A fault diagnosis method for polyester esterification stage

Abstract

This invention discloses a fault diagnosis method for the polyester esterification stage. This method combines global supervised learning and contextual metric meta-learning. Using the attribute information of a single sample and similarity information from a sample group, it first uses variational modal decomposition to obtain multi-scale data through global supervised training. Multi-scale components with fault characteristics are extracted, and multi-scale feature fusion learning is performed. Triplet loss is used to learn finer, more subtle features. A fixed multi-scale feature fusion module is then used for task meta-learning training to learn a single feature, converting the raw data of the meta-task into a basic feature space. Finally, a dimensional variational prototype module is used to adaptively measure the feature similarity of sample pairs. The statistical method of variational inference automatically learns metric scaling parameters to transform the embedding space. The method is simple and solves the fault diagnosis problem in scenarios with limited data and a fully open set.

Description

Technical Field

[0001] The invention belongs to the technical field of automatic control and relates to a fault diagnosis method in a polyester esterification stage. Background Art

[0002] In modern, complex industrial processes characterized by scale, sophistication, complexity, and intelligence, process failures can cause serious casualties and economic losses. To prevent this, a range of fault diagnosis technologies have been developed. Traditional signal processing-based methods often require excessive manual intervention and struggle to meet the diagnostic accuracy and efficiency requirements of modern equipment, driven by the large-scale and automated nature of modern equipment. With the widespread deployment of sensors and the availability of vast amounts of data, data-driven approaches have become an effective technology for industrial fault diagnosis.

[0003] In real-world industrial scenarios, due to variations in process equipment and operating conditions, new operating conditions may present fault categories that differ from the original conditions. This problem is referred to in this paper as the open-set fault diagnosis problem. Approaches to this problem fall into two main categories: one is based on discriminant models, employing a threshold classification scheme where the decision maker either rejects input samples or classifies them into known categories based on empirically set thresholds. The other is based on instance generation, which addresses the challenges posed by label space disparity in data analysis by generating domain- and category-specific data. While the open-set fault diagnosis problem accounts for the possibility of new fault types emerging during testing, in practice, many faults can be devastating and cause significant losses. Few factories allow for the occurrence of faults and the collection of samples to train fault diagnosis systems. Consequently, collecting sufficient fault data becomes challenging. Once the amount of labeled data decreases dramatically, these approaches risk performance degradation, which we refer to as the fault diagnosis problem in under-data scenarios.

[0004] To address the problem of fault diagnosis in data-poor scenarios, researchers have proposed a large number of solutions from different perspectives. One approach is to learn a general enhancement function on an auxiliary dataset based on a data augmentation method, or to directly implement an enhancement strategy on a test dataset to increase the number of training samples. However, due to the scarcity of data, the original dataset may not fully cover all possible fault conditions, thus limiting the generalization ability of the model. In addition, methods based on transfer learning can obtain transferable fault knowledge from one domain and apply it to other different but related domains. However, if there is a large difference between the source task and the fault diagnosis task, the pre-trained model may not provide effective feature representation, thereby reducing the effect of transfer learning.

[0005] In recent years, metric-based meta-learning models have received widespread attention. They achieve faster and more accurate classification in few-shot learning by learning a unified, class-independent distance feature space. Reference 1 (Reweighted Regularized Prototypical Network for Few-Shot Fault Diagnosis [J]. IEEE Transactions on Neural Networks and Learning Systems, 2022.) proposed a multi-scale dynamic fusion prototype network based on the fuzzy c-means clustering algorithm to provide a more accurate distance metric benchmark for the nearest neighbor classifier. Reference 2 (Metric-based meta-learning model for few-shot fault diagnosis under multiple limited data conditions [J]. Mechanical Systems and Signal Processing, 2021, 155: 107510.) constructed a convolutional twin neural network that accurately learns the classification boundaries between sample pair features by maximizing inter-class distance and minimizing intra-class distance. Reference 3 (TRNet: ACross-Component Few-Shot Mechanical Fault Diagnosis [J]. IEEE Transactions on Industrial Informatics, 2022) proposes a reweighted regularized prototype network using an intra-class reweighting strategy to reduce the impact of noise and outliers and obtain stable prototype estimates. However, these methods directly apply meta-learning methods to fault diagnosis in data-poor scenarios. Because the models focus on task-level learning and ignore fault feature-level learning, the diagnostic performance of these models is generally poor.

[0006] In actual industrial engineering, production line A has abundant fault samples, while production line B has a new fault category, and this fault category has a very limited number of samples. This invention is called the fault diagnosis problem in a fully open set scenario with limited data. Although metric-based meta-learning provides a solution, a large number of hard negative samples will be encountered when processing real process industry fault data. Specifically, the sample features between categories may be more similar than the sample features within the category, and the sample features within the category may be less similar than the sample features between categories. This leads to the problem of extracting poor-quality fault features and misclassification. In addition, for high-dimensional and strongly coupled process faults, using only a simple adaptive distance metric mapping will aggravate the dilemma of hard negative sample identification. Summary of the Invention

