Power equipment image processing method and system based on quantum image coding algorithm

Abstract

The application provides a power equipment image processing method and system based on a quantum image coding algorithm, comprising the following steps: S1, acquiring a multi-modal image of power equipment; S2, preprocessing the multi-modal image to obtain standardized image data; S3, based on the structural characteristics of the power equipment, adaptively dividing the standardized image data into a plurality of sub-image blocks, and mapping the pixel information of each sub-image block into quantum state information through angle coding; S4, calling a quantum image coding algorithm, initializing a quantum circuit, taking the maximum fidelity of the output state of the quantum circuit and the target sub-image block quantum state as the goal, generating an optimal quantum circuit, and extracting the hidden space quantum state features of each sub-image block based on the optimal quantum circuit; and S5, fusing the hidden space quantum state features of all the sub-image blocks in a quantum state direct product manner to construct a global feature quantum state, and inversely converting the global feature quantum state into classical enhanced image data through a quantum decoder.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically, to a method and system for image processing of power equipment based on quantum image coding algorithms. Background Technology

[0002] With the continuous expansion of the power grid and the surge in the number of power equipment, building a safe, stable, and efficient new power system has become the core goal of the industry's development. In the field of power operation and maintenance, the traditional "manual inspection-based" model is limited by factors such as high labor costs, long inspection cycles, and strong subjectivity, making it difficult to meet the urgent need for accurate identification and efficient response to massive equipment defects.

[0003] Currently, image processing technology for power equipment has gradually transitioned from traditional image processing to classic deep learning algorithms. Mainstream solutions typically utilize convolutional neural networks (CNNs) to process multimodal images (visible light, infrared, etc.) captured by drones or robots, achieving defect identification through pixel-level feature extraction. However, in practical engineering applications and complex and ever-changing operation and maintenance scenarios, existing technologies still face many insurmountable core problems:

[0004] First, the feature extraction accuracy is insufficient in complex scenarios. Power equipment is often outdoors, and inspection images are easily affected by environmental factors such as strong light reflection, rain and fog, and electromagnetic interference. Moreover, key defects such as micro-cracks in insulators, localized hot spots, and partial discharge traces inside GIS systems often exhibit minute and non-obvious features. Traditional classic algorithms mainly rely on local convolution operations, which have a limited field of view and make it difficult to capture the global correlation features between equipment structure and defects. This results in a high rate of missed and false detections of defects in interference environments.

[0005] Secondly, the generalization ability of small sample defect identification is weak. In power systems, critical faults such as transformer oil leakage and severe insulation aging are scarce samples in a "long-tail distribution". Classic deep learning algorithms usually rely on massive labeled data for driving. Due to the insufficient coverage of the distribution of critical defect data, the model often struggles to learn robust feature representations, resulting in poor generalization ability and an inability to effectively adapt to the detection needs of multiple types of equipment and multiple scenarios.

[0006] Secondly, the fusion efficiency of multimodal data is low. Modern inspection data encompasses heterogeneous modal information such as visible light, infrared thermal imaging, and acoustic signatures. Traditional data fusion methods often remain at a superficial level of splicing or simple weighting, making it difficult to deeply explore the hidden physical relationships between modalities. This not only leads to feature redundancy but also results in low processing efficiency, failing to meet the requirements of smart substations for real-time monitoring and rapid response.

[0007] Finally, high-dimensional data processing faces a computational bottleneck. With the widespread adoption of high-definition inspection equipment, the dimensionality of image data has increased dramatically. Classical algorithms experience exponentially increasing computational complexity when processing high-resolution images, resulting in high computational consumption and slow inference speed. This makes them unsuitable for edge inspection equipment with limited computing power, hindering the progress of lightweight deployment.

[0008] To address some of the aforementioned issues, some attempts have emerged in the prior art to incorporate cutting-edge computing theories. For example, Chinese Patent Publication No. CN202510448666.1 discloses a method and system for detecting defects in power equipment. This approach attempts to combine quantum neural networks with classical convolutional networks, enhancing feature representation through quantum convolution processing, and combining multi-scale feature fusion, C3STR structures, and window self-attention mechanisms to enhance semantic information and feature representation, aiming to improve detection performance on edge devices.

[0009] However, despite the introduction of quantum computing concepts into the aforementioned existing technologies to enhance feature representation, several challenges remain to be addressed in the face of ultra-high-resolution images and extremely small defects unique to power equipment. These challenges include how to further reduce the demand for qubits in quantum circuits, how to design better quantum coding strategies to reduce computational redundancy while preserving high-dimensional features, and how to more deeply utilize quantum entanglement properties to solve the global correlation problem of multimodal heterogeneous data. Therefore, there is an urgent need to develop a novel image processing method that can balance feature extraction accuracy, small-sample generalization ability, and computational efficiency. Summary of the Invention

[0010] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for image processing of power equipment based on quantum image coding algorithm.

[0011] According to one aspect of the present invention, a method for processing images of power equipment based on a quantum image coding algorithm includes:

[0012] Step S1: Acquire multimodal images of the power equipment;

[0013] Step S2: Preprocess the multimodal image to obtain standardized image data;

[0014] Step S3: Based on the structural characteristics of the power equipment, the standardized image data is adaptively divided into blocks to obtain several sub-image blocks. The pixel information of each sub-image block is mapped to quantum state information through angle encoding.

[0015] Step S4: Call the quantum image coding algorithm to initialize the quantum circuit. With the goal of maximizing the fidelity between the output state of the quantum circuit and the quantum state of the target sub-image block, generate the optimal quantum circuit. Based on the optimal quantum circuit, extract the hidden space quantum state features of each sub-image block.

