A method and system for intelligent diagnosis of krypton-xenon extraction process failure, and a storage medium

Abstract

The application provides a kind of krypton-xenon extraction process fault intelligent diagnosis method, system and storage medium, specifically, the multivariate time series data of krypton-xenon extraction process is acquired, and high-dimensional deep feature sequence is extracted by inputting multi-scale residual dense hollow convolution network;The network uses different dilated convolution with different expansion rate in different residual blocks, and fuses multi-scale time series features by dense connection;The feature sequence is weighted using channel attention and time self-attention mechanism, to obtain time-weighted features;Meanwhile, based on the covariance modeling between feature channels, the feature dimension attention score is calculated, and the variable correlation weighted feature is obtained;The time-weighted feature and the variable correlation weighted feature are first outer product fused, and the original deep feature and the time-weighted feature are second outer product fused, and the fusion results of twice are spliced to generate fault representation vector;The fault representation vector is input into the classifier, and the fault type diagnosis result is output.

Description

Technical Field

[0001] This application belongs to the field of intelligent diagnostics, and in particular relates to an intelligent diagnostic method, system and storage medium for faults in the krypton-xenon extraction process. Background Technology

[0002] Krypton-xenon extraction employs cryogenic distillation, involving multiple distillation columns, heat exchangers, and compressors—key pieces of equipment with numerous and interdependent operating parameters. In actual production, krypton-xenon extraction is prone to various malfunctions such as leakage, blockage, abnormal temperature, and pressure runaway due to fluctuations in raw material composition, equipment aging, external environmental interference, or human error. Existing fault diagnosis methods are based on mechanistic models and expert systems; however, the former faces challenges in establishing mathematical models, while the latter suffers from difficulties in building and updating its knowledge base. Convolutional neural networks are used to extract spatial features from data, or recurrent neural networks and their variants are used to detect time-series dependencies. Process data simultaneously contains short-term fluctuations caused by periodic equipment operation and long-term trends caused by slow process drift. Single-scale feature extraction networks struggle to detect these multi-scale time-series dependencies, easily missing fault symptom information. Attention mechanisms mostly focus on weighting the explicit importance of time steps or feature channels, neglecting the potential correlations between different process variables under fault conditions. The occurrence of a fault is not due to an anomaly in a single variable, but rather the result of multiple variables deviating from normal operating conditions in a coordinated manner. Failure to model the correlation patterns among these variables limits the diagnostic model's ability to identify faults. When fusing different feature views, concatenation or linear combination methods cannot fully exploit the higher-order interactions between various feature information, resulting in insufficient discriminative power in the generated fault representation vector, thus affecting the accuracy of fault classification. Summary of the Invention

[0003] This invention proposes an intelligent fault diagnosis method for the krypton-xenon extraction process, addressing the problem that existing technologies neglect the correlation between potential changes in different process variables under fault conditions and fail to fully exploit high-order interactions between various feature information. The method includes: The process involves acquiring time-series data composed of multiple process variables in the krypton-xenon extraction process; inputting the time-series data into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies; the network sets dilated convolutions with different dilation rates in different residual blocks and densely connects the output features of convolutions at each scale. By combining channel attention and temporal self-attention mechanisms, the high-dimensional deep feature sequence is weighted to detect the importance of different feature channels and the dependency relationship at different time steps, thus obtaining temporally weighted features. Based on the covariance between the feature channels of the high-dimensional deep feature sequence, a feature association model is established, and the attention score on the feature dimension is calculated. The high-dimensional deep feature sequence is then weighted to obtain variable association weighted features. The time-weighted features and the variable-associated weighted features are fused by a first outer product, and the high-dimensional deep feature sequence is fused by a second outer product with the time-weighted features. The results of the two fusions are then concatenated to generate a fault representation vector. The fault representation vector is then input into a preset classifier to output the fault type diagnosis result of the krypton-xenon extraction process.

