GIS local interference source denoising method

Abstract

The invention relates to the technical field of GIS (Geographic Information System) equipment denoising, in particular to a GIS local interference source denoising method, which comprises the following steps of: constructing a hybrid architecture of a quantum convolutional neural network and a quantum recurrent neural network, extracting features by using a variable component sub-circuit as a convolution kernel in the quantum convolutional neural network, processing time sequence data by using quantum bit entanglement in the quantum recurrent neural network, and extracting a local interference source in the quantum recurrent neural network; normalization, coding and enhancement are carried out on electrical, electromagnetic and other multi-source data, a quantum stochastic gradient descent algorithm is used for training, a multi-modal fusion network based on an attention mechanism is also constructed, features of all modals are extracted through an independent convolutional layer, then splicing and weighted fusion are carried out, and finally a quantum reinforcement learning environment is constructed. An intelligent agent strategy network and a reward function are designed based on a quantum neural network, parameters are adjusted according to quantum Q learning, quantum calculation, deep learning and a multi-modal fusion technology are integrated, the GIS local interference source denoising effect is effectively improved, and stable operation of equipment is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of GIS equipment denoising, and more specifically, to a method for denoising local interference sources in GIS. Background Art

[0002] Denoising of GIS equipment is an important technology. In the power system, the reliable operation of GIS is crucial for ensuring the stability of power transmission. However, during the operation of GIS equipment, it is often affected by local interference sources, and the noise generated by these interference sources seriously interferes with the accurate monitoring and evaluation of the equipment status.

[0003] Traditional denoising methods mainly rely on simple filtering techniques or empirical threshold settings. When facing complex and variable interference sources, it is difficult to effectively remove the partial discharge interference caused by the complex internal structure of the equipment and the electromagnetic interference coupled from the external environment only by the conventional filtering of electrical signals. At the same time, traditional methods cannot make full use of the information of multi-source data. For other modal data related to interference sources such as ultrasonic signals and chemical gas data, there are no effective integration and analysis means, resulting in inaccurate identification of interference sources and poor denoising effects. To solve this technical problem, we provide a method for denoising local interference sources in GIS. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for denoising local interference sources in GIS to solve the problems raised in the above background art.

[0005] To achieve the above object, one of the objects of the present invention is to provide a method for denoising local interference sources in GIS, including the following steps:

[0006] S1. Construct a hybrid architecture model of a quantum convolutional neural network and a quantum recurrent neural network. For the quantum convolutional neural network, a variational quantum circuit is used as a quantum convolutional kernel for quantum convolutional operations, and for the quantum recurrent neural network, a variant of the long short-term memory unit is constructed using qubit entanglement to process time series data;

[0007] S2. Normalize the multi-source data, encode it according to time series and spatial position and then input it into the model, and at the same time perform data augmentation operations. Finally, train the model using the quantum stochastic gradient descent algorithm and use the mean square error as the loss function;

[0008] S3. Construct a multi-modal fusion network based on the attention mechanism. Set independent convolutional neural network layers for each modal data to extract features, adaptively weight and fuse different modal data by learning weight vectors through the attention module, and adjust the parameters of the attention module and the structure of the feature extraction layer according to the fusion result;

[0009] S4. Construct a quantum reinforcement learning environment, with the operating state of GIS equipment, environmental parameters, and denoising model performance indicators as the state space. Design the quantum policy network and reward function of the agent based on the quantum neural network. The agent adjusts the parameters of the quantum policy network according to quantum Q-learning and reward feedback.

[0010] As a further improvement of this technical solution, when constructing the quantum convolutional neural network in S1, the specific construction method of the variational quantum circuit as the quantum convolution kernel is as follows:

[0011] Construct a basic quantum circuit unit using multiple qubits. Each qubit is initialized to the |0> state. For a variational quantum circuit with n qubits, its quantum gate operations include three rotation gates and two qubit entanglement gates;

[0012] Design the parameterized form of the quantum convolution kernel, and manipulate the quantum state and extract features by adjusting the angle parameters of the rotation gates to construct a preliminary quantum convolution kernel structure;

[0013] When performing quantum convolution operations, encode the input data into a quantum state, transform the quantum state through quantum gate operations, extract the feature information of the data, output the quantum state feature representation after quantum convolution, and then convert the quantum state back to standard data through quantum measurement operations for subsequent network layer processing.

