil tank, top oil temperature, load current, and ambient temperature, and perform data cleaning and synchronization alignment.

[0067] S2. Construct a physical field decoupling model based on bidirectional long short-term memory network (Bi-LSTM) coupled with gas state equation. Use historical normal operation data to train the model to predict theoretical oil temperature and theoretical oil pressure. Calculate the dynamic residual vector between measured and predicted values. Remove pressure fluctuations caused by normal load temperature rise and extract pure residual sequences that characterize fault features.

[0068] S3. Establish an object-oriented insulation diagnosis expert knowledge graph, transform transformer entity attributes and fault mechanisms into fuzzy production rules, and map the dynamic residual sequence obtained from the decoupled model into basic probability assignments (BPA) for different fault types through fuzzy membership functions.

[0069] S4. An improved DS evidence theory fusion algorithm based on evidence credibility correction is proposed. The algorithm calculates the conflict coefficient between evidence sources from various sensors, dynamically adjusts the weight of each evidence body, performs weighted fusion reasoning on multiple basic probability assignments generated, solves the problem of misjudgment and conflict when the fault characteristics of a single sensor are not obvious, and outputs the final fault type and confidence level.

[0070] S5. Generate multi-level early warning signals based on the fusion diagnostic results, and introduce a feedback self-evolution mechanism. Use actual operation and inspection results as verification signals to reverse the diagnostic rule thresholds in the knowledge graph and the network parameters of the prediction model, thereby realizing the self-iteration and accuracy improvement of the diagnostic logic.

[0071] In the preferred scheme, in S1, time-series operating data of distribution transformers of different regions and models are acquired through the online monitoring platform of the distribution network. Key parameters such as hydrogen in the oil, static pressure of the oil tank, top oil temperature and load current are collected. Combined with historical maintenance records, a multi-physical field insulation diagnostic database (MPID) for distribution transformers is constructed. This process covers a variety of typical states such as normal aging, partial discharge, overheating and oil circuit abnormalities. It fully considers the impact of the real operating environment such as seasonal temperature difference, load fluctuation and electromagnetic interference. Through data cleaning and time-series alignment, the problems of inconsistent sampling frequencies of different sensors and uneven distribution of fault samples are effectively solved, ensuring the integrity and generalization of the data foundation.

[0072] In the preferred embodiment, in S2, a physical-data dual-driven state decoupling model for distribution transformers (Phy-LSTM) is constructed. Its main components include a time-series prediction architecture based on a bidirectional long short-term memory network (Bi-LSTM) and a pressure correction module based on thermodynamic mechanisms. This model aims to solve the coupling problem of difficulty in distinguishing between "normal load temperature rise" and "fault-induced gas pressure increase" in single oil pressure monitoring of distribution transformers.

[0073] Preferably, a Bi-LSTM prediction network is constructed to perform high-precision inversion of the transformer top-layer oil temperature. For example... Figure 2 As shown, this network introduces a temporal attention mechanism based on a multilayer perceptron to address the dilution of key thermal features in long input sequences. Due to the significant thermal inertia and hysteresis of transformer oil temperature changes, the weights of the impact of historical load shocks on the current oil temperature exhibit a non-linear distribution. This mechanism adaptively assigns attention weights to each moment within the historical time window by calculating the energy scores of the Bi-LSTM hidden layer states, generating a context vector containing global thermal history information, thereby accurately capturing the hysteretic temperature rise features caused by sudden load changes.

[0074] Specifically, the model's input vector X t Including the ambient temperature T at the current moment amb Load current I load And the historical top oil temperature sequence T for the previous n time points top (tn:t-1). By leveraging the bidirectional transfer characteristics of Bi-LSTM, the thermal inertia of transformers and the hysteresis effect of environmental heat dissipation are fully explored.

[0075] The proposed update formula for the Bi-LSTM cell is shown below:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] In the formula, f t i t o t These represent the forget gate, input gate, and output gate, respectively; C t and h t These represent the cell state and the hidden layer state, respectively; σ is the Sigmoid activation function.

[0083] Based on this, the specific calculation process of the temporal attention mechanism is as follows:

[0084] First, calculate the hidden layer state h of the Bi-LSTM at time j. j Energy score e j :

[0085]

[0086] Subsequently, the energy score is normalized using the Softmax function to obtain the attention weight 'a' at the j-th historical moment. j :

[0087]

[0088] Finally, the hidden state at all historical moments is weighted and summed to construct the context vector c. t Then, input it into the fully connected layer to obtain the theoretical top-level oil temperature prediction value T' at the current moment. top :

[0089]

[0090] In the formula, W a U a , where v is a trainable attention weight matrix and vector.

[0091] Preferably, a pressure dynamic correction algorithm (PDMA) based on the ideal gas law is proposed, utilizing the predicted theoretical oil temperature T'.top Derivation of the theoretical static pressure P' in the oil tank theory The algorithm is based on the physical assumption that the transformer tank is a quasi-constant volume container. It takes into account the expansion effect of the insulating oil volume as the temperature changes, which compresses the space in the top air chamber and thus causes pressure changes.