[0016] Step S5: The hidden space quantum state features of all sub-image blocks are fused using the quantum state direct product method to construct a global feature quantum state. The global feature quantum state is then converted into classical enhanced image data using a quantum decoder.

[0017] Step S6: Input the classical enhanced image data into the classical deep learning model to complete the classification and localization of power equipment defects, and output the defect identification results.

[0018] Preferably, the multimodal images in step S1 include visible light images, infrared images, and partial discharge associated images.

[0019] Preferably, the preprocessing in step S2 includes:

[0020] Noise filtering: An adaptive thresholding method is used to filter noise, wherein the threshold is calculated as follows:

[0021]

[0022] In the formula, μ is the mean pixel value of the image, σ is the standard deviation of the pixels, and α and β are adjustment coefficients that are adaptively adjusted according to the modality type;

[0023] Size normalization: Images of different modalities and resolutions are uniformly scaled to a preset resolution of S×S pixels using bilinear interpolation.

[0024] Modal alignment operation: Based on the SIFT algorithm, device contour feature points of each modal image are extracted, and the RANSAC algorithm is used to achieve feature point matching and spatial registration of multimodal images, controlling the registration error within 2 pixels.

[0025] Preferably, step S3 includes:

[0026] S3.1: The standardized image data is divided into blocks to obtain a set of sub-image blocks with consistent specifications. The block size is determined uniformly based on the size of the standardized image data and the structural complexity of the power equipment.

[0027] S3.2: Normalize the pixel values ​​of each sub-image block, mapping the pixel grayscale values ​​to the [0, π] interval, wherein the normalization formula is:

[0028]

[0029] Where p is the original pixel value of the sub-image patch. , where are the minimum and maximum pixel values ​​of the sub-image block, respectively, and θ is the rotation angle after mapping;

[0030] S3.3: Encode angular information into a quantum state by rotating a single qubit around the y-axis to obtain quantum state information, wherein the encoded quantum state information satisfies:

[0031]

[0032] In the formula, For a single-qubit rotation gate around the y-axis, This represents the initial state of the qubits, where the number of qubits equals the number of pixels in the sub-image block.

[0033] Preferably, step S4 includes:

[0034] Sub-step S4.1: Set the core control parameters of the quantum image coding algorithm, and initialize the quantum circuit as a unit operator I, wherein the control parameters include the initial number of unit operators of the quantum circuit. Maximum number of unit operators Unit operator increment Number of scans N;

[0035] Sub-step S4.2: Define the fidelity function between the quantum circuit output state and the quantum state of the target sub-image block:

[0036]

[0037] in, For current quantum circuits, For current quantum circuits, For a two-qubit unit operator, The Hermitian conjugation of the quantum circuit, <0| represents the initial product state of the quantum circuit. Quantum state information for sub-image patches;

[0038] Sub-step S4.3: With the goal of maximizing the fidelity function, a forward-backward bidirectional scanning strategy is adopted to optimize each two-qubit unit operator, calculate the optimized global fidelity, and repeat N times;

[0039] Sub-step S4.4: Repeat sub-step S4.3 until the maximum number of unit operators is reached. Alternatively, if the change in the fidelity value calculated after each optimization is less than the threshold, the optimal quantum circuit can be constructed based on the final iteration result.

[0040] Sub-step S4.5: Based on the optimal quantum circuit, process the target sub-image patch. quantum state information Perform feature extraction and output the latent space quantum state features;

[0041] Sub-step S4.6: Repeat sub-steps S4.1 to S4.5 to obtain the hidden space quantum state features of all sub-image blocks.

[0042] Preferably, in sub-step S4.3, optimizing each two-qubit unit operator includes:

[0043] Construct a fidelity tensor operator, perform singular value decomposition on its matrix form, and take the unitary matrix corresponding to the largest singular value as the optimal two-qubit unit operator.

[0044] The optimal two-qubit unit operator is decomposed into a single-qubit rotation gate and a controlled NOT gate, which are used to construct quantum circuits.

[0045] Preferably, in step S4.3, the forward-backward bidirectional scanning strategy specifically includes:

[0046] Forward scan: Iterate through each unit operator in the quantum circuit in sequence, optimize the current unit operator, and update the unit operator parameters;

[0047] Reverse scan: Iterative optimization of the unit operators with updated parameters in reverse order.

[0048] Preferably, step S5 includes:

[0049] Sub-step S5.1: Fuse the latent space quantum state features of all sub-image patches through quantum state direct product. The global feature quantum state expression is:

[0050]

[0051] Where n is the number of blocks, |ψ i > represents the hidden space quantum state feature of the i-th sub-image block, and dots represents the omitted content in the middle;

[0052] Sub-step S5.2: Construct a quantum decoder, using inverse angle encoding to convert the global feature quantum state into classical image data. The decoding expression is:

[0053]

[0054] Where p is the reconstructed classical pixel value, and θ is the rotation angle corresponding to the quantum state.

[0055] Preferably, step S6 includes:

[0056] Step S6.1: Input the classic enhanced image data into the ResNet-50 classification model and the Faster R-CNN region proposal network to form an end-to-end recognition architecture;

[0057] Step S6.2: Output the defect identification result through the architecture, wherein the identification result includes the defect type, defect location coordinates and confidence parameters.

