Fault diagnosis method, device and electronic equipment of reactor system

Abstract

This invention relates to the field of nuclear industry technology, specifically to a method, apparatus, and electronic equipment for fault diagnosis of reactor systems. The method includes: acquiring operating parameter information of the reactor system within a current time window; determining an operating parameter time series matrix based on the operating parameter information; performing principal component analysis (PCA) on the operating parameter time series matrix to obtain a reconstructed time series matrix; determining an operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix; acquiring expert fault analysis text information based on the operating parameter information; and inputting the operating parameter residual map and the expert fault analysis text information into a fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model. This invention achieves accurate fault diagnosis of reactor systems, thereby ensuring the safe and stable operation of the reactor.

Description

Technical Field

[0001] This invention relates to the field of nuclear industry technology, and more specifically, to a method, apparatus, and electronic equipment for fault diagnosis of reactor systems. Background Technology

[0002] A nuclear reactor system is a device used to control and manage nuclear fission reactions, primarily for nuclear power, nuclear energy generation, scientific research, and certain industrial applications. It releases a large amount of heat energy by controlling the fission reaction of nuclear fuels such as uranium or plutonium, and converts this heat energy into electrical energy through a heat exchange system. The advantages of a reactor include high energy efficiency, the ability to generate large amounts of energy in a relatively small space, stable energy supply over long periods, no greenhouse gas production, strong environmental adaptability, and the requirement for oxygen. Therefore, it has a very broad application prospect in remote areas, deep-sea submersion, and deep space exploration.

[0003] As high-risk and highly complex nuclear energy facilities, reactors operate in environments characterized by high temperatures, high radiation, high pressures, and strongly coupled multi-physics fields. During long-term operation, factors such as equipment aging, radiation-induced drift of sensitive components, and external disturbances can all induce faults. These faults are often not caused by a single factor but are complex faults resulting from the superposition of multiple factors. Different factors mask or cancel each other out, causing anomalies to appear as weak, short-duration, and intermittent micro-disturbance signals on a single measurement channel, making them difficult to detect in a timely manner using conventional thresholding methods. If not detected and diagnosed promptly, the fault may escalate into a safety incident, even endangering personnel and environmental safety. Therefore, an advanced online fault diagnosis technology is urgently needed to ensure the safe and stable operation of reactors. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, and electronic equipment for fault diagnosis of reactor systems. It combines fault analysis by human experts with residual diagrams of reactor system operating parameters obtained based on principal component analysis, and performs fault diagnosis through a fault diagnosis model. This invention utilizes the complementarity of human prior knowledge and machine data to achieve accurate fault diagnosis of reactor systems, thereby ensuring the safe and stable operation of the reactor.

[0005] A first aspect of the present invention provides a method for fault diagnosis of a reactor system, comprising:

[0006] Obtain the operating parameter information of the reactor system within the current time window, and determine the operating parameter time series matrix based on the operating parameter information;

[0007] Based on principal component analysis, principal component extraction and data reconstruction are performed on the time series matrix of the operating parameters to obtain the reconstructed time series matrix.

[0008] Based on the operating parameter time series matrix and the reconstructed time series matrix, the operating parameter residual map is determined;

[0009] Obtain expert fault analysis text information based on the aforementioned operating parameter information from human experts;

[0010] The residual diagram of the operating parameters and the expert fault analysis text information are input into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the residual diagram of the operating parameters of the reactor system and the expert fault analysis text information.

[0011] In some embodiments, the fault diagnosis model includes an encoding layer, a feature extraction layer, a feature fusion layer, and a fault diagnosis layer connected in sequence; the step of inputting the operating parameter residual map and the expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model includes:

[0012] The residual map of the operating parameters and the text information of the expert fault analysis are input into the encoding layer. Based on the encoding layer, the residual map of the operating parameters and the text information of the expert fault analysis are encoded to obtain image encoding vector and text encoding vector respectively.

[0013] Based on the feature extraction layer, feature extraction is performed on the image encoding vector and the text encoding vector respectively to obtain image feature vector and text feature vector;

[0014] Based on the feature fusion layer, the image feature vector and the text feature vector are fused to obtain a fused feature vector;

[0015] Based on the fault diagnosis layer, fault diagnosis is performed on the fused feature vector to obtain the fault diagnosis result.

[0016] In some embodiments, the step of fusing the image feature vector and the text feature vector to obtain a fused feature vector includes:

[0017] The image feature vector is enhanced based on the text feature vector to obtain an enhanced image feature vector.

[0018] Based on the cross-attention mechanism, the enhanced image feature vector and the text feature vector are fused to obtain a preliminary fused feature vector.

[0019] Based on the pre-gating mechanism, the preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are fused to obtain the fused feature vector.

[0020] As one possible implementation, the enhancement processing of the image feature vector based on the text feature vector to obtain the enhanced image feature vector includes:

[0021] Based on the text feature vector, a scaling factor and an offset factor with the same dimension as the image feature vector are generated;

[0022] Based on the scaling factor and the offset factor, a channel-by-channel affine transformation is performed on the image feature vector to obtain the enhanced image feature vector.

[0023] As one possible implementation, the method of fusing the enhanced image feature vector and the text feature vector based on the cross-attention mechanism to obtain a preliminary fused feature vector includes:

[0024] A linear transformation is performed on the enhanced image feature vector to obtain the key matrix and the value matrix;

[0025] A linear transformation is performed on the text feature vector to obtain the query matrix;

[0026] Based on the key matrix and the query matrix, the attention score is determined;

[0027] The attention scores are normalized to obtain the attention weights;

[0028] The value matrix is ​​weighted based on the attention weights to obtain the preliminary fusion feature vector.

[0029] As one possible implementation, the method of fusing the preliminary fused feature vector, the text feature vector, and the enhanced image feature vector based on a pre-gating mechanism to obtain the fused feature vector includes:

[0030] The preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are concatenated to obtain the concatenated feature vector.

[0031] The concatenated feature vector is mapped to obtain the first channel-level gate corresponding to the preliminary fused feature vector and the second channel-level gate corresponding to the text feature vector;

[0032] The first channel-level threshold and the second channel-level threshold are normalized to obtain the first weight corresponding to the preliminary fused feature vector and the second weight corresponding to the text feature vector.

