A power transmission equipment fault diagnosis model modeling method and device, a terminal and a medium

Abstract

The application discloses a power transmission equipment fault diagnosis model modeling method and device, a terminal and a medium. The scheme of the application firstly acquires a combination of paired normal inspection image samples and degraded simulation image samples, and then according to the combination of the inspection image samples, the combination of the inspection image samples is input into an initial power transmission equipment fault diagnosis model, a backbone detection sub-model, a low-rank modulation sub-model and a gating sub-model, and a preset joint optimization objective function is combined to perform model training on the initial power transmission equipment fault diagnosis model. The scheme trains a model with gating and modulation mechanisms by acquiring a sample combination containing degraded simulation images, introduces a low-rank modulation branch in a feature extraction layer of a basic model, and adjusts the participation degree of the low-rank modulation branch according to the degradation characteristics of the input image through a learnable gating function, thereby improving the accuracy and robustness of power transmission equipment fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) inspection technology, and in particular to a modeling method, device, terminal and medium for fault diagnosis of power transmission equipment. Background Technology

[0002] With the deepening development of smart grid construction, image analysis technology for power transmission equipment based on UAV inspection has been widely applied. This technology significantly improves the efficiency of inspection operations and reduces reliance on manual labor by automatically identifying and diagnosing equipment status. To adapt to the real-time processing requirements of the terminal side, existing solutions generally adopt lightweight visual inspection models and optimize their performance in specific power transmission equipment inspection tasks through task fine-tuning strategies.

[0003] However, in real-world engineering scenarios, inspection images inevitably suffer from multiple imaging degradation effects during edge-side acquisition, wireless channel transmission, and storage. These include, but are not limited to, detail loss due to transmission compression, pixel distortion caused by environmental noise, edge blurring due to reduced resolution, and gradient anomalies induced by complex lighting conditions. These degradation factors significantly degrade image quality, and current mainstream model training methods still rely on the ideal assumption that training and test data distributions are consistent, failing to effectively model the dynamic impact of image degradation on feature space distribution. Specifically, degraded image input causes biases in the model's feature extraction process, resulting in unpredictable shifts in feature distribution and weakening the model's discriminative ability. In edge-side deployment environments, due to limited computing resources, existing models lack adaptive response mechanisms to degradation conditions. When input image quality fluctuates, internal parameters cannot be dynamically adjusted to compensate for degradation effects, ultimately leading to a systematic decline in fault diagnosis accuracy, manifested as increased false positive rates and a higher risk of missing critical defects. Summary of the Invention

[0004] This application provides a modeling method, device, terminal, and medium for fault diagnosis of power transmission equipment, which aims to achieve adaptive compensation for image degradation effects and improve the accuracy and robustness of fault diagnosis of power transmission equipment.

[0005] To achieve the above-mentioned objectives, the first aspect of this application provides a modeling method for fault diagnosis of power transmission equipment, comprising: Obtain a preset combination of inspection images, wherein the combination of inspection images includes pairs of normal inspection images and degraded simulated images. Based on the inspection image sample combination, the inspection image sample combination is input into the initial power transmission equipment fault diagnosis model. The initial power transmission equipment fault diagnosis model is trained by combining the preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model is used to extract image features from the input image and then determine the power transmission equipment fault diagnosis result based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model based on the image features and the preset gating decision logic. The low-rank modulation sub-model is used to perform directional compensation of the image features through low-rank matrix fine-tuning logic. Based on the convergence degree of the initial power transmission equipment fault diagnosis model, when the convergence degree meets the preset training termination condition, the power transmission equipment fault diagnosis model is obtained.

[0006] Preferably, the process of inputting the combined inspection image samples into the initial power transmission equipment fault diagnosis model, and training the initial power transmission equipment fault diagnosis model using a preset joint optimization objective function, specifically includes: The combined inspection image samples are input into the initial power transmission equipment fault diagnosis model; Based on the normal inspection image samples and the degraded simulation image samples included in the inspection image sample combination, and combined with the backbone detection sub-model in the initial power transmission equipment fault diagnosis model, the corresponding image features are extracted from the normal inspection image samples and the degraded simulation image samples respectively through the feature extraction logic of the backbone detection sub-model. Based on the image features, the degradation perception result of the image features is determined by the gated sub-model in the initial power transmission equipment fault diagnosis model. Then, based on the degradation perception result and the gated decision logic in the gated sub-model, the gated value corresponding to the image features is determined. The gated value is used to control the enable state of the low-rank modulation sub-model. When the gate value satisfies the triggering condition of the low-rank modulation sub-model, the image features are directionally compensated through the low-rank matrix fine-tuning logic in the low-rank modulation sub-model to obtain the compensated image features. The fault determination logic of the main detection sub-model outputs two sets of image features corresponding to the power transmission equipment fault diagnosis results. Based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model and the low-rank modulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition.

[0007] Preferably, the process further includes obtaining a preset combination of inspection image samples before: Based on a preset set of imaging degradation operators and combined with degradation operator selection weights, a degradation simulation sub-model of the input image and the degradation simulation image is constructed. The degradation operator selection weights are used to constrain the activation range of different imaging degradation operators, and each imaging degradation operator is configured with learnable control parameters for controlling degradation intensity and / or degradation form. Obtain a preset inspection image sample, and use the inspection image sample as the input image of the degradation correlation sub-model. Through the processing of the degradation correlation sub-model, obtain the degradation simulation image sample corresponding to the inspection image sample. The inspection image samples are combined with the corresponding degradation simulation image samples to obtain several sets of inspection image sample combinations.

[0008] Preferably, the degradation simulation sub-model is specifically:

[0009] In the formula, For the degraded simulated image sample, Activate weights for the degenerate operator. For the k-th imaging degradation operator, The original input inspection image sample, The learnable control parameters are those corresponding to the k-th imaging degradation operator. This is a sequence consisting of weights selected from the degradation operators corresponding to k imaging degradation operators.

[0010] Preferably, based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model and the low-rank modulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition, specifically including: Based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model, the low-rank modulation sub-model, and the degradation simulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition.

[0011] Preferably, the joint optimization objective function is specifically:

[0012] In the formula, This refers to the backbone detection model. The original input inspection image sample, Let L be the degraded simulated image sample, L be the loss function, and A and B represent the low-rank modulation parameters to be optimized in the low-rank modulation sub-model. This represents the gating parameters to be optimized in the gated sub-model. and This represents the learnable control parameters to be optimized and the weights for selecting the degradation operator in the degradation simulation sub-model.

[0013] Preferably, the imaging degradation operator includes: a compression degradation operator, a noise degradation operator, a blur degradation operator, and a gradient-adversarial degradation operator.