[0014] As a further improvement of this technical solution, when constructing the quantum recurrent neural network in S1, the method of using qubit entanglement to construct a variant of the long short-term memory unit is as follows:

[0015] Use m qubits to represent the memory state and hidden state of the long short-term memory unit. Among them, 50% of the qubits are used to store the interaction result of the current input information and the memory state, and the remaining qubits are used to control the update of the memory state and the forgetting threshold;

[0016] Implement the functions of the memory unit by designing a specific quantum gate sequence to capture and process the long-term dependence relationship of time series data.

[0017] As a further improvement of this technical solution, the method of encoding data according to time series and spatial position in S2 is as follows:

[0018] For the interference source data with time series characteristics, use the sine and cosine function encoding method to embed the time information into the data features and obtain the time variation law of the data;

[0019] For the interference source data with spatial position information, use the relative position-based encoding method to integrate the spatial position information into the data features and obtain the spatial distribution characteristics of the data.

[0020] As a further improvement of this technical solution, when constructing a multi-modal fusion network based on the attention mechanism in S3, the method of setting independent convolutional neural network layers for each modal data to extract features is as follows:

[0021] For electrical signal modal data, construct a convolutional neural network, extract its preliminary features through convolutional operations, then perform downsampling through a pooling layer to obtain secondary features, and continue to extract features through subsequent convolutional layers and pooling layers, finally obtaining the features of the electrical signal modality;

[0022] For electromagnetic radiation modal data, construct a convolutional neural network and extract the features of the electromagnetic radiation modality through different convolutional and pooling operations.

[0023] As a further improvement of this technical solution, the method of adaptively weighted fusion of different modal data by learning weight vectors through the attention module in S3 is as follows:

[0024] Concatenate the features extracted from each modal data through independent convolutional neural network layers to obtain a concatenated feature vector;

[0025] Process the concatenated feature vector through a fully connected layer to obtain an attention score vector. Finally, perform weighted fusion on the features of each modality according to the attention weights to obtain the fused features.

[0026] As a further improvement of this technical solution, the method of adaptively weighted fusion of different modal data by learning weight vectors through the attention module in S3 is as follows:

[0027] Concatenate the features extracted from each modal data through independent convolutional neural network layers to obtain a concatenated feature vector;

[0028] Process the concatenated feature vector through a fully connected layer to obtain an attention score vector. Finally, perform weighted fusion on the features of each modality according to the attention weights to obtain the fused features.

[0029] As a further improvement of this technical solution, the method of the agent adjusting the parameters of the quantum policy network according to quantum Q-learning and reward feedback in S4 is as follows:

[0030] At each time step, after the agent selects an action according to the current state, the environment will feedback a reward and the next state, and the agent uses the quantum Q-learning algorithm to update the parameters of its policy network;

[0031] First, calculate the Q-value estimation of the current state-action pair and all Q-value estimations in the next state through a quantum neural network;

[0032] Then, update the Q value according to the update formula of quantum Q learning, and adjust the parameters in the quantum policy network according to the update of the Q value through the backpropagation quantum algorithm.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] In a GIS local interference source denoising method, by constructing a hybrid architecture model of a quantum convolutional neural network and a quantum recurrent neural network, rich and accurate features are extracted using the characteristics of quantum computing, achieving the effect of enhancing the ability to process interference source data, having a significant effect of improving the denoising effect. Through the normalization, encoding of multi-source data and training with the quantum stochastic gradient descent algorithm, the efficient utilization of data and model optimization are realized, effectively improving the denoising accuracy. By constructing a multi-modal fusion network based on the attention mechanism, multi-modal data can be adaptively weighted and fused, giving full play to the advantages of each modality, further enhancing the interference source recognition and denoising ability, and ensuring the stable operation of GIS equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall work flow chart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] Please refer to Figure 1 As shown, this embodiment provides a GIS local interference source denoising method, including the following steps:

[0038] S1. Construct a hybrid architecture model of a quantum convolutional neural network and a quantum recurrent neural network. For the quantum convolutional neural network, a variational quantum circuit is used as the quantum convolutional kernel for quantum convolutional operations, and for the quantum recurrent neural network, a variant of the long short-term memory unit is constructed using qubit entanglement to process time series data.