[0033] The preliminary fused feature vector and the text feature vector are weighted and fused based on the first weight and the second weight to obtain the fused feature vector.

[0034] In some embodiments, the method further includes:

[0035] Obtain expert fault diagnosis results from human experts based on the aforementioned operating parameter information;

[0036] The fault diagnosis results and the expert fault diagnosis results are input into the integrated diagnosis model to obtain the final fault diagnosis results output by the integrated diagnosis model; the integrated diagnosis model has learned the ability to perform integrated diagnosis based on the expert fault diagnosis results and the fault diagnosis results output by the fault diagnosis model.

[0037] As an example, the integrated diagnostic model is constructed based on a feature pyramid network.

[0038] A second aspect of the present invention provides a fault diagnosis device for a reactor system, comprising:

[0039] The first acquisition module is used to acquire the operating parameter information of the reactor system within the current time window, and determine the operating parameter time series matrix based on the operating parameter information;

[0040] The second acquisition module is used to perform principal component extraction and data reconstruction on the time series matrix of the operating parameters based on principal component analysis to obtain the reconstructed time series matrix.

[0041] The determination module is used to determine the operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix;

[0042] The third acquisition module is used to acquire expert fault analysis text information based on the operating parameter information by human experts.

[0043] The diagnostic module is used to input the operating parameter residual diagram and the expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the operating parameter residual diagram and expert fault analysis text information of the reactor system.

[0044] A third aspect of the present invention provides an electronic device including a processor and a memory storing a computer program, wherein the processor executes the program to implement the fault diagnosis method for a reactor system described in the first aspect above.

[0045] The reactor system fault diagnosis method, apparatus, and electronic equipment provided by this invention have the following beneficial effects:

[0046] This invention acquires the operating parameter information of a reactor system within the current time window, determines the operating parameter time series matrix based on this information, performs principal component analysis (PCA) on the operating parameter time series matrix to obtain a reconstructed time series matrix, determines the operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix, acquires expert fault analysis text information based on the operating parameter information, and inputs the operating parameter residual map and expert fault analysis text information into a fault diagnosis model to obtain the fault diagnosis results output by the fault diagnosis model. This invention combines expert fault analysis text information and operating parameter residual map to achieve online fault diagnosis of reactor systems. By leveraging the complementarity between human prior knowledge and machine data, the accuracy of fault diagnosis can be improved, enabling timely fault diagnosis and handling, and ensuring the safe and stable operation of the reactor. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0048] Figure 1 A schematic flowchart illustrating a fault diagnosis method for a reactor system provided in an embodiment of the present invention;

[0049] Figure 2 This is an example diagram of the residual plot of the running parameters in an embodiment of the present invention, which illustrates four different specific examples (a), (b), (c), and (d).

[0050] Figure 3 This is a schematic diagram of the structure of a fault diagnosis model in an embodiment of the present invention;

[0051] Figure 4 A flowchart illustrating another method for fault diagnosis of a reactor system provided in an embodiment of the present invention;

[0052] Figure 5 A flowchart illustrating another method for fault diagnosis of a reactor system provided in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of an architecture for enhancing image feature vectors based on text feature vectors in an embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of another fault diagnosis model in an embodiment of the present invention;

[0055] Figure 8A flowchart illustrating another method for fault diagnosis of a reactor system provided in an embodiment of the present invention;

[0056] Figure 9 This is an example diagram of the architecture of the integrated diagnostic model in an embodiment of the present invention;

[0057] Figure 10 Error statistics for six fault diagnosis methods during model training and validation;

[0058] Figure 11 This is a schematic diagram of the first-order SHAP result output by the integrated diagnostic model in an embodiment of the present invention;

[0059] Figure 12 This is a schematic diagram illustrating the weight distribution of text feature vectors in the pre-gating mechanism in an embodiment of the present invention;

[0060] Figure 13 This is a schematic diagram of the structure of a fault diagnosis device for a reactor system provided in an embodiment of the present invention;

[0061] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] This invention provides a method, apparatus, and electronic device for fault diagnosis of reactor systems. By combining expert fault analysis text information and operating parameter residual diagrams, a fault diagnosis model is used to achieve real-time online diagnosis of reactor faults.

[0064] Figure 1 This is a flowchart illustrating a fault diagnosis method for a reactor system provided in an embodiment of the present invention. Figure 1 As shown, the fault diagnosis method for a reactor system according to an embodiment of the present invention includes the following steps:

[0065] S101: Obtain the operating parameter information of the reactor system within the current time window, and determine the operating parameter timing matrix based on the operating parameter information.

[0066] In some embodiments, each time window can be a preset duration, such as a 15-second window, and the current time window can be a continuous 15-second period including the current moment and previous moments. The reactor system's operating parameter information can be operating parameters related to fault diagnosis, such as thermal power, average fuel temperature, average emitter temperature, average receiver temperature, output current, output voltage, output electrical power, coolant inlet temperature, coolant outlet temperature, and average fin temperature. The reactor system's operating parameter information can be collected using measuring devices such as sensors.

[0067] As an example, if the current time window is 15 seconds long, and the sensor and other measuring devices sample every 1 second, and the operating parameter information includes 10 parameters, then the operating parameter time series matrix determined based on the operating parameter information is a 10×15 time series matrix.

[0068] S102, based on principal component analysis, performs principal component extraction and data reconstruction on the time series matrix of operating parameters to obtain the reconstructed time series matrix.

[0069] In some embodiments, principal components in the time series matrix of operating parameters can be extracted based on kernel principal component analysis (KPCA). KPCA can target nonlinear correlations in operating parameter information by mapping the original low-dimensional nonlinear data to a high-dimensional feature space through a kernel function, making the data in the high-dimensional space exhibit linear correlations. Then, principal components are extracted using linear principal component analysis, thereby indirectly capturing the nonlinear characteristics of the original data.