[0092] The specific calculation of PDMA is shown in the following formula:

[0093]

[0094] In the formula, P init and T init The tank pressure and temperature at the initial calibration time; V gas_init V is the initial gas chamber volume; oil T is the total volume of insulating oil; α is the coefficient of volumetric expansion of transformer oil; T gas (t) represents the gas chamber temperature, which is usually approximated by the predicted top oil temperature T'. top (t) or the value corrected by the heat conduction formula. This physical equation successfully quantifies and removes the "false" pressure fluctuations caused by temperature.

[0095] Preferably, combined with Figure 3 The physical field decoupling process shown defines the multi-physics dynamic residual vector R(t) as the core feature input for subsequent fault diagnosis. By calculating the deviation between the measured value and the theoretical prediction value, abnormal features caused only by internal faults (such as gas generation during discharge or local overheating of the winding) are extracted, effectively eliminating background noise interference caused by environmental and operating condition fluctuations.

[0096] The dynamic residual vector R(t) is defined as follows:

[0097]

[0098] In the formula, T meas (t) and P meas (t) represents the top oil temperature and tank pressure measured by the sensor, respectively; C H2 (t) is the measured value of hydrogen concentration, and its residual form is the first derivative, which is the gas production rate.

[0099] Preferably, to enhance the model's sensitivity to minute fault characteristics, an adaptive sliding window residual accumulation (ASW-RC) mechanism is introduced. Considering that the instantaneous values ​​of pressure and temperature deviations caused by early latent faults (such as weak partial discharges) may be overwhelmed by sensor noise, ASW-RC calculates the cumulative effect of residuals within a time window W:

[0100]

[0101] In the formula, η kThis is a time decay factor, with greater weight applied closer to the current moment. When the cumulative score S... score When the set threshold is exceeded, the subsequent expert system diagnostic process is triggered. This strategy significantly improves the system's early warning capability for low-energy faults and avoids false alarms caused by transient noise.

[0102] In the preferred embodiment, in S3, an object-oriented knowledge graph architecture for insulation diagnosis is constructed as follows: Figure 4 As shown. First, based on ontology, the knowledge architecture of the distribution transformer is defined, abstracting the physical entities and diagnostic logic into a triple form G=(E, R, A). Here, E represents the set of entities (including the transformer body, windings, insulating oil, sensors, etc.), R represents the set of relations (such as "containment", "monitoring", "causal relationship"), and A represents the set of attributes (such as rated parameters, historical thresholds).

[0103] Preferably, to establish the mapping relationship between the dynamic residual vector R(t) output by the Phy-LSTM and the specific fault type, a fuzzy membership function is used to discretize the continuous residual values ​​into semantic variables. The fault identification framework Θ={F N ,F D , F T , F O}, corresponding to normal state, discharge fault, overheating fault, and oil circuit fault, respectively. For each feature component x of the Phy-LSTM output. i (i.e. r) T ,r P , v H2 We construct a membership function for the combination of trapezoidal and Cauchy distributions to calculate the membership degree μ(x) of different fault symptoms.

[0104] With pressure residual r P For example, Figure 5 As shown; its membership function for the "high-energy discharge" symptom is defined as μ. High_P (r P )as follows:

[0105]

[0106] In the formula, λ1 and λ2 are the starting threshold and saturation threshold of the pressure residual, respectively, determined by historical statistical data in the knowledge base; c and k are the shape parameters of the control function. Through this transformation, the numerical characteristics at the physical level are converted into fuzzy trust values ​​at the logical level.

[0107] Preferably, to achieve quantitative evidence generation, a basic probability allocation (BPA) construction method based on fuzzy correlation matrices is proposed. Unlike traditional simple threshold determination, this method constructs a feature-fault correlation matrix M.REL This matrix describes the degree to which each feature supports different fault types.

[0108] For a feature vector R(t) at a certain moment, its corresponding basic probability assignment function m( Calculated using the following formula:

[0109]

[0110] In the formula, m j (F k ) represents the relationship between the j-th sensor characteristic (such as oil temperature residual) and the fault type F. k Basic level of trust; w jk For predefined association weights in the knowledge graph; w jn Ω is the association weight of the j-th feature with the n-th fault in the identification framework, which is obtained by accumulating the weighted support of this feature for all potential fault types; Ω is the uncertainty coefficient, used to characterize the cognitive uncertainty when the feature value is in the fuzzy edge range.

[0111] Preferably, a hierarchical rule inference engine is designed to handle complex combinational fault logic. The inference engine includes an atomic rule base and a composite rule base. Atomic rules handle out-of-bounds behavior with a single parameter, while composite rules handle temporal cooperative behavior with multiple parameters.

[0112] The rule storage format uses optimized production rules:

[0113]

[0114] CF stands for the credibility factor of the rule. These rules are stored in a graph database in OWL (Web Ontology Language) format, supporting millisecond-level subgraph matching and querying by the inference engine.

[0115] Through the above process, the system transforms the physical residual vector output by Phy-LSTM into a set of evidence bodies to be fused. Each evidence body contains the confidence distribution for various types of faults, providing a standardized input format for subsequent multi-source information fusion. This process not only achieves the transition from physical quantities to semantic quantities, but also preserves ambiguity information in the diagnostic process by introducing the uncertainty coefficient Ω, avoiding misjudgments caused by premature decisions.

