Industrial process fault diagnosis method based on dynamic optimal transmission graph neural network
By combining dynamic optimal transport graph neural networks and Kolmogorov-Arnold networks, the problem of poor class discrimination in fault diagnosis in industrial processes is solved, and more efficient fault identification and diagnosis are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUZHOU UNIVERSITY
- Filing Date
- 2026-05-12
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies struggle to accurately identify fault types in high-dimensional, strongly coupled, and nonlinear industrial process data. Traditional graph neural networks also struggle to explicitly constrain the organization of fault categories in fault diagnosis, resulting in poor inter-class distinguishability.
A dynamic optimal transport graph neural network is adopted to adaptively capture changes in material and energy transport methods by constructing a dynamic optimal transport graph. The Kolmogorov-Arnold network is used to map process variable features, and a feature transport loss function is added to explicitly constrain the fault category embedding space, thereby enhancing inter-class separability.
It effectively captures fault characteristics in industrial processes, with a compact fault category embedding space. Samples of the same type are more compact, and samples of different types are easier to separate, thus improving the accuracy and stability of fault diagnosis.
Smart Images
Figure CN122174055B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial process control technology and relates to an industrial process fault diagnosis method based on a modified dynamic optimal transmission graph neural network. Background Technology
[0002] Industrial process systems are typically driven by a large number of interdependent process variables, and their operating state is dynamically affected by material flow, energy exchange, and multi-level control units. In such a complex environment, once a fault occurs, it often propagates within the system in a chain reaction, causing abnormal responses in multiple variables, and in severe cases, even leading to process shutdowns or equipment damage. Therefore, accurately identifying fault types from high-dimensional, strongly coupled, and nonlinear process data and assisting operators in timely diagnosis and decision-making has always been one of the core challenges in the field of industrial intelligent monitoring.
[0003] In industrial process systems, different variables (such as temperature, pressure, and flow rate), despite having different physical dimensions, form complex interdependencies when they interact within the same production process. Most existing graph neural networks, when constructing dynamic graph structures, typically use correlations directly as edge weights, which only reflects the statistical covariance between variables. However, the internal transport paths of materials or energy within industrial systems are not always explicitly labeled in the observed data. In actual industrial plants, the transfer of materials and energy typically relies on process units such as pipes, valves, and heat exchangers. Pipelines determine the transport paths, while valves influence the degree of material distribution and transport resistance by adjusting their opening. In other words, an industrial system is essentially a dynamic transport network determined by both "path" and "resistance." A well-functioning, stable, and efficient process system often possesses a near-optimal material transport method, ensuring rational distribution of materials among process units, efficient energy utilization, and controlled disturbance propagation. Conversely, when a system malfunctions, the flow of materials or energy changes, redistributing along new paths within the faulty system. This distorts the transmission pattern, amplifying and spreading disturbances within the system, thus forming specific and identifiable fault modes. In such dynamic transmission networks, different fault modes often correspond to drastically different propagation characteristics and dynamic responses within the system. Traditional cross-entropy loss only optimizes the class probability distribution during training, but it does not explicitly constrain the organization of fault classes in the feature space. Classification models that rely solely on probability optimization are insufficient to guarantee inter-class discriminability.
[0004] The purpose of this invention is to propose an industrial process fault diagnosis method based on a dynamic optimal transmission graph neural network to solve the problems existing in the prior art, which is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] Based on this, the present invention proposes an industrial process fault diagnosis method based on dynamic optimal transport graph neural network, which can adaptively capture changes in material and energy transport modes under different fault models; generate fault features that are easier to distinguish; and its fault category embedding space is more compact and different categories are easier to separate, thus solving the problem of difficult fault differentiation and diagnosis in industry.
[0006] Specifically, the present invention is as follows: First, collect industrial process variable data and construct a dynamic optimal transmission diagram to describe the transmission modes in different industrial process faults; Then, the dynamic optimal transmission diagram is aggregated into the industrial process variable features using the information propagation mechanism to obtain industrial process variable features that contain the transmission relationships between industrial process variables; Finally, the industrial process variable features are mapped onto a Kolmogorov-Arnold network, and the nested and summed forms of multiple one-dimensional spline functions are used to approximate any multivariable continuous function, outputting multiple prediction labels.