[0014] The second aspect of this application provides a modeling apparatus for fault diagnosis of power transmission equipment, comprising: An image sample acquisition unit is used to acquire a preset inspection image sample combination, wherein the inspection image sample combination includes pairs of normal inspection image samples and degradation simulation image samples. An initial model optimization unit is used to input the inspection image sample combination into an initial power transmission equipment fault diagnosis model based on the inspection image sample combination, and to train the initial power transmission equipment fault diagnosis model in combination with a preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model is used to extract image features based on the input image, and then determine the power transmission equipment fault diagnosis result based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model based on the image features and a preset gating decision logic. The low-rank modulation sub-model is used to perform directional compensation of the image features through low-rank matrix fine-tuning logic. The model optimization termination determination unit is used to obtain the transmission equipment fault diagnosis model based on the convergence degree of the initial transmission equipment fault diagnosis model, when the convergence degree meets the preset training termination condition.

[0015] A third aspect of this application provides a modeling terminal for a fault diagnosis model of a power transmission equipment, comprising: a memory and a processor; The memory is used to store program code, which corresponds to the modeling method for fault diagnosis model of power transmission equipment as provided in the first aspect of this application. The processor is used to read and execute the program code to implement the modeling method for fault diagnosis of power transmission equipment.

[0016] The fourth aspect of this application provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores program code, which is used to be read and executed by a processor to implement the power transmission equipment fault diagnosis modeling method provided in the first aspect of this application.

[0017] As can be seen from the above technical solutions, this application has the following advantages: The proposed solution first obtains a combination of normal inspection image samples and degraded simulated image samples. Then, based on this combination, it is input into an initial power transmission equipment fault diagnosis model. The model is then trained using a preset joint optimization objective function. This initial model includes a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model extracts image features from the input image and determines the fault diagnosis result based on these features. The gating sub-model controls the enabling state of the low-rank modulation sub-model based on the image features and preset gating logic. The low-rank modulation sub-model performs directional compensation of the image features through low-rank matrix fine-tuning logic. Finally, based on the convergence degree of the initial power transmission equipment fault diagnosis model, a power transmission equipment fault diagnosis model is obtained when the convergence degree meets a preset training termination condition. This method trains the model by introducing pairs of normal and degraded simulated image samples, and combines the collaborative optimization of the backbone detection sub-model, low-rank modulation sub-model and gating sub-model. This enables the model to adaptively perceive image degradation and perform directional compensation, thus maintaining high fault diagnosis accuracy even when image quality may degrade during edge acquisition, wireless transmission and storage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of a fault diagnosis modeling method for power transmission equipment provided in this application.

[0020] Figure 2 This is a schematic diagram of the architecture of an embodiment of a fault diagnosis modeling device for power transmission equipment provided in this application.

[0021] Figure 3 This is a schematic diagram of the architecture of a modeling terminal for a fault diagnosis model of power transmission equipment provided in this application. Detailed Implementation

[0022] This application provides a modeling method, apparatus, terminal, and medium for fault diagnosis of power transmission equipment, which aims to achieve adaptive compensation for image degradation effects and improve the accuracy and robustness of fault diagnosis of power transmission equipment.

[0023] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] First, a detailed description of an embodiment of the fault diagnosis modeling method for power transmission equipment provided in this application is as follows: Please see Figure 1 The first aspect of this application provides a modeling method for fault diagnosis of power transmission equipment, including: Step 101: Obtain a preset combination of inspection image samples; The inspection image sample combination includes pairs of normal inspection image samples and degraded simulated image samples. Step 102: Based on the inspection image sample combination, input the inspection image sample combination into the initial power transmission equipment fault diagnosis model, and train the initial power transmission equipment fault diagnosis model by combining it with the preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model is used to extract image features from the input image and then determine the fault diagnosis result of the power transmission equipment based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model based on the image features and the preset gating judgment logic. The low-rank modulation sub-model is used to perform directional compensation of the image features through low-rank matrix fine-tuning logic. Step 103: Based on the convergence degree of the initial power transmission equipment fault diagnosis model, when the convergence degree meets the preset training termination condition, the power transmission equipment fault diagnosis model is obtained.

[0025] It should be noted that, for ease of understanding, some key terms in this embodiment are explained below: Inspection image sample set: refers to a set of image data used for model training, characterized in that the set contains pairs of normal inspection image samples and degraded simulated image samples. Normal inspection image samples usually refer to images acquired under ideal or standard conditions, while degraded simulated image samples are images obtained by simulating the degradation of normal images through a specific method, aiming to simulate the image quality degradation that may occur in real-world applications.

[0026] Initial transmission equipment fault diagnosis model: This refers to the baseline model used to diagnose transmission equipment faults before training begins. This model is designed to identify features related to equipment faults from images and provide diagnostic results accordingly. During training, the model's internal parameters are adjusted and optimized based on the input data and optimization objectives.

[0027] The main detection sub-model, as a core component of the initial power transmission equipment fault diagnosis model, primarily functions to receive input images and extract representative image features. Subsequently, this sub-model analyzes and judges these extracted image features to ultimately determine the fault diagnosis result of the power transmission equipment.

[0028] Low-rank modulation sub-model: This sub-model is designed to perform targeted compensation of image features. When enabled by the gated sub-model, it adjusts the image features extracted by the backbone detection sub-model by applying low-rank matrix fine-tuning logic to enhance its robustness to degraded images or improve its discriminative ability.

[0029] Gating Sub-model: This sub-model determines the enabling state of the low-rank modulation sub-model based on the image features extracted by the backbone detection sub-model and the preset gating decision logic. It functions similarly to a smart switch, dynamically controlling whether low-rank modulation compensation is needed for image features based on the degree of degradation or specific conditions.

[0030] Joint optimization objective function: This refers to a comprehensive objective function used to guide the optimization of model parameters during model training. This function typically combines multiple loss terms, such as the loss from diagnostic results, the loss from the gated sub-model, and the loss from the low-rank modulation sub-model, aiming to achieve optimal performance for the model simultaneously in multiple aspects.

[0031] Training termination criteria: These refer to a set of preset standards used to determine when to stop training during model training. These criteria may include the model's performance on the validation set no longer improving, the number of training iterations reaching its limit, and the loss value converging to a certain threshold, to ensure that the model stops training after reaching the expected performance and avoid overfitting.

[0032] More specifically, this application provides a modeling method for fault diagnosis of power transmission equipment, which aims to improve the robustness and diagnostic accuracy of the model under the condition of degradation in actual inspection images. The logical steps of this method are as follows: First, a pre-defined combination of inspection image samples needs to be obtained. This combination can be created manually by selecting normal images from a large number of inspection images and then manually performing various degradation processing on these normal images, such as adding noise, blurring, or reducing resolution using image editing software, to generate corresponding degraded simulated image samples. Subsequently, the normal image samples are paired with the corresponding degraded simulated image samples to form the inspection image sample combination.