[0039] When constructing the quantum convolutional neural network in S1, the specific construction method of using the variational quantum circuit as the quantum convolutional kernel is as follows:

[0040] First, determine the number of qubits used to construct the quantum convolutional kernel, and initialize each qubit to the |0> state, which is the basic initial state in quantum computing, indicating that the qubit is in the ground state without any superposition or entanglement of quantum information.

[0041] Construct the basic structure of a variational quantum circuit, including a series of quantum gate operations. Introduce rotation gates, which can perform rotation operations on the states of qubits. Their action forms are as follows:

[0042] where X is the Pauli X matrix and θ is the rotation angle parameter. This gate operation can rotate the qubit by an angle θ around the X-axis, thereby changing the state of the qubit.

[0043] Y is the Pauli Y matrix, and this gate operation rotates the qubit around the Y-axis.

[0044] Z is the Pauli Z matrix, which is used to rotate the state of the qubit around the Z-axis.

[0045] At the same time, use two-qubit entanglement gates to achieve entanglement between qubits. The two-qubit entanglement gate has a control qubit and a target qubit. When the control qubit is |1>, the state of the target qubit will be flipped; otherwise, the state of the target qubit remains unchanged. By reasonably combining these quantum gates, a variational quantum circuit with complex functions can be constructed, laying a foundation for subsequent feature extraction as a quantum convolution kernel.

[0046] Design the parameterized form of the quantum convolution kernel. By adjusting the angle parameter θ of the rotation gate, the manipulation of the quantum state and feature extraction are realized. They determine the rotation angle of the qubit in the initial state, thereby affecting the feature extraction ability of the quantum convolution kernel.

[0047] Then, through the two-qubit entanglement gate, entanglement between adjacent qubits is achieved, and a preliminary quantum convolution kernel structure is constructed. In this way, the states of two qubits are correlated, enhancing the feature expression ability of the quantum convolution kernel for the input data. A multi-layer quantum convolution kernel structure can be constructed as needed. Each layer optimizes the feature extraction effect on the input data by adjusting the combination of the rotation gate angle parameter and the entanglement gate, enabling the quantum convolution kernel to adapt to different input data features and denoising task requirements.

[0048] Encode the input data into a quantum state, and then input the encoded quantum state into the constructed quantum convolution kernel for quantum gate operations. Through the transformation of the quantum state by the quantum gate, the feature information of the data is extracted, and the quantum state feature representation after quantum convolution is output.

[0049] Finally, the quantum state is converted back to classical data through quantum measurement operations for subsequent network layer processing. Quantum measurement is the process of collapsing the quantum state to a classical state. Depending on the measurement basis, different measurement results can be obtained. After quantum convolutional operations, measurement operations in the computational basis are usually adopted to obtain classical measurement results. These results can be used as inputs for subsequent quantum recurrent neural networks or other network layers, thereby realizing quantum convolutional operations based on variational quantum circuits. Utilizing the parallelism of quantum computing and the superposition characteristics of quantum states, the efficiency and ability of feature extraction are improved, providing richer and more accurate feature information for subsequent interference source denoising, and enhancing the processing ability of the entire denoising model for GIS local interference source data.

[0050] When constructing a quantum recurrent neural network in S1, the method of constructing a variant of the long short-term memory unit using qubit entanglement is as follows:

[0051] Determine the number of qubits required to construct a variant of the long short-term memory unit of the quantum recurrent neural network. These qubits will be used to store and process information in time series data, including current input information, memory state, and hidden state.