[0004] Furthermore, the present invention also relates to an intelligent fault diagnosis system for the krypton-xenon extraction process, comprising the following modules: The connection module is used to acquire time series data composed of multiple process variables in the krypton-xenon extraction process; input the time series data into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies; the network sets dilated convolutions with different dilation rates in different residual blocks and densely connects the output features of convolutions at each scale. The detection module is used to combine channel attention and temporal self-attention mechanisms to weight the high-dimensional deep feature sequence, detect the importance of different feature channels and the dependency relationship at different time steps, and obtain time-weighted features. The weighting module is used to establish a feature association model based on the covariance between feature channels of the high-dimensional deep feature sequence, calculate the attention score on the feature dimension, and weight the high-dimensional deep feature sequence to obtain variable association weighted features. The output module is used to perform a first outer product fusion of the time-weighted features and the variable-associated weighted features, and to perform a second outer product fusion of the high-dimensional deep feature sequence and the time-weighted features. The results of the two fusions are then concatenated to generate a fault representation vector. The fault representation vector is then input into a preset classifier to output the fault type diagnosis result of the krypton-xenon extraction process.

[0005] This invention constructs a multi-scale residual dense dilated convolutional network to simultaneously detect temporal dependencies at different time scales during krypton-xenon extraction, extracting high-dimensional deep features from the raw data. Based on this, a channel and temporal self-attention mechanism is used to highlight information from key feature channels and important time points, establishing a correlation model based on feature covariance. This allows for the discovery of implicit correlation patterns between process variables under fault conditions. A strategy of double outer product fusion and concatenation fully explores higher-order interactions between different feature views, generating information-rich and highly discriminative fault representation vectors, thus improving the accuracy of fault diagnosis during krypton-xenon extraction. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2A schematic diagram of a multi-scale residual dense dilated convolutional network module; Figure 3 This is a schematic diagram of the channel attention mechanism structure; Figure 4 This is a schematic diagram of the temporal self-attention mechanism. Figure 5 This is the feature fusion and classification module. Detailed Implementation

[0007] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0008] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, the present invention proposes an intelligent fault diagnosis method for the krypton-xenon extraction process, such as... Figure 1 ,include: S1, Obtain time series data consisting of multiple process variables in the krypton-xenon extraction process; Input the time series data into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies; The network sets dilated convolutions with different dilation rates in different residual blocks and densely connects the output features of convolutions at each scale. Temperature, pressure, flow rate, and liquid level sensors deployed on key equipment such as the krypton-xenon extraction distillation column, compressor, and heat exchanger are used to collect process data from each measuring point in real time via a distributed control system (DCS). This data forms a data matrix with time as the sequence and multiple process variables as feature dimensions. A sliding window method is used to segment the long-term time-series data into fixed-length time-series samples. The window length L needs to be selected to cover the typical dynamic response time of the process. For example, if the sampling interval is 10 seconds, L=60 can be set, meaning each sample represents 10 minutes of process data. Z-score standardization is then applied to each sample data to eliminate dimensional differences between different variables.

[0010] The standardized data samples, with a shape of batch size × L × M, where M is the total number of process variables, are subjected to dimensionality enhancement through an initial one-dimensional convolutional layer. This layer uses 32 convolutional kernels of size 3, with a stride of 1 and padding of 1, resulting in an output shape of batch size × L × 32. The data flows through three cascaded multi-scale residual dense modules. Within each module, four parallel processing branches are set up, each containing a one-dimensional dilated convolutional layer with dilation rates of 1, 2, 4, and 8, respectively, to detect features at different time scales. Each branch's dilated convolutional layer uses 16 convolutional kernels of size 3, with corresponding padding based on its dilation rate to ensure that the length of the output feature sequence remains constant across all branches. The output feature maps of the four branches are concatenated along the channel dimension to form a feature volume that integrates multi-scale information, with 64 channels. The feature volume is input into a 1×1 convolutional layer for information integration and channel number adjustment, adjusting it to be the same as the module input channel number. The output of this layer is then element-wise added to the module input to form residual connections. The sequence obtained after all modules have processed is the high-dimensional deep feature sequence, such as... Figure 2 The high-dimensional deep feature sequence is a multi-channel feature tensor arranged in chronological order, where each time step corresponds to a feature vector composed of multi-scale convolutional features.

[0011] In an optional embodiment, the process variables include key operating parameters of the distillation column and product quality indicators.