[0033] Secondly, based on the acquired inspection image sample combination, it is input into the initial power transmission equipment fault diagnosis model, and the model is trained using a preset joint optimization objective function. During training, the initial power transmission equipment fault diagnosis model can be configured to include a basic convolutional neural network as the backbone detection sub-model, used to extract features from the input images and perform fault diagnosis. Simultaneously, a simple fully connected layer can be introduced as a gating sub-model, outputting a scalar value based on the extracted features. This scalar value is compared with a preset threshold to control the activation of the low-rank modulation sub-model. The low-rank modulation sub-model can be a small matrix multiplication layer that, when activated, performs a linear transformation on the features output by the backbone detection sub-model. The joint optimization objective function can be simply set as the cross-entropy loss of the diagnostic results, and the parameters of the backbone detection sub-model, gating sub-model, and low-rank modulation sub-model are iteratively updated using a backpropagation algorithm.

[0034] Finally, based on the convergence degree of the initial power transmission equipment fault diagnosis model, the final power transmission equipment fault diagnosis model is obtained when the convergence degree meets the preset training termination condition. The convergence degree can be judged based on the change of the loss function value during training. For example, when the decrease in the loss function value is less than a preset small threshold over several consecutive training cycles, the model is considered to have converged. At this time, training is stopped and the current model parameters are saved as the final power transmission equipment fault diagnosis model.

[0035] By employing the aforementioned method, this application effectively addresses the issues of insufficient robustness and decreased diagnostic accuracy in existing power transmission equipment fault diagnosis models when faced with degradation in actual inspection images. This method trains the model using paired normal and degraded simulated image samples, and combines this with the collaborative optimization of the backbone detection sub-model, low-rank modulation sub-model, and gating sub-model. This enables the model to adaptively perceive image degradation and perform targeted compensation. Furthermore, this scheme introduces a low-rank modulation branch into the feature extraction layer of the basic model and adaptively adjusts the participation level of the low-rank modulation branch based on the degradation characteristics of the input image using a learnable gating function. This ensures that even with potential image quality degradation during end-side acquisition, wireless transmission, and storage, high fault diagnosis accuracy can still be maintained.

[0036] In some embodiments described above in this application, a modeling method for fault diagnosis of power transmission equipment is proposed. This method acquires a combination of inspection image samples and inputs them into an initial fault diagnosis model for training. However, in the actual training process, how to finely utilize the backbone detection sub-model, low-rank modulation sub-model, and gating sub-model in the model to effectively process normal and degraded images, and to achieve targeted parameter optimization, thereby improving the model's diagnostic performance under complex degradation conditions, are technical details that need further clarification.

[0037] To address this, this application further proposes a specific process for training the initial power transmission equipment fault diagnosis model by inputting a combination of inspection image samples into the model and combining it with a pre-defined joint optimization objective function. This step aims to provide training data for the initial power transmission equipment fault diagnosis model. The combination of inspection image samples includes pairs of normal inspection image samples and degraded simulated image samples, which enables the model to learn to process both normal images and degraded images simultaneously during training, thereby enhancing its robustness.

[0038] Specifically, based on the normal inspection image samples and degraded simulated image samples included in the inspection image sample combination, and combined with the backbone detection sub-model in the initial power transmission equipment fault diagnosis model, the corresponding image features are extracted from the normal inspection image samples and degraded simulated image samples respectively through the feature extraction logic of the backbone detection sub-model. The backbone detection sub-model typically adopts a deep neural network structure, such as a convolutional neural network. Its feature extraction logic refers to the network layer performing multi-layer nonlinear transformations on the input image, thereby extracting feature representations with higher semantic information and discriminative ability from the original pixel data. By extracting features from the normal and degraded image samples separately, their respective representations in the feature space can be obtained, laying the foundation for subsequent degradation perception and compensation operations.

[0039] Subsequently, using the gated sub-model in the initial power transmission equipment fault diagnosis model, the degradation perception result of image features is determined based on the image features. Then, based on the degradation perception result and the gated decision logic in the gated sub-model, the corresponding gate value for the image features is determined. The gate value is used to control the enabling state of the low-rank modulation sub-model. The gated sub-model is designed to evaluate the degree or type of degradation of input image features. Its degradation perception result can be a continuous value representing the degradation intensity or a binary classification result indicating whether degradation exists. The gated decision logic generates a gate value based on these perception results, for example, through a preset threshold or a learned decision boundary. This gate value, as a control signal, determines whether and how the low-rank modulation sub-model needs to be activated, thereby achieving conditional processing of degraded image features.

[0040] When the gate value meets the trigger condition of the low-rank modulation sub-model, the image features are directionally compensated through the low-rank matrix fine-tuning logic within the low-rank modulation sub-model, resulting in compensated image features. The trigger condition of the low-rank modulation sub-model is typically that the gate value reaches a preset activation threshold, indicating significant degradation of the image features. The low-rank matrix fine-tuning logic refers to the fine adjustment of image features using a low-rank matrix. Low-rank matrices have fewer degrees of freedom, enabling effective feature correction with a smaller number of parameters, avoiding the introduction of excessive noise or unnecessary changes. This directional compensation aims to specifically repair or mitigate the degradation effects in image features, making them closer to the feature representation of a normal image, thereby improving the accuracy of subsequent fault diagnosis.

[0041] Next, the fault diagnosis logic of the backbone detection sub-model outputs two sets of image feature-corresponding fault diagnosis results for the power transmission equipment. The fault diagnosis logic of the backbone detection sub-model is its final classification or regression layer, responsible for outputting fault diagnosis results for the power transmission equipment based on the input image features (including the originally extracted features and the compensated features). Two sets of results are output: one based on the originally extracted image features, and the other based on the image features compensated by the low-rank modulation sub-model. This allows the model to simultaneously consider the feature performance before and after degradation during training and perform comparative learning.

[0042] Finally, based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gating sub-model and the low-rank modulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition. The loss value is an indicator that measures the difference between the model's diagnostic results and the true labels. The joint optimization objective function weights and combines the two sets of loss values ​​with possible other regularization terms. By minimizing this joint optimization objective function, the parameters of the gating sub-model and the low-rank modulation sub-model are iteratively updated using the backpropagation algorithm. This optimization strategy ensures that the gating sub-model can accurately identify degradation and control compensation, while the low-rank modulation sub-model can effectively perform directional compensation, thus enabling the entire model to achieve optimal performance when processing degraded images. The training termination condition can be that the validation set performance no longer improves, the maximum number of training epochs is reached, or the loss value falls below a certain threshold.

[0043] Through the above technical solution, this application provides a refined model training process. This process first extracts features from normal and degraded image samples using a backbone detection sub-model. Then, a gating sub-model intelligently perceives the degree of image degradation based on these features and generates a gating value according to gating decision logic. This gating value serves as a key control signal, selectively activating the low-rank modulation sub-model. When image feature degradation is detected and the trigger condition is met, the low-rank modulation sub-model performs targeted compensation for the degraded features through low-rank matrix fine-tuning logic, resulting in compensated features that are closer to the normal state. The backbone detection sub-model then outputs fault diagnosis results based on the original features and compensated features. By jointly optimizing the loss values ​​of these two sets of diagnostic results and specifically optimizing the parameters of the gating sub-model and the low-rank modulation sub-model, the model can learn how to accurately identify image degradation and efficiently and specifically correct degraded features only when necessary. This conditional compensation mechanism avoids unnecessary modification of normal image features, while significantly improving the model's fault diagnosis accuracy and robustness in the face of various degraded images, effectively solving the problem of insufficient model diagnostic performance under complex degradation conditions.

