Transformer state monitoring and data transmission method, system and device based on cloud edge collaboration and medium

Abstract

The application discloses a transformer state monitoring and data transmission method, system, device and medium based on cloud edge cooperation, belongs to the technical field of power equipment monitoring, and comprises the following steps: collecting transformer data through an edge layer; judging whether matched fault physical characteristics exist in a fault characteristic library; triggering a diagnostic rule engine to obtain a first diagnostic result when the matched fault physical characteristics exist; calling a lightweight diagnostic model to obtain a second diagnostic result; calculating the final probability of various faults according to the first diagnostic result and the second diagnostic result; determining a health state according to the final probability; determining a data transmission level according to the health state; determining data to be transmitted according to the data transmission level; predicting available bandwidth and determining a transmission mode; and transmitting the data to be transmitted from the edge layer to a cloud layer. Through the cloud edge cooperation architecture and the data hierarchical transmission mechanism, the application meets the real-time requirement of data transmission and maintains the fault diagnosis capability when communication is interrupted.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to a method, system, equipment, and medium for transformer condition monitoring and data transmission based on cloud-edge collaboration. Background Technology

[0002] Condition monitoring and fault diagnosis of power equipment are core means to ensure the safe operation of the power grid. Traditional transformer monitoring systems mainly adopt a centralized cloud processing architecture, relying on sensors to collect raw data such as vibration data, temperature data, and oil chromatography data, which are then uploaded in full to the cloud for centralized analysis via communication networks. The amount of raw data generated daily by a single 500kV transformer exceeds 2TB, far exceeding the capacity of existing communication bandwidth. Uploading all data results in more than 78% of bandwidth resources being occupied by invalid data. According to the IEEE P2030.5 standard test, the average latency of cloud analysis reaches 12 seconds, which cannot meet the millisecond-level response requirements for sudden faults such as insulation breakdown. Under extreme operating conditions, when communication is interrupted, the edge end completely loses its diagnostic capabilities. Some solutions deploy simple diagnostic rules at the edge, but the cloud only obtains binary results and cannot grasp the intermediate state information of the equipment. If traditional deep learning models are deployed, the extracted features contain a large amount of redundant information, increasing the transmission burden. How to reduce data transmission volume while ensuring diagnostic accuracy and enabling the cloud to obtain useful information has become an urgent technical challenge. Summary of the Invention

[0003] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for transformer condition monitoring and data transmission based on cloud-edge collaboration.

[0004] Therefore, the technical problem solved by this invention is: how to solve the problem of low data transmission efficiency in existing transformer condition monitoring systems, specifically manifested as severe bandwidth constraints caused by uploading all raw data, invalid data transmission consuming a large amount of bandwidth resources, cloud-based centralized analysis response delays failing to meet real-time requirements, loss of diagnostic capabilities at the edge when communication is interrupted, and the inability of edge diagnostic results to provide intermediate device status information to the cloud for trend analysis.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a transformer condition monitoring and data transmission method based on cloud-edge collaboration, comprising, Vibration data, temperature data, and oil chromatography data of the transformer are collected through the edge layer; Determine whether the vibration data, temperature data, and oil chromatography data of the transformer match the physical characteristics of the fault in the fault feature database; When a matching fault physical characteristic is found, the diagnostic rule engine is triggered to diagnose the transformer and obtain a first diagnostic result; When no matching physical characteristics of the fault are found, a lightweight diagnostic model is invoked. The vibration data, temperature data, and oil chromatography data are input into the lightweight diagnostic model for diagnosis to obtain the model probability of each type of fault. Based on the model probability, fuzzy logic reasoning is triggered to obtain the fuzzy confidence value. The fuzzy confidence value and the model probability are weighted and calculated to obtain the final probability of each type of fault, which is used as the second diagnostic result. The health status of the transformer is determined based on the first diagnostic result or the second diagnostic result, and the health status includes normal status, attention status, early warning status and fault status. The data transmission level is determined based on the health status, and the data transmission level includes Level 1, Level 2 and Level 3; The data to be transmitted is determined based on the data transmission level. When the data transmission level is Level 1, the data to be transmitted is a health index; when the data transmission level is Level 2, the data to be transmitted is a feature vector; when the data transmission level is Level 3, the data to be transmitted is a raw data fragment. Predict the available bandwidth, and determine the transmission method based on the available bandwidth and the data transmission level; The data to be transmitted is transmitted from the edge layer to the cloud layer using the transmission method described above. The cloud layer receives and stores the data to be transmitted.

[0006] As a preferred embodiment of the cloud-edge collaborative transformer condition monitoring and data transmission method described in this invention, the step of determining whether the transformer's vibration data, temperature data, and oil chromatography data contain matching fault physical characteristics in the fault feature database includes: Determine whether the vibration data of the transformer exhibits the vibration fault characteristics stored in the fault feature database; Determine whether the temperature data of the transformer exhibits temperature fault characteristics stored in the fault characteristic database; Determine whether the oil chromatographic data of the transformer exhibits the oil chromatographic fault characteristics stored in the fault characteristic database; When the vibration data, temperature data, or oil chromatography data of the transformer exhibit corresponding fault characteristics, it is determined that there are matching physical fault characteristics.

[0007] As a preferred embodiment of the cloud-edge collaborative transformer status monitoring and data transmission method described in this invention, wherein: the trigger diagnostic rule engine diagnoses the transformer to obtain a first diagnostic result, including: When the oil temperature in the transformer's temperature data exceeds a preset temperature threshold, an overload fault is determined to exist. When the vibration spectrum in the vibration data of the transformer shows a peak at a preset frequency and the total harmonic distortion rate exceeds a preset distortion rate threshold, it is determined that there is a winding loosening fault. The results of the overload fault determination and the winding loose fault determination are used as the first diagnostic results.

