Transformer cooling fan performance prediction method based on hierarchical residual network

Abstract

The present invention discloses a method for predicting the performance of a transformer cooling fan based on a hierarchical residual network, which relates to the field of equipment monitoring technology. The method comprises: obtaining time series data of the cooling fan under various degrees of performance degradation and various operating states through multiple sensors, and training a hierarchical attention residual time series network, wherein the input and output ends of the multiple time feature extraction units in the network correspond to the output end of the sensor attention module and the input end of the time attention module, respectively, and the input ends of the performance degradation prediction unit and the fault prediction unit are connected to the output end of the time attention module through a shared fully connected unit; and using the trained hierarchical attention residual time series network to obtain performance degradation prediction results and fault type prediction results based on the time series data to be diagnosed. This method can simultaneously achieve accurate prediction of the future performance of the cooling fan and accurate evaluation of the current state within a unified framework.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring, and in particular to a transformer cooling fan performance prediction method based on a hierarchical residual network. Background Art

[0002] Transformers are core components of power systems, and their operational reliability is directly linked to the safety and stability of the entire power grid. The cooling system is a critical component for ensuring the proper operation of transformers under rated load. Cooling fans, through forced air cooling, effectively control the temperature of the transformer windings and core, preventing overheating that can lead to insulation degradation, accelerated aging, and even serious failures. Therefore, effectively monitoring the cooling fan's operating status and predicting its performance degradation trends are crucial for implementing predictive maintenance, avoiding unplanned transformer outages, and ensuring a reliable power supply.

[0003] Currently, maintenance strategies for auxiliary equipment such as transformer cooling fans primarily rely on regular inspections or post-fault repairs. With the development of data-driven technologies, methods based on vibration analysis, current signal analysis, or temperature monitoring are being used for condition assessment. However, existing technologies still have the following limitations in achieving accurate long-term performance predictions:

[0004] 1) Inadequate Capture of Temporal Dependencies: Traditional data analysis methods or shallow machine learning models struggle to effectively capture the complex dynamic dependencies across different time scales in cooling fan operating data. For example, subtle vibration signature changes that signal early performance degradation may not be effectively correlated with long-term operating current and temperature drift trends, thus missing the optimal window for implementing early predictive maintenance.

[0005] 2) Single-source predictive information: Existing methods typically focus on the current assessment of "whether the state is abnormal," while treating "how future performance will evolve" as a separate task or even ignoring the latter. This separation fails to provide a coherent and comprehensive predictive view from the current state to future trends, limiting the foresight of maintenance decisions.

[0006] 3) Reliance on Complex Feature Engineering: Many data-driven methods rely heavily on domain experts to perform complex manual feature engineering to extract statistical features from raw sensor data. This process is not only time-consuming and labor-intensive, but the quality of the extracted features directly determines model performance, making it difficult to capture hidden, nonlinear performance degradation patterns in the data. Summary of the Invention

[0007] The present invention proposes a transformer cooling fan performance prediction method based on a hierarchical residual network to simultaneously achieve high-precision performance degradation prediction and accurate operating status evaluation.

[0008] An embodiment of the present invention proposes a transformer cooling fan performance prediction method based on a hierarchical residual network. Sensors are provided at multiple locations of the cooling fan. The method includes: obtaining time series data of the cooling fan under multiple performance degradation degrees and multiple operating states through multiple sensors; constructing a hierarchical attention residual time series network, and training the hierarchical attention residual time series network based on the time series data. The network includes: a sensor attention module, a time feature extraction module, a time attention module and a multi-task output module. The time feature extraction module includes multiple time feature extraction units. The multi-task output module includes a shared fully connected unit, a performance degradation prediction unit and a fault prediction unit. The input and output ends of the multiple time feature extraction units are respectively connected to the output end of the sensor attention module and the input end of the time attention module. The input ends of the performance degradation prediction unit and the fault prediction unit are connected to the output end of the time attention module through the shared fully connected unit. The trained hierarchical attention residual time series network is used to obtain performance degradation prediction results and fault type prediction results according to the time series data to be diagnosed.

[0009] In some embodiments, the sensor attention module outputs weighted features via the following formula :

[0010]

[0011]

[0012]

[0013] in, represents the attention score; , , respectively represent the learnable weights and biases of the sensor attention module; , represents the sample input to the sensor attention module; represents the time step, Indicates the number of sensor channels; represents the attention weight, ⊙ represents element-wise multiplication, Represents the exponential function.

