EIT reconstruction method and system based on cross-domain learning and physical guidance

Abstract

The invention discloses an EIT reconstruction method and system based on cross-domain learning and physical guidance. The method comprises the following steps: setting conditions; collecting voltage measurement data, and constructing a two-dimensional voltage data matrix; defining source domain data and target domain data; performing feature coding on the data of the two domains, and introducing physical bias based on a lead field similarity matrix in the coding process; and inputting the coded data into the same decoder to perform conductivity image reconstruction, and performing cross-domain consistency learning in combination with a loss function. The system comprises a condition definition module, a data preprocessing module and a model training module. By using the method and the device, the problems of reconstruction resolution reduction and structure distortion caused by insufficient measurement information can be effectively relieved, so that the image reconstruction quality under a low-electrode condition is remarkably improved. The method can be widely applied to the field of electrical impedance images.

Description

Technical Field

[0001] This invention relates to the field of electrical impedance tomography (EIT), and more particularly to an EIT reconstruction method and system based on cross-domain learning and physical guidance. Background Technology

[0002] Electrical Impedance Tomography (EIT) is an imaging method that inverts the impedance or conductivity distribution within a target region by applying boundary excitation and measuring the corresponding voltage response. EIT offers advantages such as being radiation-free, non-invasive, low-cost, having simple hardware structure, and fast measurement speed. It enables real-time monitoring of internal changes in the measured object and is therefore widely used in fields such as medical respiratory monitoring, industrial process inspection, biological tissue analysis, and flexible tactile sensing.

[0003] The EIT imaging process comprises three stages: excitation and measurement, solving the forward problem model, and reconstructing the inverse problem. Solving the inverse problem is the core challenge of EIT technology. Due to the highly nonlinear nature of the EIT inverse problem, its high-dimensional solution space, and severe ill-conditioning, image quality is strongly dependent on the completeness, accuracy, and noise sensitivity of the measurement information. In particular, the number of measuring electrodes directly determines the dimensionality of the boundary voltage data and the extent of electric field coverage, which is one of the key factors affecting reconstruction stability and resolution.

[0004] In applications such as flexible tactile sensing, electrodes are typically integrated onto the boundaries of flexible, stretchable conductive films. However, limitations in the geometry of flexible materials, available boundary area, and device fabrication processes can restrict the placement of a sufficient number of electrodes, posing a challenge to EIT reconstruction. Furthermore, from a hardware system perspective, the number of electrodes directly impacts the size of the acquisition channel and the measurement cycle. Reducing the number of electrodes can decrease system complexity, improve acquisition efficiency and power consumption, but it also weakens imaging performance.

[0005] Therefore, how to obtain high-quality reconstructed images under conditions of low electrode count is an important problem that needs to be solved in current EIT technology. Summary of the Invention

[0006] In view of this, in order to solve the technical problem that most existing EIT reconstruction methods require a large number of electrodes, resulting in excessive system complexity, the present invention proposes an EIT reconstruction method based on cross-domain learning and physical guidance, which includes the following steps: Data preprocessing: For the square EIT model, the finite element method is used to calculate the lead field similarity matrix between the excitation electrode pairs under uniform conductivity conditions. For the acquired voltage measurement data, a two-dimensional voltage data matrix is ​​constructed based on the combination relationship between the excitation electrode pairs and the measurement electrode pairs, and the spatial coordinates of each electrode pair are recorded for subsequent physical information embedding.

[0007] Model Setup: A two-dimensional voltage matrix obtained under a 16-electrode condition was used as the source domain data, and data obtained under an 8-electrode condition was used as the target domain data. Identical encoders were used to encode features from both domains, and a physical bias based on the lead field similarity matrix was introduced during the encoding process to enhance the model's ability to learn the spatial relationships between electrodes. The latent variables output by the encoder were then fed into a shared decoder to generate the reconstructed image. A latent variable constraint mechanism was used to make the target domain encoding results converge with the source domain encoding. Simultaneously, a voltage decoder was built to decode the latent variables of the target domain into a two-dimensional voltage matrix of the source domain, further strengthening cross-domain consistency.