[0058] According to another aspect of the present invention, a power equipment image processing system based on a quantum image coding algorithm includes:

[0059] Module M1: Acquires multimodal images of power equipment;

[0060] Module M2: Preprocesses the multimodal image to obtain standardized image data;

[0061] Module M3: Based on the structural characteristics of power equipment, the standardized image data is adaptively divided into several sub-image blocks, and the pixel information of each sub-image block is mapped to quantum state information through angle encoding;

[0062] Module M4: Calls the quantum image encoding algorithm to initialize the quantum circuit. With the goal of maximizing the fidelity between the output state of the quantum circuit and the quantum state of the target sub-image block, it generates the optimal quantum circuit and extracts the hidden space quantum state features of each sub-image block based on the optimal quantum circuit.

[0063] Module M5: It uses the quantum state direct product method to fuse the hidden space quantum state features of all sub-image blocks to construct a global feature quantum state, and uses a quantum decoder to inversely convert the global feature quantum state into classical enhanced image data;

[0064] Module M6: Inputs the classical enhanced image data into the classical deep learning model to complete the classification and localization of defects in power equipment and outputs the defect identification results.

[0065] Preferably, the multimodal images in module M1 include visible light images, infrared images, and partial discharge associated images.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] The algorithm is deeply adapted to the specific scenarios, addressing the challenges of minute defects in power equipment and complex environments. It optimizes the fidelity calculation and quantum gate allocation of the quantum coding algorithm. By decomposing the large graph into low-dimensional sub-blocks, the computational power requirements of quantum hardware are reduced, ensuring accurate capture of unique defects.

[0068] The collaborative design of unified segmentation and angle encoding adaptively determines the segmentation size based on the structural complexity of the equipment (such as transformers and transmission lines), avoiding fusion deviations caused by inconsistent sizes. At the same time, angle encoding is used to convert pixels into quantum rotation angles, balancing processing complexity and information integrity.

[0069] The hybrid architecture of "quantum feature extraction and classical decision making" leverages quantum direct product and entanglement properties to uncover global correlations between sub-images, overcoming the local limitations of traditional algorithms. The extracted features, after being restored through inverse angle encoding, can be seamlessly integrated with mature deep learning models (such as ResNet-50) for efficient collaboration.

[0070] The robust anti-interference mechanism of Hilbert space utilizes the probability distribution characteristics of quantum states in Hilbert space to automatically filter out noise such as electromagnetic interference and illumination changes through maximizing fidelity optimization. This makes feature extraction more robust and adaptable to harsh outdoor inspection conditions.

[0071] Significantly improve defect identification accuracy: By capturing global features through quantum superposition and entanglement, and combining block-based enhancement of local details, it can effectively identify minute hidden defects and significantly reduce the false positive and false negative rates.

[0072] Enhancing generalization ability under small sample sizes: By utilizing the exponential high-dimensional representation ability of quantum space, it can still accurately characterize the defect distribution pattern and improve model adaptability even in the absence of a large number of labeled samples (such as scarce insulation aging samples).

[0073] Efficient multimodal heterogeneous data fusion quantum entanglement can directly establish a correlation mapping between infrared temperature, visible light structure and partial discharge energy characteristics, achieving multi-dimensional information enhancement without complex preprocessing and improving overall monitoring efficiency.

[0074] Its strong environmental adaptability and stability fidelity optimization mechanism can effectively suppress interference from rain, fog, obstruction, and strong electromagnetic fields, ensuring stable output of high-quality features even in harsh environments and meeting all-weather operation and maintenance needs.

[0075] The solution boasts broad engineering application compatibility and strong versatility, supporting various devices such as transformers, GIS equipment, and insulators. It is compatible with a wide range of common inspection modes and can be flexibly deployed without redesigning the architecture. Attached Figure Description

[0076] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0077] Figure 1 This is a schematic diagram of an image processing method for power equipment based on a quantum image coding algorithm. Detailed Implementation

[0078] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0079] Example 1:

[0080] To address the technical challenges of traditional classical algorithms in power equipment image processing, such as inaccurate defect feature extraction in complex scenarios, weak generalization ability for small sample defect recognition, low efficiency of multimodal data fusion, and significant computational bottlenecks in high-dimensional data processing, this application provides a power equipment image processing scheme based on a quantum image coding algorithm (automatic quantum encoder). Through the deep integration of quantized feature extraction and block singular value decomposition techniques, high-precision defect feature capture, strong anti-interference feature enhancement, and efficient intelligent recognition of multimodal images of power equipment (visible light, infrared, partial discharge associated images, etc.) are achieved, providing technical support for the operation and maintenance of new power system equipment.

[0081] This application constructs a complete power equipment image processing system based on sub-image coding algorithms and block singular value decomposition technology, encompassing "multimodal data preprocessing → adaptive block partitioning and quantum state conversion → quantum image coding algorithm feature extraction → block feature fusion and reconstruction → intelligent defect identification." The core lies in mining global correlation features of images through the superposition and entanglement characteristics of quantum states, combined with block processing to reduce the complexity of quantum circuits, ultimately achieving a dual improvement in defect identification accuracy and processing efficiency. Figure 1 As shown, the specific technical solution is as follows:

[0082] Multimodal image preprocessing for power equipment: For the input multimodal data such as visible light, infrared, and partial discharge related images, noise filtering, size normalization and modal alignment operations are performed in sequence to eliminate environmental interference and data heterogeneity and obtain standardized image data;

[0083] Adaptive Blocking and Quantum State Transition: Based on the structural characteristics of power equipment (such as transformer partitioning, transmission line segmentation, and GIS equipment cavity division), adaptive blocking is performed to decompose large-size images into several sub-image blocks. The pixel information of each sub-image block is mapped to quantum state information through amplitude encoding, reducing the qubits requirement of a single quantum circuit and ensuring the integrity of the quantum state representation.