[0070] As one possible approach, the nuclear principal component analysis model can be trained based on the operating parameter information of the reactor system under normal conditions over a historical period to obtain a principal component subspace and inverse transformation parameters. The time series matrix of the operating parameters is then projected onto the principal component subspace and reconstructed based on the inverse transformation parameters to obtain the reconstructed time series matrix.

[0071] S103, based on the time series matrix of the operating parameters and the reconstructed time series matrix, determine the residual map of the operating parameters.

[0072] The operating parameter residual plot is determined based on the reconstructed residuals. While training a kernel principal component analysis model can reconstruct the operating parameter information of a reactor system under normal conditions, a large reconstructed residual will occur when the reactor system malfunctions. Therefore, the operating parameter residual plot can effectively characterize the fault features of the reactor system.

[0073] In some embodiments, each element in the running parameter time series matrix can be subtracted from the corresponding element in the reconstructed time series matrix, and a reconstructed residual matrix can be constructed based on the obtained difference. Each element in the reconstructed residual matrix is ​​then subjected to max-min normalization, and the elements in the normalized reconstructed residual matrix are visualized in the form of a heatmap with time as the horizontal axis and different parameters as the vertical axis, resulting in the following... Figure 2 The residual plot of the operating parameters shown is as follows. Figure 2 The illustration is based on a time series matrix with a runtime parameter dimension of 10×15, and shows schematic diagrams of four specific examples (a), (b), (c), and (d).

[0074] S104, Obtain expert fault analysis text information based on operating parameter information from human experts.

[0075] Among them, the expert fault analysis text information is a description of the nonlinear characteristics of changes in operating parameters by human experts, such as "the coolant temperature at the reactor core inlet and outlet has increased significantly, while no significant changes have been observed in thermal power, current, voltage, fin average temperature and mass flow rate".

[0076] In some embodiments, the expert fault analysis text information may be input by human experts through an interactive interface, or it may be obtained from a database used to store human expert fault analysis data, wherein the database stores human expert fault analysis text information at various times.

[0077] S105, input the residual diagram of operating parameters and expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis results output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the residual diagram of operating parameters of the reactor system and expert fault analysis text information.

[0078] In practical applications, human experts analyze the operating parameters of reactor systems for fault diagnosis. However, these analyses may be biased or inaccurate, making manual fault diagnosis unreliable. To improve the accuracy and efficiency of fault diagnosis, this invention combines residual graphs of operating parameters that characterize fault features with expert fault analysis text information. A fault diagnosis model is used to achieve online real-time fault diagnosis of reactor systems, leveraging the complementarity between human prior knowledge and machine data to enhance the accuracy of fault diagnosis.

[0079] In some embodiments, the fault diagnosis model can be a neural network model, a machine learning model, etc. The fault diagnosis model can extract fault features based on the residual graph of operating parameters and expert fault analysis text information, and then map the extracted fault features to obtain the fault diagnosis result.

[0080] As an example, the fault diagnosis model can be obtained by training a large number of historical operating parameter information of the reactor system. The training samples can include operating parameter residual map samples, expert fault analysis text information samples, and corresponding fault labels. Based on the training samples, the fault diagnosis model is trained under supervision. The loss function of the training process can be the binary cross-entropy loss function, and the expression of the loss function is as follows (1). By calculating the binary cross-entropy loss and combining it with the backpropagation algorithm, the parameters of the fault diagnosis model are gradually optimized to achieve effective learning and convergence of the model.

[0081] (1);

[0082] In the formula, These are genuine fault labels. It is the failure probability output by the model.

[0083] In some embodiments of the present invention, since multiple faults may occur simultaneously in the reactor system, the fault type label can be represented by multi-hot encoding, that is, a binary bit is assigned to each fault category, which is set to 1 if it exists and 0 otherwise.

[0084] As an example, taking the heat pipe improved TOPAZ-II reactor as an example, and diagnosing possible faults in its key components, including the sodium-potassium heat pipe, control drum shaft, cesium steam generator, and electromagnetic pump, the fault labels are shown in Table 1 below when determining training samples based on historical operating data.

[0085] Table 1 Examples of Fault Labels

[0086]

[0087] The reactor system fault diagnosis method according to embodiments of the present invention acquires the operating parameter information of the reactor system within the current time window, determines the operating parameter time series matrix based on the operating parameter information, performs principal component extraction and data reconstruction on the operating parameter time series matrix using principal component analysis to obtain a reconstructed time series matrix, determines the operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix, acquires expert fault analysis text information based on the operating parameter information, and inputs the operating parameter residual map and expert fault analysis text information into a fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model. This invention combines expert fault analysis text information and operating parameter residual map to achieve online fault diagnosis of the reactor system. Utilizing the complementarity between human prior knowledge and machine data, it can also improve the accuracy of fault diagnosis, thereby enabling timely fault diagnosis and handling, and ensuring the safe and stable operation of the reactor.

[0088] Figure 3 This is a schematic diagram of the structure of a fault diagnosis model according to an embodiment of the present invention. Figure 3 As shown, the fault diagnosis model comprises an encoding layer, a feature extraction layer, a feature fusion layer, and a fault diagnosis layer connected in sequence. Figure 4 As shown, based on Figure 3 The fault diagnosis model shown is Figure 1 The implementation process of step S105 includes the following steps:

[0089] S401, input the residual map of operating parameters and the text information of expert fault analysis into the coding layer. Based on the coding layer, encode the residual map of operating parameters and the text information of expert fault analysis respectively to obtain the image coding vector and the text coding vector.

[0090] In some embodiments, the encoding layer includes a first encoding module and a second encoding module. The first encoding module is used to encode the residual map of operating parameters, and the second encoding module is used to encode the text information of expert fault analysis. That is, the residual map of operating parameters is input to the first encoding module to obtain an image encoding vector, and the text information of expert fault analysis is input to the second encoding module to obtain a text encoding vector.

[0091] As an example, the first encoding module uses convolution and normalization to map the runtime parameter residual map into several vectors, and then superimposes positional encoding to obtain an image encoding vector with a dimension of 3×384.

[0092] As an example, the second encoding module inputs expert fault analysis text information into the BCP (Byte-level Character-Piece) segmenter, splits words and sentences into word units and converts them into indexes, and obtains a text encoding vector with a dimension of 1×128 through the embedding layer.