[0116] In the preferred embodiment, in S4, a multi-source evidence adaptive fusion model (MSE-AFM) is constructed. Its core consists of a conflict metric architecture based on evidence distance and a dynamic weighted orthogonal fusion mechanism. This model includes a conflict coefficient calculation module to quantify inconsistencies between different physical field sensors (hydrogen, oil pressure, oil temperature), and a weighted inference engine to reconstruct the basic probability assignment (BPA) when evidence sources are highly conflicting, thereby achieving accurate fault type localization.

[0117] Preferably, a Jusselme distance-based evidence reliability assessment method (RA-ED) is proposed to dynamically quantify the degree of conflict among various sensor evidence sources, effectively screening out high-confidence evidence and suppressing interference from outlier evidence. Specifically, RA-ED first treats each sensor BPA output by S3 as a point in vector space, and measures their differences (i.e., conflict coefficients) by calculating the geometric distance between evidence sources. If the evidence vector of a certain sensor (such as oil pressure) is significantly different from that of other sensors (such as hydrogen and oil temperature), it indicates that the sensor may be subject to interference or is in an atypical fault stage. In this case, its weight in the fusion process should be reduced, so that the fusion result focuses more on features with high consistency among multiple sources.

[0118] The proposed RA-ED is specifically defined as follows, which measures the difference between the i-th and j-th sources of evidence in the calculation of the conflict coefficient:

[0119]

[0120] In the formula, m i and m j , respectively, are evidence vectors from different sensors; D is the Jaccard similarity matrix, used to describe the intersection relationship of different fault sets in the identification framework. Based on distance d ij Further define the credibility weight w of the i-th piece of evidence. i :

[0121]

[0122] In the formula, 1-d ij This value represents the similarity between the i-th and j-th evidence sources, and is determined by the Jousselme distance d between them. ij The smaller the distance, the higher the similarity; 1-d kj This represents the similarity between the k-th and j-th pieces of evidence when traversing all evidence sources to calculate the total support of the system;

[0123] The dynamic weight w obtained through the above process iThe aim is to prioritize the retention of sensor information that is supported by other evidence, which contributes more to the diagnostic results due to their high consistency, thereby effectively reducing the interference of false alarms from a single sensor (such as zero-point drift of a pressure sensor) and making the fusion process dominated by high-confidence evidence.

[0124] Preferably, a weighted probabilistic orthogonal fusion strategy (WPOF) is proposed. This strategy reconstructs the evidence body by introducing a weighted averaging mechanism, followed by final inference using the Dempster orthogonal sum rule. This aims to address the "Zadeh paradox" problem that may arise in traditional DS theory when dealing with highly conflicting evidence. Its design is based on the deficiency of traditional fusion rules being overly sensitive to "veto power," and by using the aforementioned calculated weights w... i Integrate it into the evidence synthesis process, and prioritize strengthening the influence of high-quality evidence on the final judgment of the fault.

[0125] The specific calculation of WPOF in evidence fusion is shown in the following formula:

[0126] First, generate the corrected average evidence body. :

[0127]

[0128] Then, orthogonal sum rules are applied to calculate the final fault confidence m. final (A):

[0129]

[0130] In the formula, A, B, and C are subsets of the fault hypothesis set (such as discharge and overheating); K is the normalization factor, reflecting the degree of conflict after fusion. The preprocessing is equivalent to smoothing out the conflicts in the evidence space, ensuring that K<1, thereby guaranteeing the convergence and rationality of the fusion result.

[0131] Preferably, the MSE-AFM model described above achieves a complete adaptive inference process from the original BPA input, conflict distance matrix calculation, dynamic weight allocation to the final weighted fusion output. This fusion mechanism exhibits significant robustness: when there are minor conflicts in the input evidence (such as unclear early fault characteristics), the model can automatically "align" the points of each piece of evidence through weight adjustment, enhancing diagnostic confidence; while when there are serious conflicts (such as sensor faults), the model can effectively suppress interference from abnormal inputs through a low-weight allocation mechanism. Finally, the model outputs a confidence ranking of each fault type and selects the type with the highest confidence as the warning conclusion. If the maximum confidence is lower than a preset safety threshold, a "status uncertain" prompt is output to avoid misleading maintenance personnel.

[0132] In the preferred scheme, the fault warning release and knowledge base self-evolution process in S5 is as follows: First, the final fault confidence vector output by S4 is mapped into a visualized operation and maintenance decision signal. Based on the confidence value m... final The distribution is configured with a tiered early warning threshold λ, and the logic is as follows:

[0133] (i) When the confidence level of a certain type of fault is m(F) k ) >λ high At this time, a Level 1 red alert is triggered, and it is recommended to immediately shut down the power for maintenance;

[0134] (ii) When λ low <m(F k )<λ high If a Level II yellow alert is triggered, it is recommended to shorten the monitoring cycle and take oil samples for offline chromatographic analysis.

[0135] (iii) When all fault confidence levels are below λ low At this time, the system maintains a green normal state and calculates the current device's health index (HI).

[0136] Through this process, the system not only outputs the qualitative fault category, but also quantifies the risk level. The generated early warning signals are displayed in the form of trend charts and radar charts on the web or mobile terminal, intuitively reflecting the evolution trajectory of the transformer insulation status.