[0052] Initialize these qubits to specific initial states. 50% of the qubits are initialized to the |0> state, which is used to represent the initial memory state and hidden state, providing an initial information carrier for subsequent time series data processing. Design a specific sequence of quantum gates to implement the functions of the memory unit. Taking the input data of a time step as an example:

[0053] First, encode the input data into a quantum state. Then, perform an entanglement operation on the encoded input quantum state and the current memory state qubits, using a combination of quantum bit gates and rotation gates to achieve information interaction and update. For the calculation of the forget gate, apply a series of rotation gate operations to some qubits, and then obtain the quantum state representation of the forget gate through measurement. This forget gate quantum state will be used to control which information in the memory unit needs to be retained or forgotten, that is, perform selective operations on the memory state qubits to determine the degree of retention of old memory information, enhancing the ability to capture long-term dependence relationships in interference source time series data.

[0054] For the input gate and the output gate, their quantum state representations are also constructed through the same sequence of quantum gate operations and learnable parameters. The input gate determines the degree of update of the new input information in the memory cell, while the output gate controls the information output from the memory cell, which is used to generate the hidden state at the current time step and the memory state update passed to the next time step. The combination of these quantum gate operations and the parameter learning process enables the quantum recurrent neural network to adaptively process the input data at different time steps, effectively capture the dynamic changes and long-term dependence information in the time series, and provide a more powerful time series modeling ability for the subsequent interference source denoising task.

[0055] Based on the quantum state representations of the forget gate, the input gate, and the output gate, update operations are performed on the memory state qubits. Specifically, for each qubit in the memory state, through quantum gate operations, according to the control information of the forget gate and the input gate, the old memory information is selectively retained and the new input information is incorporated to achieve the update of the memory state. In this way, the quantum recurrent neural network can efficiently update the memory state at the quantum state level, maintain the long-term memory and dynamic adaptation ability for time series data, overcome the vanishing gradient or exploding gradient problems in traditional neural networks when dealing with long sequence data, better handle the long-term dependence relationship and complex dynamic changes in the GIS local interference source time series data, provide more accurate and stable time series feature information for the denoising model, and improve the denoising effect and the overall performance of the model.

[0056] Generate the hidden state at the current time step using the output gate and the updated memory state. Through specific quantum gate operations on the memory state qubits and the output gate quantum state, the memory information related to the current time step is extracted and converted into the hidden state quantum state. This hidden state will be used as the output feature at the current time step for subsequent network layer processing, enabling the model to better handle the complex changes and long-term dependence relationships in the time series interference source data, enhancing the ability to identify and remove interference noise, and providing strong technical support for the stable operation of GIS devices.

[0057] S2. Normalize the multi-source data, encode it according to the time series and spatial position, and then input it into the model. At the same time, perform data augmentation operations. Finally, train the model using the quantum stochastic gradient descent algorithm with the mean square error as the loss function.

[0058] The method of encoding the data according to the time series and spatial position in S2 is as follows:

[0059] For the interference source data with time series characteristics, the sine and cosine function encoding method is adopted. First, a timestamp is set and converted into a time encoding vector, and the dimension of the encoding vector is determined. Then, the parameters of different frequencies are calculated according to the formula. Next, the time encoding vector is calculated through the formula, and so on to calculate each element of the entire time encoding vector.

[0060] Through this encoding method, the time information is embedded into the data features, enabling the model to better capture the time variation law of the data and improving the processing ability of the time series interference source data.

[0061] For the interference source data with spatial location information, the relative position-based encoding method is adopted. The GIS device is divided into several grid regions. For the data point located at the coordinates (x, y), first, the relative position encoding component in the x direction and its square term are calculated. Similarly, the relative position encoding component in the y direction is calculated. Then, these encoding components are combined into a spatial position encoding vector.