[0008] As a preferred embodiment of the cloud-edge collaborative transformer condition monitoring and data transmission method described in this invention, the lightweight diagnostic model includes: The input layer is used to receive the fused features; The key feature extraction layer includes three convolutional units and three feature enhancement units; The feature enhancement unit adopts a multi-branch feature aggregation structure, in which each branch passes through a 1×1 convolution kernel and convolution kernels of different sizes in sequence to complete dimensionality reduction and feature extraction. The feature enhancement unit includes two channels. The first channel is sequentially processed by a 1×3 convolution and a 3×1 convolution, and the second channel is sequentially processed by a 3×1 convolution and a 1×3 convolution. The feature enhancement unit performs batch normalization at the top of each convolutional layer; The feature enhancement unit includes a residual connection branch that incorporates an attention mechanism; The last convolutional kernel of each branch of the feature enhancement unit uses dilated convolution with different dilation coefficients; The Dropout layer is used to randomly deactivate the features output by the key feature extraction layer. A global average pooling layer is used to convert the features output by the Dropout layer into a fixed-length feature vector; The output layer is used to output the probability distribution of various faults from the fixed-length feature vector through the softmax activation function.

[0009] As a preferred embodiment of the cloud-edge collaborative transformer condition monitoring and data transmission method described in this invention, the step of calling the lightweight diagnostic model involves inputting the vibration data, temperature data, and oil chromatography data into the lightweight diagnostic model for diagnosis, obtaining the model probabilities of various faults, including: Time-domain and frequency-domain features are extracted from the vibration data of the transformer; Extract the temperature change rate feature from the temperature data of the transformer; Gas concentration characteristics and gas concentration ratio characteristics were extracted from the oil chromatography data of the transformer. Time-domain features, frequency domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are time-aligned and normalized. The time-aligned and normalized time-domain features, frequency-domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are fused to generate the fused features. The fused features are input into the input layer; The fused features are enhanced by the feature enhancement unit in the key feature extraction layer to obtain enhanced features; The enhanced features are randomly deactivated through the Dropout layer to obtain the deactivated features. The deactivated features are converted into fixed-length feature vectors through the global average pooling layer. The output layer performs a softmax operation on the fixed-length feature vector to output the probability distribution of various faults. The probability distribution of the various types of faults is used as the probability of the various types of faults included in the second diagnostic result.

[0010] As a preferred embodiment of the cloud-edge collaborative transformer condition monitoring and data transmission method described in this invention, the predicted available bandwidth is expressed as follows: in, This represents the predicted available bandwidth. Represents the smoothing factor. Indicates the current bandwidth. Indicates the historical bandwidth benchmark. This represents the bandwidth of the l-th sampling point. This represents the number of sampling points in the sliding window, where l represents the sampling point number in the sliding window.

[0011] As a preferred embodiment of the cloud-edge collaborative transformer status monitoring and data transmission method of the present invention, wherein: determining the health status of the transformer based on the first diagnostic result or the second diagnostic result includes: When the final probability of all types of faults is less than the first probability threshold, the health status is determined to be a normal state. When the final probability of any type of fault is greater than or equal to the first probability threshold and less than the second probability threshold, the health status is determined to be a state of concern, and the sampling rate of the transformer's vibration data and the sampling rate of the transformer's temperature data are increased. When the final probability of any type of fault is greater than or equal to the second probability threshold and less than the third probability threshold, the health status is determined to be a warning status, and a local audible and visual alarm is triggered. When the final probability of multiple types of faults is greater than or equal to the second probability threshold, or the final probability of any type of fault is greater than or equal to the third probability threshold, the health status is determined to be a fault status, and an emergency shutdown signal is triggered.

[0012] This invention provides a transformer condition monitoring and data transmission system based on cloud-edge collaboration.

[0013] To address the aforementioned technical problems, this invention provides the following technical solution: a transformer condition monitoring and data transmission system based on cloud-edge collaboration, comprising: an edge layer for collecting vibration data, temperature data, and oil chromatography data of the transformer, and performing diagnosis through a fault feature library, a diagnostic rule engine, and a lightweight diagnostic model; The communication layer is used to determine the transmission method based on available bandwidth and data transmission level. The cloud layer is used to receive and store the data.

[0014] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the cloud-edge collaborative transformer status monitoring and data transmission method.

[0015] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned cloud-edge collaborative transformer condition monitoring and data transmission method.

[0016] The beneficial effects of this invention are as follows: By constructing a knowledge base at the edge layer that includes a fault feature library, a diagnostic rule engine, and a lightweight diagnostic model, and combining this with a dual-path collaborative diagnostic mechanism to weighted calculate the probabilities of various faults, and dynamically determining the data transmission level based on the health status, the goal of reducing data transmission volume by more than 80% is achieved. Under normal conditions, only the health index calculated based on the final probabilities of various faults is transmitted; under conditions of concern, feature vectors extracted by the lightweight model are transmitted; and under conditions of warning or fault, original data fragments are transmitted, effectively alleviating bandwidth pressure. The edge layer diagnostic response time is reduced to less than 50 milliseconds, meeting real-time requirements and maintaining fault diagnosis capabilities during communication interruptions. The lightweight diagnostic model effectively extracts key features and filters redundant information through a multi-branch feature aggregation structure, spatially separated convolutions of two channels, residual connection branches combined with an attention mechanism, and dilated convolutions with different dilation coefficients, enabling the cloud to obtain useful status information while reducing transmission volume. The use of an exponentially weighted moving average method to predict available bandwidth and dynamically select the transmission mode allows the transformer condition monitoring and data transmission method to adapt to network fluctuations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a structural diagram of a lightweight diagnostic model for a cloud-edge collaborative transformer condition monitoring and data transmission method provided in one embodiment of the present invention. Detailed Implementation