[0007] To achieve the above objectives, the present invention adopts the following technical solution, which specifically includes the following steps: Step 1: Collect industrial process variables, obtain fault data in the industrial process, divide the fault data into industrial process fault training set, industrial process fault test set, and industrial process fault verification set, and obtain the corresponding fault type training set, fault type test set, and fault type verification set.
[0008] Step 2: Normalize the fault data in the industrial process fault training set, validation set, and test set from Step 1.
[0009] As a preferred method, zero-mean normalization is used to normalize the fault data in the training, validation, and test sets of industrial process faults.
[0010] Step 3: Construct an industrial process fault diagnosis model based on a dynamic optimal transmission graph neural network; First, a dynamic optimal transmission graph is constructed to describe the transmission modes in different industrial process faults, as follows: By using a learnable weight matrix to perform a linear transformation on industrial process variables and map them to a high-dimensional space, the similarity distribution among industrial process variables can be obtained. After normalizing the similarity distribution among industrial process variables to the same metric space, the relationship between industrial process variables with higher similarity distribution is given a lower transmission cost through the inversion operation, and combined to form a cost matrix; By using a feedforward neural network to directly predict the corresponding dual variables of industrial process variables, and by using a broadcasting mechanism to extend the dual variables into the cost matrix, a dynamic optimal transmission graph is formed. Then, the dynamic optimal transmission diagram is aggregated into the industrial process variable features using the information propagation mechanism to obtain industrial process variable features that contain the transmission relationships between industrial process variables; Finally, the industrial process variable features are mapped to a Kolmogorov-Arnold network, which is more conducive to classification, and the nested and summed forms of multiple one-dimensional spline functions are used to approximate any multivariable continuous function, outputting multiple predicted labels.
[0011] Furthermore, a learnable weight matrix is used to perform a linear transformation on the industrial process variables and map them to a high-dimensional space. Dot product similarity is selected as the measure of similarity between industrial process variables, and the Softmax function is used to normalize the similarity between industrial process variables to obtain the similarity distribution between industrial process variables.
[0012] Furthermore, the similarity distribution among industrial process variables is normalized to the same metric space using a min-max method, and the relationship between industrial process variables with higher similarity distribution is given a lower transmission cost through a reversal operation, and combined to form a cost matrix.
[0013] Furthermore, to enhance the numerical stability of model training, layer normalization was added after the dual variables were predicted.
[0014] As a preferred approach, an industrial process fault diagnosis model based on a dynamic optimal transport graph neural network is built using the open-source deep learning platform PyTorch.
[0015] Step 4: Using the cross-entropy loss function and the feature transport loss function, input the industrial process fault training set and fault type training set processed in Step 2 into the industrial process fault diagnosis model constructed in Step 3 to train the model multiple times. Then, use the industrial process fault validation set and the fault type validation set to validate the trained industrial process fault diagnosis model. Save the parameters of the industrial process fault diagnosis model with the best performance in multiple training sessions.
[0016] Furthermore, the feature transport loss function specifically involves: introducing a learnable fault category embedding space; given the probability distribution of the predicted label features and the probability distribution of the true label in the fault category embedding space, constructing a feasible transport matrix from the probability distribution of the predicted label features to the probability distribution of the true label in the fault category embedding space, which satisfies the marginal constraints of the optimal transport (OT) problem; and calculating the tractable upper bound of the Wasserstein distance between the probability distribution of the predicted label features and the probability distribution of the true label in the fault category embedding space.
[0017] Step 5: Input any industrial process variable into the optimal performance industrial process fault diagnosis model saved in Step 4, and output the fault type prediction value.