[0093] S402, based on the feature extraction layer, performs feature extraction on the image encoding vector and the text encoding vector respectively to obtain the image feature vector and the text feature vector.

[0094] In some embodiments, the feature extraction layer can utilize the Transformer multi-head attention mechanism to extract features from the image encoding vector and the text encoding vector respectively, resulting in an image feature vector with a dimension of 1×768 and a text feature vector with a dimension of 1×64.

[0095] S403, based on the feature fusion layer, fuses image feature vectors and text feature vectors to obtain fused feature vectors.

[0096] In some embodiments, the feature fusion layer can be based on a cross-attention mechanism to fuse image feature vectors and text feature vectors. For example, the feature fusion layer can be a cross-attention network, which performs linear transformations on image feature vectors and text feature vectors to obtain a query matrix Q, a key matrix K, and a value matrix V; calculates an attention score based on the dot product of the query matrix Q and the key matrix K; obtains attention weights by normalizing the attention scores; and obtains the fused feature vector by performing a weighted summation operation on the attention weights and the value matrix.

[0097] In some embodiments, the feature fusion layer may first map the image feature vector and the text feature vector into feature vectors of the same dimension; based on preset weights, the two processed feature vectors are weighted to obtain a fused feature vector.

[0098] S404, based on the fault diagnosis layer, performs fault diagnosis on the fused feature vector to obtain fault diagnosis results.

[0099] In some embodiments, after model training, the fault diagnosis layer has established a mapping relationship between feature vectors and faults, thereby enabling the determination of the corresponding fault diagnosis result based on the fused feature vector output by the feature fusion layer. The fault diagnosis layer can be constructed based on a fully connected network.

[0100] According to an embodiment of the present invention, a fault diagnosis method for a reactor system includes a fault diagnosis model comprising a coding layer, a feature extraction layer, a feature fusion layer, and a fault diagnosis layer connected in sequence. Based on the coding layer, residual graphs of operating parameters and expert fault analysis text information are encoded to obtain image coding vectors and text coding vectors, respectively. Based on the feature extraction layer, features are extracted from the image coding vectors and text coding vectors to obtain image feature vectors and text feature vectors, respectively. Based on the feature fusion layer, the image feature vectors and text feature vectors are fused to obtain a fused feature vector. Based on the fault diagnosis layer, fault diagnosis is performed on the fused feature vector to obtain a fault diagnosis result. This invention can integrate residual graph features and text features from human expert fault analysis for fault diagnosis, thereby extracting more comprehensive fault features and making the fault diagnosis result more accurate.

[0101] Next, we will provide a detailed introduction to the process of fusing image feature vectors and text feature vectors in the feature fusion layer based on the fault diagnosis model.

[0102] Figure 5 This is a flowchart illustrating another method for fault diagnosis of a reactor system provided in an embodiment of the present invention. Figure 5As shown, in step S403, the process of fusing image feature vectors and text feature vectors based on the feature fusion layer to obtain fused feature vectors may include the following steps:

[0103] S501, enhances the image feature vector based on the text feature vector to obtain the enhanced image feature vector.

[0104] In some embodiments, the feature fusion layer can adjust the image feature vector based on the text feature vector to highlight information related to the text features in the image features and suppress irrelevant interference, thereby achieving image feature vector enhancement processing.

[0105] As one possible implementation, the feature fusion layer includes a lightweight neural network. The text feature vector is input into this lightweight neural network, and the output is a gated vector that matches the dimension of the image feature vector. Based on the feature fusion layer, the gated vector is activated by the sigmoid activation function so that the value of each element in the activated gated vector is between 0 and 1. Then, the activated gated vector is multiplied element by element with the image features to obtain the enhanced image feature vector.

[0106] As another possible implementation, a channel-wise affine transformation of the image feature vector can be performed based on the feature fusion layer, specifically including the following steps:

[0107] S5011 generates scaling and offset coefficients with the same dimensions as image feature vectors based on text feature vectors.

[0108] S5012 performs a channel-by-channel affine transformation on the image feature vector based on scaling and offset coefficients to obtain an enhanced image feature vector.

[0109] As one possible implementation, the feature fusion layer includes a feature enhancement unit, which may include a fully connected network and an affine transformation module. The text feature vector is input into the fully connected network to obtain scaling and offset coefficients with the same dimension as the image feature vector output by the fully connected network. The scaling, offset, and image feature vector are then input into the affine transformation module to perform a channel-by-channel affine transformation on the image feature vector based on the scaling and offset coefficients to obtain an enhanced image feature vector.

[0110] As an example, such as Figure 6 As shown, the affine transformation module can perform channel-by-channel affine transformation on the image feature vector using the following formula (2).

[0111] (2);

[0112] in, To enhance image feature vectors; This is the scaling factor; This is the offset coefficient; Image feature vector; This indicates element-wise multiplication.

[0113] In other words, the feature enhancement unit of the feature fusion layer can dynamically enhance the visual cues of the residual map based on the key phenomena or trends in the text feature vector, and suppress irrelevant interference.

[0114] S502, based on the cross-attention mechanism, fuses the enhanced image feature vector and the text feature vector to obtain a preliminary fused feature vector.

[0115] In some embodiments, the feature fusion layer includes not only feature enhancement units but also cross-attention units connected to them. Specifically, based on the cross-attention mechanism, fusing the enhanced image feature vector and the text feature vector to obtain a preliminary fused feature vector may include the following steps:

[0116] S5021 performs a linear transformation on the enhanced image feature vector to obtain the key matrix and value matrix.

[0117] S5022, perform a linear transformation on the text feature vector to obtain the query matrix.

[0118] In some embodiments, the enhanced image feature vector and text feature vector can be linearly transformed based on different fully connected networks of the cross-attention unit, uniformly mapped to a 1×96 feature vector, and a key matrix K, a value matrix V, and a query matrix Q can be generated, with the key matrix K, the value matrix V, and the query matrix Q all having a dimension of 1×96.

[0119] S5023, based on the key matrix and query matrix, determines the attention score matrix.

