Oil-immersed transformer cooling oil state on-line monitoring system and self-calibration method based on multi-mode sensing and deep learning

Abstract

The invention provides an oil-immersed transformer cooling oil state online monitoring system and a self-calibration method based on multi-modal sensing and deep learning, and belongs to the technical field of intelligent operation and maintenance of power equipment, and the system comprises an embedded sensing array unit, a self-adaptive dynamic calibration unit, an edge intelligent analysis unit and a cloud collaborative optimization platform unit. According to the invention, full-dimension sensing is realized, physical (flow velocity / temperature), chemical (gas / moisture) and mechanical (particle pollution) states of oil are covered at the same time, detection parameters are expanded to more than 12 items, self-calibration anti-interference is realized, and a long-term measurement error is controlled within + / -1.5% through reference oil sample comparison and an environment compensation algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power equipment, and in particular to an online monitoring system and a self-calibration method for the cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning. Background Art

[0002] As the core equipment of the power system, the insulation and heat dissipation performance of the oil-immersed transformer directly depends on the physical and chemical state of the cooling oil (usually mineral oil or synthetic ester oil). Cooling oil plays multiple key roles in the transformer: 1. Insulating medium: Withstands high electric field strength and prevents internal discharge; 2. Heat dissipation carrier: Dissipates heat from the windings and core through convection circulation; 3. Fault indicator: Dissolved gases (such as H2, C2H2, CO) and particulate matter in the oil can reflect early defects such as partial discharge and overheating. However, existing cooling oil status detection technology has significant defects, which seriously restricts the level of intelligent operation and maintenance of transformers:

[0003] 1. The hysteresis bottleneck of traditional offline detection: The current mainstream method relies on regular (usually every 3-6 months) oil sampling and sending it to the laboratory for gas chromatography, trace water content (Karl Fischer method) and breakdown voltage testing. The following problems exist: Response delay: It takes 3-7 days from sampling to obtaining results, and sudden faults (such as sudden changes in oil quality caused by short-circuit shocks) cannot be identified in time. Poor spatial representativeness: Single-point sampling is difficult to reflect local degradation in areas with uneven oil flow inside the transformer (such as oil cracking corresponding to winding hot spots). High labor costs: UHV substations require frequent inspections, and labor investment accounts for more than 35% of the total operation and maintenance costs.

[0004] 2. Existing online monitoring technologies lack multi-dimensional perception: Most online monitoring devices on the market are limited to single parameter detection. For thermal parameter monitoring, they only collect the top oil temperature, ignoring the impact of flow velocity distribution on heat dissipation efficiency. Gas sensors: Electrochemical sensors selectively detect H2 or CO and cannot simultaneously obtain the concentration ratios of key fault gases such as C2H2 and CH4 (IEC 60599 standard requires the detection of at least seven gases). Mechanical property monitoring gaps: There is a lack of real-time quantification of metal particles (<10μm) and fiber contamination in the oil, which accelerates the aging of insulation paper (each ppm of particle contamination shortens the paperboard life by 12%).

[0005] 3. Reliability challenges from environmental interference and sensor drift: The transformer operating environment is subject to strong electromagnetic fields (>100kV / m), mechanical vibration (50-120Hz), and a wide temperature range (-40°C to 120°C), which degrades sensor performance. Data drift: After six months of continuous operation, the MEMS flow meter's error exceeded ±8%, far exceeding the national standard (±3%). Cross-interference: H2 sensors are affected by CO and humidity, resulting in a false alarm rate of up to 15%. Calibration failure: Existing equipment lacks a self-calibration mechanism, requiring annual shutdown and disassembly for calibration, impacting power supply continuity.