[0018] As can be seen from the above technical solutions, compared with the prior art, the present invention provides an industrial process fault diagnosis method based on dynamic optimal transport graph neural network, which has the following beneficial effects: The present invention uses dynamic optimal transport graph to construct a method to describe the implicit transport paths of materials and capacity in the system under different fault states; it uses Kolmogorov-Arnold network to map process variable features to a potential representation space that is more conducive to classification; it adds feature transport loss to impose explicit constraints on the fault category embedding space, making similar samples more compact in the fault category embedding space and easier to separate different categories; compared with traditional GNN-type models, the present invention can effectively capture the correlation in industrial process data and better identify different fault states in industrial processes. Attached Figure Description
[0019] 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.
[0020] Figure 1 This is an overall flowchart of the present invention;
[0021] Figure 2 (a) shows the construction and training process of the industrial process fault diagnosis model based on the dynamic optimal transmission graph neural network; (b) shows the dynamic optimal transmission graph modeling process; (c) shows the Kolmogorov-Arnold network modeling process; and (d) shows the characteristic transport loss principle.
[0022] Figure 3 This is a flowchart of the Tennessee-Eastman process in the example;
[0023] Figure 4 This is the confusion matrix of the Tennessee-Eastman process fault classification results in the example. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.
[0025] In this application, 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. 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 limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0026] like Figure 1 As shown, an industrial process fault diagnosis method based on a dynamic optimal transport graph neural network specifically includes the following steps:
[0027] Step 1: Collect industrial process variables and obtain fault data from the industrial process. Divide the fault data into an industrial process fault training set. Industrial process failure verification set Industrial process failure test set Industrial process fault training set corresponding to fault type training set Industrial process fault verification set corresponding to fault type verification set Industrial process failure test set corresponding to failure type test set , where T represents the time step of the process variable and N represents the number of process variables;
[0028] Step 2: Process the industrial process fault training set, validation set and test set from Step 1. Use zero-mean normalization to map the industrial process fault training set to the [0,1] interval. At the same time, map the industrial process fault validation set and industrial process fault test set to the [0,1] interval according to the data distribution in the industrial process fault training set.
[0029] Step 3: Build an industrial process fault diagnosis model based on a dynamic optimal transport graph neural network using the open-source deep learning platform PyTorch.
[0030] First, a dynamic optimal transmission graph is constructed to describe the transmission patterns in different industrial process faults:
[0031] When two process variables exhibit highly similar behavior, it usually means that they share the same underlying physical mechanism or are jointly affected by the same failure factor. "Information transfer" between such node pairs is more natural, and their transmission resistance should be lower.
[0032] Based on this approach, a learnable weight matrix is used to linearly transform the industrial process variables and map them to a high-dimensional space. Dot product similarity is chosen as the measure of similarity between industrial process variables, and the Softmax function is used to normalize the similarity between the industrial process variables. The dot product similarity is as follows: ;in, For the characteristics of process variable i, The characteristics of process variable j, which is a neighbor of process variable i; Let i be the learnable weight matrix of process variable i. The learnable weight matrix of the neighboring process variable j is obtained; the similarity distribution among the industrial process variables is obtained.
[0033] The similarity distribution among industrial process variables is normalized to the same metric space using a min-max method, and the relationship between industrial process variables with higher similarity distribution is given a lower transmission cost through an inversion operation:
[0034] ; ; ;
[0035] in This represents the dot product similarity, where min is the minimum value in the array, max is the maximum value in the array, exp is the exponential calculation, and slope coefficient is the slope coefficient. Center offset coefficient Numerical stability constant , This represents the transmission cost from node i to node j; The similarity distribution is the result of min-max normalization. This is an activation function used to drastically widen the value range near 0.5, amplifying the difference in cost; the output is a cost matrix that measures the differences between industrial process variables.
[0036] The Sinkhorn algorithm in the optimal transport problem is essentially a fixed solver. The dual variables u and v must be iteratively calculated to satisfy a predefined marginal distribution. Its basic form is:
[0037] ; ; ; ;
[0038] Here, `diag()` retrieves the diagonal elements, and `u` and `v` are dual variables. For exponential calculation, C is the cost matrix. Let be the numerical stability constant, t be the iteration number, and and be given marginal distributions that satisfy . and P is the transfer matrix obtained after multiple iterations. and These are the dual variables obtained after multiple iterations;
[0039] Since the Sinkhorn algorithm requires multiple iterations to output the transfer matrix, to avoid this computationally intensive iterative process, this embodiment uses a feedforward neural network to directly predict the corresponding dual variables u and v of the industrial process variables. Furthermore, to enhance the numerical stability of the model training, layer normalization is added after predicting the dual variables u and v.