[0062] Through this encoding method, the spatial location information is integrated into the data features, enabling the model to consider the spatial distribution characteristics of the data, enhancing the processing ability of the interference source data at different positions, being able to better utilize the spatial location information for targeted denoising, improving the accuracy and effect of denoising, better identifying and removing the interference noise at different positions, ensuring the stable operation and high performance of the GIS device, and providing strong support for the reliable operation of the GIS system.

[0063] S3. Construct a multi-modal fusion network based on the attention mechanism. Set independent convolutional neural network layers for each modal data to extract features, adaptively weight and fuse different modal data through the attention module to learn the weight vector, and adjust the parameters of the attention module and the structure of the feature extraction layer according to the fusion result.

[0064] When constructing the multi-modal fusion network based on the attention mechanism in S3, the method of setting independent convolutional neural network layers for each modal data to extract features is as follows:

[0065] Construct a convolutional neural network CNN containing an L1-layer convolutional layer e , the first convolutional layer uses a convolutional kernel of k1×k1, the number is n1, and the stride is s1. For the input electrical signal data x e , perform a convolution operation through the convolutional kernel, and the weight matrix of the convolutional kernel has a shape of k1×k1×C×n1, and the shape of the bias is n1. The formula for the convolution operation is: where conv is the convolution operation, σ is the activation function. After the first convolutional layer, connect a max pooling layer with a pooling window of 2×2 and a stride of 2. For the feature map output by the convolutional layer After the max pooling operation, we obtain which is deformed into The second convolutional layer uses a convolutional kernel of k2×k2 to perform a convolutional operation on the output of the pooling layer to obtain Similarly, a max pooling layer is connected to perform downsampling on to obtain

[0066] Through the convolutional neural networks constructed respectively for the electrical signal modality and the electromagnetic radiation modality as above, the key features of each modality can be effectively extracted, providing high-quality feature information for subsequent multimodal fusion, thereby enhancing the processing ability and denoising effect of the entire multimodal fusion network based on the attention mechanism on GIS local interference source data, enabling the model to more accurately identify and remove interference noises in different modalities, and ensuring the stable operation of GIS devices and the accuracy of data.

[0067] The method for adaptively weighted fusion of different modality data by learning the weight vector through the attention module in S3 is as follows:

[0068] First, assume that we already have the features extracted from the electrical signal modality data through its independent convolutional neural network layer and the features extracted from the electromagnetic radiation modality data. These features of different modalities are concatenated in the feature dimension to obtain a concatenated feature vector.

[0069] Construct a fully connected layer with a weight matrix of W a ×m, where m is the number of attention heads, and a bias vector of b a . The concatenated feature vector is input into the fully connected layer for linear transformation to obtain an attention score vector. According to the obtained attention weight vector, the features of each modality are weighted and fused to obtain the final fused feature. Through this adaptive weighted fusion method, the model can dynamically adjust according to the importance degree of different modality data for interference source denoising, pay more attention to the modality data that contributes more to the denoising effect, and improve the effect of multimodal fusion and the performance of the denoising model.

[0070] Through the above steps, the attention module can effectively learn the weight vectors of each modality data and achieve adaptive weighted fusion of different modality data, make full use of the advantages of each modality data, improve the processing ability and denoising effect of the entire denoising model on GIS local interference sources, provide more discriminative and comprehensive feature information for subsequent model decision-making and interference source removal, enable the model to better cope with complex and variable interference source situations, and ensure the stable operation of GIS devices and high-quality monitoring of data.

[0071] S4. Construct a quantum reinforcement learning environment, using the operating state of the GIS device, environmental parameters, and the performance metrics of the denoising model as the state space. Design the quantum policy network and the reward function of the agent based on the quantum neural network. The agent adjusts the parameters of the quantum policy network according to quantum Q-learning and reward feedback.

[0072] When constructing the quantum reinforcement learning environment in S4, the method for designing the quantum policy network of the agent based on the quantum neural network is as follows:

[0073] First, determine the number of qubits used to encode the state vector. Assume the current state vector is s t , whose dimension is d. According to the relationship between the number of qubits and the state space, select the number of qubits n such that 2 n ≥d to ensure that the state information can be completely encoded. Use the amplitude encoding method to convert the state vector s t into a quantum state For each element s t (i) in the state vector, map it to the probability amplitude of the qubit, thus encoding the classical state vector into a quantum state, providing an input for subsequent quantum neural network processing.