[0014] In some embodiments, the i-th time feature extraction unit extracts the time feature by the following formula: :

[0015]

[0016]

[0017]

[0018] in, represents a one-dimensional convolutional neural network, 、 represents the convolution kernel weight used by the i-th time feature extraction unit, represents the convolution kernel size used by the i-th time feature extraction unit, 、 represents the learnable bias of the i-th temporal feature extraction unit, Indicates the number of output channels, represents the time step after convolution processing, 、 Represent the two intermediate features of the i-th time feature extraction unit, Represents 1×1 convolution, i is a positive integer less than or equal to K, and K represents the number of time feature extraction units.

[0019] In some embodiments, the temporal attention module outputs a context vector via the following formula: :

[0020]

[0021]

[0022]

[0023]

[0024] in, represents the splicing function, K represents the number of time feature extraction units, represents the time step attention score; , , respectively represent the learnable weights and biases of the temporal attention module; represents the time step weight, express No. time-step features.

[0025] In some embodiments, the shared fully connected unit performs nonlinear feature transformation by the following formula:

[0026]

[0027] in, represents the output of the shared fully connected unit, represents the linear rectification function, represents the hidden layer dimension; , , respectively represent the learnable weights and biases of the shared fully connected unit.

[0028] In some embodiments, the performance degradation prediction unit outputs the performance degradation prediction result through the following formula:

[0029]

[0030] in, , , respectively represent the learnable weights and biases of the performance degradation prediction unit; represents the performance degradation prediction result, Represents an S-shaped growth curve function.

[0031] In some embodiments, the fault prediction unit outputs a fault type prediction result using the following formula:

[0032]

[0033] in, , , respectively represent the learnable weight and bias of the fault prediction unit; Indicates the number of fault types, is the original logical value of each fault type.

[0034] In some embodiments, the loss function used to train the layered attention residual temporal network is It is obtained by the following formula:

[0035]

[0036]

[0037]

[0038] in, 、 Represent regression loss and classification loss respectively, 、 represents the hyperparameter, 、 They represent the actual degree of performance degradation and the predicted results of performance degradation respectively. Indicates the real fault type label, Represents the fault type prediction result of the i-th sample.

[0039] In some embodiments, the method further includes: generating a diagnostic report on the health status of the cooling fan according to the performance degradation prediction result and the fault type prediction result.

[0040] In some embodiments, the multiple operating conditions include: normal operation, reduced condenser water flow, condenser blockage, reduced evaporator water flow, excessive oil, presence of non-condensable gases, refrigerant leakage, refrigerant overcharge and expansion valve failure.

[0041] The transformer cooling fan performance prediction method based on a hierarchical residual network in an embodiment of the present invention uses multiple sensors to obtain time series data of the cooling fan under various performance degradation levels and operating conditions. A hierarchical attention residual time series network is constructed and trained based on the time series data. The network includes a sensor attention module, multiple time feature extraction units, a time attention module, and a multi-task output module. The time feature extraction module includes a shared fully connected unit, a performance degradation prediction unit, and a fault prediction unit. The inputs and outputs of the multiple time feature extraction units are respectively connected to the outputs of the sensor attention module and the time attention module. The inputs of the performance degradation prediction unit and the fault prediction unit are connected to the output of the time attention module via a shared fully connected unit. The trained hierarchical attention residual time series network is used to obtain performance degradation prediction results and fault type prediction results based on the time series data to be diagnosed. This method can automatically learn multi-scale dynamic features from complex sensor time series data and simultaneously achieve accurate prediction of the cooling fan's future performance and accurate assessment of its current status within a unified framework.

[0042] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a structural schematic diagram of a cooling fan according to an example of the present invention;

[0044] Figure 2 Flowchart of a method for predicting transformer cooling fan performance based on a hierarchical residual network according to an embodiment of the present invention;

[0045] Figure 3 1 is a schematic diagram of the structure of a hierarchical attention residual temporal network according to an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of input and output of a hierarchical attention residual temporal network according to an embodiment of the present invention;

[0047] Figure 5 is a schematic diagram of the relationship between the actual performance degradation level and the predicted performance degradation level of an example of the present invention;

[0048] Figure 6is a violin plot of performance degradation prediction error under various fault types according to an example of the present invention;

[0049] Figure 7 Schematic diagram of a confusion matrix for fault type prediction according to an example of the present invention. DETAILED DESCRIPTION

[0050] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0051] The following describes a transformer cooling fan performance prediction method based on a hierarchical residual network according to an embodiment of the present invention with reference to the accompanying drawings.