[0040] ; ;
[0041] Where LayerNorm is layer normalization and FFN is a feedforward network. and For learnable matrices, It is a column vector. It is a row vector;
[0042] Dynamic optimal transmission graph used to describe the implicit transmission relationships within the entire system. C is composed of The combined cost matrix, numerical stability constant Implemented using PyTorch's broadcast mechanism, ensuring that dual variables u and v are expanded and added to the list. ;Utilize information propagation mechanisms to aggregate the dynamic optimal transport graph into process variable features: ,in Here, W is the ReLU activation function, and W is the learnable matrix. For industrial process variable characteristics, Features of industrial process variables that include the transport relationships between industrial process variables;
[0043] Then, the characteristics of industrial process variables that include the transport relationships between industrial process variables will be included. Input is fed into a Kolmogorov-Arnold network that is more favorable for classification of the latent representation space, and outputs the predicted label feature y.
[0044] The Kolmogorov-Arnold network is designed based on the Kolmogorov-Arnold representation theorem, which states that any continuous multivariable function can be expressed as a superposition of multiple one-dimensional functions. It approximates any continuous multivariable function using nested and summative forms of multiple one-dimensional spline functions.
[0045] ;
[0046] ;
[0047] ;
[0048] ;
[0049] in For the Kolmogorov-Arnold network, To connect the i-th neuron in the previous layer with the j-th neuron in the next layer... Layer single-connection mapping function, For input features, for Learnable matrix, The sliu activation function is used. for Learnable matrix, For spline functions, For learnable weights, Let be the k-th order b-spline basis function of the i-th neuron;
[0050] Step 4: Process the industrial process fault training set as described in Step 2. training set of fault types Step 3 involves building an industrial process fault diagnosis model and training it multiple times.
[0051] During training, the feature transport loss with explicit constraints is applied to the embedding space to enhance the inter-class separability and intra-class compactness of the fault category embedding space. The predicted label features are transformed into class probability distributions, which are then projected into the learnable fault category embedding space and aligned with the embedding vectors of the real categories.
[0052] A learnable fault category embedding space is introduced: ;in The vector features are for fault category C, and have been normalized to row vectors;
[0053] Given the probability distribution of the predicted label features Probability distribution of true labels in the fault category embedding space Construct a feasible transport matrix from the probability distribution Q of the predicted label features to the probability distribution P of the true label in the fault category embedding space: The transmission matrix satisfies the marginal constraints of the optimal transmission problem (OT): , This means that the sum of the elements in the j-th column is 1. This means that the sum of the elements in the i-th row is 1. ;Calculate the tractable upper bound of the Wasserstein distance between the predicted label feature probability distribution and the true label probability distribution in the fault category embedding space:
[0054] ; ;
[0055] ; ; ;
[0056] Where Q is the probability distribution of the given predicted label features, and P is the probability distribution of the true label in the fault category embedding space. For the feasible transport matrix, The normalized form is 1 is an all-one matrix, and E is the identity matrix;
[0057] The final loss function consists of two parts: ;
[0058] Where CE is the cross-entropy loss function, FTL is the feature transport loss, and y is the output predicted label feature. For real labels, To adjust the ratio;
[0059] This loss function is used to measure the difference between the model's predictions and the actual values, providing direction for optimizing model parameters.
[0060] Based on the loss function size of the dynamic optimal transport graph neural network model and the industrial process fault training set... Industrial process fault training set corresponding to fault type training set The Adam optimizer is used to iteratively update the network parameters.
[0061] To assess the fault classification capability of an industrial process fault diagnosis model based on a dynamic optimal transfer graph neural network, the precision, recall, and accuracy of this model are calculated. The definitions of precision, recall, and accuracy are as follows:
[0062] ; ; ;
[0063] Among them, TP (True Positive) samples were positive and the prediction result was positive; FN (False Negative) samples were positive but the prediction result was negative; FP (False Positive) samples were negative but the prediction result was positive; TN (True Negative) samples were negative and the prediction result was negative. The higher the value of these indicators, the stronger the model's fault diagnosis ability.