[0137] Preferably, to ensure the model's adaptability to sensor drift and transformer aging during long-term operation, an online knowledge base correction strategy based on feedback error is proposed. This strategy uses actual feedback results from on-site operations and maintenance as "truth labels" to inversely update the weights of the expert system's inference rules.

[0138] Its update mechanism is shown in the following formula:

[0139]

[0140] In the formula, w is the predefined feature-fault association weight in the knowledge graph. jk η is the learning rate, which is usually taken as a minimum to ensure the stability of the update; Label is the true state confirmed by human verification (1 for confirmed, 0 for excluded); Pred is the prediction confidence of the model.

[0141] This strategy employs a closed-loop feedback mechanism, enabling the system to automatically fine-tune its internal parameters after each false alarm or missed alarm. For example, if the system issues an "overheating" warning but the site confirms it as "normal," the system will automatically reduce the contribution weight of the current temperature residual range to the overheating fault. This continuous self-correction capability allows the method proposed in this invention to become increasingly accurate with the accumulation of operational data, effectively solving the problem of traditional static threshold models failing as equipment ages.

[0142] The steps of the present invention will now be described in detail with reference to the accompanying drawings and through embodiments:

[0143] (1) Step S1: Acquisition and preprocessing of multidimensional state monitoring data

[0144] In this embodiment, multi-dimensional time-series monitoring data is first obtained by deploying an Internet of Things (IoT) sensor array on and around the distribution transformer.

[0145] Specifically, the types of data collected include:

[0146] Internal state quantity: Top layer oil temperature (T) meas ), Static pressure of the oil tank (P) meas ), Hydrogen concentration in oil (C H2 );

[0147] External environment and operating conditions: Ambient temperature (T) amb ), load current (I) load ).

[0148] Considering the inconsistent sampling frequencies of different types of sensors (e.g., oil pressure and current data are typically sampled on a second or minute basis, while hydrogen data may be sampled on an hourly basis), direct fusion would lead to timing misalignment. Therefore, this invention employs a multi-order B-spline interpolation algorithm to upsample low-frequency data and uniformly resamples all data to a standard time granularity. Simultaneously, a sliding window statistical method is used to remove instantaneous outliers caused by electromagnetic interference, constructing a time-synchronized, dimension-aligned multi-physics insulation diagnostic database (MPID) for distribution transformers, providing a high-quality data foundation for subsequent model training.

[0149] (2) Step S2: Decoupling of physical and data-driven states and feature extraction

[0150] The core of this step lies in resolving the coupling problem of distinguishing between "normal load temperature rise" and "fault-induced gas pressure increase" in distribution transformers. For example... Figure 3 As shown, this step includes three sub-modules: temperature prediction, pressure correction, and residual calculation.

[0151] 1) Constructing a Bi-LSTM temperature prediction network

[0152] like Figure 3 As shown, this invention constructs a bidirectional long short-term memory network (Bi-LSTM) that incorporates a temporal attention mechanism.

[0153] Input layer: Input vector X t Including the ambient temperature T at the current moment amb Load current I loadAnd the historical top oil temperature sequence T for the previous n time points top (tn:t-1).

[0154] Bidirectional Hidden Layer: Utilizing the forward and backward propagation characteristics of Bi-LSTM, the bidirectional dependence of transformer thermal properties is captured.

[0155] Timing Attention Layer: Since transformer oil temperature changes have significant thermal inertia and hysteresis (i.e., the current oil temperature may be most affected by a sudden load increase half an hour ago), this invention introduces an attention mechanism to adaptively assign weights to each moment within the historical time window.

[0156] The specific calculation process is as follows:

[0157] First, calculate the hidden layer state h of the Bi-LSTM at time j. j Energy score e j :

[0158]

[0159] Subsequently, the attention weight α at the j-th historical moment is obtained using the Softmax function. j :

[0160]

[0161] Finally, construct the context vector c. t It inputs into a fully connected layer and outputs the theoretical top-level oil temperature prediction value T' at the current moment. top .

[0162] 2) PDMA pressure dynamic correction algorithm

[0163] To obtain a high-precision theoretical oil temperature T' top Subsequently, based on thermodynamic principles, this invention treats the transformer tank as a quasi-constant volume container and proposes a pressure dynamic correction algorithm (PDMA). This algorithm utilizes the ideal gas law and the principle of liquid thermal expansion to derive the theoretical static pressure P' of the tank. theory :

[0164]

[0165] In the formula, P init T init This is the initial calibration value; V gas_init V oil These are the initial gas chamber volume and oil volume, respectively; V oil T represents the total volume of the insulating oil; α is the coefficient of thermal expansion of the insulating oil. gas (t) represents the gas chamber temperature, which is usually approximated by the predicted top oil temperature T'. top(t) or the value corrected by the heat conduction formula. The physical meaning of this formula is that it quantifies the "normal pressure increment" caused by the expansion of oil volume and compression of the air chamber due to the increase in temperature.