[0084] Quantum image coding algorithm feature extraction: The quantum circuit is initialized by calling the quantum image coding algorithm. With the maximization of fidelity as the objective function, the fidelity tensor matrix is ​​decomposed by singular value decomposition (SVD). The form and position of the two-qubit unit operators are optimized in turn. The optimal quantum coding circuit adapted to each sub-image block is automatically generated, the hidden space quantum state features are extracted, and redundant information and noise interference are filtered out.

[0085] Block feature fusion and reconstruction: The latent space quantum state features of all sub-image blocks are fused using the quantum state direct product method to construct a global feature quantum state. The quantum state information is then converted into classical image data through a quantum decoder to achieve defect feature enhancement and image reconstruction.

[0086] Intelligent Defect Recognition: The reconstructed and enhanced image is input into a classic deep learning model to classify and locate defects in power equipment, and output the defect type, location coordinates, and confidence score.

[0087] The specific implementation steps of this plan are as follows:

[0088] The first step is multimodal image preprocessing. Inputting multimodal images of the power equipment (visible light, infrared, and partial discharge related images) and considering the noise characteristics of different modes (e.g., environmental noise in visible light images, thermal noise in infrared images, and electromagnetic interference noise in partial discharge images), an adaptive thresholding method is used to dynamically adjust filtering parameters to filter out invalid interference information. The threshold calculation formula is as follows: ,in The average pixel value of the image. The standard deviation of pixels. , To adjust the coefficients, adaptive adjustments can be made based on the modality type, ensuring that different noise types are effectively filtered while preserving the original information of the device itself and potential defects; images of different modalities and resolutions are uniformly scaled to a preset resolution using bilinear interpolation. Pixels are used to ensure consistency between subsequent block processing and quantum coding, avoiding feature extraction deviations caused by resolution differences. The interpolation formula is:

[0089]

[0090] in For interpolated neighborhood pixel coordinates, The normalized pixel values ​​are used for interpolation, which maintains the continuity of image details during scaling and avoids distortion of defect features. Based on the SIFT algorithm, device contour feature points are extracted from each modality image. The RANSAC algorithm is used to achieve feature point matching and spatial registration of multimodal images, controlling the registration error within 2 pixels. Registration ensures that the spatial positions of devices in different modal images correspond one-to-one, enabling subsequent fusion processing to accurately associate multidimensional information of the same device region, laying a spatial consistency foundation for multimodal feature fusion.

[0091] The second step involves unified segmentation and angle-coded quantum state transformation, determining the unified segmentation size based on the image size and structural complexity of the power equipment. The specific segmentation strategy is as follows: First, calculate the number of suitable segments based on the image size to ensure that no image areas are missed or overlapped after segmentation. Then, verify the rationality of segmentation by considering the complexity of the equipment structure. For complex equipment (such as transformers and GIS equipment), the segmentation size must ensure that key component areas are not fragmented. For simple equipment (such as transmission lines and insulators), the segmentation size must balance processing efficiency and feature integrity. Finally, a unique and unified segmentation size is determined. Number of blocks This yields a set of sub-image patches with consistent specifications. The unified block design simplifies the adaptation difficulty of quantum circuits and improves the efficiency and accuracy of subsequent feature fusion.

[0092] Each sub-image block is encoded using angle coding. The pixel value is mapped to the rotation angle of the quantum state. A quantum register is used to convert classical pixel data into quantum state information. First, the pixel value is normalized, mapping the pixel grayscale value to... The interval (the angle range adapted to a single-qubit rotation gate), wherein the normalization formula is:

[0093]

[0094] in, For sub-image patches The original pixel values, These are the minimum and maximum pixel values ​​of the sub-image block, respectively. This is the rotation angle after mapping;

[0095] Angle information is encoded into a quantum state using a single-qubit rotation gate, where the encoded quantum state information satisfies:

[0096]

[0097] in, For a single-qubit rotation gate around the y-axis, The initial state of a qubit, the number of qubits (Each pixel corresponds to a rotation angle encoding of a quantum bit). This encoding method directly converts the pixel grayscale features into the phase information of the quantum state without relying on the high-dimensional quantum state amplitude distribution, which reduces the implementation difficulty of quantum circuits and ensures the complete mapping of feature information of sub-image blocks.

[0098] The third step involves feature extraction based on the quantum image coding algorithm, setting the core control parameters of the quantum image coding algorithm, including the initial number of unit operators. Maximum number of unit operators Unit operator increment Number of scans Initialize the quantum circuit as a unit operator This ensures that the initial state is free from interference from additional quantum gate operations, providing a clean initial baseline for subsequent optimization;

[0099] The objective function is the fidelity between the output state of the quantum circuit and the quantum state of the target sub-image block. The fidelity is defined as the degree of overlap between the generated state of the quantum circuit and the original quantum state of the sub-image block, expressed as: ,in This is the initial product state of the quantum circuit. For current quantum circuits, For sub-image patches quantum state information,

[0100] By maximizing this fidelity, the quantum circuit can accurately replicate the key features of the sub-image blocks, and the fidelity tensor matrix is ​​optimized. Perform singular value decomposition (SVD), the formula is as follows: Where X and Y are unitary matrices, and D is a singular value diagonal matrix, the magnitude of the singular values ​​directly reflects the importance of the corresponding features. This is achieved by maximizing... Determining the optimal two-qubit unit operator This enables local optimization of quantum circuits, allowing them to gradually adapt to the feature distribution of sub-image blocks.