[0012] Specifically, taking the krypton-xenon extraction and distillation system in the chemical process as the monitoring object, multidimensional process variables related to the operating status were selected as model inputs. Key operating parameters, such as krypton tower pressure, xenon tower temperature, feed air flow rate, lean krypton-xenon liquid level, and product extraction valve opening, reflect the operator's control actions on the distillation tower and changes in the operating conditions within the tower. Product quality indicators, such as the krypton concentration in the crude krypton fraction, the xenon concentration in the crude xenon fraction, the purity of the product krypton gas, and the purity of the product xenon gas, are core indicators for measuring the separation effect and product qualification rate of the krypton-xenon extraction process.

[0013] The process variables are constructed into a multidimensional time series matrix using data from L consecutive sampling points over a period of time. For example, if 15 variables are selected, the input data is an L-row, 15-column matrix, where each row represents a time step and each column represents a process variable. This matrix serves as the initial input to the fault diagnosis model. The model learns the data characteristics under normal operating conditions and various fault conditions by analyzing the patterns of change of the key variables over time and the relationships between the variables, thereby achieving fault identification.

[0014] This fault diagnosis neural network model is an end-to-end deep learning architecture. The input is multi-dimensional time-series data samples with fault labels, each sample being a T×M matrix, for example, 100×15. Structurally, the initial part is the aforementioned multi-scale residual dense dilated convolutional network, used to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies from the original input. The feature sequences are sequentially passed through the aforementioned channel attention mechanism module, temporal self-attention mechanism module, and feature association model module for multi-dimensional feature enhancement and information fusion, forming a comprehensive fault representation vector. The model's final end is connected to the aforementioned Softmax classifier, outputting an N-dimensional probability vector, where N is the total number of predefined fault categories, including normal operating conditions. Each element in the vector represents the predicted probability that the input sample belongs to the corresponding category. The training set consists of historical operating data, covering normal operating conditions and all known N-1 fault conditions, and is divided into training, validation, and test sets. The training process aims to minimize the difference between the predicted results and the true labels.

[0015] S2, Combining channel attention and temporal self-attention mechanisms, the high-dimensional deep feature sequence is weighted to detect the importance of different feature channels and the dependency relationship at different time steps, thus obtaining time-weighted features; Channel attention is applied to a high-dimensional deep feature sequence. A 1×C channel descriptor is obtained by global average pooling of the feature sequence. This descriptor is then input into a bottleneck structure consisting of two fully connected layers for transformation. The first fully connected layer reduces the dimension to C / r, where the reduction rate r is set to 16, and a ReLU activation function is used. The second fully connected layer restores the dimension to C and applies a Sigmoid activation function to generate weight coefficients for each feature channel. The original feature sequence is multiplied by these weights to obtain the channel-enhanced features. These channel-enhanced features are then input into a temporal self-attention module. A query matrix Q, a key matrix K, and a value matrix V are generated through three linear transformations. The dot product of Q and the transpose of K is calculated and scaled and normalized using Softmax to obtain the attention weights between time steps. These weights are multiplied by the V matrix to obtain the temporally weighted features.

[0016] In an optional embodiment, the step of combining channel attention and temporal self-attention mechanisms to weight the high-dimensional deep feature sequence and detect the importance of different feature channels and the dependency relationship at different time steps includes: The channel attention mechanism generates weights for each feature channel by globally compressing and transforming the high-dimensional deep feature sequence, and then weights the original high-dimensional deep feature sequence. The temporal self-attention mechanism detects dependencies between time steps by linearly mapping feature sequences to queries, keys, and values, calculating the similarity between queries and keys to obtain attention scores, and weighting values ​​based on the scores.

[0017] The channel attention mechanism's model structure borrows from the classic Squeeze-and-Excitation network module. For an input high-dimensional deep feature sequence, where the sequence dimension is the time step T multiplied by the number of feature channels C, a global average pooling layer is used to average all feature values ​​of each channel in the time dimension, thereby compressing the T-dimensional information into a single value. The T×C feature map is then transformed into a 1×C channel descriptor vector, which contains the global receptive field information of each feature channel.