[0166] 3) Generation of dynamic residual vectors

[0167] like Figure 3 As shown in the output section, the measured values ​​are differiated from the theoretical predictions to obtain the multiphysics dynamic residual vector R(t):

[0168]

[0169] in:

[0170] r T (t)=T meas -T' top Temperature residual indicates abnormal heat dissipation system or overheating of windings;

[0171] r P (t)=P meas -P' theory Pressure residual indicates non-thermal internal gas generation pressure or oil circuit anomalies.

[0172] v H2 (t): Hydrogen production rate, characterizing the degree of insulation decomposition.

[0173] This process successfully eliminated background noise caused by environmental and load fluctuations.

[0174] (3) Step S3: Construction of expert knowledge graph and fuzzy mapping

[0175] 1) Knowledge Graph Architecture

[0176] like Figure 4 As shown, this invention constructs an object-oriented knowledge graph for insulation diagnosis experts. This graph is based on an ontology definition and is described using a triple G=(E, R, A):

[0177] Entity set E includes the core entity "distribution transformer", component entities "oil tank", "insulating oil" and "winding", sensor entities "pressure sensor", "temperature sensor" and "hydrogen sensor", and fault entities "discharge fault" and "overheating fault", etc.

[0178] The relation set R clearly defines the physical connections and logical derivations.

[0179] This map provides structured prior knowledge support for subsequent diagnostic reasoning.

[0180] 2) Fuzzy membership mapping

[0181] To transform continuous physical residuals into logical-level trust levels, this invention defines a fault identification framework Θ={F N , F D , F T , F O (These correspond to normal, discharge, overheating, and oil circuit failure, respectively).

[0182] Combination Figure 5 The fuzzy membership function diagram shown illustrates that for each feature component (using pressure residual r), P For example, construct the membership function μ(x) of the ascending semi-trapezoidal distribution:

[0183] Normal region (0-λ1): when λ P When μ < λ1, μ = 0, indicating no risk of failure;

[0184] Fuzzy transition region (λ1-λ2): As the residual increases, the membership degree increases linearly or nonlinearly;

[0185] High-energy discharge region (≥λ2): when r P When ≥λ2, μ=1, indicating that a fault has occurred.

[0186] This transformation generates the basic probability assignment (BPA) for subsequent fusion.

[0187] (4) Step S4: Adaptive Fusion of Multi-Source Evidence (MSE-AFM)

[0188] To address the potential zero-point drift or failure issues of a single sensor, this step utilizes a multi-source evidence adaptive fusion model (MSE-AFM) for fault-tolerant inference.

[0189] 1) Evidence reliability assessment (RA-ED)

[0190] First, the BPAs generated by each sensor are treated as vectors, and the conflict coefficient d between the i-th evidence source and the j-th evidence source is calculated using the Jousselme distance. ij :

[0191]

[0192] Based on this distance, the credibility weight w for each piece of evidence is calculated. i If a sensor (such as oil pressure) is located at a large distance from other sensors (hydrogen, oil temperature) (i.e., there is severe conflict), its weight w i This will be significantly reduced, thereby suppressing interference from outlier data.

[0193] 2) Weighted Probability Orthogonal Fusion (WPOF)

[0194] Using weight wi Generate the corrected average body of evidence Then, the Dempster orthogonal sum rule is applied for final fusion:

[0195]

[0196] This mechanism ensures that the system can still output robust diagnostic results even in the event of partial sensor failure.

[0197] (5) Step S5: Hierarchical early warning and knowledge base self-evolution

[0198] 1) Tiered early warning

[0199] Based on the final confidence level m after fusion final The system automatically triggers multi-level warnings:

[0200] Level 1 Warning (Red): Extremely high confidence level, immediate power outage and maintenance recommended;

[0201] Level 2 Warning (Yellow): High confidence level, oil sample analysis is recommended;

[0202] Normal state (green): Calculate the health index.

[0203] 2) Feedback self-evolution mechanism

[0204] This is the intelligent manifestation of the invention. The system uses the actual feedback results from on-site operation and maintenance (Labels, such as diagnostic results from core inspection) as "truth values" to assign association weights w in the S3 knowledge graph. jk Perform reverse correction:

[0205]

[0206] The present invention achieves significant improvements through the above-mentioned technical means, and its advantages are as follows:

[0207] (1) Fault characteristics under multi-physics field conditions are decoupled, and the false alarm problem under heavy overload conditions is completely solved. The shortcomings of the existing technology are: the existing technology cannot distinguish between the pressure rise caused by normal thermal expansion and contraction and the pressure rise caused by fault gas generation.

[0208] This invention introduces a Phy-LSTM decoupling model. First, by using a Bi-LSTM network combined with a temporal attention mechanism, the nonlinear hysteresis effect of load current and ambient temperature on the top oil temperature is accurately captured, resulting in a high-precision "theoretical oil temperature". Then, the PDMA algorithm (based on the ideal gas law) is used to convert this theoretical oil temperature into "theoretical thermally induced pressure".

[0209] Final Result: By subtracting the theoretical thermally induced pressure from the measured pressure, the system directly eliminates the pressure background value caused by normal operating conditions (such as peak summer load) at the physical level. The remaining dynamic residual is caused only by internal faults (such as gas generation or local overheating). Therefore, regardless of load fluctuations, the system will not falsely report due to thermal expansion and contraction, achieving high reliability under all operating conditions.