[0044] In some embodiments described above, this application proposes to improve the diagnostic capabilities of a power transmission equipment fault diagnosis model by acquiring a preset combination of inspection image samples and inputting them into the model for training. However, in practical applications, the preset degradation simulation image samples may not be able to fully cover the types and intensities of imaging degradation that may occur in power transmission equipment under various complex operating conditions, thus limiting the model's generalization ability and robustness when faced with unknown or diverse degradation situations.

[0045] To address this, this application further proposes that before obtaining a preset combination of inspection image samples, a degradation simulation sub-model is first constructed based on a preset set of imaging degradation operators and degradation operator selection weights. The degradation operator selection weights constrain the activation range of different imaging degradation operators, and each imaging degradation operator is configured with learnable control parameters to control degradation intensity and / or degradation form. Subsequently, the preset inspection image samples are obtained and used as input images to the degradation simulation sub-model. Through processing by the degradation simulation sub-model, degradation simulation image samples corresponding to the inspection image samples are obtained. Finally, the inspection image samples and their corresponding degradation simulation image samples are combined to obtain several sets of inspection image sample combinations.

[0046] Specifically, the imaging degradation operator set refers to a collection of mathematical models or algorithms used to simulate various degradation phenomena that images may encounter during acquisition, transmission, or storage. These operators can simulate various image quality degradation conditions such as blurring, noise, compression artifacts, and changes in brightness or contrast. Degradation operator selection weights are used to quantify and control the contribution or activation probability of different degradation operators in the imaging degradation operator set when generating degraded simulated image samples. By adjusting these weights, flexible control over combinations of different degradation types can be achieved, thereby generating diverse degradation scenarios. The degradation simulation sub-model is a functional module whose function is to generate degraded simulated image samples with specific degradation characteristics based on the original input inspection image samples, combined with a preset imaging degradation operator set, degradation operator selection weights, and learnable control parameters. This sub-model can simulate various degradation processes that images may undergo in the real world. Constraining the activation range of different imaging degradation operators means ensuring, through specific mechanisms (e.g., limiting the weights or parameter ranges of degradation operators), that the intensity or frequency of each imaging degradation operator is within a reasonable and controllable range when generating simulated degradation image samples. This avoids generating unrealistic or extremely degraded images, thus guaranteeing the effectiveness of the simulated degradation samples. Each imaging degradation operator is configured with learnable control parameters to control the degradation intensity and / or degradation form. These parameters are variables associated with each imaging degradation operator, used to finely adjust the specific performance of the degradation effect simulated by that operator. For example, for the blur operator, the learnable control parameters can control the size or type of the blur kernel; for the noise operator, they can control the variance or distribution of the noise. These parameters are optimizable during model training and can be adjusted through backpropagation to better match the actual degradation patterns. The preset inspection image samples refer to the raw, generally considered high-quality or unaffected by significant degradation image data collected during the inspection of power transmission equipment. These samples serve as the basis for generating simulated degradation images. Degraded simulated image samples are images with specific degradation characteristics generated by the degradation simulation sub-model based on the original inspection image samples. These samples appear in pairs with the original inspection image samples and are used to train the power transmission equipment fault diagnosis model, enabling it to identify and handle faults occurring under different degradation conditions. Inspection image sample combinations consist of paired data consisting of a normal inspection image sample and its corresponding degraded simulated image sample. This combination format provides paired inputs for training the power transmission equipment fault diagnosis model, allowing the model to learn to perform accurate fault diagnosis even in the presence of image degradation.

[0047] More specifically, this step constructs and introduces a differentiable imaging degradation modeling mechanism to simulate various imaging degradation processes that may occur during the acquisition, transmission, and storage of images from power transmission equipment at the end-side, thereby providing controllable complex scenario inputs for subsequent robust fine-tuning. The differentiable imaging degradation modeling module consists of multiple cascaded degradation operator layers, each of which is a differentiable operator. Its degradation form and intensity are controlled by corresponding learnable parameters. Through a learnable operator selection mechanism, different degradation operators are adaptively combined, thereby introducing representative imaging quality degradation while ensuring semantic consistency, and optimizing it through backpropagation during joint training.

[0048] (1) Set of differentiable imaging degradation operators: Suppose the original input image of the power transmission equipment is represented as:

[0049] Where H represents the image height, W represents the image width, and C represents the number of image channels. Construct a set of imaging degradation operators:

[0050] in, This represents the k-th imaging degradation operator. These are learnable control parameters corresponding to the degradation operator, used to control the degradation intensity or degradation form. In one specific embodiment, the imaging degradation operator includes a compression degradation operator, a noise degradation operator, a blur degradation operator, and a gradient-adversarial degradation operator.

[0051] Compression degradation operators, noise degradation operators, and blur degradation operators are used to simulate degradation in conventional image transmission or acquisition. The degradation form or intensity of each operator can be determined by learnable parameters. The control mechanism, which can be optimized via backpropagation during training, is simplified here; specific formulas are not provided for brevity. The gradient adversarial degradation operator is used to simulate adversarial perturbations that strongly interfere with the model's discrimination. It generates perturbation images based on gradient information from the current model parameters to characterize the impact of extreme degradation conditions on the model's robustness in complex scenarios. Its implementation is as follows:

[0052] Where sign(.) is the sign function. The original fault labels are the actual fault labels for the power transmission equipment images. This is the output of the current basic fine-tuning model. These are the model parameters for the main detection sub-model. This represents the gradient of the loss function L with respect to the input image.

[0053] (2) Degeneracy operator selection mechanism based on learnable parameters: To achieve adaptive modeling of degradation types, a set of learnable degradation operator selection parameters is introduced:

[0054] in, This represents the selection weight parameter for the k-th degenerate operator. The selection parameter is initialized to a preset range in the early stage of training to constrain the initial activation range of different degenerate operators.

[0055] The selection parameters are mapped to degenerate operator activation weights using a normalized mapping function:

[0056] in, In an alternative implementation, the Sigmoid mapping function can also be used to independently activate each degradation operator.

[0057] (3) Unified representation and joint optimization of imaging degradation modeling: Based on the aforementioned set of degradation operators and operator selection mechanism, the degradation modeling process of the input image x can be expressed as follows:

[0058] in, The image is a degraded simulation result. This represents the activation weight of the k-th degenerate operator. and All parameters are learnable and are jointly trained with subsequent low-rank modulation fine-tuning parameters during model fine-tuning. Through end-to-end backpropagation optimization, the generated degraded images not only retain the original semantics but also pose an effective challenge to the model in the feature space, thereby improving the model's robustness to complex degraded scenes.