[0008] Inference stage: Using only the encoder and image reconstruction decoder of the target domain, feature encoding is performed on the input data under 8-electrode conditions to generate a conductivity image, achieving high-resolution reconstruction under low electrode number conditions.

[0009] In addition to the above methods, the present invention also proposes an EIT reconstruction system based on cross-domain learning and physical guidance, which includes a condition definition module, a data preprocessing module, a model training module, and a model application module.

[0010] Based on the above scheme, this invention provides an EIT reconstruction method and system based on cross-domain learning and physical guidance. It fully utilizes the advantages of simulated data and proposes a cross-domain learning mechanism. The rich measurement information from a high-electrode-count configuration (source domain) guides the encoding process of a low-electrode-count configuration (target domain), significantly alleviating the problem of insufficient data dimensionality caused by a low number of electrodes. This allows for high-quality imaging while reducing the number of electrodes and the size of the acquisition channels, thereby reducing hardware implementation costs and signal acquisition time, and improving the system's real-time performance and energy efficiency. Furthermore, the similarity matrix of the lead field calculated under uniform conductivity is used as the physical bias of the encoder's attention mechanism. Simultaneously, spatial location information is introduced into the encoding features, enabling the model to learn the spatial coupling relationship and electric field distribution patterns between electrode pairs. This effectively reduces the sensitivity of the deep learning model to changes in measurement patterns, improving the physical consistency and robustness of the reconstruction results. Attached Figure Description

[0011] Figure 1 This is a flowchart of the steps of an EIT reconstruction method based on cross-domain learning and physical guidance according to the present invention; Figure 2This is a schematic diagram of the two-dimensional square impedance model and measurement method of the source domain (16 electrodes) and target domain (8 electrodes) in an embodiment of the present invention; Figure 3 This is a schematic diagram of the data acquisition and preprocessing process for touch simulation of the electrical impedance model in an embodiment of the present invention; Figure 4 This is a structural diagram of the EIT reconstruction model based on cross-domain learning and physical guidance in an embodiment of the present invention; Figure 5 This is a visualization of the reconstruction results of some samples in the embodiments of the present invention using only 8-electrode data, only 16-electrode data, and the model in this example. Detailed Implementation

[0012] In addition to the technical issues mentioned in the background, traditional reconstruction methods exist for EIT reconstruction under conditions of limited electrode numbers. These include linearization methods such as NOSER and WLS, iterative regularization methods such as TV regularization or gradient optimization, and nonlinear iterative methods such as Gauss-Newton and Levenberg-Marquardt. These methods generally rely on high-dimensional boundary measurement data to ensure the discriminability of the Jacobian matrix and the stability of the regularization solution. When the number of electrodes decreases, the boundary measurement information is insufficient, the rank of the inversion matrix decreases, and the electric field coverage area is incomplete. This leads to a decrease in image resolution, an increase in artifacts, and a significant increase in sensitivity to noise, making it difficult to obtain reliable imaging results in scenarios with a low number of electrodes.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0015] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0016] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0017] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0018] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0019] Reference Figure 1 This is a flowchart illustrating an optional example of the EIT reconstruction method based on cross-domain learning and physical guidance proposed in this invention. This method can be applied to computer devices, and the EIT reconstruction method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Construct the conditions for the source domain dataset (16 electrodes) and the conditions for the target domain dataset (8 electrodes).

[0020] Step S2: For the square EIT models of the source and target domains respectively, the finite element method is used to calculate the similarity matrix of the conduction field between the excitation electrode pairs under the condition of uniform conductivity.

[0021] Step S3: Construct a corresponding two-dimensional voltage data matrix based on the measured voltage values ​​at the boundary between the source and target domains.

[0022] Step S4: Introduce a physical bias based on the lead field similarity matrix to encode the two-dimensional voltage data matrix and obtain the encoded data.

[0023] Step S5: Input the encoded data into the same decoder to reconstruct the conductivity image, and combine the loss function to perform cross-domain consistency learning.

[0024] Step S6: Using only the feature encoder of the target domain and the shared image reconstruction decoder, inference is performed on the two-dimensional voltage data matrix acquired under the 8-electrode condition.

