Intelligent diagnosis method for thermal abnormality of inverter

Abstract

The invention relates to the field of machine learning and deep learning, and discloses an intelligent diagnosis method for thermal anomalies of an inverter, which comprises the following steps: acquiring multi-dimensional operation data of the inverter, constructing a topological graph, mapping physical quantities into graph nodes, and defining connecting edges; constructing a graph neural network model, and training by adopting a self-supervision strategy of random mask reconstruction; the method comprises the steps of obtaining operation data in real time and recognizing working conditions, controlling a fan to execute rotating speed perturbation in a steady state and extracting transient response characteristics to be injected into a graph structure, conducting self-adaptive normalization on a reconstruction residual error based on a volatility index in a dynamic state, and judging abnormity based on the processed residual error. The binary problems of feature covering in a steady state and environmental noise interference in a dynamic state are solved.

Description

Technical Field

[0001] This invention relates to an intelligent diagnostic method for inverter thermal anomalies, belonging to the fields of machine learning and deep learning technologies. Background Technology

[0002] In current photovoltaic power generation and energy storage systems, the thermal stability of the inverter is directly related to the overall operational safety and lifespan of the station. Existing thermal protection schemes are mostly based on deterministic physical thresholds or static temperature rise models, determining the state by monitoring whether sensor readings reach upper limits or the temperature rise rate is abnormal. While rule-based monitoring logic is feasible under laboratory conditions, its ability to detect early potential problems faces fundamental limitations in the long-term operation of actual outdoor stations. Specifically, existing technologies introduce algorithmic models to assist in diagnosis, but their judgment strategies are difficult to break free from dependence on fixed parameters and cannot adapt to the characteristic evolution throughout the equipment's lifecycle. For example, Chinese invention patent CN118174623A discloses a thermal event diagnostic device for a main drive inverter and... The method, including the motor drive system of the device, introduces a junction temperature estimation model and diagnoses anomalies by verifying whether the difference between the junction temperature estimated by the model and the sensor sampling temperature exceeds a predetermined tolerance range. However, it is essentially a passive defense based on static deviation. Such methods rely on a preset fixed tolerance threshold, assuming that the thermal resistance characteristics of the device are constant. They ignore the gradual drift of the physical coupling relationship caused by dust accumulation on the heat sink, drying of the heat transfer medium, or aging of the device during long-term operation of the inverter. When early slight thermal resistance degradation occurs, the resulting minute temperature deviation is often masked within a wide preset tolerance range, causing the system to fail to identify potential structural hazards before the absolute temperature reaches the hard alarm line.

[0003] In natural environments, the thermal performance of inverters is a product of strong nonlinear coupling of multiple variables, including load, ambient temperature, irradiance, and the state of the heat dissipation system. Traditional threshold methods struggle to eliminate the masking effect of operating conditions. Under low load or low temperature conditions, even if early thermal resistance degradation occurs, such as dust accumulation in the air duct or aging of the heat transfer medium, the absolute temperature may still be below the alarm threshold, leading to undetected hidden dangers. To address this issue, the industry has attempted to introduce data-driven machine learning algorithms for predictive maintenance. However, conventional supervised learning schemes rely on training with a large amount of historical data labeled with faults. In industrial settings, the vast majority of operating data consists of unlabeled normal samples, while fault samples are scarce and costly to obtain. This results in label-dependent algorithms facing difficulties in cold start and limited generalization capabilities.