[0006] 4. Data siloing and insufficient intelligent diagnostic capabilities: Even with the deployment of multiple sensors, existing systems still suffer from data fragmentation. Temperature, gas, and flow rate data are stored independently, and a multi-parameter correlation model has not been established. Shallow algorithm limitations: Threshold-based alarm rule engines (e.g., "H2 > 150ppm triggers an alert") cannot distinguish between gas production patterns associated with overheating (200-300°C) and arcing (>700°C). Lack of predictive capabilities: It is impossible to construct a dynamic model of oil aging based on historical data, making it difficult to provide a quantitative basis for optimizing oil change intervals (the current common practice of fixed oil change intervals results in over-maintenance of over 30%). Summary of the Invention

[0007] The purpose of the present invention is to provide an online monitoring system and self-calibration method for the cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning, so as to solve the technical problems mentioned in the background technology.

[0008] A closed-loop system consisting of a multimodal sensor network, edge intelligent analysis, and cloud-based self-calibration is constructed to break through the limitations of single parameter detection and achieve real-time, multi-dimensional, self-correcting monitoring of the cooling oil status of oil-immersed transformers.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] An online monitoring system for the cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning includes an embedded sensor array unit, an adaptive dynamic calibration unit, an edge intelligent analysis unit, and a cloud-based collaborative optimization platform unit. The embedded sensor array unit is connected to the adaptive dynamic calibration unit, which is arranged on the oil-immersed transformer. The adaptive dynamic calibration unit is connected to the edge intelligent analysis unit, and the edge intelligent analysis unit is connected to the cloud-based collaborative optimization platform unit. The embedded sensor array unit is used to collect status data of the oil-immersed transformer, the adaptive dynamic calibration unit is used to automatically calibrate the collected data, and the edge intelligent analysis unit is used to identify the collected data through a graph neural network, and then construct a knowledge graph library.

[0011] Furthermore, the embedded sensing array unit includes a thermal flow coupling sensor module, a dielectric chemical composite sensor module and an optical turbidity sensor module. The thermal flow coupling sensor module integrates a miniature PT100 temperature probe and a MEMS ultrasonic flowmeter, which are embedded in the inner wall of the oil channel to synchronously monitor the oil temperature, flow velocity and turbulence characteristics. The dielectric chemical composite sensor module uses an interdigitated electrode array to detect the dielectric constant of the oil to reflect the trace content, and detects the concentration of H2, CO and C2H2 fault gases through the nano-gas sensitive film ZnO / rGO. The optical turbidity sensor module uses the laser scattering principle to quantify the pollution level of metal debris and fibers in suspended particles in the oil, and then outputs the specific turbidity.

[0012] Furthermore, the adaptive dynamic calibration unit includes a reference oil sample tank and an environmental compensation analysis and processing module. The reference oil sample tank contains mineral oil. The micro pump is started regularly to inject the reference oil into the sensor cavity to eliminate the sensor zero drift. The environmental compensation analysis and processing module models the coupling interference of temperature, humidity and electromagnetic fields on the sensor output based on the LSTM network, and corrects the original data in real time.

[0013] Furthermore, the edge intelligent analysis unit includes a source data fusion module, a degradation trend prediction module and a fault tracing module. The source data fusion module establishes a spatiotemporal correlation model of oil temperature, flow rate, dielectric properties, gas concentration and turbidity through the graph neural network (GNN). The degradation trend prediction module uses the Transformer timing model and combines the transformer load historical data to predict the oil breakdown voltage attenuation curve. The fault tracing module builds a knowledge graph library. When an abnormal C2H2 / H2 ratio is detected, it automatically associates the potential causes of partial discharge or arc faults.

[0014] Furthermore, the cloud-based collaborative optimization platform unit includes a blockchain evidence storage module and a federated learning model update module. The blockchain evidence storage module is used to encrypt and store oil status data to ensure that the inspection report cannot be tampered with and meets the audit requirements of the power industry. The federated learning model update module aggregates the monitoring data of several transformers, dynamically optimizes the edge AI algorithm, and solves the small sample cold start problem.

[0015] A self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning, the method comprising the following steps:

[0016] Step 1: Initialize and calibrate several modal sensors;

[0017] Step 2: Online monitoring and dynamic priority scheduling;

[0018] Step 3: Anti-interference compensation and credibility assessment;

[0019] Step 4: Edge intelligent reasoning and decision tree generation;

[0020] Step 5: Closed-loop verification and self-assessment.