[0120] As one possible implementation, the attention score matrix can be determined by calculating the dot product of the query matrix Q and the key matrix K based on the cross attention unit. The dimension of the attention score matrix is ​​as shown in the following equation (3).

[0121] (3);

[0122] in, This is the attention score matrix; The key matrix; denoted as the dimension of the key matrix.

[0123] S5024, normalize the attention score matrix to obtain the attention weight matrix.

[0124] As an example, a cross-attention unit may include a Softmax function, which normalizes each element in the attention score matrix to obtain an attention weight matrix.

[0125] S5025 performs weighted operations on the value matrix based on the attention weight matrix to obtain a preliminary fused feature vector.

[0126] As an example, the value matrix can be weighted and summed based on the attention weight matrix to obtain a preliminary fused feature vector.

[0127] S503, based on a pre-gating mechanism, performs fusion processing on the initial fused feature vector, text feature vector, and enhanced image feature vector to obtain a fused feature vector.

[0128] In other words, in addition to feature enhancement units and cross-attention units connected to them, the feature fusion layer may also include pre-gating units. The pre-gating units are used to perform fusion processing on the preliminary fused feature vector, text feature vector and enhanced image feature vector based on the pre-gating mechanism to obtain the fused feature vector.

[0129] As one possible implementation, based on a pre-gating mechanism, the initial fused feature vector, text feature vector, and enhanced image feature vector are fused to obtain the fused feature vector. The implementation steps include:

[0130] S5031, the initial fused feature vector, text feature vector and enhanced image feature vector are concatenated to obtain the concatenated feature vector.

[0131] As an example, the concatenated feature vector can have a dimension of 3×96.

[0132] S5032, the concatenated feature vectors are mapped to obtain the first channel-level gate vector corresponding to the preliminary fused feature vector and the second channel-level gate vector corresponding to the text feature vector.

[0133] S5033, normalize the first channel-level gating vector and the second channel-level gating vector to obtain the first weight vector corresponding to the preliminary fused feature vector and the second weight vector corresponding to the text feature vector.

[0134] As an example, based on the Softmax function in the pre-gated unit, the first-channel-level gating vector and the second-channel-level gating vector are normalized so that the weights of both paths are greater than 0 and add up to 1.

[0135] S5034, the preliminary fused feature vector and the text feature vector are weighted and fused based on the first weight vector and the second weight vector to obtain the fused feature vector.

[0136] In some embodiments, the pre-gating unit may include a lightweight fully connected network, into which the concatenated feature vectors are input to obtain a first-channel-level gating vector and a second-channel-level gating vector. The principle of the lightweight fully connected network is shown in equations (4) and (5) below.

[0137] (4);

[0138] (5);

[0139] in, This is the first-channel level gate vector; This is the second-channel level gate vector; The concatenated feature vector; , , , All of these are model parameters.

[0140] As an example, the fusion module of the pre-gated unit can perform weighted fusion of the preliminary fusion feature vector and the text feature vector based on the first weight vector and the second weight vector to obtain the fusion feature vector. The specific implementation process is shown in the following formula (6).

[0141] (6);

[0142] in, To fuse feature vectors; This is for the initial fusion of feature vectors; For text feature vectors; This is the first-channel level gate vector; This is the second-channel level gate vector; Layer normalization is performed to eliminate numerical scaling fluctuations.

[0143] In some embodiments of the present invention, the feature fusion layer injects textual information into image features through feature enhancement processing, then uses a cross-attention mechanism to allow the textual features to extract relevant information from the enhanced image feature vector for fusion, and finally uses a pre-gating mechanism to fuse the textual feature vector and the initially fused feature vector. In this way, the fused feature vector contains both image and textual information, and the two interact deeply, thereby better representing the fault characteristics of the reactor system.

[0144] Figure 7 This is a schematic diagram of another fault diagnosis model in an embodiment of the present invention. Figure 7As shown, the encoding layer in the fault diagnosis model first encodes the residual map of the running parameters and the text information of the expert fault analysis, respectively, to obtain an image encoding vector with a dimension of 3×384 and a text encoding vector with a dimension of 1×128. The feature extraction layer uses the Transformer multi-head attention mechanism to extract features from the image encoding vector and the text encoding vector, respectively, to obtain an image feature vector with a dimension of 1×768 and a text feature vector with a dimension of 1×64. The feature fusion layer first performs an affine transformation on the image feature vector based on the text feature vector, and then uses a linear transformation to map the text feature vector and the enhanced image feature vector to a unified dimension of 1×96. After a cross-attention mechanism, a preliminary fused feature vector is obtained. Then, the concatenated feature vector is passed through a pre-gating mechanism to obtain a gated vector. The text feature vector and the preliminary fused feature vector are then fused to obtain a fused feature vector of 1×96. Based on the fault diagnosis layer, the fused feature vector is mapped through a fully connected network and then normalized using a sigmoid function to obtain the output fault diagnosis result. Among them, the fault diagnosis result can be the fault probability distribution result, and the fault corresponding to the highest probability value is identified as the fault currently existing in the reactor system.

[0145] The fault diagnosis method for reactor systems according to embodiments of the present invention, through feature enhancement processing, cross-attention mechanism and pre-gating mechanism, aligns and deeply fuses the text vector corresponding to the expert fault analysis text information with the image feature vector corresponding to the system operating parameter residual map, thereby improving the characterization capability of fault features and further improving the accuracy of fault diagnosis.

[0146] Since human experts also have sufficient experience in fault diagnosis, the present invention also provides the following embodiments to further improve the accuracy of fault diagnosis.

[0147] Figure 8 This is a flowchart illustrating another method for fault diagnosis of a reactor system provided in an embodiment of the present invention. Figure 8 As shown, the method includes the following steps:

[0148] S801 acquires the operating parameter information of the reactor system within the current time window, and determines the operating parameter timing matrix based on the operating parameter information.

[0149] S802, based on principal component analysis, performs principal component extraction and data reconstruction on the time series matrix of operating parameters to obtain the reconstructed time series matrix.

[0150] S803 determines the residual map of operating parameters based on the operating parameter time series matrix and the reconstructed time series matrix.