[0074] The quantum convolutional layer uses a variational quantum circuit as the convolutional kernel. As mentioned before, the variational quantum circuit consists of multiple qubits and quantum gates. Determine the structure of the quantum convolutional kernel. Each qubit is parameterized by a rotation gate, and adjacent qubits are entangled through qubit gates to construct a quantum convolutional kernel structure with certain feature extraction capabilities.

[0075] Input the encoded quantum state into the quantum convolutional layer. For the convolution operation of the quantum convolutional kernel on the quantum state, it is similar to the sliding window method of classical convolution. Apply the quantum gate operations of the quantum convolutional kernel at different positions of the quantum state. Specifically, for each k×k local region of the quantum state, the quantum gates in the quantum convolutional kernel operate on these qubits in sequence. According to the action principle and parameter settings of the quantum gates, change the state of the qubits, thereby extracting the local feature information of the state and obtaining a new quantum state which represents the feature quantum state after quantum convolution. In this process, the operations of the quantum gates are carried out according to the parameterized form of the variational quantum circuit. By performing such a sequence of quantum gate operations on the local region of the quantum state, the local feature extraction of the state is realized, obtaining a more discriminative quantum state feature representation and providing richer information for subsequent processing.

[0076] Connect a quantum fully-connected layer after the quantum convolutional layer. The quantum fully-connected layer is also composed of multiple qubits and quantum gates, and its role is to further integrate and process the quantum state features output by the quantum convolutional layer to obtain the quantum state representation of the action probability distribution.

[0077] The weights of the quantum fully-connected layer are implemented through learnable parameterized quantum gates. Suppose the input quantum state of the quantum fully-connected layer is The output quantum state is For each output qubit, entanglement and information transfer are carried out with the input qubits through a series of rotation gate operations. Through such quantum gate operations, the local features extracted by the quantum convolutional layer are integrated and transformed to obtain the quantum state related to the action probability distribution It contains the probability information of different actions in the probability amplitudes and entanglement relationships of the quantum state, providing a basis for subsequent quantum measurement and action selection.

[0078] Through the quantum measurement operation, the quantum state of the action probability distribution is converted into a classical action probability distribution p(a t |s t ), for the quantum state After measurement, the probability |b of each state |i> is obtained i | 2 , and these probability values constitute the action probability distribution p(a t |s t ). The agent selects the action a according to this probability distribution t . In this way, the agent balances between exploring new actions and exploiting existing experience, continuously learning and optimizing its strategy to adapt to different environmental states and task requirements. Under the guidance of the policy network based on the quantum neural network, it makes better action selections, so that in the environment of GIS local interference source denoising, it can effectively explore and utilize various strategies, improve the denoising effect and the overall performance of the system, providing an innovative and efficient method for solving complex interference source problems and promoting the application and development of quantum reinforcement learning in related fields.

[0079] In S4, the method for the agent to adjust the parameters of the quantum policy network according to quantum Q-learning and reward feedback is as follows:

[0080] Design a quantum neural network structure, whose input is the quantum state representation of the current state-action pair (s t , a t ) and the quantum state representation of the next state s t+1 This quantum neural network can be composed of multiple quantum layers, including quantum convolutional layers and quantum fully connected layers, to extract state- and action-related features and output a quantum state representation of the Q value.

[0081] Input into the quantum neural network, and after a series of quantum gate operations, finally output the quantum state of the Q value estimate for the current state-action pair where n is the number of qubits related to the Q value representation, and q i，t is the probability amplitude of the quantum state.

[0082] For the next state s t+1 , it is necessary to calculate the Q value estimates corresponding to all possible actions, combine with the quantum state representations of each possible action, and then input them into the quantum neural network. Repeat the above quantum gate operations and measurement processes to obtain the Q value estimate Q(s t+1 , a) for each action in state s t+1 .