[0101] The optimal unit operator Decomposed into single-qubit rotating gates Controlled-NOT (CNOT) gates, single-qubit rotation gates, are used to control the state of a single qubit, enabling fine-tuning of its characteristics. Their matrix expression is:

[0102]

[0103] Controlled-NOT gates (CNOT) are used to construct entanglement relationships between qubits, enabling the capture of correlations in features. Their matrix expression is as follows:

[0104]

[0105] Based on the above quantum gates, aptor image blocks are constructed. Optimal quantum circuit The structure and parameters of the quantum circuit are entirely adaptively determined by the features of the sub-image blocks, eliminating the need for a pre-set fixed circuit template and ensuring adaptability to different defect features. This allows for the application of quantum circuits to... Perform feature extraction and output latent space quantum state features. This feature can condense the key information of sub-image patches and filter out redundancy and noise;

[0106] Repeat the above fidelity calculation, SVD decomposition, and quantum gate allocation process, according to The number of unit operators is gradually increased in sequence, while N forward-backward scan optimizations are performed. The forward scan optimizes each unit operator sequentially from the beginning of the quantum circuit, and the backward scan optimizes from the end of the quantum circuit. Through bidirectional iteration, the coordination and adaptation of each part of the quantum circuit are ensured, and the ability to capture defect features is continuously improved until the preset maximum number of unit operators or the fidelity convergence threshold is reached. At this time, the quantum circuit has fully learned the feature rules of the sub-image blocks and can output stable and accurate latent space features.

[0107] The fourth step is block feature fusion and image reconstruction, which uses the quantum state direct product method to fuse the latent space quantum state features of all sub-image blocks. Construct the global characteristic quantum state, expressed as:

[0108]

[0109] This fusion method can fully preserve the independence of the features of each sub-image block, while establishing the correlation between sub-image blocks through quantum entanglement, fully exploring the distribution pattern and spatial correlation of global defect features, and avoiding the loss or redundant superposition of feature information in traditional fusion methods.

[0110] Construct a quantum decoder and use inverse amplitude encoding to convert global characteristic quantum states. Converting to classical image data, the decoding process strictly corresponds to the inverse operation of the encoding process, ensuring that quantum state features can be accurately mapped back to classical pixel space. The expression is:

[0111]

[0112] in, For the reconstructed classical pixel values, the key defect features condensed in the quantum state are amplified and restored to the classical image through decoding operations, thereby enhancing the defect features and making small, inconspicuous defect features easier to identify in the reconstructed image. At the same time, noise information remaining in the encoding and fusion process is filtered out, improving the signal-to-noise ratio of the image.

[0113] The fifth step is intelligent defect recognition. The reconstructed and enhanced image is input into a classic deep learning model (such as ResNet series, RCNN, etc.). The model extracts local texture and detail features of the image through convolutional layers, reduces feature dimensionality and retains key information by pooling layers, integrates global features by fully connected layers, determines the defect category by a softmax classifier, and accurately locates the defect in the image by bounding box regression module. The combination of classic model and quantum feature enhancement not only takes advantage of the advantages of quantum computing in high-dimensional feature extraction, but also leverages the mature experience of classic deep learning in defect classification and localization.

[0114] Output defect identification results, including defect type (such as insulator crack, localized overheating, GIS partial discharge traces, oil leakage, etc.) and location coordinates. and confidence parameters The confidence level parameter reflects the reliability of the identification results, providing decision-making reference for operation and maintenance personnel, assisting them in judging the urgency and priority of the defects, and ultimately achieving accurate identification of power equipment defects and efficient operation and maintenance response.

[0115] Example 2

[0116] The power equipment image processing system in this embodiment adopts a modular design, including a preprocessing module, a unified segmentation and angle encoding module, a quantum image encoding feature extraction module, a fusion reconstruction module, and a defect identification module. These modules are sequentially connected through data interfaces, forming a closed-loop processing system from multimodal image input to defect identification result output. The functional boundaries of each module are clear, and the data flow is well-defined, ensuring the overall processing efficiency and scalability.

[0117] 1. Preprocessing module

[0118] The core function of the preprocessing module is to eliminate the heterogeneity and noise interference of multimodal images, providing a standardized data foundation for subsequent quantum processing. The specific implementation is as follows:

[0119] Noise Filtering: An adaptive threshold filtering algorithm is employed to dynamically adjust filtering parameters to suit the noise type, taking into account the noise characteristics of different image modalities. For Gaussian noise in visible light images, thermal noise in infrared images, and electromagnetic pulse noise in partial discharge images, the filtering threshold is uniformly calculated using the following formula:

[0120]

[0121] in,

[0122] The average pixel value of the image. The standard deviation is the pixel value. Adjustment coefficients are applied to images under different lighting conditions. This ensures that different types of noise are effectively suppressed while preserving detailed information about defect features. The core advantage of this threshold calculation method lies in its ability to dynamically adapt to differences in brightness and contrast of different images through a weighted combination of pixel mean and standard deviation, avoiding over-filtering or incomplete filtering problems caused by a fixed threshold.

[0123] Size normalization: Bilinear interpolation is used to uniformly scale all input images to size. Pixels (S is the preset resolution, such as...) The interpolation process is achieved through the following formula:

[0124]

[0125] in, For interpolating neighborhood pixel coordinates, the weight function Bilinear interpolation, by linearly weighting the gray values ​​of four adjacent pixels, can maintain the continuity of image gray levels during scaling, avoid jagged distortion at defect edges, and ensure that the normalized image can still accurately reflect the key features such as the shape and size of the defect, providing a consistent data foundation for subsequent quantum coding and feature extraction.