[0025] In some feasible embodiments, step S1 specifically includes: A finite element model for electrical impedance imaging was established using a two-dimensional square field. Sixteen electrodes (source domain model) and eight electrodes (target domain model) were evenly spaced at the field boundary. The measurement method employed an "adjacent excitation, adjacent measurement" strategy. Adjacent electrodes formed electrode pairs; each time, one set of electrode pairs was selected to inject excitation current, and the remaining electrode pairs were used as measurement electrode pairs. The corresponding voltage measurements were recorded. Figure 2 As shown, by sequentially cyclically exciting the source domain model, 208 boundary voltage measurements can be obtained. Due to the electrical symmetry of the model, the voltage values ​​obtained are the same when the excitation electrode pair and the measurement electrode pair are interchanged, resulting in 104 effective independent voltage measurements. The target domain model, on the other hand, can obtain 40 boundary voltage measurements, of which 20 are effective independent measurements.

[0026] The finite element model is initially set to have a uniform conductivity distribution, and the initial boundary voltage measurements are obtained. Subsequently, several circular impedance anomaly regions of varying numbers and sizes were randomly generated in the model to simulate local impedance disturbances caused by contact changes or tissue differences in the actual environment, and the corresponding voltage measurements were obtained. .

[0027] The method for generating circular impedance anomaly regions is as follows: A circular region is generated based on a preset radius by randomly selecting a center position within the field. A radial impedance distribution is then established within this region, where the impedance value gradually changes outwards from the center along the radius, forming a radial gradient distribution that more closely approximates actual physical characteristics. By limiting the maximum number, minimum radius, and edge spacing of the anomaly regions, overlapping or exceeding boundaries is avoided, ensuring the rationality and controllability of the generated data samples. Finally, an EIT image reconstruction dataset is constructed using the impedance anomaly region distribution and its corresponding voltage measurement results. A one-to-one correspondence exists between the samples in the source domain model and the target domain model, meaning the circular impedance anomaly regions in the samples are consistent.

[0028] In some feasible embodiments, step S2 specifically includes: First, the area to be imaged is denoted as region. The region was discretized using the finite element method to obtain... Each finite element element. For each group of adjacent electrode excitation modes... By solving the forward problem of electrical impedance tomography, the potential distribution in the region under this excitation mode is obtained. In its continuous form, it satisfies the following elliptic equation: in, This represents the conductivity distribution within the region.

[0029] Under finite element discretization conditions, for any triangular element Its three vertex coordinates are denoted as The corresponding node potential value is By employing a linear shape function, the potential inside the element is approximated as a linear function, with its gradient being constant within the element. The gradient of the potential (i.e., the lead field) on this element can be expressed as: in, This is a matrix composed of the geometric positions of the unit vertices. Using the above relationships, for all units within the region... The corresponding incentive modes can be obtained in sequence. lead field set .

[0030] Furthermore, the excitation modes of the two sets of adjacent electrodes were calculated. and Based on the similarity in electric field distribution, the energy inner product of the lead fields is defined as: Combining finite element analysis, the above integral is numerically approximated at the element level, namely: in The number of excitation modes of adjacent electrodes.

[0031] To eliminate the impact of overall energy differences in lead fields across different excitation modes on similarity metrics, this embodiment normalizes the matrix: Further combinations of excitation modes for all adjacent electrodes Calculating the above energy inner product yields the lead field similarity matrix. : Indicates electrode pair and The energy product of the guide field Indicates the area to be imaged. Indicates electrode pair After the excitation current in the region The internal conductive field, Indicates electrode pair After the excitation current in the region The lead field formed inside.

[0032] In some feasible embodiments, step S3 specifically includes: The difference between the voltage measurement under abnormal conditions and the voltage measurement V0 under the initial uniform field condition V1 is taken as the voltage change caused by impedance anomaly. To better capture the spatial distribution characteristics and electrode geometry in impedance imaging, the effective voltage measurement values ​​are mapped and constructed into a two-dimensional voltage data matrix according to the correspondence between the excitation electrode pair and the measurement electrode pair. The dimension of the two-dimensional voltage data matrix of the source domain model is... The target domain is ,like Figure 3 As shown.