[0151] S804: Obtain expert fault analysis text information based on operating parameter information from human experts.

[0152] S805 inputs the residual diagram of operating parameters and expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis results output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the residual diagram of operating parameters of the reactor system and expert fault analysis text information.

[0153] S806, obtain expert fault diagnosis results from human experts based on operating parameter information.

[0154] In this context, expert fault diagnosis results can be inferences about faults in various components of the system made by human experts based on operating parameter information. There can be multiple human experts, and for fault predictions from multiple experts, the principle of majority rule can be followed to determine the expert fault diagnosis result.

[0155] In some embodiments, expert fault diagnosis results can be input by relevant personnel through an interactive interface, or retrieved from an expert fault diagnosis database through relevant interfaces. In other words, human experts can make fault predictions for various components in real time based on the reactor system's operating parameters. Because human experts possess a certain level of professional knowledge and experience, their fault diagnosis results can provide the model with a clearer understanding of fault characteristics. However, human experts can also make mistakes. Therefore, to improve diagnostic accuracy, an integrated diagnostic model can be used to combine expert fault diagnosis results with the fault diagnosis results output by the fault diagnosis model, outputting a high-confidence fault probability distribution.

[0156] As an example, there can be multiple human experts, such as seven. These human experts perform the following fault diagnosis based on operational parameter information:

[0157] Expert 1: "Preliminary assessment suggests that the problem lies solely with the heat pipes in the radiator."

[0158] Expert 2: "Preliminary assessment suggests that the problem lies solely with the heat pipes in the radiator."

[0159] Expert 3: "Preliminary assessment suggests that the heat pipe in the radiant radiator is malfunctioning, and the cesium vapor generator and electromagnetic pump may also be faulty."

[0160] Expert 4: "Preliminary assessment suggests that the problem lies only with the heat pipes in the radiator, and the electromagnetic pump may also be faulty."

[0161] Expert 5: "Preliminary assessment suggests that the problem lies only with the heat pipes in the radiator, while the cesium vapor generator may also be malfunctioning."

[0162] Expert 6: "Preliminary assessment suggests that the problem lies solely with the heat pipes in the radiator."

[0163] Expert 7: "Preliminary assessment suggests that the problem lies solely with the heat pipes in the radiator."

[0164] Based on the above fault diagnoses from human experts and adhering to the principle of majority rule, the expert fault diagnosis is determined to be a heat pipe fault.

[0165] S807 inputs the fault diagnosis results output by the fault diagnosis model and the expert fault diagnosis results into the integrated diagnosis model to obtain the final fault diagnosis results output by the integrated diagnosis model; the integrated diagnosis model has learned the ability to perform integrated diagnosis based on the expert fault diagnosis results and the fault diagnosis results output by the fault diagnosis model.

[0166] In some embodiments, the integrated diagnostic model can first combine the fault diagnosis results output by the fault diagnosis model and the expert fault diagnosis results, and then output the fault probability of each component of the system through feature extraction and mapping, which is the final fault diagnosis result.

[0167] Among them, such as Figure 9 As shown, the integrated diagnostic model can be constructed based on a feature pyramid network. It uses a diagnostic matrix composed of the fault diagnosis results output by the fault diagnosis model and the expert fault diagnosis results, employing bidirectional feature transfer via bottom-up, top-down, and lateral connections to generate multi-scale features with strong semantic information. Then, by convolutional operations on the extracted features and passing them through a sigmoid layer, the probability distribution of fault diagnosis for each subsystem is output, which is the final fault diagnosis result. In this embodiment, as... Figure 9 As shown, taking a convolutional layer with different scale extraction capabilities from P1 to P5 as an example, P1 has the largest scale and P5 has the smallest scale.

[0168] In some embodiments, the ensemble diagnostic model can be trained together with the fault diagnosis model, or it can be trained separately. The training samples for the ensemble diagnostic model include expert fault diagnosis result samples and fault diagnosis result samples output by the fault diagnosis model, along with corresponding fault labels. The training process can employ supervised training, and the loss function can be the binary cross-entropy loss function.

[0169] To verify the effectiveness of the reactor system fault diagnosis method of this invention, the inventors tested six fault diagnosis methods: (1) fault diagnosis based solely on text features corresponding to expert fault analysis text information through a model; (2) fault diagnosis based solely on image features corresponding to the residual map of operating parameters through a model; (3) fault diagnosis based on the fused features through multimodal fusion of text features and image features; (4) fault diagnosis based on the fused features through multimodal fusion of text features and image features using a cross-attention mechanism; (5) fault diagnosis based on the fault diagnosis model using the enhanced processing + cross-attention mechanism + pre-gating mechanism of this invention; and (6) fault diagnosis using the integrated diagnosis model of this invention. Figure 10 As shown, by statistically analyzing the errors of the above six fault diagnosis methods during model training and model validation, it was found that compared with using text feature vectors or image feature vectors alone for fault diagnosis, multimodal fusion fault diagnosis can significantly reduce training and validation errors. Furthermore, compared with fusion methods that only apply cross-attention mechanisms, the multimodal fusion mechanism of enhancement processing + cross-attention mechanism + pre-gating mechanism provided by this invention can further improve diagnostic accuracy. In addition, compared with fault diagnosis models alone, the validation error of the integrated diagnostic model of this invention is less than 2%, further improving diagnostic accuracy.

[0170] Regarding the diagnostic method in (6) above, that is, the fault diagnosis method of the integrated diagnostic model provided in the embodiments of the present invention, the following is combined with Figure 11 , Figure 12 Further explanation will follow.