[0004] Therefore, how to break free from the dependence on fault label data, construct a diagnostic mechanism that adaptively learns the physical coupling law from unlabeled normal data, and maintains high robustness under the full-cycle aging characteristics of equipment and complex dynamic and steady-state operating conditions, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A smart diagnostic method for inverter thermal anomalies, comprising the following steps: Construct topology graph data, obtain multidimensional operating time series data of inverter, map each dimension of physical quantity in multidimensional operating time series data to nodes of graph structure, and define connection edges between nodes according to the correlation between physical quantities. Self-supervised model training: Construct a graph neural network model and train the graph neural network model using a self-supervised learning strategy. The self-supervised learning strategy includes randomly selecting target nodes in the mask graph structure in each training iteration, with the optimization objective being to minimize the difference between the reconstructed value and the true value of the target node. This drives the graph neural network model to learn the coupled feature representations between each node, resulting in a trained graph neural network model. The system performs differential reasoning based on operating conditions, acquires real-time operating data of the inverter, and identifies the operating conditions. When the operating condition is steady-state, it controls the inverter's cooling fan to perform speed micro-perturbation operation, extracts the transient response features of temperature parameters to the speed micro-perturbation operation, and injects the transient response features as temporary nodes into the graph structure for the trained graph neural network model to calculate the reconstruction residual. The calculated reconstruction residual is defined as the first reconstruction residual. When the operating condition is dynamic, it uses the trained graph neural network model to calculate the reconstruction residual of the current graph structure, which is defined as the second reconstruction residual. Then, it calculates the volatility index of the graph nodes representing the input excitation within a preset time window, generates a dynamic residual tolerance coefficient based on the volatility index, and uses the dynamic residual tolerance coefficient to normalize the second reconstruction residual. Anomaly detection is based on the statistical distribution characteristics of the first reconstructed residual or the normalized second reconstructed residual. When the residual value deviates from the preset confidence interval, a thermal anomaly diagnostic signal is generated.

[0006] Preferably, in the step of constructing the topology graph data, the connection edges between nodes are defined by a parameterized adjacency matrix, which is a learnable parameter in the graph neural network model; the self-supervised model training step further includes: using historical running data to jointly optimize the parameterized adjacency matrix, so that the graph neural network model adaptively adjusts the values ​​of each element in the parameterized adjacency matrix according to the goal of minimizing reconstruction differences, and dynamically represents the evolution of the coupling strength between nodes; the anomaly detection step further includes: monitoring the weight distribution change of the parameterized adjacency matrix, and when the first reconstruction residual or the second reconstruction residual exceeds the threshold and the weight distribution change rate is lower than the preset drift rate threshold, confirming the generation of the thermal anomaly diagnostic signal.

[0007] Preferably, the step of extracting the transient response characteristics of temperature parameters to speed micro-disturbance operation specifically includes: issuing a speed control command to the cooling fan with a sinusoidal modulation signal superimposed with a preset frequency; collecting the junction temperature data of the inverter's insulated gate bipolar transistor as a temperature parameter; calculating the phase lag parameter and amplitude-frequency response ratio parameter of the junction temperature data relative to the sinusoidal modulation signal; and using the phase lag parameter and amplitude-frequency response ratio parameter as transient response characteristics.

[0008] Preferably, in the step of normalizing the second reconstruction residual output by the graph neural network model using the dynamic residual tolerance coefficient, the normalized diagnostic score is calculated using the following formula: ,in, The second reconstruction residual after normalization. The original reconstruction residual is the output of the graph neural network model. As a volatility indicator, This is the preset adjustment coefficient.

[0009] Preferably, the steps of calculating the first reconstruction residual or the second reconstruction residual using the graph neural network model specifically include: keeping the randomly deactivated units in the graph neural network model in an active state; performing multiple forward inferences on the running data at the same time to obtain multiple reconstruction values; calculating the variance of the multiple reconstruction values ​​and defining the variance as a model uncertainty index; calculating the difference between the mean of the multiple reconstruction values ​​and the true value as the basic residual, and using the model uncertainty index to weight the basic residual to obtain the first reconstruction residual or the second reconstruction residual.

[0010] Preferably, the step of generating thermal anomaly diagnostic signals further includes: comparing the model uncertainty index with a preset confidence threshold; when the model uncertainty index is higher than the confidence threshold, determining that the current operating condition is a long-tailed unknown operating condition, and suppressing the generation of thermal anomaly diagnostic signals or reducing the priority of thermal anomaly diagnostic signals.

[0011] Preferably, the step of calculating the volatility index of the graph nodes representing the input excitation within a preset time window specifically includes: selecting DC-side voltage nodes and DC-side current nodes as graph nodes of the input excitation; collecting time series data of the graph nodes of the input excitation within the preset time window; calculating the ratio of the standard deviation to the mean of the time series data, and determining the ratio as the volatility index.