[0052] In an embodiment of the present invention, sensors are installed at multiple locations on the cooling fan to collect operational data, such as motor current, drive-end vibration, net oil pressure, and evaporator refrigerant pressure. The cooling fan may be a transformer cooling fan in a power system. It should be noted that as a rotating machine, monitoring the motor current and drive-end vibration of the cooling fan can be used to directly assess the cooling fan's health.

[0053] For example, the structure of the cooling fan is as follows: Figure 1 As shown in the figure, it includes condensers, evaporators, compressors, steam heat exchangers, multiple valves, multiple flow meters, multiple pumps, and other devices. Taking the deployment of sensors at seven locations on the cooling fan as an example, the collected data includes:

[0054] 1) Net Oil Pressure (PO_net), which indicates the effective pressure of the chiller's lubricating oil system, is usually the difference between the oil pump outlet pressure and the compressor crankcase pressure;

[0055] 2) Oil Sump Temperature (TO_sump), indicating the temperature of the lubricating oil stored in the oil sump;

[0056] 3) Oil Feed Pressure (PO_feed), which indicates the pressure in the lubricating oil supply line;

[0057] 4) Temperature Condenser Approach (TCA) is an indicator used to measure the heat exchange efficiency of the condenser. It refers to the difference between the saturated condensing temperature of the refrigerant in the condenser and the cooling water outlet temperature.

[0058] 5) Pressure Refrigerant Evaporator (PRE), which indicates the pressure of the refrigerant in the evaporator;

[0059] 6) Temperature Refrigerant Evaporator (TRE), which indicates the temperature of the refrigerant in the evaporator;

[0060] 7) Unit Status: Indicates whether the chiller is running, shut down, or in a specific operating mode.

[0061] Figure 2 This is a flow chart of a transformer cooling fan performance prediction method based on a hierarchical residual network according to an embodiment of the present invention.

[0062] like Figure 2 As shown in Figure 2, the transformer cooling fan performance prediction method based on the hierarchical residual network includes:

[0063] S21, obtaining time series data of the cooling fan under various performance degradation degrees and various operating states through multiple sensors.

[0064] Illustratively, the plurality of operating conditions include: normal operation, reduced condenser water flow, condenser blockage, reduced evaporator water flow, excessive oil, presence of non-condensable gases, refrigerant leakage, refrigerant overcharge, and expansion valve failure.

[0065] Specifically, the acquired time series data forms a dataset for subsequent training of a hierarchical attention residual time series network. This dataset, derived from a full-lifecycle simulation testbed for transformer cooling fans, encompasses various operating states, ranging from healthy to varying degrees of degradation. After acquiring the time series data, to improve training efficiency, it is cleaned to address missing or outliers. Subsequently, the time series data is annotated with corresponding operating state labels (i.e., different fault types) and performance degradation degree labels (i.e., numerically encoding the operating state and performance degradation degree) based on experimental records or maintenance standards. Finally, the multidimensional time series data is normalized, and a sliding window method (e.g., a sliding window of 10 time steps) is used to segment the continuous time series data sequence into samples with fixed time steps. This data is then formatted into samples suitable for input into a time series deep learning model (i.e., the subsequent hierarchical attention residual time series network).

[0066] The performance degradation degree label can take values ​​in the range of 0 to 1, for example, expressed as a percentage of the remaining effective life or the decay rate of a key performance parameter, ranging from 0 to 1, where 0 represents complete health and 1 represents complete failure.

[0067] S22, construct a hierarchical attention residual temporal network and train the hierarchical attention residual temporal network based on temporal data.

[0068] Among them, such as Figure 3 As shown, the hierarchical attention residual timing network includes: a sensor attention module, a parallel time feature extraction module, a time attention module and a multi-task output module. The parallel time feature extraction module includes multiple time feature extraction units. The multi-task output module includes a shared fully connected unit, a performance degradation prediction unit and a fault prediction unit. The input and output ends of the multiple time feature extraction units are respectively connected to the output end of the sensor attention module and the input end of the time attention module. The input ends of the performance degradation prediction unit and the fault prediction unit are connected to the output end of the time attention module through a shared fully connected unit.