[0171] Figure 11 The first-order SHAP (Shapley Additive Explanations) results output by the integrated diagnostic model in this embodiment of the invention are shown below. Taking the first-order SHAP calculation results of heat pipe failure as an example, the horizontal axis represents the sequence number, and the vertical axis represents the normalized first-order SHAP value. From the overall sorting, sequences 1-4 represent the machine diagnostic SHAP values ​​(diagnostic probability distribution / contribution value) for the four components. Sequence numbers 5-32 are grouped into sets of four, each corresponding to an expert's diagnostic results for the four components. Therefore, there are a total of 7 experts providing 7 sets of expert diagnostic results. Figure 11It can be seen that the features with the highest contribution values ​​are mainly concentrated in the expert diagnosis probability distribution of numbers 5-32, and their SHAP absolute values ​​are significantly higher than those of the machine diagnosis probability distribution of numbers 1-4. This indicates that under the current model structure and sample distribution conditions, the diagnostic network relies more on the fault judgment information given by human experts when making output decisions, rather than relying entirely on the initial prediction results from the machine. Furthermore, from the perspective of contribution amplitude distribution, expert diagnosis features not only dominate in quantity, but also have a larger range of SHAP value fluctuations, indicating that these features have a stronger ability to regulate the model output.

[0172] Figure 12 This is the weight distribution of text feature vectors in the pre-gating mechanism in this embodiment of the invention, where the horizontal and vertical axes both represent indices. It can be viewed as representing a 1×96 matrix as a 6×16 matrix. Figure 12 It can be seen that the weight of text feature vectors is mostly greater than 0.5, indicating that text input information provided by human experts in the relevant field plays a dominant role in the fault diagnosis process; at the same time, the weight of local regions is relatively low, indicating that although image modal features do not play a dominant role, they still play an important constraining and auxiliary role.

[0173] Combination such as Figure 11 and Figure 12 It is evident that the diagnosis and analysis of human experts contribute significantly to the final fault diagnosis results, thus better demonstrating the significance of integrating the fault diagnosis of human experts with the fault diagnosis output of the model.

[0174] To achieve the above embodiments, the present invention also provides a fault diagnosis device for a reactor system.

[0175] Figure 13 This is a schematic diagram of a fault diagnosis device for a reactor system provided in an embodiment of the present invention. Figure 13 As shown, the reactor system fault diagnosis device of this embodiment includes a first acquisition module 1301, a second acquisition module 1302, a determination module 1303, a third acquisition module 1304, and a diagnosis module 1305. Wherein,

[0176] The first acquisition module 1301 is used to acquire the operating parameter information of the reactor system within the current time window, and determine the operating parameter timing matrix based on the operating parameter information;

[0177] The second acquisition module 1302 is used to perform principal component extraction and data reconstruction on the time series matrix of the running parameters based on principal component analysis to obtain the reconstructed time series matrix.

[0178] The determination module 1303 is used to determine the operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix;

[0179] The third acquisition module 1304 is used to acquire expert fault analysis text information based on the operating parameter information by human experts;

[0180] The diagnostic module 1305 is used to input the operating parameter residual diagram and the expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the operating parameter residual diagram and expert fault analysis text information of the reactor system.

[0181] In some embodiments, the fault diagnosis model includes an encoding layer, a feature extraction layer, a feature fusion layer, and a fault diagnosis layer connected in sequence; the diagnosis module 1305 is specifically used for:

[0182] The residual map of the operating parameters and the text information of the expert fault analysis are input into the encoding layer. Based on the encoding layer, the residual map of the operating parameters and the text information of the expert fault analysis are encoded to obtain image encoding vector and text encoding vector respectively.

[0183] Based on the feature extraction layer, feature extraction is performed on the image encoding vector and the text encoding vector respectively to obtain image feature vector and text feature vector;

[0184] Based on the feature fusion layer, the image feature vector and the text feature vector are fused to obtain a fused feature vector;

[0185] Based on the fault diagnosis layer, fault diagnosis is performed on the fused feature vector to obtain the fault diagnosis result.

[0186] In some embodiments, the diagnostic module 1305 is further configured to:

[0187] The image feature vector is enhanced based on the text feature vector to obtain an enhanced image feature vector.

[0188] Based on the cross-attention mechanism, the enhanced image feature vector and the text feature vector are fused to obtain a preliminary fused feature vector.

[0189] Based on the pre-gating mechanism, the preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are fused to obtain the fused feature vector.

[0190] In some embodiments, the diagnostic module 1305 is further configured to:

[0191] Based on the text feature vector, a scaling factor and an offset factor with the same dimension as the image feature vector are generated;

[0192] Based on the scaling factor and the offset factor, a channel-by-channel affine transformation is performed on the image feature vector to obtain the enhanced image feature vector.

[0193] In some embodiments, the diagnostic module 1305 is further configured to:

[0194] A linear transformation is performed on the enhanced image feature vector to obtain the key matrix and the value matrix;

[0195] A linear transformation is performed on the text feature vector to obtain the query matrix;

[0196] Based on the key matrix and the query matrix, determine the attention score matrix;

[0197] The attention score matrix is ​​normalized to obtain the attention weight matrix;

[0198] The value matrix is ​​weighted based on the attention weight matrix to obtain the preliminary fusion feature vector.

[0199] In some embodiments, the diagnostic module 1305 is further configured to:

[0200] The preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are concatenated to obtain the concatenated feature vector.

[0201] The concatenated feature vectors are mapped to obtain the first channel-level gate vector corresponding to the preliminary fused feature vector and the second channel-level gate vector corresponding to the text feature vector.

[0202] The first channel-level gate vector and the second channel-level gate vector are normalized to obtain the first weight vector corresponding to the preliminary fused feature vector and the second weight vector corresponding to the text feature vector.

[0203] The preliminary fused feature vector and the text feature vector are weighted and fused based on the first weight vector and the second weight vector to obtain the fused feature vector.

[0204] In some embodiments, the diagnostic module 1305 is further configured to:

[0205] Obtain expert fault diagnosis results from human experts based on the aforementioned operating parameter information;

[0206] The fault diagnosis results and the expert fault diagnosis results are input into the integrated diagnosis model to obtain the final fault diagnosis results output by the integrated diagnosis model; the integrated diagnosis model has learned the ability to perform integrated diagnosis based on the expert fault diagnosis results and the fault diagnosis results output by the fault diagnosis model.

[0207] As an example, the integrated diagnostic model is constructed based on a feature pyramid network.

[0208] It should be noted that the explanations and descriptions in the above embodiments of the fault diagnosis method for reactor systems can also be applied to the fault diagnosis device for reactor systems in the embodiments of this application, and will not be repeated here.