[0019] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for transformer condition monitoring and data transmission based on cloud-edge collaboration, including: Step 1: Collect vibration data, temperature data, and oil chromatography data of the transformer through the edge layer; Step 2: Determine whether the vibration data, temperature data, and oil chromatography data of the transformer match the physical characteristics of the fault in the fault feature database; Step 3: When a matching fault physical characteristic is found, the diagnostic rule engine is triggered to diagnose the transformer and obtain the first diagnostic result; Step 4: When no matching physical characteristics of the fault exist, the lightweight diagnostic model is invoked. The vibration data, temperature data, and oil chromatography data are input into the lightweight diagnostic model for diagnosis to obtain the model probability of each type of fault. Based on the model probability, fuzzy logic reasoning is triggered to obtain the fuzzy confidence value. The fuzzy confidence value and the model probability are weighted and calculated to obtain the final probability of each type of fault, which is used as the second diagnostic result. Step 5: Determine the health status of the transformer based on the first diagnostic result or the second diagnostic result. The health status includes normal status, attention status, early warning status, and fault status. Step 6: Determine the data transmission level based on the health status, whereby the data transmission levels include Level 1, Level 2, and Level 3; Step 7: Determine the data to be transmitted based on the data transmission level. When the data transmission level is Level 1, the data to be transmitted is a health index; when the data transmission level is Level 2, the data to be transmitted is a feature vector; when the data transmission level is Level 3, the data to be transmitted is a raw data fragment. Step 8: Predict the available bandwidth, and determine the transmission method based on the available bandwidth and the data transmission level; Step 9: Transmit the data to be transmitted from the edge layer to the cloud layer using the transmission method described above; Step 10: The cloud layer receives and stores the data to be transmitted.

[0021] It should be noted that the method in this embodiment is applicable to online fault diagnosis of oil-immersed transformers and is deployed at the substation edge gateway. It can collect and process three core data types of transformers in real time: vibration, temperature, and oil chromatography. First, based on industry guidelines, a structured fault feature library containing "fault type - multi-dimensional feature thresholds" is constructed. At the same time, the collected data is preprocessed, including removing power frequency interference from vibration signals, eliminating outliers in temperature data, and correcting oil chromatography feature parameters, providing standardized input for subsequent diagnosis.

[0022] The diagnostic process employs a hierarchical decision-making logic: preprocessed data is matched against a fault feature library. If the vibration, temperature, and oil chromatography characteristics of a certain type of fault fully meet the threshold conditions, the diagnostic rule engine is triggered, outputting a first diagnostic result based on the equipment's physical operation logic rules. This result, based on explicit domain knowledge, possesses 100% diagnostic confidence. If the data does not fully match any fault type in the feature library, the process enters the lightweight model and fuzzy inference collaborative diagnosis stage. First, the lightweight model with grouped dilated convolution outputs the model probabilities of various faults. Then, based on the degree to which the features deviate from the threshold, fuzzy logic inference is constructed to calculate the fuzzy confidence value. After weighted fusion, a second diagnostic result is generated, clarifying the fault type, final probability, and handling recommendations.

[0023] Specifically, the edge layer includes sensors and an edge gateway. The sensors are used to collect vibration data, temperature data, and oil chromatography data of the transformer, and then upload the collected data to the edge gateway.

[0024] Vibration data is acquired using accelerometers with a sampling rate of 1 kHz to 10 kHz. The accelerometers are installed at key locations on the transformer body to monitor vibration signals in real time during transformer operation.

[0025] Temperature data is acquired via thermocouples or infrared sensors at a sampling rate of 1 Hz. Thermocouples or infrared sensors are used to monitor the oil temperature of the transformer and temperature changes in other critical components.

[0026] Oil chromatographic data were obtained through dissolved gas analysis (DGA), with sampling frequency of daily or weekly. DGA is used to determine the concentrations of dissolved gases such as hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6) in transformer oil.

[0027] The edge gateway receives vibration data, temperature data, and oil chromatography data uploaded by sensors, and analyzes and processes the data.

[0028] Step 2 involves determining whether the transformer's vibration data, temperature data, and oil chromatography data match any fault physical characteristics in the fault feature database, including: Determine whether the vibration data of the transformer exhibits the vibration fault characteristics stored in the fault feature database; Determine whether the temperature data of the transformer exhibits temperature fault characteristics stored in the fault characteristic database; Determine whether the oil chromatographic data of the transformer exhibits the oil chromatographic fault characteristics stored in the fault characteristic database; When the vibration data, temperature data, or oil chromatography data of the transformer exhibit corresponding fault characteristics, it is determined that there are matching physical fault characteristics.

[0029] Specifically, the fault feature database stores the physical characteristics of common transformer faults.

[0030] The vibration fault features stored in the fault feature library include: vibration spectrum shift features in specific frequency bands, such as a sudden increase in amplitude in the 650Hz to 1.2kHz frequency band; and abnormal vibration energy ratio features at specific frequencies, such as vibration energy ratios at 100Hz and 200Hz exceeding the baseline value by more than 30%.

[0031] The fault feature library stores temperature fault features including combinations of oil temperature and load rate, such as oil temperature exceeding 85°C and transformer load rate exceeding 90%.

[0032] The fault features stored in the fault feature library include combinations of specific gas concentrations, such as acetylene (C2H2) concentration greater than 5 μL / L and hydrogen (H2) concentration greater than 150 μL / L in oil chromatographic data.

[0033] The edge gateway compares the collected transformer vibration data, temperature data, and oil chromatography data with the vibration fault features, temperature fault features, and oil chromatography fault features stored in the fault feature database to determine whether the corresponding physical fault characteristics are present. When the transformer's vibration data, temperature data, or oil chromatography data show corresponding fault characteristics, it is determined that a matching physical fault characteristic exists.

[0034] In step 3, the diagnostic rule engine is triggered to diagnose the transformer and obtain the first diagnostic result, including: When the oil temperature in the transformer's temperature data exceeds a preset temperature threshold, an overload fault is determined to exist. When the vibration spectrum in the vibration data of the transformer shows a peak at a preset frequency and the total harmonic distortion rate exceeds a preset distortion rate threshold, it is determined that there is a winding loosening fault. The results of the overload fault determination and the winding loose fault determination are used as the first diagnostic results.

[0035] Specifically, the diagnostic rule engine diagnoses the transformer based on industry-standard decision tree logic. When step 2 determines that a matching fault physical characteristic exists, the diagnostic rule engine is triggered. Industry standards can refer to IEC60599.

[0036] When an overload fault is detected, the diagnostic rule engine extracts the oil temperature from the transformer's temperature data and determines whether the oil temperature exceeds a preset temperature threshold. In one embodiment, the preset temperature threshold is 85°C. Furthermore, the diagnostic rule engine also obtains the transformer's load rate and determines whether the transformer's load rate exceeds a preset load rate threshold. In one embodiment, the preset load rate threshold is 90%. When both the oil temperature and the load rate exceed the preset load rate threshold, the diagnostic rule engine determines that the transformer has an overload fault, and the overload fault determination result is an overload risk.