[0059] In this way, the imaging degradation modeling module can adaptively select degradation types and intensities that better match the characteristics of real complex scenes during the training process, thereby improving the model's generalization ability and robustness to real degradation inputs.

[0060] It should be noted that the aforementioned imaging degradation operators include compression degradation operators, noise degradation operators, blurring degradation operators, and gradient adversarial degradation operators. Among them, the compression degradation operator, noise degradation operator, and blurring degradation operator are used to simulate degradation in conventional image transmission or acquisition. The degradation form or intensity of each operator can be determined by learnable parameters. It can be controlled and optimized through backpropagation during training.

[0061] Specifically, imaging degradation operators are mathematical models, algorithms, or functions used to simulate the quality degradation that may occur in images during acquisition, transmission, and storage. Their purpose is to enable power transmission equipment fault diagnosis models to learn robust characteristics for identifying faults under different degrees of degradation by introducing controlled degradation.

[0062] Compression degradation operators are designed to simulate the quality loss of images during transmission or storage caused by lossy compression (such as JPEG, WEBP, etc.), including blockiness and loss of detail. They can be implemented by applying standard lossy compression algorithms and adjusting their compression quality parameters to simulate different degrees of compression degradation. For example, by setting the JPEG quality factor from high to low, different scenarios ranging from slight to severe compression can be simulated.

[0063] Noise degradation operators are used to simulate random pixel value variations introduced into images during acquisition or transmission due to environmental interference (such as sensor noise and transmission channel noise). They can be implemented by adding common noise models such as Gaussian noise, salt-and-pepper noise, and Poisson noise. For example, Gaussian noise intensity can be controlled by specifying the mean and variance; salt-and-pepper noise can be controlled by specifying the noise density.

[0064] Blur degradation operators simulate image detail blurring caused by factors such as camera defocusing, motion blur, and atmospheric turbulence. This can be achieved by applying convolution kernels such as Gaussian blur kernels, motion blur kernels, and defocus blur kernels. For example, Gaussian blur can be controlled by adjusting the standard deviation of the Gaussian kernel; motion blur can be simulated by specifying the direction and length of motion.

[0065] Gradient adversarial degradation operators are a more advanced and targeted method of degradation simulation, designed to mislead the model through tiny, imperceptible perturbations. This degradation is typically generated based on the model's gradient information to maximize the likelihood of misjudgment, thereby improving the model's robustness against potential adversarial attacks. It can be implemented using gradient-based attack methods (such as FGSM and PGD), for example, by calculating the gradient of the loss function with respect to the input image and making small adjustments to pixel values ​​along the gradient direction to generate adversarial examples. This operator can simulate malicious attacks or image distortion under extreme conditions.

[0066] Each type of imaging degradation operator is equipped with learnable control parameters. These parameters are used to dynamically adjust the degradation intensity and / or form of the corresponding degradation operator. These parameters can be scalars, vectors, or matrices, and they serve as inputs to the internal functions of the degradation operator, controlling the degradation effect. For example, for Gaussian noise, It could be the standard deviation of the noise; for JPEG compression, These can be quality factors. During model training, these learnable control parameters α_K are optimized through backpropagation. Backpropagation is a commonly used machine learning model training method used to update model parameters based on the gradient of the loss function to minimize the loss function. In each training iteration, the loss between the model output and the true label is calculated, then the gradient of the loss with respect to all learnable parameters is calculated using the chain rule, and the parameters are updated in the reverse direction of the gradient using an optimizer (such as SGD, Adam, etc.).

[0067] This application introduces a degradation simulation sub-model, which dynamically generates diverse and realistic degradation simulation image samples based on a preset set of imaging degradation operators and learnable control parameters. This approach overcomes the limitation of insufficient coverage of preset degradation samples in traditional methods, allowing the power transmission equipment fault diagnosis model to encounter a wider range of degradation scenarios during training. In particular, by introducing degradation operator selection weights and learnable control parameters, the degradation simulation sub-model can adaptively adjust the degradation intensity and form, thereby generating more challenging or realistic degradation samples. This effectively improves the generalization ability and robustness of the power transmission equipment fault diagnosis model to images with different degrees and types of degradation when facing complex and changing actual operating conditions, ensuring the accuracy and reliability of the diagnostic results.

[0068] In some embodiments, when the steps of this method include a degenerate simulation sub-model, this application further proposes to optimize the parameters of the gated sub-model and the low-rank modulation sub-model based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results and in conjunction with the joint optimization objective function, until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition. Specifically, this includes: optimizing the parameters of the gated sub-model, the low-rank modulation sub-model, and the degenerate simulation sub-model based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results and in conjunction with the joint optimization objective function, until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition.

[0069] Among them, parameter optimization of the degradation simulation sub-model refers to dynamically adjusting the learnable control parameters within the degradation simulation sub-model during model training, based on the diagnostic performance of the backbone detection sub-model on normal inspection image samples and degradation simulation image samples. The degradation operator selects weight μ. The role of the degradation simulation sub-model is to generate corresponding degradation simulation image samples based on the original input inspection image sample x. By optimizing its parameters, the degradation simulation sub-model can generate more challenging samples that more closely resemble complex real-world degradation scenarios, thereby effectively improving the backbone detection sub-model's ability to learn more robust and generalizable features. Specifically, parameter optimization is typically achieved through the backpropagation algorithm. In each training iteration, the gradient of the loss function with respect to the parameters of the degradation simulation sub-model is calculated based on the diagnostic loss of the backbone detection sub-model on normal and degraded samples. Subsequently, the optimizer updates the parameters of the degradation simulation sub-model based on these gradients. This mechanism allows the degradation simulation sub-model to dynamically adjust its degradation generation strategy, for example, by increasing the simulation of a certain type or intensity of degradation that the diagnostic model currently performs poorly, or by adjusting the combined weights of different degradation operators to continuously provide beneficial training data for the diagnostic model.

[0070] By incorporating the degradation simulation sub-model into the parameter optimization process of the joint optimization objective function through the above technical solution, the degradation simulation sub-model can co-evolve with the gating sub-model and the low-rank modulation sub-model. Specifically, during training, the degradation simulation sub-model is no longer a fixed sample generator, but rather dynamically adjusts its strategy for generating degradation samples based on the diagnostic performance of the backbone detection sub-model. When the backbone detection sub-model performs poorly under certain degradation conditions, the optimization process prompts the degradation simulation sub-model to generate more such challenging degradation samples, thereby forcing the backbone detection sub-model to learn more robust feature representations. This joint optimization mechanism ensures that the degradation simulation sub-model can continuously provide "hard samples" for the diagnostic model, effectively improving the generalization ability and diagnostic accuracy of the power transmission equipment fault diagnosis model when facing various complex and unknown degradation situations, and avoiding performance bottlenecks caused by inconsistencies between degradation sample generation and the diagnostic model training objectives.