[0033] In some feasible embodiments, step S4 specifically includes: For the input two-dimensional voltage data matrix, this invention treats the measurement electrode pair data corresponding to each excitation current mode as a token. The source domain data contains 16 tokens, and the target domain data contains 8 tokens. Taking the target domain as an example, the feature vector of each token is: To enable the encoder to learn the physical spatial relationships between different excitation modes and capture the periodic characteristics of the electrode distribution, this invention introduces Fourier position encoding for each token. The specific process is as follows: Number the tokens according to the adjacent order of the electrode pairs: Set 8 frequencies, from low to high: Calculate the first according to Fourier encoding The location code of each token: By combining the measured voltage features and the location encoding features, a token input is obtained. All tokens are concatenated to form a feature matrix. .

[0034] Furthermore, to enhance the model's ability to learn the physical coupling relationship of electrodes, this invention introduces a lead field similarity matrix as a physical bias in the Transformer attention calculation. To improve the model's adaptability to real data, the similarity matrix... Add learnable correction quantity to the basis The final bias matrix is ​​obtained as follows: In the For the input two-dimensional matrix in the coding layer After normalization, we get: Calculate the attention term: After introducing the physical bias, we get: The updated attention output is: in This is the learnable parameter matrix.

[0035] The updated attention distribution is weighted and fused with the input features of the encoding layer to participate in the update of the encoded representation.

[0036] The single-layer output is obtained after processing by the feedforward network. In this embodiment, both the source and target domains adopt a 4-layer coding structure.

[0037] The encoder output is pooled to obtain a global feature vector, and then a mean vector is generated through linear mapping. With the variance logarithm vector Then, latent variables are generated through reparameterized sampling. : In some feasible embodiments, step S5 specifically includes: Latent variables obtained by reparameterized sampling First, it is mapped to a dimension of A two-dimensional matrix is ​​then generated by a continuously upsampled two-dimensional convolutional layer, resulting in a matrix of size [missing information]. The conductivity image.

[0038] The source and target domains use the same decoder for image reconstruction. This shared decoder ensures that the mapping from latent variables to the image space remains consistent between the two domains, thereby promoting the convergence of encoded latent variables under different electrode numbers within the same latent space and achieving cross-domain consistency learning. Through this mechanism, the latent variables in the target domain can fully inherit the high-information structure provided by the latent space in the source domain, resulting in higher-quality reconstruction.

[0039] To further promote the convergence of target domain features to the source domain distribution, this invention constructs a voltage decoder that decodes the latent variables of the target domain into a two-dimensional voltage matrix of the source domain. This decoder consists of two layers of two-dimensional transposed convolutions and constructs a cross-domain constraint mechanism through supervised reconstruction of source domain voltage data, thereby enhancing cross-domain consistency.

[0040] To ensure the quality of the reconstructed images, mean square error constraints are applied to the reconstructed images in both the source and target domains. Let the reconstructed images generated in the source and target domains be... , Image reconstruction and labeling using MSE constraints Consistency: Apply cross-domain feature consistency constraints to latent variables to ensure that latent variables in the target domain are consistent. To the source domain Approach: Simultaneously, the source domain voltage matrix output by the voltage decoder... Establish cross-domain reconstruction consistency constraints: Therefore, the total loss during the model training process is: in, These are adjustable weights used to balance cross-domain alignment constraints, image reconstruction constraints, and voltage reconstruction constraints. In this embodiment, parameters are set. .

[0041] Based on steps S4 and S5, the model flow built in this example is as follows: Figure 4 As shown in the diagram, during the model training phase, a five-fold cross-validation method was used to divide the dataset. All samples were divided into five equal parts. Each time, three parts were selected as the training set, one as the validation set, and one for inference. This process was repeated five times to improve the model's generalization ability under different data distributions and effectively avoid overfitting. Finally, the model parameters that performed best on the validation set were selected as the fixed model for the inference phase.