[0021] Furthermore, the specific process of step 1 is as follows: three groups of sensor nodes are installed at the high-voltage winding outlet, radiator inlet and bottom of the oil tank of the transformer oil channel. Each group of nodes integrates temperature, flow rate, dielectric, gas and turbidity sensors. The three-dimensional coordinates (x, y, z) of each node are recorded by a laser rangefinder, and an oil flow path topology map is established. New oil is injected when the transformer is shut down, and the static calibration mode is started to maintain the oil temperature at 40℃±0.5℃ and the flow rate at 0m / s. The baseline value of each sensor is recorded. If the deviation of a sensor exceeds ±2%, the automatic cleaning program is triggered and the ultrasonic vibrator clears the bubbles on the probe surface.

[0022] Furthermore, the specific process of step 2 is as follows: in steady-state operation, when the load rate is <60%, the default sampling frequency is 1 Hz for temperature and flow rate, and the sampling frequency is 0.2 Hz for gas and turbidity. When the change rate of any parameter exceeds the threshold, the sampling frequency is automatically increased to 10 Hz for 30 seconds. The oil group migration time is calculated based on the oil flow velocity and the node spacing, and the upstream and downstream sensor data are timestamped and compensated to ensure the spatiotemporal synchronization of multiple parameter data of the same oil group.

[0023] Furthermore, the specific process of step 3 is: establish the sensor error transfer function: ΔS=K1×Tamb+K2×sin(2πf vid t)+K3×B2EM, where K1, K2, K3 are calibration coefficients determined by experiments, f vid is the frequency, according to the sensor historical error σ and the current environmental interference intensity I env , calculate the real-time credibility weight:

[0024]

[0025] Low-weight data W < 0.5 will be marked as suspicious, triggering redundant sensor cross-validation.

[0026] Furthermore, the specific process of step 4 is as follows: several levels of anomaly detection, a detection rule engine, hard threshold alarms, a response time of less than 1 second, time series model detection, an LSTM network predicting parameter trends for the next hour, triggering an alert if the predicted value exceeds the threshold confidence interval P>95%, knowledge graph detection, matching arc discharge patterns in the fault database when C2H2 / H2>2 and oil temperature>85°C, generating treatment suggestions, and using reinforcement learning to dynamically adjust diagnostic model parameters based on the feedback from operation and maintenance personnel on the alarm;

[0027] The specific process of step 5 is as follows: Dynamic oil change strategy generation: The cloud receives the oil degradation curves of several transformers, identifies common patterns through cluster analysis, and issues regional optimization strategies. The high humidity in coastal areas causes the BDV degradation rate of transformers to accelerate by 15%, shortening the oil change cycle from 5 years to 4.3 years. The fault handling knowledge base is updated through crowdsourcing. After handling the alarm event, the operation and maintenance personnel submit the fault cause and handling record via mobile terminals. The cloud uses natural language processing technology to extract key information and update the fault and feature mapping relationship in the knowledge graph. The edge unit synchronizes the updates every 24 hours.

[0028] Virtual fault injection is automatically started every month to verify whether the system response meets expectations. The C2H2 concentration in the simulated oil increases stepwise to 50ppm, with alarm delay less than 3 seconds and fault type identification accuracy greater than 90%. The system also generates a health self-assessment report every quarter, including: sensor health, model prediction accuracy, and recommended maintenance items.

[0029] The present invention has the following beneficial effects due to the adoption of the above technical solution:

[0030] The present invention has full-dimensional perception: it covers the physical (flow rate / temperature), chemical (gas / moisture), and mechanical (particle contamination) states of the oil simultaneously, and the detection parameters are expanded to more than 12 items. It has self-calibration and anti-interference: through reference oil sample comparison and environmental compensation algorithm, the long-term measurement error is controlled within ±1.5% (traditional technology ±5%). Predictive maintenance: it can warn of the risk of oil breakdown 72 hours in advance, with an accuracy rate of >92% (compared with laboratory chromatographic analysis that lags behind by more than 7 days). BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a block diagram of the system principle of the present invention;

[0032] Figure 2 This is a block diagram of the embedded sensor array unit module of the present invention;

[0033] Figure 3 This is a block diagram of the adaptive dynamic calibration unit module of the present invention;

[0034] Figure 4 This is a block diagram of the edge intelligent analysis unit module of the present invention;

[0035] Figure 5 This is a block diagram of the unit modules of the cloud collaborative optimization platform of the present invention;

[0036] Figure 6 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.