[0210] (2) Advantage 2: Significantly improves the sensitivity to detecting subtle latent faults. Corresponding to the disadvantages of existing technologies: Existing technologies are masked by environmental baseline values ​​and are difficult to detect early faults with small pressure increments.

[0211] Because this invention uses dynamic residual vectors instead of absolute values, background noise has been eliminated. Even small pressure increments caused by faults (e.g., 2 kPa) will appear as significant non-zero values ​​in the residual sequence. Furthermore, the ASW-RC (Adaptive Sliding Window Residual Accumulation) mechanism introduced in S2 can amplify small, persistent deviations through integration.

[0212] Final result: The system is able to keenly detect weak fault signals submerged in environmental fluctuations, achieving "zero missed detection" of latent hidden dangers such as early partial discharge.

[0213] (3) Advantage 3: It has strong anti-interference and fault tolerance capabilities, and solves the problem of misjudgment caused by the drift of a single sensor. Corresponding to the disadvantages of existing technologies: Existing technologies lack a fusion mechanism, and the failure of a single sensor will lead to system paralysis or false alarm.

[0214] This invention constructs an MSE-AFM multi-source evidence adaptive fusion model. Before making a decision, the system first calculates the Jousselme distance (conflict coefficient) between the three evidence sources: oil temperature, oil pressure, and hydrogen. Once a sensor data point (such as a sudden increase in oil pressure) is found to have a serious logical conflict with other sensors (normal temperature and hydrogen levels), the algorithm automatically determines that the sensor is "untrustworthy" and significantly reduces its weight through a dynamic weighting mechanism.

[0215] Final result: Even if a cheap sensor experiences zero-point drift or damage, the system can still output correct diagnostic results by relying on other healthy sensors, greatly reducing the dependence on hardware stability and making it suitable for large-scale, low-cost deployment in distribution networks.

[0216] (4) Advantage 4: It endows the diagnostic model with the ability to adapt and evolve throughout its entire life cycle. In contrast, the existing technology has the disadvantage that the model parameters are fixed and the accuracy decreases as the equipment ages.

[0217] This invention establishes a feedback self-evolution mechanism. Each time an on-site inspection occurs (whether confirming or ruling out a diagnosis), the system obtains a true label. The algorithm calculates the deviation between the predicted result and the true label, and uses gradient descent to inversely correct the association weights in the knowledge graph.

[0218] Final result: The model can automatically adapt to the changes in physical characteristics of the transformer as it ages (such as poor heat dissipation and changes in airtightness), so that the diagnostic accuracy does not decrease but increases over time, achieving true intelligent maintenance-free operation.

[0219] There are alternative solutions that can achieve the same purpose of the invention.

[0220] (1) Explanation of alternative solutions:

[0221] Yes, to broaden the scope of protection, the following alternatives can be included:

[0222] (2) Specific replacement content:

[0223] 1) Alternative Prediction Models: The Bi-LSTM network in S2 can be replaced with GRU (Gated Recurrent Unit), Transformer Encoder, TCN (Temporal Convolutional Network), or RNN (Recurrent Neural Network). Although the accuracy may differ slightly, the underlying principle of "predicting oil temperature using a time-series model" remains the same.

[0224] 2) Replacement of physical equations: The "ideal gas law" in PDMA can be replaced by the more complex "van der Waals equation" or "empirical fitting polynomial" to describe the relationship between temperature and pressure.

[0225] 3) Alternatives to fusion algorithms: The improved DS evidence theory in S4 can be replaced by Bayesian Network, Fuzzy Integral, or Voting Logic.

[0226] 4) Input parameter substitution: In addition to oil temperature, oil pressure, and hydrogen, "vibration signal", "acoustic signal" or "micro water content" can be added as inputs to expand the fusion to more dimensions.

[0227] Explanation of related terms

[0228] (1) Physics-Data Dual Driven: This refers to the fusion of white-box models based on physical mechanisms (such as the ideal gas law) and black-box models based on deep learning (such as neural networks), which utilizes the feature extraction capabilities of data and follows physical conservation laws to improve the generalization ability and interpretability of the model.

[0229] (2) Bi-LSTM (Bidirectional Long Short-Term Memory): A bidirectional long short-term memory network, an improved recurrent neural network (RNN) that processes time series data through two state layers, forward and backward, and can effectively capture the historical dependence and future trend of transformer thermal characteristics.

[0230] (3) Temporal Attention Mechanism: A deep learning weighted algorithm used to adaptively assign weights to different historical moments in a long sequence of inputs in order to capture the nonlinear hysteresis effect of transformer load change on the current oil temperature.

[0231] (4) PDMA (Pressure Dynamic Modification Algorithm): The pressure dynamic modification algorithm proposed in this invention is based on the thermodynamic principle of a quasi-constant volume container, which quantifies and eliminates the non-fault pressure component caused by the thermal expansion of insulating oil volume.

[0232] (5) MSE-AFM (Multi-Source Evidence Adaptive Fusion Model): The multi-source evidence adaptive fusion model proposed in this invention dynamically adjusts the fusion weights of each sensor data based on the degree of conflict measured by the evidence distance.