[0064] During each training session, the industrial process failure verification set is used. The input is fed into a trained industrial process fault diagnosis model, which outputs a set of predicted values. , the predicted value set Fault type validation set The model is compared and validated, and the parameters of the industrial process fault diagnosis model with the best performance in multiple training iterations are automatically saved.
[0065] Step 5: Input any industrial process variable data into the optimal industrial process fault diagnosis model and output the predicted values. This enables the diagnosis of different types of faults in industrial processes.
[0066] To demonstrate the effectiveness of the above method, the following is combined with... Figure 3 The Tennessee-Eastman process is illustrated to demonstrate the fault diagnosis performance of this invention. The Tennessee-Eastman process, jointly developed by Eastman Chemical Company and the University of Tennessee, aims to construct complex, realistic, yet reproducible industrial chemical processes. Due to its complexity, numerous variables, and strong coupling, the Tennessee-Eastman process serves as a benchmark platform in the field of process control. The entire industrial process of the Tennessee-Eastman process mainly involves realizing the following reaction equations: A, C, D, and E are gaseous raw materials, while G, H, and F are liquid products.
[0067] The Tennessee-Eastman process has 22 measurement variables (XMEAS 1-22) and 9 control variables (XMV 1-4, 6-8, 10-11). The specific descriptions of the measurement process variables are shown in Table 1, and the descriptions of the control variables are shown in Table 2.
[0068] Table 1: Description of Tennessee-Eastman process measurement variables
[0069] XMEAS(1) Feed A (flow 1) <![CDATA[Km 3 / h]]> XMEAS (2) D feed (flow 2) Kg / h XMEAS (3) Feed E (flow 3) Kg / h XMEAS (4) Total feed (flow 4) <![CDATA[Km 3 / h]]> XMEAS (5) Recirculation flow (flow 8) <![CDATA[Km 3 / h]]> XMEAS (6) Reactor feed rate (flow rate 6) <![CDATA[Km 3 / h]]> XMEAS (7) reactor pressure kPa (table value) XMEAS (8) Reactor grade % XMEAS (9) Reactor temperature ℃ XMEAS (10) Emission rate (flow 9) <![CDATA[Km 3 / h]]> XMEAS (11) Separator temperature ℃ XMEAS (12) Separator liquid level % XMEAS (13) Separator pressure kPa (table value) XMEAS (14) Separator bottom flow rate (flow 10) <![CDATA[m 3 / h]]> XMEAS (15) Stripping Tower Grade % XMEAS (16) stripping tower pressure kPa (table value) XMEAS (17) Flow rate at the bottom of the stripper tower (flow 11) <![CDATA[m 3 / h]]> XMEAS (18) stripping tower temperature ℃ XMEAS (19) Stripping tower flow rate Kg / h XMEAS (20) Compressor power kW XMEAS (21) Reactor cooling water outlet temperature ℃ XMEAS (22) Separator cooling water outlet temperature ℃
[0070] Table 2: Description of Tennessee-Eastman process control variables
[0071] XMV(1) D Feed rate (flow 2) XMV(2) Feed rate (flow 3) XMV (3) A. Feed rate (flow 1) XMV (4) Total feed volume (flow 4) XMV (6) Drain valve (flow 9) XMV (7) Separator tank liquid flow rate (flow 10) XMV (8) Liquid product flow rate of stripper (flow 11) XMV (10) Reactor cooling water flow rate XMV (11) Condenser cooling water flow rate
[0072] And 28 types of faults. Since the change curves of faults 3, 9 and 16 are similar to those of normal operation, they were excluded from this experiment; while faults 18, 21 and 23 were omitted because their unknown fault samples proved to be difficult to distinguish; therefore, only 22 fault categories as shown in Table 3 were selected.