[0038] like Figure 1 As shown, an online monitoring system for the cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning includes an embedded sensor array unit, an adaptive dynamic calibration unit, an edge intelligent analysis unit and a cloud-based collaborative optimization platform unit. The embedded sensor array unit is connected to the adaptive dynamic calibration unit, which is arranged on the oil-immersed transformer, the adaptive dynamic calibration unit is connected to the edge intelligent analysis unit, and the edge intelligent analysis unit is connected to the cloud-based collaborative optimization platform unit. The embedded sensor array unit is used to collect status data of the oil-immersed transformer, the adaptive dynamic calibration unit is used to automatically calibrate the collected data, and the edge intelligent analysis unit is used to identify the collected data through a graph neural network, and then construct a knowledge graph library.

[0039] In the embodiment of the present invention, Figure 2 As shown in the figure, the embedded sensing array unit includes a thermal-fluid coupling sensor module, a dielectric-chemical composite sensor module and an optical turbidity sensor module. The thermal-fluid coupling sensor module integrates a miniature PT100 temperature probe and a MEMS ultrasonic flowmeter, which are embedded in the inner wall of the oil channel to synchronously monitor the oil temperature, flow velocity and turbulence characteristics. The dielectric-chemical composite sensor module uses an interdigital electrode array to detect the dielectric constant of the oil to reflect the trace content, and detects the concentration of H2, CO and C2H2 fault gases through the nano-gas-sensitive film ZnO / rGO. The optical turbidity sensor module uses the laser scattering principle to quantify the pollution level of metal debris and fibers in the suspended particles in the oil, and then outputs the specific turbidity.

[0040] In the embodiment of the present invention, Figure 3 As shown in the figure, the adaptive dynamic calibration unit includes a reference oil sample tank and an environmental compensation analysis and processing module. The reference oil sample tank contains mineral oil. The micro pump is started regularly to inject the reference oil into the sensor cavity to eliminate the zero drift of the sensor. The environmental compensation analysis and processing module models the coupling interference of temperature, humidity and electromagnetic field on the sensor output based on the LSTM network, and corrects the original data in real time.

[0041] In the embodiment of the present invention, Figure 4As shown, the edge intelligent analysis unit includes a source data fusion module, a degradation trend prediction module and a fault tracing module. The source data fusion module establishes a spatiotemporal correlation model of oil temperature, flow rate, dielectric properties, gas concentration and turbidity through the graph neural network GNN. The degradation trend prediction module uses the Transformer timing model and combines the transformer load historical data to predict the oil breakdown voltage attenuation curve. The fault tracing module builds a knowledge graph library. When an abnormal C2H2 / H2 ratio is detected, it automatically associates the potential causes of partial discharge or arc fault.

[0042] In the embodiment of the present invention, Figure 5 As shown in the figure, the cloud-based collaborative optimization platform unit includes a blockchain evidence storage module and a federated learning model update module. The blockchain evidence storage module is used to encrypt and store oil status data to ensure that the inspection report cannot be tampered with and meets the audit requirements of the power industry. The federated learning model update module aggregates the monitoring data of several transformers, dynamically optimizes the edge AI algorithm, and solves the small sample cold start problem.

[0043] A self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning, the method comprising the following steps:

[0044] Step 1: Multimodal sensor initialization and spatial calibration

[0045] 1. Sensor spatial topology deployment

[0046] Three sets of sensor nodes were installed at key locations in the transformer's oil channels (the high-voltage winding outlet, the radiator inlet, and the bottom of the oil tank). Each set of nodes integrated temperature, flow velocity, dielectric, gas, and turbidity sensors. A laser rangefinder recorded the three-dimensional coordinates (x, y, z) of each node, creating a topological map of the oil flow path (e.g., "Node A → Node B: distance 1.2 m, oil flow direction 30°").