[0083] Update the Q value according to the update formula of quantum Q-learning. First, obtain the current reward and Q(s t , a t ) calculated through the above steps. Then, calculate the updated Q value according to the update formula.

[0084] To adjust the parameters of the quantum policy network through the backpropagation quantum algorithm, it is necessary to define a loss function. Use the mean squared error loss function, that is where Q(s t , a t ) new is the updated Q value, and Q(s t , a t ) is the Q value estimate currently output by the quantum neural network. This loss function measures the difference between the updated Q value and the current estimate value, and adjusts the parameters of the quantum policy network by minimizing this loss function.

[0085] Based on the gradient calculation of the quantum state, similar to the backpropagation algorithm of classical neural networks, in the quantum environment, it is necessary to calculate the gradients of the loss function with respect to each parameter in the quantum neural network. According to the calculated gradients, use an optimization algorithm to update the parameters in the quantum policy network. By continuously repeating the above processes of Q value calculation, update, and parameter adjustment of the backpropagation quantum algorithm, the agent can gradually learn the optimal policy, enabling it to make better action choices when facing the GIS local interference source denoising environment, improving the denoising effect and the overall performance of the system.

[0086] Through the above steps, the agent can effectively utilize the quantum Q-learning algorithm to update the parameters of its policy network, optimize its decision-making strategy during continuous interaction and learning processes, better cope with various situations in the GIS local interference source denoising environment, and achieve efficient interference source removal and system performance improvement.

[0087] The above-mentioned GIS local interference source denoising method constructs a hybrid architecture of a quantum convolutional neural network and a quantum recurrent neural network. The former uses a variational quantum circuit as a convolution kernel to extract features, and the latter uses quantum bit entanglement to process time-series data. Then, after normalizing, encoding, and enhancing multi-source data such as electrical and electromagnetic data, it is trained using the quantum stochastic gradient descent algorithm. A multi-modal fusion network based on the attention mechanism is also constructed. After each modality extracts features through an independent convolutional layer, they are concatenated and weighted for fusion. Finally, a quantum reinforcement learning environment is constructed. Based on the quantum neural network, an agent policy network and a reward function are designed, and the parameters are adjusted relying on quantum Q-learning. By integrating quantum computing, deep learning, and multi-modal fusion technologies, the denoising effect of GIS local interference sources is effectively improved, and the stable operation of the equipment is ensured.

[0088] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for denoising local interference sources in GIS, characterized in that: It includes the following steps: S1. Construct a hybrid architecture model of a quantum convolutional neural network and a quantum recurrent neural network. In the quantum convolutional neural network, a variational quantum circuit is used as a quantum convolution kernel for quantum convolution operations. In the quantum recurrent neural network, a variant of the long short-term memory unit is constructed using qubit entanglement to process time series data; S2. Normalize multi-source data, encode it according to time series and spatial position, and then input it into the model. At the same time, perform data augmentation operations. Finally, train the model using the quantum stochastic gradient descent algorithm, and use the mean squared error as the loss function; S3. Construct a multi-modal fusion network based on the attention mechanism. Set independent convolutional neural network layers for each modal data to extract features. Learn weight vectors through the attention module to adaptively weight and fuse different modal data, and adjust the parameters of the attention module and the structure of the feature extraction layer according to the fusion results; S4. Construct a quantum reinforcement learning environment, use the operating state of GIS equipment, environmental parameters, and the performance indicators of the denoising model as the state space, design the quantum policy network and reward function of the agent based on the quantum neural network, and the agent adjusts the parameters of the quantum policy network according to quantum Q learning and reward feedback.