[0071] Building upon this foundation, to enhance the robustness of the power transmission equipment fault diagnosis model under complex degradation conditions, a gating sub-model, a low-rank modulation sub-model, and a degradation simulation sub-model were introduced, and the parameters of these sub-models were optimized. However, in actual training, effectively coordinating the learning objectives among the backbone detection model, the gating sub-model, the low-rank modulation sub-model, and the degradation simulation sub-model to ensure that the model achieves good diagnostic performance on both normal and degraded samples, while avoiding the model getting trapped in local optima or experiencing convergence difficulties during optimization, remains a problem that needs to be solved.

[0072] In this regard, this application further proposes that the joint optimization objective function is specifically as follows:

[0073] In the formula, As the backbone detection model, The original input inspection image sample, For the degraded simulated image samples, L is the loss function, and A and B represent the low-rank modulation parameters to be optimized in the low-rank modulation sub-model. This represents the gating parameters to be optimized in the gating sub-model. and This represents the learnable control parameters to be optimized and the weights for selecting the degradation operator in the degradation simulation sub-model.

[0074] The joint optimization objective function aims to simultaneously optimize the parameters of multiple sub-models by minimizing a comprehensive loss. Specifically, This indicates that the optimization process will target the low-rank modulation parameters A and B to be optimized in the low-rank modulation sub-model, the gating parameter g to be optimized in the gating sub-model, and the learnable control parameter α and the weights of the degradation operator to be optimized in the degradation simulation sub-model. This process is carried out. The joint optimization of these parameters ensures that the various sub-models can work together to improve the overall model performance. This represents the expectation for the input inspection image sample x, which means that the optimization process considers the average performance across the entire input data distribution, thus giving the model better generalization ability. This indicates the backbone detection model. The loss value (.) is obtained by processing the original input inspection image sample `x`. This loss ensures that the model can accurately diagnose faults when processing normal, non-degraded images. Indicates the backbone detection model (.) Degraded simulated image samples The loss value obtained after processing. This part of the loss forces the model to learn how to process and diagnose images under various degradation conditions, thereby enhancing the model's robustness. L(.) is the loss function, and its specific form can be selected according to the task type. For example, for classification tasks, it can be the cross-entropy loss function; for regression tasks, it can be the mean squared error loss function. Its role is to quantify the difference between the model's predicted results and the true labels. The low-rank modulation parameters A and B are the core components of the low-rank modulation sub-model. They perform targeted compensation of image features through low-rank matrix fine-tuning logic to adapt to different degradation conditions. The gating parameter g is the key to the gating sub-model. It dynamically controls the enabling state of the low-rank modulation sub-model based on image features and gating decision logic to achieve adaptive feature compensation. Learnable control parameters Weights for degeneracy operators These are key to the degradation simulation sub-models, enabling them to generate diverse and challenging degradation image samples, thus providing rich learning data for model training.

[0075] More specifically, this invention introduces a degradation-aware gated low-rank modulation fine-tuning into the feature extraction network of a lightweight detection model, which is used to achieve adaptive compensation for degraded image features while maintaining the original discriminative ability of the backbone model.

[0076] (1) General form of gated low-rank modulation fine-tuning of degradation perception: Suppose the fine-tuned backbone detection model consists of L cascaded sub-feature mapping layers, expressed as:

[0077] in, For the first Each sub-feature mapping layer, where x is the input original image of the power transmission equipment. This indicates that the original parameters of the backbone model are frozen during fine-tuning. " is the symbol for cascading connections in a network.

[0078] In any sub-layer of the backbone model In this context, a low-rank modulation branch is introduced, which is expressed as follows:

[0079] in, and For the low-rank modulation matrix pair of this layer, This is the feature dimension of this layer. For the rank dimension of low-rank modulation, we have , For the first Low-rank modulation increment parameters of the layer.

[0080] To achieve adaptive modulation for degradation sensing, this invention introduces a gating mechanism, expressed as follows:

[0081] in, The backbone detection model is in the layer above (i.e., the first layer) Output features of the layer Indicates the first The learnable mapping function corresponding to the layer, This represents the activation function. For the first The gating activation signal for the low-rank branch of the layer. In one implementation, In another implementation, to enable independent gating at each level, multiple sub-layers share the same gating signal.

[0082] In summary, the first The forward computation form of each sub-layer is represented as follows:

[0083] in, The first gated low-rank modulation branch was introduced into the backbone detection model. Sub-layers, This is a global or hierarchical scaling factor used to balance the relative influence of the original convolutional branch and the low-rank modulation branch.

[0084] In this embodiment, by jointly optimizing the degradation operator control parameters in the above steps and the low-rank modulation parameters and gating parameters in this step, conditional and controllable modulation for robust low-rank modulation fine-tuning is achieved. During the training process, the original image x and its corresponding degradation image are compared... The joint optimization objective is:

[0085] in, A backbone detection model with gated low-rank modulation branches was introduced. Let L be the degraded image obtained by modeling the degradation of the input image x, and let L be the loss function. and These are the low-rank modulation parameters and gating parameters. and The degradation operators defined in the above steps can be learned as control parameters and degradation operator selection parameters. During training, the original parameters of the backbone detection model... Keep frozen, parameters Joint optimization, which combines the degradation process with the model fine-tuning process, can further improve the robustness of the resulting model, thereby enhancing the accuracy of fault diagnosis for power transmission equipment in practical applications.

[0086] More specifically, the loss function L( ) is the task loss function used to constrain the difference between the model output and the target label. Its specific form can be set according to the type of power transmission equipment fault diagnosis task.

[0087] In a preferred embodiment, when the power transmission equipment fault diagnosis task is a classification or state recognition task, the loss function L( The cross-entropy loss function is used, and its expression is:

[0088] In the formula, Indicates the number of fault categories or equipment status categories. Indicates the input sample at the th Real labels on the class, Indicates the backbone detection model Inspection image samples of the original input Belongs to the The predicted probability of a class.

[0089] In another preferred embodiment, when the power transmission equipment fault diagnosis task is a target detection, defect location, or area discrimination task, the loss function L( The location regression loss function is used to constrain the difference between the model-predicted fault location and the actual labeled location. Its expression is:

[0090] In the formula, This indicates the number of faulty targets or defective regions participating in the training. Indicates the first The true location parameters of a faulty target or defective area. Indicates the backbone detection model For the first Predicted location parameters for a target or region express Norm, used to measure the regression error between predicted and actual locations.

[0091] Through the above structural design, the low-rank modulation branch only affects the gating parameters. Activation affects the feature mapping, thereby enabling conditional and controllable modulation of the feature extraction process.

[0092] Furthermore, in terms of gating parameters When convergence reaches a small value, the impact of low-rank branches on output features becomes negligible, at which point the model is equivalent to the original network structure; while when When activated, the low-rank branch adaptively modifies the feature map for a specific input distribution or task condition.