[0042] In some feasible embodiments, step S6 specifically includes: During the inference phase, only the feature encoder of the target domain (low electrode configuration) and the shared image reconstruction decoder are used to infer the two-dimensional voltage data matrix acquired under 8-electrode conditions. First, the input voltage matrix is ​​divided into multiple tokens according to the excitation mode, and the spatial relationship between electrode pairs is established through the position encoding module. Then, the latent features are extracted by the target domain dedicated encoder and latent variable representations are generated. Finally, the latent variables are input into the shared image decoder to complete the reconstruction of the conductivity distribution.

[0043] Since the target domain latent variable distribution has been fully aligned to the source domain latent space through cross-domain consistency constraints and physical bias injection during the training phase, the model can generate conductivity images with high resolution, low artifacts and high stability by relying only on 8 electrode input data during the inference phase, effectively compensating for the lack of information caused by the low number of electrodes.

[0044] Based on the overall process of the above method, the present invention also provides a set of data examples: To verify the effectiveness of the model proposed in this invention, this embodiment further compares the reconstruction performance of three types of models: a model trained based on a single 16-electrode input; a model trained based on a single 8-electrode input; and an 8-electrode cross-domain reconstruction model proposed in this invention. The three types of models are systematically compared and verified from both quantitative and qualitative perspectives using image reconstruction metrics (such as Peak Signal-to-Noise Ratio (PSNR), SSIM, and RMSE) and visualized reconstruction results. The table below shows the quantitative comparison of each metric. Figure 5 The reconstruction results of some samples are shown in the comparison, and the results show that the cross-domain reconstruction method proposed in this invention can significantly improve the image reconstruction quality under low electrode conditions.

[0045] Table 1 Performance Comparison of Three Types of Models In summary, this invention constructs a cross-domain learning framework that utilizes the highly informative voltage features contained in the source domain (high electrode count) to guide feature encoding and latent space learning in the target domain (low electrode count), achieving feature transfer across electrode count conditions. This "cross-electrode count domain" transfer method is proposed for the first time in the EIT low-electrode imaging task, effectively alleviating the problems of reduced reconstruction resolution and structural distortion caused by insufficient measurement information, thereby significantly improving image reconstruction quality under low electrode conditions.

[0046] This invention introduces the lead field similarity matrix calculated based on a uniform conductance model as a physical bias into the attention mechanism of the encoding layer. This physical constraint guides the distribution of attention weights between different excitation modes, enabling the model to explicitly learn the spatial coupling relationship and electric field distribution between electrode pairs. By integrating physical information into the feature encoding process, the model not only achieves higher stability but also possesses good physical interpretability, which helps enhance its generalization ability to real EIT data.

[0047] This invention constructs a voltage decoder to decode latent variables in the target domain into a two-dimensional voltage matrix in the source domain, achieving cross-domain reconstruction supervision in the voltage domain. This supervision, together with conductivity image reconstruction supervision, forms a dual reconstruction constraint. This dual-domain supervision mechanism not only enhances the cross-domain consistency of the latent variable space but also improves the fitting ability of the latent space to the voltage-image relationship, thereby further improving the stability of cross-domain learning and the reliability of the reconstruction results.

[0048] An EIT reconstruction system based on cross-domain learning and physical guidance includes: The condition definition module is used to execute step S1; The data preprocessing module is used to perform steps S2 and S3; The model training module is used to perform steps S4 and S5; The model application module is used to execute step S6.

[0049] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0050] An EIT reconstruction device based on cross-domain learning and physical guidance: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements an EIT reconstruction method based on cross-domain learning and physical guidance as described above.

[0051] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0052] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement an EIT reconstruction method based on cross-domain learning and physical guidance as described above.

[0053] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0054] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for EIT reconstruction based on cross-domain learning and physical guidance, characterized in that, Includes the following steps: Set the first and second conditions; Voltage measurement data are collected under the first condition and the second condition respectively, and a two-dimensional voltage matrix is ​​constructed based on the combination relationship between the excitation electrode pair and the measurement electrode pair; Source domain data and target domain data are constructed based on the aforementioned two-dimensional voltage matrix; A physical bias based on the lead field similarity matrix is ​​introduced to encode the source domain data and the target domain data to obtain encoded data; The encoded data is input into the same decoder for conductivity image reconstruction. Cross-domain consistency learning is performed using a loss function to obtain the trained encoder and image reconstruction decoder.

2. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 1, characterized in that, The step of acquiring voltage measurement data under the first and second conditions respectively, and constructing a two-dimensional voltage matrix based on the combination relationship between the excitation electrode pair and the measurement electrode pair, specifically includes: The finite element method was used to calculate the similarity matrix of the lead field between the excitation electrode pairs under uniform conductivity conditions. Voltage measurement data are collected under the first and second conditions, and a two-dimensional voltage data matrix is ​​constructed based on the combination relationship between the excitation electrode pair and the measurement electrode pair. Record the spatial position information of each electrode pair.

3. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 2, characterized in that, The lead field similarity matrix is ​​represented as follows: in, Represents the lead field similarity matrix. express Normalization process, Indicates the number of adjacent electrode pairs. Indicates electrode pair and The energy product of the guide field Indicates the area to be imaged. Indicates electrode pair After the excitation current in the region The internal conductive field, Indicates electrode pair After the excitation current in the region The lead field formed inside.

4. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 2, characterized in that, The step of constructing source domain data and target domain data based on the two-dimensional voltage matrix specifically includes: The two-dimensional voltage matrix constructed under the first condition is used as the source domain data; The two-dimensional voltage matrix constructed under the second condition is used as the target domain data.

5. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 4, characterized in that, The step of performing positional encoding on the source domain data and the target domain data to obtain the data to be encoded specifically includes: The electrode pairs in the source domain data and target domain data are numbered in adjacent order to obtain the electrode pair sequence information; Based on the sequence number information of the electrode pair and the set frequency, the position code is calculated according to the Fourier code to obtain the position code feature; The source domain data and the target domain data are concatenated with the corresponding position encoding features to obtain the data to be encoded.

6. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 4, characterized in that, The process of introducing a physical bias based on the lead field similarity matrix during encoding specifically includes: A learnable correction factor is added to the lead field similarity matrix to obtain the physical bias matrix; The physical bias matrix is ​​added to the calculated result of the attention term, and the attention distribution is updated. The updated attention distribution is weighted and fused with the input features of the encoding layer to participate in the update of the encoded representation.

7. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 4, characterized in that, Also includes: The encoded data is pooled to obtain a global feature vector; A linear mapping is applied to the global feature vector to generate a mean vector and a logarithm vector of variance; Latent variables are generated through reparameterized sampling based on the mean vector and the logarithm of variance vector.

8. The EIT reconstruction method based on cross-domain learning and physical guidance according to claim 4, characterized in that, The loss function includes: in, For adjustable weights, This indicates a consistency constraint on cross-domain reconstruction results. This indicates a cross-domain reconstruction input consistency constraint. This indicates implicit variable constraints. Represents the MSE loss function. Represents the source domain reconstructed image. Represents the image reconstructed from the target domain. Indicates the actual label, This represents a two-dimensional voltage matrix in source domain form obtained by decoding the latent variables of the target domain. A two-dimensional voltage matrix representing the actual source domain data. Represents the mean vector of the source domain. Represents the mean vector of the target domain. express Norm.

9. An EIT reconstruction system based on cross-domain learning and physical guidance, characterized in that, include: The condition definition module is used to set the first and second conditions; The data preprocessing module collects voltage measurement data under the first and second conditions respectively, and constructs a two-dimensional voltage matrix based on the combination relationship between the excitation electrode pair and the measurement electrode pair; and constructs source domain data and target domain data based on the two-dimensional voltage matrix. The model training module introduces a physical bias based on the lead field similarity matrix to encode the source domain data and the target domain data to obtain encoded data. The encoded data is then input into the same decoder for conductivity image reconstruction. Cross-domain consistency learning is performed using a loss function to obtain the trained encoder and image reconstruction decoder.

10. An EIT reconstruction device based on cross-domain learning and physical guidance, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the EIT reconstruction method based on cross-domain learning and physical guidance as described in any one of claims 1-8.

Images