[0073] Table 3: Description of Tennessee-Eastman Process Failure Types
[0074] 1 A / C changes, B remains unchanged (Flow 4) Step 2 A / C remains constant, B varies (Flow 4) Step 4 Reactor cooling water inlet temperature change Step 5 Condenser cooling water inlet temperature variation Step 6 A. Feed loss (flow 1) Step 7 Pressure loss of material C (flow 4) Step 8 Compositional changes of materials A, B, and C (Flow 4) random 10 Temperature change of material C (flow 2) random 11 Reactor cooling water inlet temperature random 12 Condenser cooling water inlet temperature random 13 Reaction dynamics Slow offset 14 Reactor cooling water valve Viscous 15 Condenser cooling water valve Viscous 17 unknown unknown 19 unknown unknown 20 unknown unknown 22 Reactor cooling system failure unknown 24 Condenser cooling system failure unknown 25 Feed pipe blockage unknown 26 Discharge pipe blockage unknown 27 unknown unknown 28 unknown unknown
[0075] The specific prediction steps are as follows:
[0076] Step 1: For the Tennessee-Eastman process, 154,000 industrial process fault data were collected, of which the first 70% were used as the training set, the middle 20% as the validation set, and the last 10% as the test set.
[0077] Step 2: Process the industrial process fault training set, validation set and test set from Step 1. Use zero-mean normalization to map the industrial process fault training set to the [0,1] interval. At the same time, map the industrial process fault validation set and industrial process fault test set to the [0,1] interval according to the data distribution in the industrial process fault training set.
[0078] Step 3: Build an industrial process fault diagnosis model based on a dynamic optimal transport graph neural network using the open-source deep learning platform PyTorch.
[0079] First, a dynamic optimal transport graph is constructed, mapping all time-dimensional features of each industrial process variable to a continuous high-dimensional space, with the embedding layer size set to embedding=16; slope coefficient... Center offset coefficient Numerical stability constant The dynamic optimal transport graph is aggregated into industrial process variable features using information propagation mechanisms, outputting industrial process variable features that include the transport relationships between industrial process variables. ;
[0080] Then, the characteristics of industrial process variables that include the transport relationships between industrial process variables will be included. The input is fed into the Kolmogorov-Arnold network, and the output is the predicted label y; the hidden layer size of the Kolmogorov-Arnold network is [31,63,31,22].
[0081] Step 4: Using the training set data of the Tennessee-Eastman process, compare the predicted label y with the true label y. The model's prediction loss is calculated by inputting the loss function constructed from the cross-entropy loss function and the feature transport loss, and the model parameters are optimized. The hyperparameters of the training model are set as follows: batchsize=256, epoch=50, learning rate=0.005.
[0082] After every 5 training iterations, the model is validated using the Tennessee-Eastman process validation set and the fault type validation set data, and the model parameters with the best performance after multiple training iterations are saved.
[0083] Step 5: Input any Tennessee-Eastman process variable into the industrial process fault diagnosis model and output the model's predicted fault type value.
[0084] To verify the prediction performance of the method of this invention, the classification results of the fault diagnosis model built by supervised autoencoder network (SAE), semi-supervised stacked autoencoder network (SSAE), graph attention neural network (GAT), graph convolutional network (GCN), graph attention neural network-V2 (GAT-V2), supervised graph temporal graph convolutional network (TGCN-S), and deep learning-based dynamic graph (DGDL) were compared. Since the supervised autoencoder network and semi-supervised stacked autoencoder network have different structures from other comparative models, the batch size of the stacked supervised autoencoder network and semi-supervised stacked autoencoder network was set to 128. The test set data was input into the seven comparative networks mentioned above and the method described in this invention. The fault detection effects of the eight methods are shown in Table 4. All results came from the same host equipped with an Intel Core i5-14600KF 5.30GHz CPU and a GeForce GTX 5060-Ti 8G GPU, running in Python 3.9 and PyTorch 1.13 environment. At the same time, all experiments used the same five random seeds, and each random seed was used for three experiments. The average of the results from multiple experiments was taken.