[0069] Specifically, a hierarchical attention residual temporal network can be constructed based on the feature dimensions and number of label categories of the preprocessed time series data. The dataset obtained in step S21 is then divided into independent training, validation, and test sets in a ratio of 6:2:2. During training, a multi-task learning strategy is adopted, setting a composite loss function consisting of a weighted combination of a regression loss (such as mean squared error) for performance degradation prediction and a classification loss (such as categorical cross entropy) for fault type prediction. Gradient descent optimizers such as Adam are used for end-to-end training. The training set is used directly for model parameter learning, adjusting weights through forward and backpropagation to establish a mapping between input and output. The validation set monitors the training process, prevents overfitting (e.g., through early stopping), and adjusts hyperparameters. The test set provides an unbiased evaluation of the network's final performance, ensuring the generalizability of the results.

[0070] For example, during the training process, an early stopping mechanism is used to monitor the compound loss value on the validation set, and the training is terminated early when the loss no longer improves, and the model with the best performance is saved to prevent overfitting.

[0071] S23, using the trained hierarchical attention residual time series network to obtain performance degradation prediction results and fault type prediction results based on the time series data to be diagnosed.

[0072] In this embodiment, the partial implementation process of step S22 and the implementation process of step S23 are as follows: Figure 4 As shown in the figure, the core diagnostic model is a hierarchical attention residual time series network. The input of this model is the time series data of the multi-dimensional sensors of the transformer cooling fan. The model has two outputs, output 1 and output 2. Output 1 is a regression task that outputs the performance degradation prediction result, which represents the quantitative severity level; output 2 is a classification task that outputs the fault type prediction result, which represents the most likely fault type.

[0073] Specifically, the optimal hierarchical attention residual time series network trained and saved in step S22 is deployed in the application environment. In actual use, the time series data to be diagnosed (which can be real-time or historical) is converted using the same preprocessing and windowing methods and then input into the hierarchical attention residual time series network deployed in the application environment. Through a single forward propagation calculation, this network simultaneously outputs two prediction results: one is a quantitative assessment score prediction of the degree of cooling fan performance degradation (e.g., obtained by denormalizing the sigmoid output of the regression head); the other is the most likely predicted fault type and its confidence level (e.g., calculated by applying the Softmax function to the raw logical value output by the classification head).

[0074] In some embodiments of the present invention, the method further includes: generating a diagnostic report on the health status of the cooling fan according to the performance degradation prediction result and the fault type prediction result.

[0075] Thus, a comprehensive diagnostic report on the health status of the cooling fan can be obtained, which helps the staff to fully understand the status of the cooling fan.

[0076] In some embodiments, the sensor attention module outputs weighted features via the following formula :

[0077] (1)

[0078] (2)

[0079] (3)

[0080] in, represents the attention score; , , respectively represent the learnable weights and biases of the sensor attention module; , represents the sample input to the sensor attention module; represents the time step, Indicates the number of sensor channels; represents the attention weight, ⊙ represents element-wise multiplication, Represents the exponential function.

[0081] Specifically, see Figure 3The sensor attention module consists of a fully connected layer and a Softmax activation function layer connected in sequence. After the time series samples obtained from the time series data are input, they are first passed through a fully connected layer and a Softmax activation function layer to dynamically weight different sensor channels (features) to learn which sensor information contributes more to the diagnosis under the current working conditions, thereby adaptively focusing on key features. The fully connected layer processes the above formula (1) to obtain the calculated attention score , and then the attention weight is obtained by processing the above formula (2) through the Softmax activation function layer Afterwards, With time series samples Weighted to obtain weighted features . Thus, a sensor feature with greater contribution can be obtained.

[0082] In some embodiments, the i-th time feature extraction unit extracts the time feature by the following formula: :

[0083] (4)

[0084] (5)

[0085] (6)

[0086] in, represents a one-dimensional convolutional neural network, 、 represents the convolution kernel weight used by the i-th time feature extraction unit, represents the convolution kernel size used by the i-th time feature extraction unit, 、 represents the learnable bias of the i-th temporal feature extraction unit, Indicates the number of output channels, represents the time step after convolution processing (affected by stride and padding), 、 Represent the two intermediate features of the i-th time feature extraction unit, Represents 1×1 convolution, i is a positive integer less than or equal to K, and K represents the number of time feature extraction units.