[0037] When a winding loosening fault is detected, the diagnostic rule engine performs spectral analysis on the transformer's vibration data to extract vibration spectrum information. It determines whether a peak appears in the vibration spectrum at a preset frequency and calculates the total harmonic distortion (THD). In one embodiment, the preset frequency is 650Hz. In another embodiment, referring to the benchmark requirements for transformer vibration harmonics in "GB / T1094.10-2003 Power Transformers Part 10: Sound Level Measurement" and "Southern Power Grid Transformer Condition-Based Maintenance Regulations (Q / CSG1201007-2018)," the THD benchmark threshold for 35-500kV oil-immersed transformers is 4%-6%, and the preset distortion threshold is 5%, which is a commonly used fault determination threshold. When a peak appears in the vibration spectrum at the preset frequency and the THD exceeds the preset distortion threshold, the diagnostic rule engine determines that the transformer has a winding loosening fault.

[0038] The diagnostic rule engine outputs the results of overload fault determination and winding loose fault determination as the first diagnostic result. The first diagnostic result includes the determination of the fault type, such as overload fault or winding loose fault.

[0039] In step 4, the lightweight diagnostic model includes: The input layer is used to receive the fused features; The key feature extraction layer includes three convolutional units and three feature enhancement units; The feature enhancement unit adopts a multi-branch feature aggregation structure, in which each branch passes through a 1×1 convolution kernel and convolution kernels of different sizes in sequence to complete dimensionality reduction and feature extraction. The feature enhancement unit includes two channels. The first channel is sequentially processed by a 1×3 convolution and a 3×1 convolution, and the second channel is sequentially processed by a 3×1 convolution and a 1×3 convolution. The feature enhancement unit performs batch normalization at the top of each convolutional layer; The feature enhancement unit includes a residual connection branch that incorporates an attention mechanism; The last convolutional kernel of each branch of the feature enhancement unit uses dilated convolution with different dilation coefficients; The Dropout layer is used to randomly deactivate the features output by the key feature extraction layer. A global average pooling layer is used to convert the features output by the Dropout layer into a fixed-length feature vector; The output layer is used to output the probability distribution of various faults from the fixed-length feature vector through the softmax activation function.

[0040] Specifically, the lightweight diagnostic model adopts a five-layer architecture, consisting of an input layer, a key feature extraction layer, a Dropout layer, a global average pooling layer, and an output layer.

[0041] The input layer is used to receive fused features. In one embodiment, the input layer receives 128-dimensional fused features, of which the first 64 dimensions are vibration data features, including time-domain peak values ​​(1-8 dimensions), kurtosis (9-16 dimensions), RMS (17-24 dimensions), frequency-domain FFT spectral principal components (25-48 dimensions), and wavelet packet energy (49-64 dimensions); the middle 32 dimensions are temperature data features, including hourly averages (65-80 dimensions) and 24-hour temperature gradients (81-96 dimensions); and the last 32 dimensions are oil chromatography data features, including the concentrations of hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6) (97-101 dimensions), IEC three ratios (102-116 dimensions), and Rogers ratios (117-128 dimensions).

[0042] The key feature extraction layer includes convolutional unit 1, feature enhancement unit 1, convolutional unit 2, feature enhancement unit 2, convolutional unit 3, and feature enhancement unit 3. For example... Figure 1As shown, convolutional units and feature enhancement units are connected alternately. The fused features are fed from the input layer to convolutional unit 1 of the key feature extraction layer. The output of convolutional unit 1, after batch normalization and Leaky ReLU activation, is fed to feature enhancement unit 1. The output of feature enhancement unit 1 is superimposed with the output of the residual connection branch and fed to convolutional unit 2. The output of convolutional unit 2, after batch normalization and Leaky ReLU activation, is fed to feature enhancement unit 2. The output of feature enhancement unit 2, after superimposing with the output of the residual connection branch, is fed to convolutional unit 3. The output of convolutional unit 3, after batch normalization and Leaky ReLU activation, is fed to feature enhancement unit 3. The output of feature enhancement unit 3 serves as the output of the key feature extraction layer and is fed to the Dropout layer.

[0043] Convolutional unit 1 contains 32 3×3 dilated convolutional kernels with a dilation rate r in the height direction. h =1, void ratio r in the width direction w =1. Hollow-direction void ratio r of convolutional unit 2. h =2, void ratio r in the width direction w =2. Hollow-direction void ratio r of convolutional unit 3. h =3, void ratio r in the width direction w =3.

[0044] The convolutional unit uses dilated convolution, which effectively expands the receptive field, reduces feature information loss, and has lower computational cost, facilitating lightweight model deployment. The formula for dilated convolution is: in, For the output feature map in The value of the position, The corresponding height direction of the output feature map, This corresponds to the width direction of the output feature map. It involves double summation at all positions of the convolution kernel. K h The height of the convolution kernel. K w The width of the convolution kernel. It is the position index of the convolution kernel in the height direction. It is the position index of the convolution kernel in the width direction. The void ratio in the height direction. The void ratio is in the width direction. The convolutional kernel is learned during model training. Position weight, Given the input feature map, =1,2,..., , =1,2,..., , Values ​​= 1, 2, ... , Values ​​== 1, 2, ..., .

[0045] Batch normalization is applied to the output of each convolutional unit to achieve a uniform data distribution and avoid gradient vanishing. The batch normalization process includes: calculating the mean μ and variance σ² of the mini-batch data; standardizing the input; and applying scaling and bias, including the following steps: Standardize the input: In the formula, To prevent small constants from being divided by zero, The input data is the output of each convolutional unit. , This represents the mean of the current small batch of data. The variance of the current small batch of data. This is the output value after standardization.

[0046] Apply scaling and bias: In the formula, Scaling parameters for training and learning, To train and learn bias parameters, This is the output value after scaling and biasing.