[0012] Preferably, in the step of constructing the topology graph data, the nodes of the graph structure include strongly correlated nodes, excitation nodes, and environment nodes; the strongly correlated nodes include at least the junction temperature of the insulated gate bipolar transistor, the capacitor temperature, and the heat sink temperature; the excitation nodes include at least the DC side voltage, the DC side current, and the AC side power; and the environment nodes include at least the ambient temperature inside the chassis and the fan speed feedback signal.

[0013] Preferably, in the self-supervised model training step, the strategy for at least one target node in the random mask graph structure specifically includes: calculating the degree centrality index of each node in the graph structure; dynamically adjusting the masking probability of each node based on the degree centrality index, wherein nodes with higher degree centrality index have lower masking probability, so that the graph neural network model can make more use of the contextual information of edge nodes for feature reconstruction.

[0014] Preferably, after the step of generating the thermal anomaly diagnostic signal, the method further includes performing a thermal anomaly root cause localization step: extracting the attention weight values ​​of each connection edge connected to the target node in the graph neural network model; comparing the attention weight values ​​with a preset baseline weight distribution; identifying connection edges whose attention weight value changes exceed a preset localization threshold, and marking the physical components corresponding to the neighboring nodes connected to the connection edge as the associated root cause components causing the thermal anomaly.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the intelligent diagnosis of inverter thermal anomalies, a graph topology structure mapping the correlation of multidimensional physical quantities is constructed. A graph neural network driven by a random masking self-supervised learning strategy is adopted to learn the implicit nonlinear coupling laws between sensor nodes. During the training phase, the model reconstructs the features of the masked nodes using the context information of neighboring nodes. Without the need for fault labels, a high-dimensional structural consistency benchmark representing the normal physical laws of the equipment is established. Real-time diagnosis uses the cross-derivative relationship to reconstruct the logic and keenly capture the small structural damages related to voltage, current and temperature. Under low load or low temperature conditions where the absolute temperature value has not triggered the threshold, the early thermal resistance degradation hidden anomalies are detected by identifying the failure of physical coupling laws. This realizes the transformation from a threshold-based passive defense to a law-based proactive diagnosis method.

[0016] 2. Configure the adjacency matrix of node connection relationships as learnable parameters of the graph neural network. During the backpropagation process, the matrix is ​​jointly optimized and dynamically updated based on the objective of minimizing the reconstruction error. The parameterized graph structure evolution mechanism continuously tracks and internalizes the drift of physical coupling strength caused by device aging, contact surface oxidation, or dust accumulation. The weight distribution between nodes is automatically calibrated during long-term operation. The structure's adaptive capability distinguishes between gradual feature shifts caused by normal aging of equipment and sudden structural damage caused by faults. This avoids the problem of increased false alarm rate caused by feature distribution drift in fixed topology models throughout their entire life cycle.

[0017] 3. During the online inference phase, the randomly deactivated units are kept active. The variance of the reconstructed values ​​is calculated by multiple random forward propagations of the input data at the same time. This quantifies the uncertainty of the model's perception of the current operating conditions. The uncertainty index is mapped to the residual judgment weight factor using the Bayesian approximation idea. The confidence of the input data is dynamically evaluated. The probability weighted logic automatically suppresses the generation of false high residuals when the model faces long-tail normal operating conditions or transient noise in the environment, ensuring that the diagnostic system generates abnormal signals when the consistency of physical laws is substantially destroyed. Attached Figure Description