[0087] Specifically, see Figure 3 , the attention-weighted features Input to K ( Figure 3Taking K=2 as an example, parallel one-dimensional convolutional neural network (1D Convolutional Neural Network, 1D-CNN) branches (i.e., temporal feature extraction units) are constructed. Each branch uses convolution kernels of different sizes (taking two branches as an example, the convolution kernel sizes are 3 and 5 respectively) to capture short-term local patterns and medium- and long-term dynamic trends in the data respectively. Residual connections are used within each branch to add the input of the branch to the output after two layers of convolution, which helps to train deeper networks and effectively learn fault features. For the i-th branch, the convolution operations of (4) and (5) are performed in sequence, and then the residual connection is performed through (6).

[0088] It should be noted that in the above formula (6) The role of the 1×1 convolution is to ensure that the dimensions of the input and output match.

[0089] In some embodiments, the temporal attention module outputs a context vector via :

[0090] (7)

[0091] (8)

[0092] (9)

[0093] (10)

[0094] in, represents the splicing function, K represents the number of time feature extraction units, represents the time step attention score; , , respectively represent the learnable weights and biases of the temporal attention module; represents the time step weight, express No. time-step features.

[0095] Specifically, see Figure 3 The temporal attention module includes a splicing layer, a fully connected layer, and a Softmax activation function layer connected in sequence. The feature maps from the K scale branches are spliced ​​and fused through the splicing layer using the above formula (7), and then the time step attention score is calculated through the fully connected layer using the above formula (8) , calculate the time step weight through the Softmax activation function layer , we can learn which time points are most critical for judging the occurrence and development of faults in the current time window. Then, the above formula (10) will be and Perform weighted summation to achieve weighted summation of features at different time steps, and obtain a fixed-dimensional context vector that highly condenses key spatiotemporal information. .

[0096] In some embodiments, the shared fully connected unit performs nonlinear feature transformation by the following formula:

[0097] (11)

[0098] in, represents the output of the shared fully connected unit, represents the linear rectification function, represents the hidden layer dimension; , , respectively represent the learnable weights and biases of the shared fully connected unit.

[0099] In some embodiments, the performance degradation prediction unit outputs the performance degradation prediction result through the following formula:

[0100] (12)

[0101] in, , , respectively represent the learnable weights and biases of the performance degradation prediction unit; represents the performance degradation prediction result, Represents an S-shaped growth curve function.

[0102] In some embodiments, the fault prediction unit outputs a fault type prediction result using the following formula:

[0103] (13)

[0104] in, , , respectively represent the learnable weights and biases of the fault prediction unit; Indicates the number of fault types, is the original logical value of each fault type.

[0105] Specifically, see Figure 3 , the context vector C is input into a shared fully connected unit and subjected to the final nonlinear feature transformation according to Equation (11). Subsequently, two independent output heads are separated from the shared fully connected unit: the performance degradation prediction head (regression head) and the operating status assessment head (classification head), corresponding to the two diagnostic tasks respectively.

[0106] The regression head, consisting of a fully connected layer and a Sigmoid (or linear) activation function layer, outputs a continuous value between 0 and 1 according to Equation (12), representing the normalized future performance indicator or degree of degradation, which serves as the main output of the present invention. The classification head, consisting of a fully connected layer, outputs raw logical values ​​(logits) representing various operating states (healthy, slightly degraded, etc.) according to Equation (13). These serve as auxiliary outputs of the present invention and are used to qualitatively evaluate the current state.

[0107] For example, see Figure 3 After obtaining H, the shared fully connected unit can also be processed by a regularization layer. This regularization layer follows the ReLU activation function layer and randomly discards some neuron outputs (such as setting them to zero), forcing the network to not rely on specific neurons, thereby suppressing overfitting.

[0108] In some embodiments, the loss function used to train the layered attention residual temporal network is It is obtained by the following formula:

[0109] (14)

[0110] (15)

[0111] (16)

[0112] in, 、 Represent regression loss and classification loss respectively, 、 represents the hyperparameter, 、 They represent the actual degree of performance degradation and the predicted results of performance degradation respectively. Indicates the real fault type label, Represents the fault type prediction result of the i-th sample.