[0233] (6) Basic Probability Assignment (BPA): A core concept in DS evidence theory, used to quantify the degree to which a piece of evidence (such as pressure residual) supports a specific hypothesis (such as discharge fault).

[0234] (7) Physical principle: Charles's Law (Isochoric Process), that is, when the volume is constant, the pressure of a certain mass of gas is directly proportional to the thermodynamic temperature (P1 / T1= P2 / T2). This is the physical basis of the PDMA algorithm of this invention.

[0235] (8) Equipment target: This invention is mainly aimed at oil-immersed distribution transformers (either fully sealed or with a breather), which are the most common equipment in the power distribution network.

Claims

1. A method for multi-dimensional sensing of insulation status and intelligent fault early warning of distribution transformers, characterized in that: First, a temperature prediction model based on Bi-LSTM and temporal attention mechanism is constructed. This model decouples and separates temperature rise and pressure fluctuations under normal operating conditions using the thermodynamic equation of state, extracting dynamic residual vectors that characterize the essence of the fault. Second, an object-oriented fault diagnosis knowledge graph is constructed, mapping residual features to fuzzy evidence bodies. Finally, an improved DS evidence theory is used to adaptively weight and fuse multi-source heterogeneous data, and the knowledge base rules are corrected online based on an operation and maintenance feedback mechanism, achieving closed-loop evolution of diagnostic logic and high-precision early warning. The process includes the following steps: S1. Construct a multi-dimensional condition monitoring dataset for distribution transformers, acquire time-series monitoring data on hydrogen concentration in oil, static pressure in oil tank, top oil temperature, load current, and ambient temperature, and perform data cleaning and synchronization alignment. S2. Construct a physical field decoupling model based on bidirectional long short-term memory network coupled with gas state equation, train the model using historical normal operation data to predict theoretical oil temperature and theoretical oil pressure, calculate the dynamic residual vector between measured and predicted values, remove pressure fluctuations caused by normal load temperature rise, and extract pure residual sequences that characterize fault features. S3. Establish an object-oriented insulation diagnosis expert knowledge graph, transform transformer entity attributes and fault mechanisms into fuzzy production rules, and map the dynamic residual sequence obtained from the decoupled model into basic probability assignments for different fault types through fuzzy membership functions. S4. An improved DS evidence theory fusion algorithm based on evidence credibility correction is proposed. The algorithm calculates the conflict coefficient between evidence sources from various sensors, dynamically adjusts the weight of each evidence body, performs weighted fusion reasoning on multiple basic probability assignments generated, solves the problem of misjudgment and conflict when the fault characteristics of a single sensor are not obvious, and outputs the final fault type and confidence level. S5. Generate multi-level early warning signals based on the fusion diagnostic results, and introduce a feedback self-evolution mechanism. Use actual operation and inspection results as verification signals to reverse the diagnostic rule thresholds in the knowledge graph and the network parameters of the prediction model, thereby realizing the self-iteration and accuracy improvement of the diagnostic logic.

2. The method for multi-dimensional perception of insulation status and intelligent fault early warning of distribution transformers according to claim 1, characterized in that: In step S1, time-series operation data of distribution transformers of different regions and models are acquired through the online monitoring platform of the distribution network. This includes time-series monitoring data of hydrogen concentration in oil, static pressure in the oil tank, top oil temperature, load current, and ambient temperature. A multi-physics field insulation diagnosis database for distribution transformers is constructed by combining historical maintenance records. The database covers typical states of normal aging, partial discharge, overheating, and abnormal oil circuits. It considers the influence of seasonal temperature differences, load fluctuations, and electromagnetic interference in the actual operating environment. Through data cleaning and time-series alignment, the problems of inconsistent sampling frequencies of different sensors and uneven distribution of fault samples are solved, ensuring the integrity and generalizability of the data foundation.

3. The method for multi-dimensional perception of insulation status and intelligent fault early warning of distribution transformers according to claim 1, characterized in that: In step S2, a state decoupling model for distribution transformers based on a dual physical-data driven approach is constructed. Its main components include a time series prediction architecture based on a bidirectional long short-term memory network (Bi-LSTM) and a pressure correction module based on thermodynamic mechanisms. The specific implementation of the distribution transformer state decoupling model is as follows: S201. A Bi-LSTM prediction network is constructed to perform high-precision inversion of the top oil temperature of the transformer. A time-series attention mechanism based on a multilayer perceptron is introduced. By calculating the energy score of the hidden layer state of the Bi-LSTM, attention weights are adaptively assigned to each moment in the historical time window to generate a context vector containing global thermal history information, which accurately captures the hysteresis temperature rise characteristics caused by load mutation. S202 proposes a pressure dynamic correction algorithm based on the ideal gas law. It uses the predicted theoretical oil temperature to derive the theoretical static pressure of the oil tank. The algorithm is based on the physical assumption that the transformer oil tank is a quasi-constant volume container. By quantifying the expansion effect of the insulating oil volume with temperature change, it quantifies the normal pressure fluctuation caused by temperature and separates it from the measured pressure.