[0018] The channel descriptor vector is fed into a bottleneck structure consisting of two fully connected layers for transformation. The first fully connected layer reduces the C-dimensional vector to C / r dimensions, where r is the reduction rate, and uses the ReLU activation function to add non-linearity. The second fully connected layer restores the vector to C dimensions and applies the Sigmoid activation function. This process enables the model to learn the dependencies between channels and outputs a 1×C weight vector, where each value is between 0 and 1, representing the importance of the corresponding feature channel. The weight vector is then multiplied at the channel level with the original high-dimensional deep feature sequence to recalibrate the features of each channel, thereby enhancing the expression of key feature channels and suppressing secondary feature channels, such as... Figure 3 .

[0019] The self-attention mechanism is implemented based on the self-attention module of the standard Transformer model. For a feature sequence with an input dimension of T×C, it is processed through three linear transformation layers, i.e., three weight matrices. , , The weight matrices are then mapped to generate a query matrix Q, a key matrix K, and a value matrix V, respectively. All three matrices have a dimension of T×d, where d is the feature dimension after mapping.

[0020] By calculating the dot product of the query matrix Q and the transpose of the key matrix K, a T×T attention score matrix is ​​obtained. Each element in this matrix represents the degree of attention from one time step to another in the sequence. For training stability, each element of this score matrix is ​​divided by a scaling factor, typically the square root of the mapping dimension d. The Softmax function is applied to each row of the score matrix to normalize the scores into a probability distribution, yielding the attention weight matrix. Multiplying the T×T attention weight matrix by the value matrix V yields the output sequence, as shown below. Figure 4 The feature vector at each time step of the output sequence is a weighted sum of the value vectors V of all time steps in the original sequence. The weights are the calculated attention scores, which enables the model to detect long-distance dependencies between any two time steps in the sequence.

[0021] S3. Based on the covariance between the feature channels of the high-dimensional deep feature sequence, establish a feature association model, calculate the attention score on the feature dimension, and weight the high-dimensional deep feature sequence to obtain variable association weighted features. The high-dimensional deep feature sequence extracted from the convolutional network is transposed, with the feature channel dimension as rows and the time step dimension as columns. Based on this transposed matrix, a covariance matrix is ​​calculated. Each element of this covariance matrix represents the linear correlation strength between two corresponding feature channels; this is the feature association model. By normalizing the covariance matrix between feature channels, the correlation strength between channels is transformed into attention weights for the corresponding feature dimensions, reflecting the degree of association between variables. To obtain the overall importance of each feature channel, each row of the covariance matrix is ​​summed, resulting in a C-dimensional vector. Each element represents the sum of the covariances of that channel with all other channels, reflecting the association strength of that channel in the entire feature system. The Softmax function is applied to this C-dimensional vector, transforming it into a probability distribution, i.e., an attention score vector. This attention score vector is then multiplied element-wise with the original high-dimensional deep feature sequence, thereby scaling the features of each channel according to its overall correlation with other channels, generating variable association weighted features.

[0022] In an optional embodiment, the step of establishing a feature association model based on the covariance between feature channels of the high-dimensional deep feature sequence, calculating attention scores on the feature dimensions, and weighting the high-dimensional deep feature sequence to obtain variable association weighted features includes: Based on the high-dimensional deep feature sequence Calculate the covariance matrix between feature channels Where T is the time step and C is the number of feature channels; By applying the Softmax function to the row sum of the covariance matrix, attention score vectors are generated along the feature dimensions. The calculation formula is: in It is a vector consisting entirely of 1s; Using the attention score vector Channel-weighted features H of high-dimensional deep features are obtained by performing channel-weighted features related to variables. , where ⊙ denotes element-wise multiplication. express The transpose of .

[0023] Specifically, given a high-dimensional feature sequence H of dimension T×C extracted by a deep network, the sequence is centered by subtracting the mean of each channel. The covariance matrix is ​​then calculated. The matrix is ​​a C×C square matrix, where the element in the i-th row and j-th column represents the covariance between the i-th and j-th feature channels over all T time steps. The matrix represents the pairwise linear correlation between all feature channels.