[0047] The activation function used is the LeakyReLU function. The LeakyReLU function effectively solves the problem of gradient not updating when the neuron's input value is negative. It outputs the entire value when the input value is positive, and when negative input values ​​occur during backpropagation, it retains some negative information by setting smaller weights. This negative information is then used to calculate the gradient update information, effectively eliminating the jagged gradient direction phenomenon. The formula for calculating the LeakyReLU function is: In the formula, This is the output of the LeakyReLU function neuron after activation. The input value to the neuron, i.e., the output value after scaling and biasing. , For a negative slope, if is only a conditional statement.

[0048] Feature enhancement units such as Figure 1As shown, it includes a batch normalization module, an activation function, and a residual module. Feature enhancement units 1, 2, and 3 have the same structure. The feature enhancement units enrich semantic information by using multi-branch feature aggregation. Each branch first undergoes dimensionality reduction through a 1×1 convolution kernel, and then uses convolution kernels of different sizes to extract features. To reduce computation, the 5×5 convolution kernels in the branches are replaced with two cascaded 3×3 convolutions. Spatial separation convolution is also used, with two channels. The first channel passes through a 1×3 convolution followed by a 3×1 convolution, and the second channel passes through a 3×1 convolution followed by a 1×3 convolution.

[0049] The feature enhancement unit performs batch normalization at the top of each convolutional layer to accelerate model convergence. The feature enhancement unit includes residual connection branches incorporating attention mechanisms to improve diagnostic performance and effectively extract key information from fused features. The last convolutional kernel of each branch employs dilated convolutions with different dilation coefficients, widening the receptive field while maintaining a constant number of parameters. This adapts to the lightweight deployment requirements of edge gateways, ensuring that feature extraction diversity is increased while controlling computational load during transformer state data feature extraction, thus guaranteeing the timeliness and accuracy of edge layer diagnostics.

[0050] The residual module employs a deep residual shrinking network, which integrates a deep residual network, an attention mechanism, and a soft thresholding function. The residual network addresses the training difficulty of deep networks by introducing a residual learning framework, enabling the network to facilitate information transfer and flow by learning the residual relationship between input and output. The soft thresholding function, as an effective nonlinear mapping method, demonstrates excellent performance in signal denoising and sparse representation. The attention mechanism allows the network to focus on key features in the input data, thereby improving the model's resistance to noise and outliers. The residual module effectively extracts useful information from the fused features.

[0051] The formula for calculating deep residual shrinkage networks is: In the formula, Given the input feature map, The output of the deep residual shrinking network, This represents a three-dimensional tensor, from which features are extracted using operations such as convolution on the feature map. This is a threshold vector that depends on the feature distribution of the current batch of data. The more dispersed the feature distribution, the larger the corresponding value; the more concentrated the feature distribution, the smaller the corresponding value. This is a soft threshold function.

[0052] For each batch of training data, calculate the standard deviation of each dimension of the fusion features, and take 1.2 times the standard deviation as the threshold vector corresponding to that dimension. This allows the soft threshold function to effectively filter out noise features.

[0053] The output of the key feature extraction layer is fed into the Dropout layer. The Dropout layer is used to set the output of neurons in the key feature extraction layer to zero with a certain probability, reducing the interdependence between neurons and thus eliminating overfitting during model training. The calculation formula for the Dropout layer is as follows: During the training phase, it is represented as: in, This is the output of the Dropout layer. This is the input to the Dropout layer, i.e., the output of the neurons in the key feature extraction layer. For a with Matrices of the same shape have elements that follow a Bernoulli distribution, and ⊙ denotes element-wise multiplication.

[0054] During the testing phase, it is represented as: in, The scaling factor is the probability that a neuron is discarded (output set to 0). This is because only a few neurons are used during training. A proportion of neurons participate in the calculation, and during testing, all neurons participate, so... Scaling the input ensures that the expected output matches the training phase.

[0055] The output of the Dropout layer is fed into the global average pooling layer, which performs dimensionality reduction on the output of the Dropout layer. The formula for calculating the global average pooling layer is: in, y c This represents the output of the global draw pooling operation. The first feature map of the input Output data at position (i,j) in each channel. It is the position index of the convolution kernel in the height direction. It is the index of the convolution kernel in the width direction, and its value ranges from 1 to the width of the feature map. ; The height of the feature map, This represents the width of the feature map.

[0056] The output layer is used to map the output of the global flat pooling layer to a high-dimensional space, and then use a softmax classifier to classify the faults to obtain the probability distribution of each type of fault.

[0057] Furthermore, in step 4, the lightweight diagnostic model is invoked, and the vibration data, temperature data, and oil chromatography data are input into the lightweight diagnostic model for diagnosis, obtaining the model probabilities of various faults, including: Time-domain and frequency-domain features are extracted from the vibration data of the transformer; Extract the temperature change rate feature from the temperature data of the transformer; Gas concentration characteristics and gas concentration ratio characteristics were extracted from the oil chromatography data of the transformer. Time-domain features, frequency domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are time-aligned and normalized. The time-aligned and normalized time-domain features, frequency-domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are fused to generate the fused features. The fused features are input into the input layer; The fused features are enhanced by the feature enhancement unit in the key feature extraction layer to obtain enhanced features; The enhanced features are randomly deactivated through the Dropout layer to obtain the deactivated features. The deactivated features are converted into fixed-length feature vectors through the global average pooling layer. The output layer performs a softmax operation on the fixed-length feature vector to output the probability distribution of various faults. The probability distribution of the various types of faults is used as the probability of the various types of faults included in the second diagnostic result.

[0058] Specifically, time-domain features are extracted from the transformer vibration data, including peak value, kurtosis, and RMS. Frequency-domain features are extracted from the transformer vibration data, including FFT spectrum principal components and wavelet packet energy.

[0059] Temperature change rate features are extracted from the transformer temperature data. These features include hourly average and 24-hour temperature gradient.

[0060] Gas concentration characteristics and gas concentration ratio characteristics were extracted from the oil chromatography data of the transformer. Gas concentration characteristics included the concentrations of hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6). Gas concentration ratio characteristics were calculated using the IEC three-ratio method and the Rogers ratio method.