[0126] Modal alignment: Device contour feature points are extracted from each modal image based on the SIFT (Scale Invariant Feature Transform) algorithm. Feature point detection is achieved through Gaussian difference pyramid and keypoint localization. Feature descriptors are represented by 128-dimensional vectors. Feature point matching is performed using the RANSAC (Random Sample Consensus) algorithm to eliminate mismatched points. The loss function for the matching process is:

[0127]

[0128] in, Modal The The coordinate vector of each feature point Given the homography matrix, the optimal registration parameters are solved by minimizing the loss function. The SIFT algorithm possesses scale invariance and rotation invariance, effectively handling differences in shooting angle and distance between multimodal images. The RANSAC algorithm, through random sampling and consistency verification, can eliminate more than 90% of mismatched points, ensuring registration accuracy. After registration, the feature point error between different modal images is controlled within 2 pixels, ensuring accurate spatial correspondence of the same device region in multimodal images, laying the foundation for subsequent fusion and association of multimodal features.

[0129] 2. Unified Segmentation and Angle Encoding Module

[0130] The unified segmentation and angle encoding module realizes unified segmentation and angle encoding quantum state conversion of images. The core is to convert high-dimensional image data into low-dimensional quantum state information that can be processed by quantum circuits. At the same time, unified segmentation simplifies the subsequent processing flow. The specific implementation is as follows:

[0131] Unified Blocking: Determining a unique and unified block size based on the image size and structural complexity of power equipment. The block decision-making logic is as follows:

[0132] Image size adaptation: based on the normalized image resolution. Calculate the number of blocks Ensure that a is an integer factor of S to avoid missing or overlapping image regions after segmentation;

[0133] Structural complexity verification: For structurally complex equipment (such as transformers and GIS equipment), verify whether the block size 'a' can ensure that the critical component area is not interrupted by the block boundary, thus ensuring the integrity of the defect features; for structurally simple equipment, verify whether the block size 'a' can fully cover the area where defects may occur while taking into account processing efficiency.

[0134] Angle-coded quantum state conversion: Angle coding is used to convert classical pixel data of sub-image blocks into quantum state information. The specific steps are as follows:

[0135] Pixel angle mapping: for each sub-image patch raw pixel values Normalization is performed to map pixel grayscale values ​​to the range of \([0,\pi]\) (adapting to the angle range of a single-qubit rotating gate), as shown in the formula:

[0136]

[0137] in, These are the minimum and maximum pixel values ​​of the sub-image block, respectively. The mapped rotation angle ensures that the relative differences in pixel grayscale are fully preserved, while adapting to the angle operation range of the quantum rotating gate.

[0138] Quantum state encoding: using a single qubit to rotate around the y-axis via a gate The angle information is encoded into a quantum state, and the encoded quantum state information satisfies:

[0139]

[0140] in, It is a single-qubit rotation gate. This represents the initial state of a qubit. Number of qubits. (Each pixel corresponds to a rotation angle encoding of a quantum bit). This encoding method directly converts the pixel grayscale features into the phase information of the quantum state without relying on the high-dimensional quantum state amplitude distribution, which reduces the implementation difficulty of quantum circuits and ensures the complete mapping of feature information of sub-image blocks.

[0141] 3. Quantum Image Encoding Feature Extraction Module

[0142] The quantum image coding feature extraction module is the core of this system. It achieves accurate extraction of defect features through quantum circuit optimization, as detailed below:

[0143] Parameter initialization: Set the core control parameters of the quantum image coding algorithm. The parameter values ​​are determined based on a balance between the feature complexity of the power equipment image and the computational efficiency of the quantum circuit.

[0144] Initial number of unit operators (L is the number of qubits) to ensure that the initial circuit has basic feature extraction capabilities, and to avoid the initial circuit being too simple, which would lead to slow optimization convergence;

[0145] Maximum number of unit operators This value corresponds to the maximum possible combination of two-qubit unit operators, avoiding computational redundancy and overfitting caused by excessive circuit complexity;

[0146] Unit operator increment The number of unit operators is increased gradually to balance optimization efficiency and feature capture capability, and to avoid optimization oscillations caused by adding too many operators at once.

[0147] The number of scans, N, is determined based on extensive simulation verification to ensure that the circuit fully optimizes convergence through multiple forward and backward scans, achieving a balance between convergence performance and computational cost.

[0148] Initialize the quantum circuit as a unit operator That is, all quantum gates are Identity operations, and the initial output state of the circuit is This provides a clean initial baseline for subsequent optimization.

[0149] Fidelity optimization: The goal is to maximize the fidelity between the output state of the quantum circuit and the quantum state of the sub-image block. Fidelity is defined as the degree of overlap between the generated state of the quantum circuit and the original quantum state of the sub-image block, directly reflecting the quantum circuit's ability to replicate the features of the original image. The expression is:

[0150] in, For current quantum circuits, For a two-qubit unit operator, This represents the Hermitian conjugation of a quantum circuit. To maximize F, it is necessary to pair each unit operator individually. Optimization can be achieved through the following steps:

[0151] Constructing the fidelity tensor operator: Defining the fidelity tensor operator ,in unit operator The complement of the action subsystem (i.e., the subsystem composed of other qubits besides the action qubits). unit operator The quantum state after the circuit on the right is applied, unit operator The quantum state after the circuit on the left is applied. This tensor operator condenses the unit operator. The characteristic correlation information of the preceding and following circuits is the core basis for optimizing the unit operator;

[0152] Singular Value Decomposition (SVD): Matrix form of fidelity tensor operators (The dimension is 4×4, corresponding to the state space of a two-qubit system) Perform SVD decomposition:

[0153]

[0154] Where X and Y are 4×4 unitary matrices, It is a singular value diagonal matrix. The magnitude of singular values ​​directly reflects the importance of the corresponding feature components;

[0155] Determining the optimal unit operator: The goal of unit operator optimization is to maximize... (The trace function reflects the similarity between two matrices), substituting it yields... Since D is a diagonal matrix and its singular values ​​are non-increasing, the optimal unit operator for the trace function to reach its maximum value is: .