[0209] Figure 14 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 14 As shown, the electronic device may include a processor 1410, a communication interface 1420, a memory 1430, and a communication bus 1440, wherein the processor 1410, the communication interface 1420, and the memory 1430 communicate with each other via the communication bus 1440. The processor 1410 can call the computer program in the memory 1430 to execute the steps of the reactor system fault diagnosis method provided in the above embodiment.

[0210] Furthermore, the logical instructions in the aforementioned memory 1430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0211] On the other hand, embodiments of the present invention also provide a computer program product, the computer program product including a computer program, the computer program being able to be stored on a computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to perform the steps of the fault diagnosis method for the reactor system provided in the above embodiments.

[0212] On the other hand, embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being used to cause a processor to execute the fault diagnosis method for the reactor system provided in the above embodiments.

[0213] The non-transitory computer-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0214] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0215] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0216] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for a reactor system, characterized in that, include: Obtain the operating parameter information of the reactor system within the current time window, and determine the operating parameter time series matrix based on the operating parameter information; Based on principal component analysis, principal component extraction and data reconstruction are performed on the time series matrix of the operating parameters to obtain the reconstructed time series matrix. Based on the operating parameter time series matrix and the reconstructed time series matrix, the operating parameter residual map is determined; Obtain expert fault analysis text information based on the aforementioned operating parameter information from human experts; The residual diagram of the operating parameters and the expert fault analysis text information are input into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the residual diagram of the operating parameters of the reactor system and the expert fault analysis text information. The fault diagnosis model comprises an encoding layer, a feature extraction layer, a feature fusion layer, and a fault diagnosis layer connected in sequence; the step of inputting the operating parameter residual map and the expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model includes: The residual map of the operating parameters and the text information of the expert fault analysis are input into the encoding layer. Based on the encoding layer, the residual map of the operating parameters and the text information of the expert fault analysis are encoded to obtain image encoding vector and text encoding vector respectively. Based on the feature extraction layer, feature extraction is performed on the image encoding vector and the text encoding vector respectively to obtain image feature vector and text feature vector; Based on the feature fusion layer, the image feature vector and the text feature vector are fused to obtain a fused feature vector; Based on the fault diagnosis layer, fault diagnosis is performed on the fused feature vector to obtain the fault diagnosis result; The process of fusing the image feature vector and the text feature vector to obtain a fused feature vector includes: The image feature vector is enhanced based on the text feature vector to obtain an enhanced image feature vector. Based on the cross-attention mechanism, the enhanced image feature vector and the text feature vector are fused to obtain a preliminary fused feature vector. Based on a pre-gating mechanism, the preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are fused to obtain the fused feature vector. The method based on a pre-gating mechanism fuses the initial fused feature vector, the text feature vector, and the enhanced image feature vector to obtain the fused feature vector, including: The preliminary fused feature vector, the text feature vector, and the enhanced image feature vector are concatenated to obtain the concatenated feature vector. The concatenated feature vectors are mapped to obtain the first channel-level gate vector corresponding to the preliminary fused feature vector and the second channel-level gate vector corresponding to the text feature vector. The first channel-level gate vector and the second channel-level gate vector are normalized to obtain the first weight vector corresponding to the preliminary fused feature vector and the second weight vector corresponding to the text feature vector. The preliminary fused feature vector and the text feature vector are weighted and fused based on the first weight vector and the second weight vector to obtain the fused feature vector.

2. The method according to claim 1, characterized in that, The step of enhancing the image feature vector based on the text feature vector to obtain an enhanced image feature vector includes: Based on the text feature vector, a scaling factor and an offset factor with the same dimension as the image feature vector are generated; Based on the scaling factor and the offset factor, a channel-by-channel affine transformation is performed on the image feature vector to obtain the enhanced image feature vector.

3. The method according to claim 1, characterized in that, The method based on cross-attention mechanism, which fuses the enhanced image feature vector and the text feature vector to obtain a preliminary fused feature vector, includes: A linear transformation is performed on the enhanced image feature vector to obtain the key matrix and the value matrix; A linear transformation is performed on the text feature vector to obtain the query matrix; Based on the key matrix and the query matrix, determine the attention score matrix; The attention score matrix is ​​normalized to obtain the attention weight matrix; The value matrix is ​​weighted based on the attention weight matrix to obtain the preliminary fusion feature vector.

4. The method according to any one of claims 1-3, characterized in that, Also includes: Obtain expert fault diagnosis results from human experts based on the aforementioned operating parameter information; The fault diagnosis results and the expert fault diagnosis results are input into the integrated diagnosis model to obtain the final fault diagnosis results output by the integrated diagnosis model; the integrated diagnosis model has learned the ability to perform integrated diagnosis based on the expert fault diagnosis results and the fault diagnosis results output by the fault diagnosis model.

5. The method according to claim 4, characterized in that, The integrated diagnostic model is constructed based on a feature pyramid network.

6. A fault diagnosis device for a reactor system, characterized in that, The fault diagnosis device, applicable to the method according to any one of claims 1-5, comprises: The first acquisition module is used to acquire the operating parameter information of the reactor system within the current time window, and determine the operating parameter time series matrix based on the operating parameter information; The second acquisition module is used to perform principal component extraction and data reconstruction on the time series matrix of the operating parameters based on principal component analysis to obtain the reconstructed time series matrix. The determination module is used to determine the operating parameter residual map based on the operating parameter time series matrix and the reconstructed time series matrix; The third acquisition module is used to acquire expert fault analysis text information based on the operating parameter information by human experts. The diagnostic module is used to input the operating parameter residual diagram and the expert fault analysis text information into the fault diagnosis model to obtain the fault diagnosis result output by the fault diagnosis model; wherein, the fault diagnosis model has learned the ability to perform fault diagnosis on the reactor system based on the operating parameter residual diagram and expert fault analysis text information of the reactor system.

7. An electronic device, characterized in that, It includes a processor and a memory storing a computer program, wherein the processor executes the program to implement the fault diagnosis method for the reactor system according to any one of claims 1-5.

Images