[0047] 2. Sensor baseline calibration

[0048] With the transformer shut down, inject new oil (compliant with GB 2536) and initiate static calibration mode: maintain the oil temperature at 40°C ± 0.5°C and the flow rate at 0 m / s. Record the baseline values ​​of each sensor (e.g., dielectric constant ε_baseline = 2.21, turbidity NTU_baseline = 0.3). If a sensor deviation exceeds ±2% (e.g., a gas sensor H2 reading > 5 ppm), an automatic cleaning procedure is triggered: a micro-ultrasonic vibrator removes air bubbles from the probe surface.

[0049] Step 2: Online monitoring and dynamic priority scheduling

[0050] 1. Dynamic frequency adjustment for data acquisition, steady-state operation: When the load rate is <60%, the default sampling frequency is 1Hz (temperature / flow rate) and 0.2Hz (gas / turbidity). Transient event triggering: When the rate of change of any parameter exceeds the threshold (such as ΔT / Δt>5℃ / min), the sampling frequency is automatically increased to 10Hz (all parameters) for 30 seconds.

[0051] 2. Spatiotemporal alignment of multimodal data: Based on the oil flow velocity (v) and the node spacing (L), the oil swarm migration time (Δt = L / v) is calculated, and the upstream and downstream sensor data are timestamped and compensated (e.g., Δt = 2.1 seconds from node A to node B) to ensure spatiotemporal synchronization of multi-parameter data for the same oil swarm.

[0052] Step 3: Anti-interference compensation and credibility assessment

[0053] 1. Modeling of environmental coupling interference and establishing the sensor error transfer function: ΔS=K1×Tamb+K2×sin(2πf vid t)+K3×B2EM where K1, K2, and K3 are calibration coefficients determined by experiments (e.g., when the vibration frequency ∫vid=100Hz, K2=0.03).

[0054] 2. Data credibility weight allocation: Calculate the real-time credibility weight based on the sensor's historical error (σ) and the current environmental interference intensity (I_env):

[0055] Low-weight data (W<0.5) will be marked as “suspicious”, triggering redundant sensor cross-validation.

[0056] Step 4: Edge Intelligent Reasoning and Decision Tree Generation

[0057] 1. Multi-level anomaly detection: Level 1 (rule engine): Hard threshold alarms (e.g., H₂ > 150 ppm, particulate matter > 10 ppm) with a response time of < 1 second. Level 2 (time series model): An LSTM network predicts parameter trends for the next hour. If the predicted value exceeds the threshold confidence interval (P > 95%), an alert is triggered. Level 3 (knowledge graph): When C₂H₂ / H₂ > 2 and the oil temperature > 85°C, the fault database matches the "arcing" pattern and generates a recommended action (e.g., "Stop the machine and inspect the casing connection").

[0058] 2. A self-optimizing diagnostic model that uses reinforcement learning (DQN algorithm) to dynamically adjust diagnostic model parameters (such as reducing the initial weight coefficient of the gas sensor) based on the operator's feedback on the alarm (false positives / missed negatives).

[0059] Step 5: Cloud-edge collaborative optimization and policy delivery

[0060] 1. Dynamic oil change strategy generation: The cloud receives oil degradation curves (BDV vs. time) from multiple transformers, identifies common patterns through cluster analysis (e.g., high humidity in coastal areas causes the BDV degradation rate to accelerate by 15%), and issues regional optimization strategies (e.g., shortening the oil change interval from 5 years to 4.3 years).

[0061] 2. The fault handling knowledge base is updated through crowdsourcing. After handling an alarm event, operations and maintenance personnel submit the fault cause and handling record (e.g., "Code F003: Confirmed poor contact of tap changer") via mobile terminals. The cloud uses NLP technology to extract key information and update the fault-feature mapping relationship in the knowledge graph. Edge units synchronize updates every 24 hours.