[0156] Quantum gate allocation: the operator that allocates the optimal two-qubit unit. The decomposition is a product of elementary quantum gates, based on the general decomposition theorem for unitary operators of two qubits. This ensures that the decomposed quantum gates can be physically implemented on quantum hardware. The decomposition form is as follows:

[0157]

[0158] Iterative optimization: An optimization strategy of alternating forward and backward scanning is adopted to ensure the coordinated adaptation of operators in each unit of the quantum circuit. The specific process is as follows:

[0159] Forward scan: Following the order m=1→M, sequentially scan each unit operator. Perform fidelity tensor operator construction, SVD decomposition, optimal operator determination and quantum gate allocation operations, and update unit operator parameters;

[0160] Reverse scanning: Repeat the above operation in the order of m=M→1, optimize each unit operator in reverse, and use the optimization results of the later circuit to correct the parameters of the earlier circuit to improve the overall coordination of the circuit.

[0161] Iterative loop: Repeat the forward-backward scanning process N times. Calculate the global fidelity F after each scan. The iteration can be terminated early when the change in F is less than a preset threshold.

[0162] After iteration, the quantum circuit Converging to the optimal state, outputting the hidden space quantum state characteristics. This feature condenses the key defect information of sub-image patches and filters out redundancy and noise.

[0163] 4. Integration and Reconstruction Module

[0164] The fusion and reconstruction module realizes the fusion of block quantum features and classical image reconstruction. The core is to integrate the local features of each sub-image block into global features and restore the classical image. The specific implementation is as follows:

[0165] Feature fusion: The latent space features of all sub-image patches are fused using a quantum state direct product approach. Construct the global characteristic quantum state, expressed as:

[0166]

[0167] Where n is the number of blocks. Quantum state direct product is a natural way to combine multiple subsystems in a quantum system, preserving the independence of each sub-image block's features while establishing global correlations between them through quantum entanglement—the quantum state features of each sub-image block are correlated with the features of other sub-image blocks through entanglement, thereby uncovering the distribution patterns of global defect features (such as the spatial extension of defects on the device, the positional relationships between multiple defects, etc.). Due to the unified block design, the quantum state dimensions of each sub-image block are consistent, eliminating the need for additional dimension adaptation operations during the fusion process, significantly improving fusion efficiency and avoiding feature misalignment caused by inconsistent block specifications. For example, when... Each sub-image block corresponds to When the quantum state has 100 qubits, the global characteristic quantum state can fully carry the high-dimensional information of the global defect characteristics, which is difficult to achieve with traditional classical fusion methods.

[0168] Inverse angle encoding is used to convert global feature quantum states into classical image data. The decoding process is a strict inverse operation of the angle encoding process, ensuring that quantum state features can be accurately mapped back to classical pixel space.

[0169] 5. Defect Identification Module

[0170] The defect identification module classifies and locates defects based on classic deep learning models. Its core principle is leveraging the maturity of classic models in classification and localization tasks to process quantum-enhanced image features. ResNet-50 is used as the core classification model, which addresses the vanishing gradient problem in deep networks through residual connections, effectively extracting high-level semantic features of the image. Simultaneously, a Region Proposal Network (RPN) from Faster R-CNN is introduced to achieve defect localization, forming an end-to-end architecture of "feature extraction - region proposal - classification and localization". The model input is the reconstructed enhanced image (S×S), and the output is the defect category probability distribution (corresponding to various types of power equipment defects), bounding box coordinates (corresponding to the defect's location in the image), and confidence score (corresponding to the reliability of the identification result).

[0171] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0172] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for image processing of power equipment based on a quantum image coding algorithm, characterized in that, include: Step S1: Acquire multimodal images of the power equipment; Step S2: Preprocess the multimodal image to obtain standardized image data; Step S3: Based on the structural characteristics of the power equipment, the standardized image data is adaptively divided into blocks to obtain several sub-image blocks. The pixel information of each sub-image block is mapped to quantum state information through angle encoding. Step S4: Call the quantum image encoding algorithm to initialize the quantum circuit. With the goal of maximizing the fidelity between the output state of the quantum circuit and the quantum state of the target sub-image block, generate the optimal quantum circuit. Based on the optimal quantum circuit, extract features from the quantum state information of each sub-image block and output the hidden space quantum state features. Step S5: The hidden space quantum state features of all sub-image blocks are fused using the quantum state direct product method to construct a global feature quantum state. The global feature quantum state is then converted into classical enhanced image data using a quantum decoder. Step S6: Input the classical enhanced image data into the classical deep learning model to complete the classification and localization of power equipment defects, and output the defect identification results.

2. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, The multimodal images in step S1 include visible light images, infrared images, and partial discharge associated images.

3. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, The preprocessing in step S2 includes: Noise filtering: An adaptive thresholding method is used to filter noise, wherein the threshold is calculated as follows: In the formula, μ is the mean pixel value of the image, σ is the standard deviation of the pixels, and α and β are adjustment coefficients that are adaptively adjusted according to the modality type; Size normalization: Images of different modalities and resolutions are uniformly scaled to a preset resolution of S×S pixels using bilinear interpolation. Modal alignment operation: Based on the SIFT algorithm, device contour feature points of each modal image are extracted, and the RANSAC algorithm is used to achieve feature point matching and spatial registration of multimodal images, controlling the registration error within 2 pixels.

4. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, Step S3 includes: S3.1: The standardized image data is divided into blocks to obtain a set of sub-image blocks with consistent specifications. The block size is determined uniformly based on the size of the standardized image data and the structural complexity of the power equipment. S3.2: Normalize the pixel values ​​of each sub-image block, mapping the pixel grayscale values ​​to the [0, π] interval, wherein the normalization formula is: Where p is the original pixel value of the sub-image patch. , where are the minimum and maximum pixel values ​​of the sub-image block, respectively, and θ is the rotation angle after mapping; S3.3: Encode angular information into a quantum state by rotating a single qubit around the y-axis to obtain quantum state information, wherein the encoded quantum state information satisfies: In the formula, For a single-qubit rotation gate around the y-axis, This represents the initial state of the qubits, where the number of qubits equals the number of pixels in the sub-image block.

5. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, Step S4 includes: Sub-step S4.1: Set the core control parameters of the quantum image coding algorithm, and initialize the quantum circuit as a unit operator I, wherein the control parameters include the initial number of unit operators of the quantum circuit. Maximum number of unit operators Unit operator increment Number of scans N; Sub-step S4.2: Define the fidelity function between the quantum circuit output state and the quantum state of the target sub-image block: in, For current quantum circuits, For current quantum circuits, For a two-qubit unit operator, The Hermitian conjugation of the quantum circuit, <0| represents the initial product state of the quantum circuit. Quantum state information for sub-image patches; Sub-step S4.3: With the goal of maximizing the fidelity function, a forward-backward bidirectional scanning strategy is adopted to optimize each two-qubit unit operator, calculate the optimized global fidelity, and repeat N times; Sub-step S4.4: Repeat sub-step S4.3 until the maximum number of unit operators is reached. Alternatively, if the change in the fidelity value calculated after each optimization is less than the threshold, the optimal quantum circuit can be constructed based on the final iteration result. Sub-step S4.5: Based on the optimal quantum circuit, process the target sub-image patch. quantum state information Perform feature extraction and output the latent space quantum state features; Sub-step S4.6: Repeat sub-steps S4.1 to S4.5 to obtain the hidden space quantum state features of all sub-image blocks.

6. The image processing method for power equipment based on quantum image coding algorithm according to claim 5, characterized in that, In sub-step S4.3, optimizing each two-qubit unit operator includes: Construct a fidelity tensor operator, perform singular value decomposition on its matrix form, and take the unitary matrix corresponding to the largest singular value as the optimal two-qubit unit operator. The optimal two-qubit unit operator is decomposed into a single-qubit rotation gate and a controlled NOT gate, which are used to construct quantum circuits.

7. The image processing method for power equipment based on quantum image coding algorithm according to claim 5, characterized in that, In step S4.3, the forward-backward bidirectional scanning strategy specifically includes: Forward scan: Iterate through each unit operator in the quantum circuit in sequence, optimize the current unit operator, and update the unit operator parameters; Reverse scan: Iterative optimization of the unit operators with updated parameters in reverse order.

8. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, Step S5 includes: Sub-step S5.1: Fuse the latent space quantum state features of all sub-image patches through quantum state direct product. The global feature quantum state expression is: Where n is the number of blocks, |ψ i > represents the hidden space quantum state feature of the i-th sub-image block, and dots represents the omitted content in the middle; Sub-step S5.2: Construct a quantum decoder, using inverse angle encoding to convert the global feature quantum state into classical image data. The decoding expression is: Where p is the reconstructed classical pixel value, and θ is the rotation angle corresponding to the quantum state. These are the minimum and maximum pixel values ​​of the sub-image block, respectively.

9. The image processing method for power equipment based on quantum image coding algorithm according to claim 1, characterized in that, Step S6 includes: Step S6.1: Input the classic enhanced image data into the ResNet-50 classification model and the Faster R-CNN region proposal network to form an end-to-end recognition architecture; Step S6.2: Output the defect identification result through the architecture, wherein the identification result includes the defect type, defect location coordinates and confidence parameters.

10. A power equipment image processing system based on a quantum image coding algorithm, characterized in that, include: Module M1: Acquires multimodal images of power equipment; Module M2: Preprocesses the multimodal image to obtain standardized image data; Module M3: Based on the structural characteristics of power equipment, the standardized image data is adaptively divided into several sub-image blocks, and the pixel information of each sub-image block is mapped to quantum state information through angle encoding; Module M4: Calls the quantum image encoding algorithm, initializes the quantum circuit, generates the optimal quantum circuit with the goal of maximizing the fidelity between the output state of the quantum circuit and the quantum state of the target sub-image block, and extracts features of the quantum state information of each sub-image block based on the optimal quantum circuit, and outputs the hidden space quantum state features. Module M5: It uses the quantum state direct product method to fuse the hidden space quantum state features of all sub-image blocks to construct a global feature quantum state, and uses a quantum decoder to inversely convert the global feature quantum state into classical enhanced image data; Module M6: Inputs the classical enhanced image data into the classical deep learning model to complete the classification and localization of defects in power equipment and outputs the defect identification results.

Images