[0024] To obtain the overall importance of each feature channel, the covariance matrix is... Multiply by a C×1 vector of all ones. Summate each row of the covariance matrix, resulting in a C-dimensional vector where each element represents the sum of the covariances of that channel with all other channels, reflecting the correlation strength of that channel in the entire feature system. Apply the Softmax function to this C-dimensional vector, transforming it into a probability distribution, i.e., the attention score vector. The Softmax function amplifies the weights of channels with strong correlations while compressing the weights of channels with weak correlations. The resulting attention score vector is then used. The original high-dimensional deep feature sequence H is weighted. The eigenvalues ​​of all T time steps of the i-th channel of H are multiplied by a fraction. In this way, feature channels that are highly correlated with most other variables are given high weights, thereby strengthening the expression of key variable features.

[0025] As another optional embodiment of the present invention, the feature association model can be implemented using the following matrix attention method: Based on the high-dimensional deep feature sequence H∈ Calculate the covariance matrix Σ∈ between feature channels Apply the Softmax function to each row of the covariance matrix Σ to generate an attention score matrix A ∈ [feature dimension]. The attention score matrix A is multiplied by the original high-dimensional deep feature sequence H to obtain the variable-association weighted features. This approach enables more precise modeling of the interactions between different feature channels, allowing each new channel's features to incorporate information from other related channels, thus achieving more complex feature interaction modeling.

[0026] S4, perform a first outer product fusion of the time-weighted features and the variable-associated weighted features, and perform a second outer product fusion of the high-dimensional deep feature sequence and the time-weighted features. Then, concatenate the results of the two fusions to generate a fault representation vector. Input the fault representation vector into a preset classifier and output the fault type diagnosis result of the krypton-xenon extraction process.

[0027] For the first outer product fusion, the feature vectors of the time-weighted feature sequence and the variable-association-weighted feature sequence at each time step are multiplied by an outer product to generate a series of interaction matrices. Then, the interaction matrices of all time steps are summed or averaged to obtain a fused two-dimensional feature map, which is then flattened into a one-dimensional vector. For the second outer product fusion, the same method is used: the original high-dimensional deep feature sequence and the time-weighted feature sequence are multiplied by an outer product, pooled, and flattened at each time step to obtain another one-dimensional vector. The two one-dimensional vectors generated by outer product fusion are concatenated along the feature dimension to obtain a fault representation vector. The time-weighted features are used to characterize the significance of different time steps and feature channels in the fault change process, while the variable-association-weighted features are used to represent the correlation and dependence between process variables. Outer product fusion can explicitly model the high-order interaction relationship between the time dimension and the variable dimension, where the introduction of time-weighted features and the original high-dimensional features enhances the complementary expressive power of information at different levels. Concatenating the results of different outer product fusions can maintain the complementarity of various features while avoiding the limitations of a single fusion method on the expression of fault features.

[0028] The classifier is preferably a multilayer perceptron, which contains one or more hidden layers, each consisting of several neurons and a ReLU activation function. The input layer of the classifier receives the fault representation vector, and the output layer is a fully connected layer with the number of neurons equal to the total number of predefined fault categories plus one normal state category. The output layer uses a Softmax activation function to transform the network's raw output into a probability value corresponding to each category. The category with the highest probability value is output as the diagnostic result, such as normal, blockage fault, or leakage fault. Figure 5 .

[0029] In an optional embodiment, the preset classifier is a Softmax classifier, used to map the fault representation vector to N predefined fault categories.

[0030] At the end of the fault diagnosis model, a Softmax classifier is deployed to perform the classification task. The core of this classifier is a fully connected layer, also known as a linear layer, followed by a Softmax activation function. A fault representation vector, with dimensions D, is received from the upstream part of the model after sufficient feature extraction and fusion. This D-dimensional vector is then input into the fully connected layer. The weight matrix of this fully connected layer has dimensions D×N, and the bias vector has dimensions N, where N is the predefined total number of fault categories, for example, N=5, representing normal operating conditions, fault 1, fault 2, fault 3, and fault 4, respectively.