[0062] Step 6: Closed-loop verification and system self-assessment

[0063] 1. Virtual injection test

[0064] Virtual fault injection (e.g., simulating a step increase in the C2H2 concentration in the oil to 50 ppm) is automatically initiated every month to verify whether the system response meets expectations (alarm delay < 3 seconds, fault type identification accuracy > 90%).

[0065] 2. Health Self-Assessment Report

[0066] The system generates a quarterly self-assessment report, including: sensor health (such as "the performance of the vibration compensation module has dropped by 12%"); model prediction accuracy (such as "the BDV prediction error has increased from ±3.2% to ±4.1%"); and recommended maintenance items (such as "replace the gas-sensitive film of node B").

[0067] Matters not covered by the present invention are known technologies.

[0068] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An online monitoring system for cooling oil in oil-immersed transformers based on multimodal sensing and deep learning, featuring: It includes an embedded sensor array unit, an adaptive dynamic calibration unit, an edge intelligent analysis unit and a cloud-based collaborative optimization platform unit. The embedded sensor array unit is connected to the adaptive dynamic calibration unit, and the embedded sensor array unit is set on the oil-immersed transformer. The adaptive dynamic calibration unit is connected to the edge intelligent analysis unit, and the edge intelligent analysis unit is connected to the cloud-based collaborative optimization platform unit; the embedded sensor array unit is used to collect status data of the oil-immersed transformer, the adaptive dynamic calibration unit is used to automatically calibrate the collected data, and the edge intelligent analysis unit is used to identify the collected data through a graph neural network, and then build a knowledge graph library.

2. The online monitoring system for cooling oil status of oil-immersed transformers based on multimodal sensing and deep learning according to claim 1 is characterized by: The embedded sensing array unit includes a thermal-fluid coupling sensor module, a dielectric-chemical composite sensor module, and an optical turbidity sensor module. The thermal-fluid coupling sensor module integrates a miniature PT100 temperature probe and a MEMS ultrasonic flowmeter, is embedded in the inner wall of the oil channel, and synchronously monitors the oil temperature, flow velocity, and turbulence characteristics. The dielectric-chemical composite sensor module uses an interdigitated electrode array to detect the dielectric constant of the oil to reflect the trace content, and detects the concentration of H2, CO, and C2H2 fault gases through the nano-gas-sensitive film ZnO / rGO. The optical turbidity sensor module uses the laser scattering principle to quantify the pollution level of metal debris and fibers in suspended particles in the oil, and then outputs the specific turbidity.

3. The online monitoring system for cooling oil status of oil-immersed transformers based on multimodal sensing and deep learning according to claim 1 is characterized in that: The adaptive dynamic calibration unit includes a reference oil sample tank and an environmental compensation analysis and processing module. The reference oil sample tank contains mineral oil. The micro pump is started regularly to inject the reference oil into the sensor cavity to eliminate the sensor zero drift. The environmental compensation analysis and processing module uses the LSTM network to model the coupling interference of temperature, humidity and electromagnetic fields on the sensor output, and corrects the original data in real time.

4. The online monitoring system for cooling oil status of oil-immersed transformers based on multimodal sensing and deep learning according to claim 1 is characterized in that: The edge intelligent analysis unit includes a source data fusion module, a degradation trend prediction module and a fault tracing module. The source data fusion module establishes a spatiotemporal correlation model of oil temperature, flow rate, dielectric properties, gas concentration and turbidity through the graph neural network (GNN). The degradation trend prediction module uses the Transformer timing model and combines the transformer load historical data to predict the oil breakdown voltage attenuation curve. The fault tracing module builds a knowledge graph library. When an abnormal C2H2 / H2 ratio is detected, it automatically associates the potential causes of partial discharge or arc faults.

5. The online monitoring system for cooling oil status of oil-immersed transformers based on multimodal sensing and deep learning according to claim 1 is characterized in that: The cloud collaborative optimization platform unit includes a blockchain evidence module and a federated learning model update module. The blockchain evidence module is used to encrypt and store oil status data. To ensure that the inspection report cannot be tampered with and meets the audit requirements of the power industry, the federated learning model update module aggregates the monitoring data of several transformers, dynamically optimizes the edge AI algorithm, and solves the small sample cold start problem.