[0018] Figure 1 This is a flowchart of the inverter thermal anomaly diagnosis based on active excitation and dynamic suppression mechanism of the present invention; Figure 2 This is a graph showing the time drift of the parameterized adjacency matrix weights, which characterize the physical coupling evolution of this invention. Figure 3 This is a time sequence diagram of the self-supervised training interaction of the graph neural network based on the random mask reconstruction strategy of this invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below. It should be noted that the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0020] This invention discloses an intelligent diagnostic method for inverter thermal anomalies. It constructs a topology graph data structure characterizing the inverter's thermophysical state. The processor acquires multidimensional time-series data of the inverter during operation. This multidimensional data covers strongly correlated nodes such as IGBT junction temperature, capacitor temperature, and heatsink temperature; excitation nodes such as DC-side voltage, DC-side current, AC-side power, and modulation ratio; and environmental nodes such as chassis ambient temperature and fan speed feedback. These physical quantities are mapped to node feature vectors in a graph neural network. Simultaneously, based on the energy flow or heat conduction logic between physical quantities, connection edges between nodes are defined, or a learnable parameterized adjacency matrix is ​​initialized to implicitly characterize the relationships between nodes. To assess correlation strength, a lightweight graph neural network model, such as GraphSAGE or Graph Attention Network, is constructed. This model is pre-trained using a self-supervised learning strategy. The training process does not rely on manually labeled fault tags. Through a random masking mechanism, some node features in the graph structure are randomly set to zero or replaced with mask labels in each iteration. This forces the model to utilize the contextual information of the remaining neighboring nodes, aggregate features through graph convolution operations, and predict the original values ​​of the masked nodes. This minimizes the reconstruction error as the loss function, driving the model to learn the deep interrelationships between various physical quantities through the backpropagation algorithm. In other words, the normal thermodynamic equations of the inverter are internalized in the weights of the neural network.

[0021] After model training is complete, the system enters the online inference phase, collecting inverter operating data in real time and inputting it into the model. At this time, the processor identifies the current operating condition. If it is determined to be a steady-state condition, it triggers active feature enhancement logic, controlling the cooling fan to superimpose a micro-perturbation signal of preset frequency and amplitude on the current speed reference. Simultaneously, it collects transient response data of the temperature sensor to this perturbation, calculates dynamic thermal impedance characteristics such as phase lag or amplitude-frequency response ratio, and dynamically injects these characteristics as temporary graph nodes into the current graph structure to enrich the feature dimensions under low-load conditions. If it is determined to be a high-volatility dynamic condition, it calculates the volatility index of the input excitation node within the sliding window, and uses this index to generate residuals through a monotonically increasing mapping function. The tolerance coefficient is then used to perform full or masked inference on the current or enhanced graph structure using a trained graph neural network. The consistency residual between the model reconstruction value and the sensor measurement value of key nodes such as IGBT temperature is calculated. For dynamic operating conditions, the residual tolerance coefficient is used to normalize the original residual to suppress spurious biases introduced by environmental noise. At the same time, the system keeps the Dropout layer in the model active during the inference phase and performs multiple forward propagations on the same input data to obtain the distribution variance of the reconstruction value. This variance is used as a quantitative indicator of the model's cognitive uncertainty. Finally, based on the weighted residual value and uncertainty indicator, a thermal anomaly diagnostic signal is generated when the statistical distribution deviates from the confidence interval based on normal data.

[0022] Example 1: In this example, among various application scenarios of machine learning and deep learning, a typical photovoltaic power station desert high dust accumulation scenario is selected. In this scenario, a string inverter operates for a long time in an environment with large diurnal temperature differences and continuous micro-dust deposition. Over time, a thin layer of sand and dust gradually adheres to the surface of the heat sink fins, causing a slight nonlinear increase in heat dissipation resistance. However, due to the low irradiance period in winter, the overall load rate of the inverter is low. Even though the heat dissipation performance has degraded, the absolute temperature reading of the IGBT is still far below the 90-degree Celsius threshold set by traditional over-temperature protection. At this time, the intelligent diagnostic system deployed in this invention begins to intervene. The system converts the collected multi-dimensional data stream, including voltage, current, power, and temperature at various points, into graph structure data and inputs it into a graph attention network model that has been pre-trained with unlabeled historical normal data. During the inference process, the model uses the learned normal thermodynamics... By studying coupling weights, the system attempts to reconstruct the IGBT junction temperature using current low-power input and ambient temperature data. Since the model's internal parameters are fixed under the heat transfer law of a clean heatsink, the predicted theoretical junction temperature is lower than the actual measured junction temperature that has risen due to dust accumulation. Although both are within the safe range in absolute value, the reconstruction residuals between them show a statistical deviation. The system captures this structural anomaly that violates the physical coupling law and, combined with the low volatility operating conditions at the time, confirms that the high residuals are not due to environmental interference. Without triggering any hard temperature thresholds, the system issues an early warning signal to the operation and maintenance center indicating a decrease in the cooling system's performance. After the operation and maintenance personnel conduct on-site verification and dust removal according to the guidance, the reconstruction residuals immediately return to the normal Gaussian distribution range near zero. This process vividly demonstrates how the present invention, through a self-supervised graph mask reconstruction mechanism, achieves a keen perception of the structural destruction of the correlation between physical variables at the data level.