[0031] The fully connected layer performs a linear transformation on the input D-dimensional vector, outputting an N-dimensional original score vector, also known as logits. Each element in this vector corresponds to an unnormalized score for a fault category. The Softmax activation function is applied to this N-dimensional score vector. The Softmax function transforms the original scores into a probability distribution, calculated using the following formula: ,in For the original score of class j, The predicted probability output for class j is an N-dimensional probability vector, where each element has a value between 0 and 1, and the sum of all elements is 1. Each element represents the probability that the input sample belongs to the corresponding fault category. The model outputs the category with the highest probability value as the diagnostic result.

[0032] In the second embodiment, the present invention also proposes an intelligent fault diagnosis system for the krypton-xenon extraction process, comprising the following modules: The connection module is used to acquire time series data composed of multiple process variables in the krypton-xenon extraction process; input the time series data into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies; the network sets dilated convolutions with different dilation rates in different residual blocks and densely connects the output features of convolutions at each scale. The detection module is used to combine channel attention and temporal self-attention mechanisms to weight the high-dimensional deep feature sequence, detect the importance of different feature channels and the dependency relationship at different time steps, and obtain time-weighted features. The weighting module is used to establish a feature association model based on the covariance between feature channels of the high-dimensional deep feature sequence, calculate the attention score on the feature dimension, and weight the high-dimensional deep feature sequence to obtain variable association weighted features. The output module is used to perform a first outer product fusion of the time-weighted features and the variable-associated weighted features, and to perform a second outer product fusion of the high-dimensional deep feature sequence and the time-weighted features. The results of the two fusions are then concatenated to generate a fault representation vector. The fault representation vector is then input into a preset classifier to output the fault type diagnosis result of the krypton-xenon extraction process.

[0033] In an optional embodiment, the process variables include key operating parameters of the distillation column and product quality indicators.

[0034] In an optional embodiment, the step of combining channel attention and temporal self-attention mechanisms to weight the high-dimensional deep feature sequence and detect the importance of different feature channels and the dependency relationship at different time steps includes: The channel attention mechanism generates weights for each feature channel by globally compressing and transforming the high-dimensional deep feature sequence, and then weights the original high-dimensional deep feature sequence. The temporal self-attention mechanism detects dependencies between time steps by linearly mapping feature sequences to queries, keys, and values, calculating the similarity between queries and keys to obtain attention scores, and weighting values ​​based on the scores.

[0035] In an optional embodiment, the step of establishing a feature association model based on the covariance between feature channels of the high-dimensional deep feature sequence, calculating attention scores on the feature dimensions, and weighting the high-dimensional deep feature sequence to obtain variable association weighted features includes: Based on the high-dimensional deep feature sequence Calculate the covariance matrix between feature channels Where T is the time step and C is the number of feature channels; By applying the Softmax function to the row sum of the covariance matrix, attention score vectors are generated along the feature dimensions. The calculation formula is: in It is a vector consisting entirely of 1s; Using the attention score vector Channel-weighted features H of high-dimensional deep features are obtained by performing channel-weighted features related to variables. , where ⊙ denotes element-wise multiplication. express The transpose of .

[0036] In an optional embodiment, the preset classifier is a Softmax classifier, used to map the fault representation vector to N predefined fault categories.