6. The self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning according to claim 1 is characterized in that: The method comprises the following steps: Step 1: Initialize and calibrate several modal sensors; Step 2: Online monitoring and dynamic priority scheduling; Step 3: Anti-interference compensation and credibility assessment; Step 4: Edge intelligent reasoning and decision tree generation; Step 5: Closed-loop verification and self-assessment.

7. The self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning according to claim 1 is characterized in that: The specific process of step 1 is as follows: three sets of sensor nodes are installed at the high-voltage winding outlet, radiator inlet and bottom of the oil tank of the transformer oil channel. Each set of nodes integrates temperature, flow rate, dielectric, gas and turbidity sensors. The three-dimensional coordinates (x, y, z) of each node are recorded by a laser rangefinder, and an oil flow path topology map is established. New oil is injected into the transformer when it is shut down. The static calibration mode is started, the oil temperature is maintained at 40℃±0.5℃ and the flow rate is 0m / s. The baseline value of each sensor is recorded. If the deviation of a sensor exceeds ±2%, the automatic cleaning program is triggered, and the ultrasonic vibrator removes bubbles on the probe surface.

8. The self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning according to claim 1 is characterized in that: The specific process of step 2 is as follows: in steady-state operation, when the load rate is <60%, the default sampling frequency is 1Hz for temperature and flow rate, and the sampling frequency is 0.2Hz for gas and turbidity. When the change rate of any parameter exceeds the threshold, the sampling frequency is automatically increased to 10Hz for 30 seconds. The oil group migration time is calculated based on the oil flow velocity and the node spacing, and the upstream and downstream sensor data are timestamped and compensated to ensure the spatiotemporal synchronization of multiple parameter data of the same oil group.

9. The self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning according to claim 1 is characterized in that: The specific process of step 3 is: establish the sensor error transfer function: ΔS=K1×Tamb+K2×sin(2πf vid t)+K3×B2EM, where K1, K2, K3 are calibration coefficients determined by experiments, f vid is the frequency, according to the sensor historical error σ and the current environmental interference intensity I env , calculate the real-time credibility weight: Low-weight data W < 0.5 will be marked as suspicious, triggering redundant sensor cross-validation.

10. The self-calibration method for an online monitoring system for cooling oil status of an oil-immersed transformer based on multimodal sensing and deep learning according to claim 1 is characterized in that: The specific process of step 4 includes: several layers of anomaly detection, a detection rule engine, hard threshold alarms, a response time of less than 1 second, time series model detection, an LSTM network predicting parameter trends for the next hour, triggering an alert if the predicted value exceeds the threshold confidence interval P>95%, knowledge graph detection, matching arc discharge patterns in the fault database when C2H2 / H2>2 and the oil temperature>85°C, generating treatment recommendations, and using reinforcement learning to dynamically adjust diagnostic model parameters based on the operator's feedback on the alarm. The specific process of step 5 is as follows: Dynamic oil change strategy generation: The cloud receives the oil degradation curves of several transformers, identifies common patterns through cluster analysis, and issues regional optimization strategies. The high humidity in coastal areas causes the BDV degradation rate of transformers to accelerate by 15%, shortening the oil change cycle from 5 years to 4.3 years. The fault handling knowledge base is updated through crowdsourcing. After handling the alarm event, the operation and maintenance personnel submit the fault cause and handling record via mobile terminals. The cloud uses natural language processing technology to extract key information and update the fault and feature mapping relationship in the knowledge graph. The edge unit synchronizes the updates every 24 hours. Virtual fault injection is automatically started every month to verify whether the system response meets expectations. The C2H2 concentration in the simulated oil increases stepwise to 50ppm, with alarm delay less than 3 seconds and fault type identification accuracy greater than 90%. The system also generates a health self-assessment report every quarter, including: sensor health, model prediction accuracy, and recommended maintenance items.

Images