[0061] The extracted time-domain features, frequency-domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are fused to generate fused features. After the fused features are input to the input layer, they sequentially pass through a key feature extraction layer, a Dropout layer, a global average pooling layer, and an output layer, outputting the probability distributions of various fault types. These probability distributions serve as the probabilities of various fault types included in the second diagnostic result.

[0062] It should be noted that before calling the lightweight diagnostic model for diagnosis, preprocessing of the transformer's vibration data, temperature data, and oil chromatography data is also included.

[0063] Specifically, data preprocessing includes time alignment, normalization, outlier handling, and feature fusion.

[0064] In time alignment, vibration data, temperature data, and oil chromatography data with different sampling frequencies are aligned according to timestamps. Since the sampling rate of vibration data is 1kHz to 10kHz, the sampling rate of temperature data is 1Hz, and the sampling frequency of oil chromatography data is daily or weekly, the sampling frequencies of the three types of data are different. Time alignment is required by interpolation or downsampling to ensure that the data with different sampling frequencies are consistent in the time dimension.

[0065] In normalization, the characteristics of vibration data and temperature data are standardized. Normalization methods include Min-Max standardization or Z-Score standardization. Min-Max standardization scales the data to a specified interval, while Z-Score standardization transforms the data into a distribution with a mean of 0 and a standard deviation of 1.

[0066] In outlier handling, the 3σ rule or the Isolation Forest method is used to remove outliers caused by sensor malfunctions. The 3σ rule, based on the principle of normal distribution, considers data deviating from the mean by more than three standard deviations as outliers. Isolation Forest is a tree-based anomaly detection method that identifies outliers by randomly partitioning the feature space.

[0067] In feature fusion, fused features are constructed as inputs to a lightweight diagnostic model. Feature fusion integrates features from vibration data, temperature data, and oil chromatography data; for example, it fuses the principal components of the vibration spectrum with the rate of temperature change to generate fused features.

[0068] Furthermore, in step 4, fuzzy logic inference is triggered based on the model probability to obtain a fuzzy confidence value; the fuzzy confidence value is then weighted and calculated with the model probability to obtain the final probability of various faults, which serves as the second diagnostic result, specifically including: S401: Call the lightweight diagnostic model to obtain the model probability of various faults.

[0069] The lightweight diagnostic model outputs the probability distribution of various faults, represented as follows: ,in This represents the probability distribution of various types of faults output by the model. Representing categories 1, 2, and so on. Model probability of class of faults This represents the total number of fault types.

[0070] It should be noted that the lightweight diagnostic model is a prediction model trained on historical data. Although it can effectively capture data patterns, it still has certain uncertainties in practical applications, mainly manifested in data distribution deviation and limited generalization ability.

[0071] S402: Based on the model probability, trigger fuzzy logic reasoning to obtain the fuzzy confidence value.

[0072] Fuzzy logic reasoning is a reasoning method based on expert experience and fuzzy rules, capable of handling uncertain and fuzzy information. This invention's fuzzy logic reasoning takes model probabilities, transformer operating parameters, and sensor data features as input, performs reasoning through a fuzzy rule base, and outputs fuzzy confidence values ​​for various fault types. (Fuzzy confidence values) This represents the confidence level of the fault based on fuzzy logic reasoning, and its value range is... .

[0073] S403: The fuzzy confidence value and the model probability are weighted and calculated to obtain the final probability of each type of fault.

[0074] For a certain type of fault, its final probability The calculation formula is: in, The final probability of a certain type of failure. This represents the fuzzy confidence value for this type of fault, obtained through fuzzy logic reasoning. This refers to the model probability of this type of fault output by the lightweight diagnostic model. As the first weight, It is the second weight, and .

[0075] In this invention, the first weight Second weight The determination of the optimal weight configuration is based on a validation set tuning method using historical fault data. By testing the diagnostic accuracy of different weight combinations on historical fault data, the weight combination with the highest accuracy is selected as the preferred weight configuration.

[0076] Historical fault data of transformers were collected to construct a validation dataset. Each sample in the validation dataset contains monitoring data at the time of the fault, fuzzy confidence values ​​output by fuzzy logic inference, the output probability of the fault by the lightweight diagnostic model, and the true label of the actual fault.

[0077] Set multiple sets of weight combinations to be tested. ,satisfy In one embodiment, combination 1 is selected: Combination 2: Combination 3: Combination 4: Combination 5: .

[0078] For each sample in the validation dataset, calculate the final probability using the corresponding weight combination. The system then determines whether a sample has malfunctioned based on a fault determination threshold. For each weight combination, the diagnostic accuracy on the validation dataset is calculated.

[0079] In one embodiment, the test results for the five weight combinations are shown in Table 1: Table 1. Accuracy Test Results of Weighted Combinations

[0080] Based on the test results shown in Table 1, weight combination 3 has the highest accuracy; therefore, combination 3 is selected as the preferred weight configuration. In the preferred embodiment, the first weight... Set to 0.7, second weight Set to 0.3.

[0081] In one embodiment, the lightweight diagnostic model outputs a model probability of 0.8 for overload faults, and the fuzzy confidence value of the overload fault obtained from fuzzy logic inference is 0.7. Based on the optimized weight configuration, the final probability of the overload fault is: S404: Use the final probability of each type of fault as the second diagnostic result.

[0082] The second diagnostic result includes the final probability distribution of various faults, which is used for subsequent health status determination, data transmission level determination, and alarm triggering.

[0083] When step 2 determines that a matching physical fault characteristic exists, step 3 triggers the diagnostic rule engine to obtain the first diagnostic result. The first diagnostic result is a clear fault type determination based on industry standards. In this case, the first diagnostic result is directly used as the final diagnostic result, without the need for fuzzy logic reasoning and weighted calculation.