[0093] (2) Gated low-rank adaptive modulation structure for convolutional neural networks: In a preferred embodiment, the target lightweight model is a convolutional neural network structure. Addressing the issues of concentrated parameters, strong spatial locality, and difficulty in directly adapting to standard fully connected low-rank modulation fine-tuning structures in convolutional operators, this invention further proposes a variant of a low-rank adaptive modulation structure for convolutional layers to achieve efficient and controllable modulation of the convolutional feature mapping process.

[0094] Specifically, for any sublayer in a convolutional neural network Without changing the original convolutional layer parameters Under this premise, the present invention introduces a parallel low-rank convolutional branch in this layer and adjusts its intensity through the proposed gating mechanism. The overall mapping relationship is defined as follows:

[0095] in, Indicates the first The input feature map for the convolutional operation. Indicates the first The output feature map of the convolutional operation. Indicates the first The original convolution mapping function of the layer, its parameters Keep frozen during training; and They represent the first The first and second convolutional kernels in the low-rank convolutional modulation branch constitute a low-rank decomposition of the original convolutional weight space. Their kernel size is smaller than or equal to the original kernel size, thus significantly reducing the number of trainable parameters. For the first Low-rank modulation increment parameters of the layer. Indicates the use of convolution kernel For input features Perform convolution operations. This is a convolution operation.

[0096] Accordingly, after completing the differentiable imaging degradation modeling module and the degradation-aware low-rank modulation fine-tuning, a robust and lightweight power transmission equipment fault diagnosis model is obtained. For any input image x, the model first obtains intermediate feature representations through the feature extraction backbone network, and then adaptively adjusts the activation level of the low-rank modulation branch according to the degradation features of the input image, finally outputting the corresponding fault detection result.

[0097] In one specific implementation, the overall mapping relationship of the model can be represented as: This represents an input image of power transmission equipment. A gated low-rank modulation branch was introduced into the fine-tuned detection model. This indicates the frozen backbone network parameters. This represents the low-rank modulation increment parameter introduced through the low-rank modulation branch. For gating parameters, This represents the fault detection results output by the model.

[0098] When the input image is a raw image with a low degree of degradation, the gating parameters The adaptive value is relatively small, and the model mainly relies on the backbone parameters. To complete the discrimination, the original detection performance is maintained; when the input image is a degraded image, the gating parameters are adjusted. The adaptive value is larger, and the low-rank modulation branch participates in feature modeling to achieve targeted compensation for degraded features.

[0099] This embodiment uses various differentiable degradation operators to parametrically model the degradation form and intensity of the input image, and jointly optimizes the degradation process and the model fine-tuning process. Through the above methods, the fine-tuned model can output stable and reliable power transmission equipment fault diagnosis results for both the original image and images in complex degradation scenarios under end-side deployment conditions, thereby improving the robustness and practicality of power transmission equipment detection in complex application environments.

[0100] The above is a detailed description of an embodiment of a fault diagnosis modeling method for power transmission equipment provided in this application. The following is a detailed description of related embodiments of a fault diagnosis modeling device, terminal, and storage medium for power transmission equipment provided in this application.

[0101] Please see Figure 2 This application provides an embodiment of a fault diagnosis modeling device for power transmission equipment, the device comprising: The image sample acquisition unit 201 is used to acquire a preset inspection image sample combination, wherein the inspection image sample combination includes a pair of normal inspection image samples and degradation simulation image samples. The initial model optimization unit 202 is used to input the inspection image sample combination into the initial power transmission equipment fault diagnosis model according to the inspection image sample combination, and to train the initial power transmission equipment fault diagnosis model in combination with the preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model and a gating sub-model. The backbone detection sub-model is used to extract image features from the input image and then determine the fault diagnosis result of the power transmission equipment based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model according to the image features and the preset gating decision logic. The low-rank modulation sub-model is used to perform directional compensation of image features through low-rank matrix fine-tuning logic. The model optimization termination determination unit 203 is used to obtain the power transmission equipment fault diagnosis model based on the convergence degree of the initial power transmission equipment fault diagnosis model. When the convergence degree meets the preset training termination condition, the model is obtained.

[0102] In practical implementation, the image sample acquisition unit first selects high-quality normal inspection image samples from historical inspection data, and generates corresponding degraded simulated image samples through a preset set of imaging degradation operators, forming paired sample combinations to simulate image quality degradation that may occur in real-world scenarios. The initial model optimization unit inputs this sample combination into the initial power transmission equipment fault diagnosis model. The backbone detection sub-model processes the normal and degraded simulated image samples sequentially, extracting their respective image features. Based on the extracted image features, the gating sub-model calculates the degradation perception result through a preset gating decision logic and outputs the corresponding gating value. This gating value dynamically determines whether the low-rank modulation sub-model is enabled. When the gating value meets the trigger condition, the low-rank modulation sub-model uses low-rank matrix fine-tuning logic to perform directional compensation on the image features to correct the feature shift caused by image degradation. The model optimization termination decision unit continuously monitors the change in the loss value during training. When the loss value converges to a preset threshold or reaches the maximum number of iterations, training is terminated and the final power transmission equipment fault diagnosis model is output. Through the above technical solution, the device can effectively cope with various degradation interferences introduced during the deployment and transmission of inspection images at the end side, ensuring that the model maintains stable feature extraction capability and high-precision fault diagnosis performance for degraded images in practical applications.

[0103] Furthermore, such as Figure 3 As shown, this application proposes a modeling terminal for power transmission equipment fault diagnosis, comprising a memory 33 and a processor 31. The memory 33 and processor 31 are connected via a communication bus 34. The memory is a hardware component in the terminal used for persistent or temporary storage of data and program code. It may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). In the power transmission equipment fault diagnosis modeling terminal, the memory is mainly used to store program code, i.e., the instruction set for implementing the aforementioned power transmission equipment fault diagnosis modeling method, as well as various types of data generated during method execution, such as intermediate data, training samples, model parameters, and optimization objective functions. The capacity and read / write speed of the memory directly affect the efficiency of model training and the overall performance of the terminal.

[0104] The processor is the core computing unit of the terminal, responsible for interpreting and executing the program code stored in memory, as well as performing various data processing and computation tasks. The processor can be one or more central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), or dedicated neural network processors (NPUs), etc. In the power transmission equipment fault diagnosis modeling terminal, the processor drives the entire modeling method by executing program code, including image feature extraction, collaborative work between sub-models, loss calculation, parameter backpropagation, and optimization.

[0105] This application provides a computer-readable storage medium containing program code, which is read and executed by a processor to implement the above-mentioned modeling method for fault diagnosis of power transmission equipment.

[0106] Specifically, a computer-readable storage medium refers to any type of physical device capable of storing digital data or computer program instructions. Its function is to provide persistent storage space for program code, ensuring that the code is not lost after power failure and can be read by the processor when needed. Common computer-readable storage media include, but are not limited to: hard disk drives, solid-state drives, read-only memory, random access memory, flash memory, and various forms of non-volatile memory. In this application, the medium is used to store program code implementing a modeling method for fault diagnosis of power transmission equipment.