[0113] Specifically, the present invention employs a multi-task learning strategy to perform end-to-end training on a hierarchical attention residual temporal network. A combined loss function (Equation (14)) is defined as the weighted sum of the mean squared error loss for performance degradation prediction (Equation (15)) and the categorical cross entropy loss for fault type prediction (Equation (16)). This combined loss function is minimized using an optimizer such as Adam. An early stopping mechanism is introduced during training to monitor the loss of the validation set. If the loss no longer decreases within a preset period, training is terminated early, and the weights of the model with the best performance on the validation set are saved, effectively preventing model overfitting.

[0114] After training the Hierarchical Attention Residual Time Series Network, it is deployed. In practical applications, real-time cooling fan sensor data streams can be input, and the deployed Hierarchical Attention Residual Time Series Network can simultaneously output quantitative predictions of future performance and an assessment of the current operating status, providing comprehensive and accurate data support for developing condition-based maintenance (CBM) and predictive maintenance (PdM) strategies.

[0115] The following combination Figure 5-Figure 7 The beneficial effects of the transformer cooling fan performance prediction method based on the hierarchical residual network according to the embodiment of the present invention are described as follows:

[0116] 1) Achieved high-precision performance degradation prediction

[0117] The present invention innovatively designs the performance degradation assessment as a regression task. Through experiments, the scatter plot of the actual value of performance degradation VS the predicted value is as follows: Figure 5 As shown in Figure 2, the performance degradation prediction error distribution under each fault type is as follows: Figure 6 See Figure 5 The Hierarchical Attention Residual Temporal Network achieved an R² score of 0.826 between its predictions and the true values ​​for performance degradation levels (on a 0-1 scale), indicating that the model's predictions can explain over 82% of the variation in the true performance degradation level. Furthermore, the root mean squared error (RMSE) and mean absolute error (MAE) were reduced to approximately 0.13 and 0.10, respectively (calculated on a 0-1 scale). This means that the average prediction error is less than one level, demonstrating that the proposed method can accurately and reliably quantify the severity of a fault. Figure 6 The results shown prove that the method of the present invention can accurately and reliably quantitatively predict the long-term performance evolution trend of the cooling fan.

[0118] 2) Provides an accurate assessment of the current state

[0119] The present invention can simultaneously complete the two tasks of performance degradation prediction and fault classification assessment in a unified model. The experimental results based on the public ASHRAE Project 1043-RP dataset are as follows: Figure 7The results show that the proposed method achieved an overall accuracy of 94% in fault type recognition across nine operating conditions (eight faults and one normal state). The F1 scores for key faults such as "non-condensable gas," "condenser blockage," and "refrigerant overcharge" reached 0.99 or higher, and the F1 scores for most fault types exceeded 0.88, demonstrating extremely high recognition accuracy and strong generalization capabilities.

[0120] In summary, the transformer cooling fan performance prediction method based on a hierarchical residual network in the present invention improves diagnostic sensitivity through multi-scale feature extraction. Specifically, by using parallel residual time series modules with different receptive fields, it can effectively capture the entire process of fault evolution, from weak short-term signals to significant long-term trends. This makes the model more likely to identify anomalies in the early stages of fault occurrence compared to single-scale models, thus facilitating predictive maintenance. An integrated multi-task diagnostic framework enhances information dimensionality. Specifically, the present invention innovatively employs an end-to-end multi-task learning framework to simultaneously implement performance degradation prediction and fault type classification within a unified model. This design allows two related tasks to share the underlying time series feature representation, enabling the model to exploit the inherent correlation between performance degradation severity and fault type. Through collaborative training, the model not only improves the prediction accuracy of a single task but, more importantly, provides a more comprehensive diagnostic view encompassing both qualitative and quantitative dimensions. Compared to traditional single-task diagnostic models, this provides richer and more valuable information support for power system operation and maintenance decision-making.

[0121] It should be noted that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic device), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). Furthermore, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing it in another suitable manner if necessary, and then storing it in a computer memory.