[0023] Example 2: This example aims to verify the performance advantages of the method of the present invention compared with the traditional static threshold method and conventional supervised learning method under different operating conditions by comparing experimental data. A full-condition simulation test platform for inverters, including a programmable DC power supply, a power grid simulator, and an environmental temperature chamber, was built. A string inverter with a rated power of 100kW was selected as the test object. By artificially controlling the fan speed decay, covering the heat sink surface with standard dust nets of different thicknesses, and simulating power mutation sequences caused by cloud cover, a test set was constructed containing various typical scenarios such as normal operation, early dust accumulation, fan aging, and dynamic cloud cover. The graph mask-based autoencoder diagnostic model of the present invention was combined with traditional logic based on a fixed temperature threshold, and... A supervised classification model based on conventional LSTM is deployed in the same edge computing unit for synchronous monitoring. The traditional threshold is set at 95 degrees Celsius, while the anomaly judgment threshold of this invention is a standardized residual exceeding 3 times the standard deviation. 1000 sample points are collected for each working condition for statistical analysis. The focus is on examining the detection rate (recall) of each method under early weak faults and the false alarm rate (False Alarm Rate) under severe dynamic working conditions. The key experimental data are shown in the table below. The data shows that the detection rate of this invention reaches 98.5% under early dust accumulation conditions, far exceeding the 0% of the traditional threshold method. The false alarm rate under dynamic cloud cover conditions is only 0.2%, which is better than the 12.4% of the conventional LSTM model.

[0024] Table 1: Comparison of Diagnostic Performance of the Invention Method and Existing Technologies under Different Working Conditions Example 3: This example combines Figures 1 to 3 This section describes an intelligent diagnostic method for inverter thermal anomalies, such as... Figure 1 As shown, the process begins with the multi-dimensional operational data acquisition stage, which collects data such as IGBT junction temperature, DC voltage, DC current, and environmental nodes. It then enters the topology graph data construction stage, mapping physical quantities to nodes and defining connection edges. Next, the system performs real-time operational condition identification, distinguishing between steady-state and dynamic operating conditions. If identified as a steady-state condition, the process enters the active excitation feature enhancement branch, controlling the fan speed with micro-perturbations and extracting transient response features to inject into temporary nodes. If identified as a dynamic condition, the process enters the residual adaptive normalization branch, calculating the volatility index and generating a dynamic residual tolerance coefficient. The processing results of these two branches are fed into the Graph Neural Network (GNN) inference module to calculate the first or second reconstruction residual and uncertainty index. In the step of weighting the basic residual using the model uncertainty index, the processor executes a dynamic suppression procedure based on prediction confidence. The system calculates the final first or second reconstruction residual according to the nonlinear suppression formula. : ,in, The arithmetic mean of the reconstructed values ​​output by multiple forward inferences under random deactivation conditions is used to construct a graph neural network. To enable sensors to collect the true values ​​of physical quantities in real time, To quantify the variance of multiple forward inference results, the model's understanding of the current operating conditions is uncertain. The preset uncertainty suppression coefficient has a range of values. to The specific value depends on the inherent noise level of the sensor. The calculation logic introduces a variance term into the denominator to establish a negative correlation constraint between the residual value and the model uncertainty. When the model encounters unseen long-tail conditions or high noise inputs that cause the prediction variance to increase, it automatically reduces the current residual weight to prevent false alarms triggered by model cognitive bias. At the same time, the system includes a self-supervised model training and parameterized adjacency matrix joint optimization loop to dynamically characterize the evolution of the coupling strength between nodes and update the model weights. Finally, based on the statistical distribution characteristics and the confidence interval deviation, anomaly judgment is made, and a thermal anomaly diagnostic signal containing root cause localization information is generated.