[0107] Program code refers to a set of instructions written in one or more programming languages ​​that can be understood and executed by a computer processor. In this application, the program code specifically includes all the logical steps and algorithms for implementing the aforementioned modeling method for power transmission equipment fault diagnosis, such as: the logic for acquiring combinations of inspection image samples, the logic for constructing a degradation simulation sub-model, the logic for training the initial power transmission equipment fault diagnosis model (including feature extraction, gating determination, low-rank compensation, loss calculation, and parameter optimization), and the logic for determining the training termination condition. After compilation or interpretation, this code can guide the processor to complete the entire modeling process.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the terminals, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed terminals, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0110] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0111] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0112] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0113] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0114] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A modeling method for fault diagnosis of power transmission equipment, characterized in that, include: Obtain a preset combination of inspection images, wherein the combination of inspection images includes pairs of normal inspection images and degraded simulated images. Based on the inspection image sample combination, the inspection image sample combination is input into the initial power transmission equipment fault diagnosis model. The initial power transmission equipment fault diagnosis model is trained by combining the preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model is used to extract image features from the input image and then determine the power transmission equipment fault diagnosis result based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model based on the image features and the preset gating decision logic. The low-rank modulation sub-model is used to perform directional compensation of the image features through low-rank matrix fine-tuning logic. Based on the convergence degree of the initial power transmission equipment fault diagnosis model, when the convergence degree meets the preset training termination condition, the power transmission equipment fault diagnosis model is obtained.

2. The modeling method for fault diagnosis of power transmission equipment according to claim 1, characterized in that, The process of inputting the combined inspection image samples into the initial power transmission equipment fault diagnosis model and training the initial power transmission equipment fault diagnosis model using a preset joint optimization objective function specifically includes: The combined inspection image samples are input into the initial power transmission equipment fault diagnosis model; Based on the normal inspection image samples and the degraded simulation image samples included in the inspection image sample combination, and combined with the backbone detection sub-model in the initial power transmission equipment fault diagnosis model, the corresponding image features are extracted from the normal inspection image samples and the degraded simulation image samples respectively through the feature extraction logic of the backbone detection sub-model. Based on the image features, the degradation perception result of the image features is determined by the gated sub-model in the initial power transmission equipment fault diagnosis model. Then, based on the degradation perception result and the gated decision logic in the gated sub-model, the gated value corresponding to the image features is determined. The gated value is used to control the enable state of the low-rank modulation sub-model. When the gate value satisfies the triggering condition of the low-rank modulation sub-model, the image features are directionally compensated through the low-rank matrix fine-tuning logic in the low-rank modulation sub-model to obtain the compensated image features. The fault determination logic of the main detection sub-model outputs two sets of image features corresponding to the power transmission equipment fault diagnosis results. Based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model and the low-rank modulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition.

3. The modeling method for fault diagnosis of power transmission equipment according to claim 2, characterized in that, Before obtaining the preset combination of inspection image samples, the following steps are also included: Based on a preset set of imaging degradation operators and combined with degradation operator selection weights, a degradation simulation sub-model of the input image and the degradation simulation image is constructed. The degradation operator selection weights are used to constrain the activation range of different imaging degradation operators, and each imaging degradation operator is configured with learnable control parameters for controlling degradation intensity and / or degradation form. Obtain a preset inspection image sample, and use the inspection image sample as the input image of the degradation correlation sub-model. Through the processing of the degradation correlation sub-model, obtain the degradation simulation image sample corresponding to the inspection image sample. The inspection image samples are combined with the corresponding degradation simulation image samples to obtain several sets of inspection image sample combinations.

4. The modeling method for fault diagnosis of power transmission equipment according to claim 3, characterized in that, The specific degradation simulation sub-model is as follows: In the formula, For the degraded simulated image sample, Activate weights for the degenerate operator. For the k-th imaging degradation operator, The original input inspection image sample, The learnable control parameters are those corresponding to the k-th imaging degradation operator. This is a sequence consisting of weights selected from the degradation operators corresponding to k imaging degradation operators.

5. The modeling method for fault diagnosis of power transmission equipment according to claim 3, characterized in that, Based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model and the low-rank modulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination conditions, specifically including: Based on the loss values ​​of the two sets of power transmission equipment fault diagnosis results, and combined with the joint optimization objective function, the parameters of the gated sub-model, the low-rank modulation sub-model, and the degradation simulation sub-model are optimized until the convergence of the initial power transmission equipment fault diagnosis model meets the preset training termination condition.

6. The modeling method for fault diagnosis of power transmission equipment according to claim 5, characterized in that, The joint optimization objective function is specifically as follows: In the formula, This refers to the backbone detection model. The original input inspection image sample, Let L be the degraded simulated image sample, L be the loss function, and A and B represent the low-rank modulation parameters to be optimized in the low-rank modulation sub-model. This represents the gating parameters to be optimized in the gated sub-model. and This represents the learnable control parameters to be optimized and the weights for selecting the degradation operator in the degradation simulation sub-model.

7. The modeling method for fault diagnosis of power transmission equipment according to claim 3, characterized in that, The imaging degradation operators include: compression degradation operator, noise degradation operator, blur degradation operator, and gradient adversarial degradation operator.

8. A modeling device for fault diagnosis of power transmission equipment, characterized in that, include: An image sample acquisition unit is used to acquire a preset inspection image sample combination, wherein the inspection image sample combination includes pairs of normal inspection image samples and degradation simulation image samples. An initial model optimization unit is used to input the inspection image sample combination into an initial power transmission equipment fault diagnosis model based on the inspection image sample combination, and to train the initial power transmission equipment fault diagnosis model in combination with a preset joint optimization objective function. The initial power transmission equipment fault diagnosis model includes: a backbone detection sub-model, a low-rank modulation sub-model, and a gating sub-model. The backbone detection sub-model is used to extract image features based on the input image, and then determine the power transmission equipment fault diagnosis result based on the image features. The gating sub-model is used to control the enabling state of the low-rank modulation sub-model based on the image features and a preset gating decision logic. The low-rank modulation sub-model is used to perform directional compensation of the image features through low-rank matrix fine-tuning logic. The model optimization termination determination unit is used to obtain the transmission equipment fault diagnosis model based on the convergence degree of the initial transmission equipment fault diagnosis model, when the convergence degree meets the preset training termination condition.

9. A modeling terminal for fault diagnosis of power transmission equipment, characterized in that, include: Memory and processor; The memory is used to store program code, which corresponds to the modeling method for fault diagnosis model of power transmission equipment as described in any one of claims 1 to 7; The processor is used to read and execute the program code to implement the modeling method for fault diagnosis of power transmission equipment.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that is read and executed by a processor to implement the power transmission equipment fault diagnosis modeling method as described in claims 1 to 7.

Images