[0122] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0123] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0124] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0125] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0126] In the present invention, unless otherwise specified or limited, the terms "installed," "connected," "connect," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integration; mechanical connection, electrical connection; direct connection, or indirect connection through an intermediate medium; internal communication between two components, or interaction between two components, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0127] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediary. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or diagonally above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0128] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A transformer cooling fan performance prediction method based on a hierarchical residual network, characterized in that: The cooling fan is provided with sensors at multiple locations, and the method includes: Acquiring time series data of the cooling fan under various degrees of performance degradation and various operating states through the plurality of sensors; Constructing a hierarchical attention residual temporal network and training the hierarchical attention residual temporal network based on the temporal data, wherein the network comprises: a sensor attention module, a temporal feature extraction module, a temporal attention module, and a multi-task output module, wherein the temporal feature extraction module comprises a plurality of temporal feature extraction units, and the multi-task output module comprises a shared fully connected unit, a performance degradation prediction unit, and a fault prediction unit, wherein the input and output ends of the plurality of temporal feature extraction units are respectively connected to the output end of the sensor attention module and the input end of the temporal attention module, and the input ends of the performance degradation prediction unit and the fault prediction unit are connected to the output end of the temporal attention module through the shared fully connected unit; The trained hierarchical attention residual time series network is used to obtain performance degradation prediction results and fault type prediction results based on the time series data to be diagnosed; The sensor attention module outputs weighted features through the following formula : ； in, represents the attention score; , respectively represent the learnable weights and biases of the sensor attention module; , represents the sample input to the sensor attention module; represents the time step, Indicates the number of sensor channels; represents the attention weight, represents element-wise multiplication, represents the exponential function; No. The time feature extraction unit extracts the time feature through the following formula : ; in, represents a one-dimensional convolutional neural network, Indicates the The convolution kernel weights used by the temporal feature extraction unit, Indicates the The convolution kernel size used by the temporal feature extraction unit, Indicates the The learnable bias of the temporal feature extraction unit, Indicates the number of output channels, represents the time step after convolution processing, Respectively represent Two intermediate features of the temporal feature extraction unit, represents 1×1 convolution, is a positive integer less than or equal to K, where K represents the number of time feature extraction units.

2. The transformer cooling fan performance prediction method based on hierarchical residual network according to claim 1 is characterized in that: The temporal attention module outputs the context vector by : ; in, represents the concatenation function, represents the number of temporal feature extraction units, represents the time step attention score; , respectively represent the learnable weights and biases of the temporal attention module; represents the time step weight, express No. time-step features.

3. The transformer cooling fan performance prediction method based on hierarchical residual network according to claim 2 is characterized in that: The shared fully connected unit performs nonlinear feature transformation by the following formula: ； in, represents the output of the shared fully connected unit, represents the linear rectification function, represents the hidden layer dimension; , respectively represent the learnable weights and biases of the shared fully connected unit.

4. The transformer cooling fan performance prediction method based on hierarchical residual network according to claim 3 is characterized in that: The performance degradation prediction unit outputs the performance degradation prediction result through the following formula: ； in, , respectively represent the learnable weights and biases of the performance degradation prediction unit; represents the performance degradation prediction result, Represents an S-shaped growth curve function.

5. The transformer cooling fan performance prediction method based on hierarchical residual network according to claim 3 is characterized in that: The fault prediction unit outputs the fault type prediction result through the following formula: ； in, , respectively represent the learnable weight and bias of the fault prediction unit; Indicates the number of fault types, is the original logical value of each fault type.

6. The transformer cooling fan performance prediction method based on hierarchical residual network according to claim 1 is characterized in that: The loss function used to train the layered attention residual temporal network It is obtained by the following formula: ； in, Represent regression loss and classification loss respectively, represents the hyperparameter, They represent the actual degree of performance degradation and the predicted results of performance degradation respectively. Indicates the real fault type label, Indicates the Fault type prediction results for samples.

7. The transformer cooling fan performance prediction method based on a hierarchical residual network according to any one of claims 1 to 6, characterized in that: The method further comprises: A diagnosis report on the health status of the cooling fan is generated according to the performance degradation prediction result and the fault type prediction result.

8. The transformer cooling fan performance prediction method based on a hierarchical residual network according to any one of claims 1 to 6, characterized in that: The multiple operating states include: normal operation, reduced condenser water flow, condenser blockage, reduced evaporator water flow, excessive oil, presence of non-condensable gas, refrigerant leakage, refrigerant overcharge and expansion valve failure.

Images