[0025] like Figure 2 As shown, this figure depicts the trend curve of connection weights changing with operating time. The horizontal axis represents operating time in weeks, ranging from 0 to 100, and the vertical axis represents connection weights, ranging from 0.60 to 0.95. The figure contains three main trend lines, representing the connection weights of fan speed and IGBT temperature, DC voltage and AC power, and ambient temperature and capacitor temperature, respectively. These three curves exhibit a monotonically decreasing non-linear characteristic over time. For example, the weight of fan speed and IGBT temperature gradually decreases from an initial 0.85 to 0.63 in the 100th week, intuitively reflecting the invention's continuous optimization of the parameterized adjacency matrix to dynamically fit the physical coupling strength decay process caused by equipment aging. Figure 3 As shown in the figure, this diagram illustrates in detail the specific steps of self-supervised model training in a temporal interactive manner. It involves four main interactive objects: training data, mask trainer, graph neural network, and optimizer. The process begins with inputting historical normal data into the mask trainer. The trainer sequentially performs operations such as calculating the degree centrality index of nodes, dynamically adjusting the node masking probability, and randomly masking target nodes. The processed mask graph structure is then fed into the graph neural network. The network layer reconstructs the masked nodes using the neighbor node information and calculates the reconstruction error. Immediately afterwards, the optimizer performs backpropagation based on this error to update the weights and simultaneously updates the parameterized adjacency matrix. Finally, the training result is returned to complete the deep learning of the coupling relationship between physical quantities.

[0026] Example 4: In this implementation, in the inverter control system, the cooling fan typically uses PWM signals for open-loop or closed-loop speed regulation. When the input feature detection module of the graph neural network model identifies that the inverter is in a steady-state low-load mode with a power fluctuation rate of less than 5% for more than 5 minutes, the diagnostic logic immediately takes over the fan control, generating a sinusoidal modulation signal superimposed on the current reference duty cycle. The frequency of this signal is set to a low-frequency band between 0.1Hz and 0.5Hz, and the amplitude is controlled within ±5% to ±10% of the reference speed to ensure that it does not cause severe oscillations in IGBT temperature, thereby endangering device safety. At the same time, the data stream of the IGBT junction temperature sensor is synchronously acquired using a high-frequency sampling interrupt. Through the sliding window discrete Fourier transform (DFT) or correlation analysis algorithm, the fundamental phase difference and amplitude attenuation ratio of the temperature response signal relative to the fan speed excitation signal are calculated in real time. When the transient response characteristics of the temperature parameter to the speed micro-perturbation operation are extracted, the processor executes the sliding window discrete Fourier transform (DFT) signal processing procedure, and the system establishes a length coverage to Each disturbance cycle uses a first-in-first-out (FIFO) data buffer to synchronously store the fan speed command sequence and the IGBT junction temperature sampling sequence. During the calculation process, a Hanning window function is applied to the data in the buffer to suppress spectral leakage, targeting the fundamental frequency of the fan speed micro-perturbation. Perform single-point DFT operations to analyze the speed and temperature signals respectively. Complex vector at frequency.

[0027] The phase lag parameter is determined by calculating the difference in arguments of two complex vectors, and the amplitude-frequency response ratio parameter is determined by calculating the ratio of magnitudes of the two complex vectors. The frequency domain extraction path utilizes the orthogonality principle to accurately separate the active thermal response signal at a specific frequency from the broadband background temperature noise caused by load fluctuations, ensuring the reproducibility of feature data under low signal-to-noise ratio conditions. The phase difference physically corresponds to the lag time constant of the heat conduction path, and the amplitude ratio corresponds to the dynamic thermal resistance. These two parameters are extremely sensitive to changes in contact thermal resistance on the heat dissipation path, such as the drying of thermal grease, but cannot be observed in steady-state DC signals. After the calculated phase lag value and amplitude-frequency response ratio are quantified, they are used as two newly added virtual node features. They are connected to the fan speed node and IGBT temperature node in the original graph structure through predefined graph connection rules. The updated graph data is input into the GNN model for inference. At this time, the model no longer relies on the static temperature value, but combines the dynamic impedance spectrum characteristics of the system for comprehensive judgment, achieving a deep understanding of the health status of the heat dissipation system under steady state.