[0085] Table 4. Fault detection results of eight methods
[0086]
[0087] As shown in Table 4, the method of this invention achieves better precision and recall than the baseline model in predicting multiple fault categories in the Tennessee-Eastman process. Precision is improved by 30.41%–67.24% and recall by 51.10%–77.21% compared to the baseline model, respectively. This indicates that the method of this invention has stronger discriminative ability and more stable generalization performance. Compared to autoencoder models, it performs better in processing industrial process fault data because it can effectively capture the spatial correlation between nodes using a graph structure. Compared to graph neural network models, it can learn material and energy transfer patterns under different fault states, and the samples of the same type are more compact and samples of different types are easier to separate in the fault category embedding space. This improves fault diagnosis performance and outputs results that are closer to the actual fault situation.
[0088] Specific fault classifications are as follows: Figure 4As shown, the horizontal axis represents the labels predicted by the model based on the test set data in the Tennessee-Eastman process, and the vertical axis represents the true labels corresponding to the test set data in the Tennessee-Eastman process. The vast majority of the data is concentrated on the diagonal of the confusion matrix, which indicates that the labels predicted by the model match the true labels, and only a very small number of labels are misclassified. The method of this invention demonstrates performance far superior to the benchmark model.
[0089] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for fault diagnosis of industrial processes based on a dynamic optimal transport graph neural network, characterized in that, Specifically as follows: First, industrial process variable data are collected to construct a dynamic optimal transmission diagram to describe the transmission modes in different industrial process faults, specifically: By using a learnable weight matrix to perform a linear transformation on industrial process variables and map them to a high-dimensional space, the similarity distribution among industrial process variables can be obtained. After normalizing the similarity distribution among industrial process variables to the same metric space, the relationship between industrial process variables with higher similarity distribution is given a lower transmission cost through the inversion operation, and combined to form a cost matrix; By using a feedforward neural network to directly predict the corresponding dual variables of industrial process variables, and by using a broadcasting mechanism to extend the dual variables into the cost matrix, a dynamic optimal transmission graph is formed. Then, the dynamic optimal transmission diagram is aggregated into the industrial process variable features using the information propagation mechanism to obtain industrial process variable features that contain the transmission relationships between industrial process variables; Finally, the industrial process variable features are mapped onto a Kolmogorov-Arnold network, and the nested and summed forms of multiple one-dimensional spline functions are used to approximate any multivariable continuous function, outputting multiple prediction labels.
2. The industrial process fault diagnosis method based on dynamic optimal transport graph neural network as described in claim 1, characterized in that, The industrial process variables are linearly transformed and mapped to a high-dimensional space using a learnable weight matrix. Dot product similarity is selected as the measure of similarity between industrial process variables, and the Softmax function is used to normalize the similarity between industrial process variables to obtain the similarity distribution between industrial process variables.
3. The industrial process fault diagnosis method based on dynamic optimal transport graph neural network as described in claim 1, characterized in that, The similarity distribution among industrial process variables is normalized to the same metric space using a min-max method, and the relationship between industrial process variables with higher similarity distribution is given a lower transmission cost through a reversal operation, and the two are combined to form a cost matrix.
4. The industrial process fault diagnosis method based on dynamic optimal transport graph neural network as described in claim 1, characterized in that, Layer normalization was added after the dual variables were predicted.
5. The industrial process fault diagnosis method based on dynamic optimal transport graph neural network as described in claim 1, characterized in that, Zero-mean normalization was used to normalize the industrial process variable data.
6. The industrial process fault diagnosis method based on dynamic optimal transport graph neural network as described in claim 1, characterized in that, The dynamic optimal transport graph neural network is trained using the cross-entropy loss function and the feature transport loss function; The feature transport loss function is as follows: a learnable fault category embedding space is introduced; given the probability distribution of the predicted label features and the probability distribution of the true label in the fault category embedding space, a feasible transport matrix is constructed from the probability distribution of the predicted label features to the probability distribution of the true label in the fault category embedding space, which satisfies the marginal constraints of the optimal transport problem; and the tractable upper bound of the Wasserstein distance between the probability distribution of the predicted label features and the probability distribution of the true label in the fault category embedding space is calculated.