2. The method for denoising local interference sources of GIS according to claim 1, characterized in that: When constructing the quantum convolutional neural network in S1, the specific construction method of using the variational quantum circuit as the quantum convolution kernel is as follows: Use multiple qubits to construct a basic quantum circuit unit, and each qubit is initialized to the 0 state. For a variational quantum circuit with n qubits, its quantum gate operations include three rotation gates and two qubit entanglement gates; Design the parameterized form of the quantum convolution kernel, and manipulate the quantum state and extract features by adjusting the angle parameters of the rotation gates to construct a preliminary quantum convolution kernel structure; When performing quantum convolution operations, encode the input data into a quantum state, transform the quantum state through quantum gate operations, extract the feature information of the data, output the quantum state feature representation after quantum convolution, and then convert the quantum state back to standard data through quantum measurement operations for subsequent network layer processing.

3. A method for denoising local interference sources in GIS according to claim 2, characterized in that: When constructing the quantum recurrent neural network in S1, the method of constructing a variant of the long short-term memory unit using qubit entanglement is as follows: Use m qubits to represent the memory state and hidden state of the long short-term memory unit. Among them, 50% of the qubits are used to store the interaction results of the current input information and the memory state, and the remaining qubits are used to control the update of the memory state and the forgetting threshold; Realize the functions of the memory unit by designing a specific quantum gate sequence to capture and process the long-term dependence relationship of time series data.

4. A method for denoising local interference sources in GIS according to claim 3, characterized in that: The method of encoding data according to time series and spatial position in S2 is as follows: For interference source data with time series characteristics, use the sine and cosine function encoding method to embed time information into the data features to obtain the time variation law of the data; For interference source data with spatial position information, use the relative position-based encoding method to integrate spatial position information into the data features to obtain the spatial distribution characteristics of the data.

5. A method for denoising local interference sources in GIS according to claim 4, characterized in that: When constructing the multi-modal fusion network based on the attention mechanism in S3, the method of setting independent convolutional neural network layers for each modal data to extract features is as follows: For the electrical signal modal data, a convolutional neural network is constructed to extract its preliminary features through convolutional operations, then downsampling is performed through a pooling layer to obtain secondary features, and then features are continuously extracted through subsequent convolutional layers and pooling layers, finally obtaining the features of the electrical signal modality; For the electromagnetic radiation modal data, a convolutional neural network is constructed to extract the features of the electromagnetic radiation modality through different convolutional and pooling operations.

6. A method for denoising local interference sources in GIS according to claim 5, characterized in that: In S3, the method of learning weight vectors through the attention module to adaptively weight and fuse different modal data is as follows: The features extracted from each modal data through independent convolutional neural network layers are concatenated to obtain a concatenated feature vector; The concatenated feature vector is processed through a fully connected layer to obtain an attention score vector. Finally, the features of each modality are weighted and fused according to the attention weights to obtain the fused features.

7. A method for denoising local interference sources in GIS according to claim 6, characterized in that: In S4, when constructing the quantum reinforcement learning environment, the method of designing the quantum policy network of the agent based on the quantum neural network is as follows: A quantum neural network is used to construct the policy network of the agent, where the input of the quantum neural network is the current state vector, which is encoded into a quantum state; The policy network consists of multiple quantum layers, including quantum convolutional layers and quantum fully connected layers. The quantum convolutional layer uses a variational quantum circuit as a convolutional kernel to perform convolutional operations on the quantum state to extract local features of the state, and then further processes through the quantum fully connected layer to output the quantum state representation of the action probability distribution; The quantum state of the action probability distribution is converted into an action probability distribution through a quantum measurement operation, and the agent selects an action according to this probability distribution.

8. A method for denoising local interference sources of GIS according to claim 7, characterized in that: In S4, the method for the agent to adjust the parameters of the quantum policy network according to quantum Q-learning and reward feedback is as follows: At each time step, after the agent selects an action according to the current state, the environment will feedback a reward and the next state, and the agent uses the quantum Q-learning algorithm to update the parameters of its policy network; First, calculate the Q-value estimate of the current state-action pair and all Q-value estimates in the next state through the quantum neural network; Then, update the Q-value according to the update formula of quantum Q-learning, and adjust the parameters in the quantum policy network according to the update of the Q-value through the backpropagation quantum algorithm.

Images