[0028] Example 5: This example illustrates how the present invention, through a joint optimization mechanism of parameterized adjacency matrices, avoids using fixed 0 / 1 binary adjacency matrices to hard-encode the connections between nodes in the initial stage of constructing a graph neural network. Instead, it defines a floating-point matrix matching the dimension of the number of nodes as learnable parameters. The initial values ​​of the elements in this matrix can be set based on prior physical knowledge, but gradient updates are fully open during subsequent training. As the inverter is put into operation, the equivalent series resistance (ESR) of internal components such as electrolytic capacitors will increase over time, and the convective heat transfer coefficient of the heat sink will change due to surface oxidation. These physical changes will lead to various monitoring variables... The coupling strength between quantities undergoes a slow structural drift. During the online operation phase, this invention, for example, uses the accumulated normal operation data from the past week that is judged to be without alarms to initiate a background fine-tuning training process once a week. With the goal of minimizing reconstruction error, the gradient descent algorithm is used to update the weights of the parameterized adjacency matrix by a small amplitude without changing the weights of the feature extraction layer of the main neural network. During the joint optimization and online update of the parameterized adjacency matrix, the system executes the maximum cumulative drift constraint procedure to distinguish between normal physical aging of equipment and structural damage caused by faults. The system stores the initial adjacency matrix, which has been factory-calibrated or manually confirmed, in non-volatile memory. In each subsequent fine-tuning iteration, for any weight element in the current adjacency matrix... The processor calculates the weights relative to the initial weights. Absolute drift amount, used to determine whether the drift amount satisfies the constraints. ,in, The preset physical aging tolerance threshold has a range of values. to If the drift exceeds the threshold, the optimizer will forcibly truncate the weight update magnitude. Within the boundary range, the procedure ensures that the model adapts to slow and finite physical parameter drift caused by heat sink surface oxidation or thermal grease drying, and refuses to adapt to drastic topology changes caused by sudden failures such as fan stoppage or complete blockage of air ducts, ensuring diagnostic robustness under long-term operation. This allows the graph topology structure inside the model to dynamically track and approximate the current real physical thermal network of the inverter. When slow parameter drift occurs due to aging, the weight distribution of the adjacency matrix will evolve smoothly to offset the baseline drift of the reconstruction residual and avoid false alarms. When sudden failures such as fan stoppage or foreign object obstruction occur, the system can generate high residuals and accurately trigger alarms because the speed of physical relationship destruction far exceeds the update rate of the adjacency matrix.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent diagnosis method for inverter thermal abnormalities, characterized in that, The method comprises the following steps: constructing topology graph data, obtaining multi-dimensional running time series data of the inverter, mapping each dimension physical quantity in the multi-dimensional running time series data as a node of the graph structure, and defining the connection edges between the nodes according to the correlation between the physical quantities; training a self-supervised model, constructing a graph neural network model, and training the graph neural network model by using a self-supervised learning strategy, wherein the self-supervised learning strategy comprises randomly masking target nodes in the graph structure in each training iteration, and minimizing the difference between the reconstructed value and the true value of the target nodes as an optimization objective, so as to drive the graph neural network model to learn the coupling feature representation between the nodes, and obtain the trained graph neural network model; performing working condition differentiation reasoning, and obtaining the running data of the inverter and identifying the running working condition in real time; when the running working condition is a steady-state working condition, performing a speed micro-disturbance operation on the heat dissipation fan of the inverter, extracting the transient response characteristics of the temperature parameters to the speed micro-disturbance operation, injecting the transient response characteristics as temporary nodes into the graph structure for the trained graph neural network model to calculate a reconstructed residual error, and defining the calculated reconstructed residual error as a first reconstructed residual error; when the running working condition is a dynamic working condition, calculating a reconstructed residual error of the current graph structure by using the trained graph neural network model, defining the reconstructed residual error as a second reconstructed residual error, calculating a fluctuation rate index of the graph nodes representing the input excitation within a preset time window, generating a dynamic residual error tolerance coefficient based on the fluctuation rate index, and normalizing the second reconstructed residual error by using the dynamic residual error tolerance coefficient; abnormality determination, generating a thermal abnormality diagnosis signal when the residual value deviates from a preset confidence interval based on the statistical distribution characteristics of the first reconstructed residual error or the normalized second reconstructed residual error.

2. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, In the step of constructing topology graph data, the connection edges between the nodes are defined by a parameterized adjacency matrix, and the parameterized adjacency matrix is a learnable parameter in the graph neural network model; The step of training a self-supervised model further comprises: jointly optimizing the parameterized adjacency matrix by using historical running data, so that the graph neural network model adaptively adjusts the values of the elements in the parameterized adjacency matrix according to the objective of minimizing the reconstruction difference, and dynamically represents the evolution of the coupling strength between the nodes; and the step of abnormality determination further comprises: monitoring the weight distribution change of the parameterized adjacency matrix, and confirming the generation of the thermal abnormality diagnosis signal when the first reconstructed residual error or the second reconstructed residual error exceeds a threshold value and the weight distribution change rate is lower than a preset drift rate threshold.

3. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, The step of extracting the transient response characteristics of the temperature parameters to the speed micro-disturbance operation specifically comprises: issuing a speed control instruction of a sine wave modulation signal superimposed with a preset frequency to the heat dissipation fan; collecting the insulated gate bipolar transistor junction temperature data of the inverter as the temperature parameters; calculating the phase lag parameter and the amplitude-frequency response ratio parameter of the junction temperature data relative to the sine wave modulation signal; and taking the phase lag parameter and the amplitude-frequency response ratio parameter as the transient response characteristics.

4. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, In the step of normalizing the second reconstruction residual output by the graph neural network model by using a dynamic residual tolerance coefficient, the normalized diagnosis score is calculated by the following formula: wherein, is the normalized second reconstruction residual, is the original reconstruction residual output by the graph neural network model, is the volatility index, is a preset adjustment coefficient.

5. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, The step of calculating the first reconstruction residual or the second reconstruction residual by using the graph neural network model specifically comprises: keeping random inactive units in the graph neural network model in an activated state; performing multiple forward inferences on the running data at the same time to obtain multiple reconstruction values; calculating the variance of the multiple reconstruction values, and defining the variance as a model uncertainty indicator; calculating the difference between the mean of the multiple reconstruction values and the true value as a basic residual, and weighting the basic residual by using the model uncertainty indicator to obtain the first reconstruction residual or the second reconstruction residual.

6. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 5, characterized in that, The step of generating the thermal anomaly diagnosis signal further comprises: comparing the model uncertainty indicator with a preset confidence threshold; when the model uncertainty indicator is higher than the confidence threshold, determining that the current working condition is a long-tail unknown working condition, and inhibiting the generation of the thermal anomaly diagnosis signal or reducing the priority of the thermal anomaly diagnosis signal.

7. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, The step of calculating the fluctuation rate indicator of the graph nodes representing the input excitation within the preset time window specifically comprises: selecting the DC side voltage node and the DC side current node as the graph nodes of the input excitation; collecting time series data of the graph nodes of the input excitation within the preset time window; calculating the ratio of the standard deviation to the mean of the time series data, and determining the ratio as the fluctuation rate indicator.

8. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, In the step of constructing the topological graph data, the nodes of the graph structure include strong correlation nodes, excitation nodes and environment nodes; the strong correlation nodes at least include the junction temperature of the insulated gate bipolar transistor, the capacitor temperature and the heat sink temperature; the excitation nodes at least include the DC side voltage, the DC side current and the AC side power; and the environment nodes at least include the internal environment temperature of the case and the fan speed feedback signal.

9. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, In the step of training the self-supervised model, the strategy of randomly masking at least one target node in the graph structure specifically comprises: calculating the degree centrality indicator of each node in the graph structure; dynamically adjusting the masking probability of each node according to the degree centrality indicator, wherein the node with a higher degree centrality indicator has a lower masking probability, so that the graph neural network model utilizes more context information of the edge node for feature reconstruction.

10. The intelligent diagnosis method for thermal abnormality of an inverter according to claim 1, characterized in that, After the step of generating the thermal anomaly diagnosis signal, a step of performing thermal anomaly root cause positioning is further included: extracting the attention weight values of each connection edge connected to the target node in the graph neural network model; comparing the attention weight values with a preset reference weight distribution; identifying the connection edge whose attention weight value changes by more than a preset positioning threshold, and marking the physical components corresponding to the neighbor nodes connected by the connection edge as the associated root cause components causing the thermal anomaly.

Images