[0084] Step 5: Determine the health status of the transformer based on the first diagnostic result or the second diagnostic result. The health status includes normal status, attention status, early warning status, and fault status, including: When the final probability of all types of faults is less than the first probability threshold, the health status is determined to be a normal state. When the final probability of any type of fault is greater than or equal to the first probability threshold and less than the second probability threshold, the health status is determined to be a state of concern, and the sampling rate of the transformer's vibration data and the sampling rate of the transformer's temperature data are increased. When the final probability of any type of fault is greater than or equal to the second probability threshold and less than the third probability threshold, the health status is determined to be a warning status, and a local audible and visual alarm is triggered. When the final probability of multiple types of faults is greater than or equal to the second probability threshold, or the final probability of any type of fault is greater than or equal to the third probability threshold, the health status is determined to be a fault status, and an emergency shutdown signal is triggered.

[0085] In one embodiment, based on experience, the first probability threshold is 30%, the second probability threshold is 60%, and the third probability threshold is 90%. When the final probability of all faults is less than 30%, the health status is determined to be normal, and the transformer needs to continue monitoring. When the final probability of any type of fault is greater than or equal to 30% and less than 60%, the health status is determined to be a state of concern, and the sampling rates of the transformer's vibration data and temperature data are increased, for example, increasing the sampling rate of vibration data from 5kHz to 10kHz and the sampling rate of temperature data from 1min to 10s. When the final probability of any type of fault is greater than or equal to 60% and less than 90%, the health status is determined to be a warning state, triggering a local audible and visual alarm. At this time, the transformer may have a fault, and data related to this fault type needs to be closely monitored. When the final probability of multiple types of faults is greater than or equal to 60%, or the final probability of any type of fault is greater than or equal to 90%, the health status is determined to be a fault state, triggering an emergency shutdown signal, requiring the transformer to be shut down for maintenance.

[0086] In determining the data transmission level based on the health status, the communication layer uses communication technologies such as 5G / fiber optics, while the cloud layer is used to collect data uploaded from various edge layers and manage this data.

[0087] Because the amount of data generated by transformers every day is large, far exceeding the communication bandwidth capacity of the communication layer, according to actual power grid measurements, 78% of the bandwidth of this data is consumed in invalid data transmission, thus reducing the data transmission efficiency. Therefore, in order to improve the data transmission efficiency, a bandwidth-adaptive data hierarchical transmission mechanism is adopted. This mechanism achieves the goal of reducing the transmission volume by more than 80% through dynamic data hierarchical strategies and intelligent transmission scheduling algorithms.

[0088] Data transmission level is determined based on health status. When the health status is normal, the data transmission level is Level 1. When the health status is in a state of concern, the data transmission level is Level 2. When the health status is in a state of alert or fault, the data transmission level is Level 3. The types of data transmitted are shown in Table 2. Table 2 Comparison of Data Transmission Levels

[0089] Predicted available bandwidth, denoted as, in, This represents the predicted available bandwidth. This represents the smoothing factor, with a value range of 0.6 to 0.8. This represents the current bandwidth, which can be measured instantaneously using the most recent probe (such as ICMP packet round-trip time). This represents the historical bandwidth baseline, which is the moving average of bandwidth over a past period (e.g., 30 minutes). This represents the bandwidth of the l-th sampling point. This represents the number of sampling points in the sliding window, where l represents the sampling point number in the sliding window.

[0090] As shown in Table 3, select the corresponding transmission method based on the available bandwidth.

[0091] Table 3 Comparison of Transmission Methods

[0092] Fault status data has the highest priority and can preempt bandwidth for transmission, followed by warning status data, and data under attention and normal status are transmitted in queue order. When bandwidth is insufficient, historical redundant data under normal status is discarded first.

[0093] Specifically, the edge layer transmits the data to be transmitted from the edge layer to the cloud layer via a transmission method. The edge layer and the cloud layer use the HTTP / 2 communication protocol. The data transmitted by the edge layer is in JSON format and includes information such as device number, data acquisition time, health status, final probability of various faults, and the data to be transmitted.

[0094] The cloud layer receives data transmitted from the edge layer and stores it according to device coding and date rules. The cloud layer establishes a relational database to store basic device information, health status statistics, and fault records.

[0095] Therefore, by building a knowledge base, the system can maintain the diagnostic capability (response <50ms) for more than 15 types of faults (e.g., winding deformation, multi-point grounding of iron core, deterioration of insulating oil, excessive dielectric loss of bushing, stuck on-load tap changer, cooling system failure, partial discharge, arc discharge in oil, loose mechanical parts, poor contact of disconnecting switch, etc.) during network interruption; and by adopting a bandwidth adaptive data hierarchical transmission mechanism, the transmission volume is effectively reduced by more than 80%.

[0096] Example 2 is an embodiment of the present invention. This embodiment provides a transformer condition monitoring and data transmission system based on cloud-edge collaboration, including: an edge layer for collecting vibration data, temperature data and oil chromatography data of the transformer, and performing diagnosis through a fault feature library, a diagnostic rule engine and a lightweight diagnostic model; The communication layer is used to determine the transmission method based on available bandwidth and data transmission level. The cloud layer is used to receive and store the data.

[0097] This embodiment also provides an electronic device applicable to the cloud-edge collaborative transformer condition monitoring and data transmission method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the cloud-edge collaborative transformer condition monitoring and data transmission method proposed in the above embodiment.

[0098] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the cloud-edge collaborative transformer condition monitoring and data transmission method proposed in the above embodiments.

[0099] The storage medium proposed in this embodiment and the method for implementing transformer status monitoring and data transmission based on cloud-edge collaboration proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0100] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for transformer condition monitoring and data transmission based on cloud-edge collaboration, characterized in that: include, Vibration data, temperature data, and oil chromatography data of the transformer are collected through the edge layer; Determine whether the vibration data, temperature data, and oil chromatography data of the transformer match the physical characteristics of the fault in the fault feature database; When a matching fault physical characteristic is found, the diagnostic rule engine is triggered to diagnose the transformer and obtain a first diagnostic result; When no matching physical characteristics of the fault exist, the lightweight diagnostic model is invoked, and the vibration data, temperature data, and oil chromatography data are input into the lightweight diagnostic model for diagnosis to obtain the model probabilities of various faults; based on the model probabilities, fuzzy logic reasoning is triggered to obtain fuzzy confidence values; The fuzzy confidence value and the model probability are weighted and calculated to obtain the final probability of each type of fault, which is used as the second diagnostic result. The health status of the transformer is determined based on the first diagnostic result or the second diagnostic result, and the health status includes normal status, attention status, early warning status and fault status. The data transmission level is determined based on the health status, and the data transmission level includes Level 1, Level 2 and Level 3; The data to be transmitted is determined based on the data transmission level. When the data transmission level is Level 1, the data to be transmitted is a health index; when the data transmission level is Level 2, the data to be transmitted is a feature vector; when the data transmission level is Level 3, the data to be transmitted is a raw data fragment. Predict the available bandwidth, and determine the transmission method based on the available bandwidth and the data transmission level; The data to be transmitted is transmitted from the edge layer to the cloud layer using the transmission method described above. The cloud layer receives and stores the data to be transmitted.

2. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 1, characterized in that: The determination of whether the transformer's vibration data, temperature data, and oil chromatography data match the fault physical characteristics in the fault feature database includes: Determine whether the vibration data of the transformer exhibits the vibration fault characteristics stored in the fault feature database; Determine whether the temperature data of the transformer exhibits temperature fault characteristics stored in the fault characteristic database; Determine whether the oil chromatographic data of the transformer exhibits the oil chromatographic fault characteristics stored in the fault characteristic database; When the vibration data, temperature data, or oil chromatography data of the transformer exhibit corresponding fault characteristics, it is determined that there are matching physical fault characteristics.

3. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 2, characterized in that: The trigger diagnostic rule engine diagnoses the transformer and obtains a first diagnostic result, including: When the oil temperature in the transformer's temperature data exceeds a preset temperature threshold, an overload fault is determined to exist. When the vibration spectrum in the vibration data of the transformer shows a peak at a preset frequency and the total harmonic distortion rate exceeds a preset distortion rate threshold, it is determined that there is a winding loosening fault. The results of the overload fault determination and the winding loose fault determination are used as the first diagnostic results.

4. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 3, characterized in that: The lightweight diagnostic model includes: The input layer is used to receive the fused features; The key feature extraction layer includes three convolutional units and three feature enhancement units; The feature enhancement unit adopts a multi-branch feature aggregation structure, in which each branch passes through a 1×1 convolution kernel and convolution kernels of different sizes in sequence to complete dimensionality reduction and feature extraction. The feature enhancement unit includes two channels. The first channel is sequentially processed by a 1×3 convolution and a 3×1 convolution, and the second channel is sequentially processed by a 3×1 convolution and a 1×3 convolution. The feature enhancement unit performs batch normalization at the top of each convolutional layer; The feature enhancement unit includes a residual connection branch that incorporates an attention mechanism; The last convolutional kernel of each branch of the feature enhancement unit uses dilated convolution with different dilation coefficients; The Dropout layer is used to randomly deactivate the features output by the key feature extraction layer. A global average pooling layer is used to convert the features output by the Dropout layer into a fixed-length feature vector; The output layer is used to output the probability distribution of various faults from the fixed-length feature vector through the softmax activation function.

5. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 4, characterized in that: The lightweight diagnostic model is invoked by inputting the vibration data, temperature data, and oil chromatography data into the model for diagnosis, obtaining the model probabilities of various faults, including: Time-domain and frequency-domain features are extracted from the vibration data of the transformer; Extract the temperature change rate feature from the temperature data of the transformer; Gas concentration characteristics and gas concentration ratio characteristics were extracted from the oil chromatography data of the transformer. Time-domain features, frequency domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are time-aligned and normalized. The time-aligned and normalized time-domain features, frequency-domain features, temperature change rate features, gas concentration features, and gas concentration ratio features are fused to generate the fused features. The fused features are input into the input layer; The fused features are enhanced by the feature enhancement unit in the key feature extraction layer to obtain enhanced features; The enhanced features are randomly deactivated through the Dropout layer to obtain the deactivated features. The deactivated features are converted into fixed-length feature vectors through the global average pooling layer. The output layer performs a softmax operation on the fixed-length feature vector to output the probability distribution of various faults. The probability distribution of the various types of faults is used as the probability of the various types of faults included in the second diagnostic result.

6. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 5, characterized in that: The predicted available bandwidth is expressed as follows: in, This represents the predicted available bandwidth. Represents the smoothing factor. Indicates the current bandwidth. Indicates the historical bandwidth benchmark. This represents the bandwidth of the l-th sampling point. This represents the total number of sampling points in the sliding window, and l represents the sampling point number in the sliding window.

7. The transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in claim 6, characterized in that: Determining the health status of the transformer based on the first diagnostic result or the second diagnostic result includes: When the final probability of all types of faults is less than the first probability threshold, the health status is determined to be a normal state. When the final probability of any type of fault is greater than or equal to the first probability threshold and less than the second probability threshold, the health status is determined to be a state of concern, and the sampling rate of the transformer's vibration data and the sampling rate of the transformer's temperature data are increased. When the final probability of any type of fault is greater than or equal to the second probability threshold and less than the third probability threshold, the health status is determined to be a warning status, and a local audible and visual alarm is triggered. When the final probability of multiple types of faults is greater than or equal to the second probability threshold, or the final probability of any type of fault is greater than or equal to the third probability threshold, the health status is determined to be a fault status, and an emergency shutdown signal is triggered.

8. A transformer condition monitoring and data transmission system based on cloud-edge collaboration, employing the transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in any one of claims 1 to 7, characterized in that, include: The edge layer is used to collect vibration data, temperature data, and oil chromatography data of the transformer, determine whether the vibration data, temperature data, and oil chromatography data of the transformer have matching physical fault features in the fault feature library, trigger the diagnostic rule engine to obtain a first diagnostic result when a matching physical fault feature is found, call the lightweight diagnostic model to obtain a second diagnostic result, obtain the final probability of various types of faults based on the first diagnostic result and the second diagnostic result, and determine the health status and data transmission level based on the final probability of various types of faults. The communication layer is used to determine the data to be transmitted based on the data transmission level, predict the available bandwidth, and determine the transmission method based on the available bandwidth and the data transmission level. The cloud layer is used to receive and store the data to be transmitted.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the transformer condition monitoring and data transmission method based on cloud-edge collaboration as described in any one of claims 1 to 7.

Images