4. The method for multi-dimensional sensing of insulation status and intelligent fault early warning of distribution transformers according to claim 1, characterized in that: In step S3, an object-oriented insulation diagnosis expert knowledge graph is constructed, transforming transformer entity attributes and fault mechanisms into fuzzy production rules. This step is specifically implemented as follows: S301, establish the mapping relationship from the dynamic residual vector R(t) to the specific fault type, and use the fuzzy membership function to discretize the continuous residual values ​​into semantic variables. For each feature component, construct the membership function μ(x) of the combination of trapezoidal and Cauchy distributions. For example, the membership function of pressure residual is defined as follows: ； The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows: In the formula, r P λ1 and λ2 are the starting threshold and saturation threshold of the pressure residual, respectively, which are determined by historical statistical data in the knowledge base; c and k are the shape parameters of the control function; through this transformation, the numerical characteristics at the physical level are converted into fuzzy trust values ​​at the logical level. S302 proposes a basic probability allocation construction method based on fuzzy correlation matrices. This method constructs a feature-fault correlation matrix describing the degree of support of each feature quantity for different fault types, and calculates the basic probability allocation function m( ): ； The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows: In the formula, m j (F k ) represents the relationship between the j-th sensor feature and the fault type F. k Basic level of trust; w jk For predefined association weights in the knowledge graph; w jn Ω is the association weight of the j-th feature with the n-th fault in the identification framework, which is obtained by accumulating the weighted support of this feature for all potential fault types; Ω is the uncertainty coefficient, used to characterize the cognitive uncertainty when the feature value is in the fuzzy edge range; S303, design a hierarchical rule inference engine to handle combinational fault logic. The inference engine includes an atomic rule base for handling single-parameter out-of-bounds behavior and a composite rule base for handling multi-parameter temporal cooperative behavior. The rule storage format adopts the following optimized production rules: ； The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows: The above formula describes the i-th production rule in the inference engine, specifically used for logical discrimination of high-energy arc faults; in the formula, the logical operator ∧ represents the AND relation, meaning that all conditions of the antecedent must be satisfied simultaneously to trigger inference; r P isHigh and v H2 "is High" indicates that both the pressure residual characteristic and the hydrogen production rate characteristic after fuzzification are in a significantly high semantic state; "CF=0.95" on the arrow represents the confidence factor of this rule as 0.95, meaning that under the condition of simultaneous pressure change and hydrogen production surge, the system determines the current discharge fault type F with a 95% confidence level. D The reasoning logic for this high-energy arcing accurately corresponds to the physical mechanism by which high-energy discharge causes the insulating oil to decompose violently and generate shock waves.

5. The method for multi-dimensional perception of insulation status and intelligent fault early warning of a distribution transformer according to claim 1, characterized in that: In step S4, a multi-source evidence adaptive fusion model is constructed, which includes a conflict coefficient calculation module and a weighted inference engine; the specific implementation is as follows: S401 proposes an evidence reliability assessment method based on Jousselme distance, which calculates the distance d between the i-th evidence source and the j-th evidence source. ij and the credibility weight w of the i-th piece of evidence i ； ； The meanings of the letters in the formula are as follows: In the formula, m i and m j , , respectively, are evidence vectors from different sensors; D is the Jaccard similarity matrix, used to describe the intersection relationship of different fault sets in the identification framework; n represents the total number of sensor evidence sources participating in the fusion; 1-d ij This value represents the similarity between the i-th and j-th evidence sources, and is determined by the Jousselme distance d between them. ij The smaller the distance, the higher the similarity; 1-d kj This represents the similarity between the k-th and j-th pieces of evidence when traversing all evidence sources to calculate the total support of the system; S402 proposes a weighted probability orthogonal fusion strategy, utilizing the weights w obtained in S401. i Generate the corrected average body of evidence The Dempster orthogonal sum rule is applied to calculate the final failure confidence level m. final (A): ； ； The meanings of the letters in the formula are as follows: In the formula, A, B, and C are subsets of the fault hypothesis set; K is the normalization factor, reflecting the degree of conflict after fusion; through The preprocessing is equivalent to smoothing out the conflicts in the evidence space, ensuring that K<1, thereby guaranteeing the convergence and rationality of the fusion result.

6. The method for multi-dimensional perception of insulation status and intelligent fault early warning of a distribution transformer according to claim 1, characterized in that: In step S5, the fault warning release and knowledge base self-evolution are specifically implemented as follows: S501 maps the final fault confidence vector output by S4 to an operation and maintenance decision signal, based on the confidence value m. final The distribution sets a tiered early warning threshold λ. When the fault confidence level is greater than λ... high A level one red alert is triggered at this time, due to λ low With λ high A level 2 yellow alert is triggered when the temperature is below λ. low Maintain a green, normal state at all times; S502 proposes an online knowledge base correction strategy based on feedback error. It utilizes actual feedback results from on-site operations and maintenance as truth labels, and updates the weights of the expert system's inference rules in reverse according to the following formula: ； The Chinese meanings of the letters, letter combinations, and symbols in the formula are as follows: In the formula, w is the predefined feature-fault association weight in the knowledge graph. jk η is the learning rate, which is usually taken as a minimum to ensure the stability of the update; Label is the true state confirmed by human verification, where 1 is confirmed and 0 is excluded; Pred is the prediction confidence of the model.

Images