[0037] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0038] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0039] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0040] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0041] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for intelligent fault diagnosis in the krypton-xenon extraction process, characterized in that, Includes the following steps: A time-series data consisting of multiple process variables in the krypton-xenon extraction process is obtained. This time-series data is generated by real-time data collection from temperature, pressure, flow rate, and liquid level sensors deployed on key equipment such as the krypton-xenon extraction distillation column, compressor, and heat exchanger via a distributed control system (DCS). The data is then used to form a data matrix with time as the sequence and multiple process variables as feature dimensions. This time-series data is input into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies. The network uses dilated convolutions with different expansion rates within different residual blocks and densely connects the output features of convolutions at each scale. By combining channel attention and temporal self-attention mechanisms, the high-dimensional deep feature sequence is weighted to detect the importance of different feature channels and the dependencies between different time steps, resulting in time-weighted features. Specifically, the weighting of the high-dimensional deep feature sequence by combining channel attention and temporal self-attention mechanisms to detect the importance of different feature channels and the dependencies between different time steps includes: the channel attention mechanism performing global information compression and transformation on the high-dimensional deep feature sequence to generate weights for each feature channel, and then weighting the original high-dimensional deep feature sequence; the temporal self-attention mechanism linearly mapping the weighted feature sequence to queries, keys, and values, calculating the similarity between queries and keys to obtain an attention score, and then weighting the values ​​based on the score to detect dependencies between time steps. Based on the covariance between the feature channels of the high-dimensional deep feature sequence, a feature association model is established, and the attention score on the feature dimension is calculated. The high-dimensional deep feature sequence is then weighted to obtain variable association weighted features. The time-weighted features and the variable-associated weighted features are fused by a first outer product, and the high-dimensional deep feature sequence is fused by a second outer product with the time-weighted features. The results of the two fusions are then concatenated to generate a fault representation vector. The fault representation vector is then input into a preset classifier to output the fault type diagnosis result of the krypton-xenon extraction process.

2. The method according to claim 1, characterized in that, The process variables include key operating parameters of the distillation column and product quality indicators.

3. The method according to claim 1, characterized in that, The feature association model is established based on the covariance between feature channels of the high-dimensional deep feature sequence, and attention scores are calculated on the feature dimensions. The high-dimensional deep feature sequence is then weighted to obtain variable association weighted features, including: Based on the high-dimensional deep feature sequence Calculate the covariance matrix between feature channels Where T is the time step and C is the number of feature channels; By applying the Softmax function to the row sum of the covariance matrix, attention score vectors are generated along the feature dimensions. The calculation formula is: ； in It is a vector consisting entirely of 1s; Using the attention score vector Channel-weighted features H of high-dimensional deep features are obtained by performing channel-weighted features related to variables. , where ⊙ denotes element-wise multiplication. express The transpose of .

4. The method according to any one of claims 1-3, characterized in that, The preset classifier is a Softmax classifier, which is used to map the fault representation vector to N predefined fault categories.

5. A fault diagnosis system for the krypton-xenon extraction process, characterized in that, Includes the following modules: A connection module is used to acquire time-series data composed of multiple process variables in the krypton-xenon extraction process. This time-series data is generated by real-time acquisition of process data from temperature, pressure, flow rate, and liquid level sensors deployed on key equipment such as the krypton-xenon extraction distillation column, compressor, and heat exchanger via a distributed control system (DCS). The data is then used to form a data matrix with time as the sequence and multiple process variables as feature dimensions. The time-series data is input into a multi-scale residual dense dilated convolutional network to extract high-dimensional deep feature sequences containing multi-scale temporal dependencies. The network uses dilated convolutions with different expansion rates within different residual blocks and densely connects the output features of convolutions at each scale. The detection module is used to combine channel attention and temporal self-attention mechanisms to weight the high-dimensional deep feature sequence, detect the importance of different feature channels and the dependency relationship at different time steps, and obtain time-weighted features. The weighting module is used to establish a feature association model based on the covariance between feature channels of the high-dimensional deep feature sequence, calculate attention scores in the feature dimension, and weight the high-dimensional deep feature sequence to obtain variable association weighted features. Specifically, the weighting of the high-dimensional deep feature sequence by combining channel attention and temporal self-attention mechanisms to detect the importance of different feature channels and the dependencies between different time steps includes: the channel attention mechanism generating weights for each feature channel by performing global information compression and transformation on the high-dimensional deep feature sequence, and then weighting the original high-dimensional deep feature sequence; the temporal self-attention mechanism linearly mapping the weighted feature sequence to queries, keys, and values, calculating the similarity between queries and keys to obtain attention scores, and weighting the values ​​according to the scores to detect dependencies between time steps. The output module is used to perform a first outer product fusion of the time-weighted features and the variable-associated weighted features, and to perform a second outer product fusion of the high-dimensional deep feature sequence and the time-weighted features. The results of the two fusions are then concatenated to generate a fault representation vector. The fault representation vector is then input into a preset classifier to output the fault type diagnosis result of the krypton-xenon extraction process.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-4.

Images