[0007] The purpose of the present invention is to solve the problems existing in the prior art and provide a fault diagnosis method for the polyester esterification stage. Specifically, the present invention uses VMD technology (variable mode decomposition) to perform modal decomposition on raw data to obtain multi-scale data; trains a multi-scale feature fusion module through global supervision to complete multi-scale data fault feature fusion, and uses a hard negative sample recognition module to reduce the problem of hard sample difficulty in recognition, so as to more effectively learn subtle features; fixes the multi-scale feature fusion module and shares it in the meta-task training stage, so that the raw data of the scenario task is converted into a basic feature space; uses variational inference to convert the embedding space into a more realistic data space, and adaptively learns prototype vectors to measure the feature similarity of fault sample pairs.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A method for diagnosing a fault in a polyester esterification stage comprises the following steps:

[0010] (1) Collect data during the polyester esterification stage of production line B during the actual production process to obtain fault sample B to be diagnosed;

[0011] (2) Experts and technicians perform fault diagnosis on the fault sample B to be diagnosed, and divide the fault sample B to be diagnosed into the diagnosed fault sample B1 with known fault type and the fault sample B2 with unknown fault type. All B1 are combined into the support set S * , S * There are C* fault types in it;

[0012] (3) Using VMD technology to decompose B1 and B2, a multi-scale feature with a scale number of H is obtained;

[0013] (4) S * The samples with the same fault type are classified into one category, and the fault prototype τ of the c*th category is calculated. c* (i.e., cluster center), c*=1,2...,C*;

[0014]

[0015] Where, Stands for S * There are C* samples of fault types in Stands for S * The number of samples of C* fault types in Stands for S * The D-dimensional feature vector of the sample in, Stands for S * The fault type labels of the samples in Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing;

[0016] (5) Calculation With τ c* Euclidean distance represents the D-dimensional feature vector of B2, Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing;

[0017] (6) Calculate the probability that the fault type of B2 belongs to the c* class Represents the model's prediction Label, α represents the dimensionality scaling parameter of the dimensional variation prototype module, τ u* Represents fault prototypes of other categories except category c*;

[0018] (7) The fault type corresponding to the maximum probability is taken as the fault type of B2;

[0019] The multi-scale feature fusion module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, the second pooling layer, the third one-dimensional convolution layer and the fourth one-dimensional convolution layer connected in sequence; the first one-dimensional convolution layer uses H convolution kernels of the same size to fuse and superimpose multi-scale features of scale number H into a single-scale feature map; the purpose of not setting a pooling layer after the third and fourth one-dimensional convolution layers is to retain sufficient information for the subsequent dimensional variation prototype module to perform convolution operations;

[0020] The dimensional variation prototype module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, and the second pooling layer connected in sequence;

[0021] The training method for the multi-scale feature fusion module and the dimensional variational prototype module is as follows: for the fault samples of production line A, the fault type of the source domain data is first identified through global supervised training. Then, the multi-scale feature fusion module is fixed for task meta-learning training to learn a feature representation so that the same type of faults are closer to different types of faults in the embedding space. Finally, the dimensional variational prototype module is used to adaptively measure the feature similarity of sample pairs, and the metric scaling parameters are automatically learned through the statistical method of variational inference to transform the embedding space. Production lines A and B are two production lines with the same process flow and are operated simultaneously.

[0022] As the preferred technical solution:

[0023] In the above-mentioned method for fault diagnosis in the polyester esterification stage, in step (2), the process of the experts and technicians performing fault diagnosis on the fault sample B to be diagnosed is as follows: the experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault based on the process parameters and the fault phenomenon, compare them with the known fault database, and give the fault type based on professional knowledge and the specific production situation.

[0024] In the above-mentioned fault diagnosis method for the polyester esterification stage, in step (3), when B1 is decomposed using VMD technology, the number of modes is set to 3. When the original signal is decomposed into 4 modes, the center frequency distance between two adjacent modal components is the closest, which can be considered as over-decomposition of the mode, so the number of modes is set to 3.

[0025] As described above, in a fault diagnosis method for the polyester esterification stage, the kernel size in the first one-dimensional convolution layer of the multi-scale feature fusion module is 64×1×3, the step size is 2×1, the number of kernels is 16, and the activation function is ReLU; the kernel size in the first pooling layer is 2×1, the step size is 2×1, the number of kernels is 16, and the activation function is Max; the kernel size in the second one-dimensional convolution layer is 3×1, the step size is 2×1, the number of kernels is 32, and the activation function is ReLU; the kernel size in the second pooling layer is 2×1, the step size is 2×1, the number of kernels is 32, and the activation function is Max; the kernel size in the third one-dimensional convolution layer is 2×1, the step size is 2×1, the number of kernels is 64, and the activation function is ReLU; the kernel size in the fourth one-dimensional convolution layer is 2×1, the step size is 2×1, the number of kernels is 64, and the activation function is ReLU.

[0026] The present invention first decomposes the signal into different modes and treats it as multi-scale data. It then performs feature extraction at multiple scales only in the first layer. The features are then fused and the signal is treated as a single-scale signal for feature extraction. Large-scale convolution kernels first extract features from different scales, which helps to obtain more general features of the fault. After feature fusion, smaller convolution kernels are used to extract finer-grained fault feature information, which helps to extract overall features.

[0027] As described above, in a fault diagnosis method for the polyester esterification stage, the kernel size in the first one-dimensional convolution layer of the dimensional variational prototype module is 2×1, the step size is 2×1, the number of kernels is 64, and the activation function is ReLU; the kernel size in the first pooling layer is 2×1, the step size is 2×1, the number of kernels is 64, and the activation function is Max; the kernel size in the second one-dimensional convolution layer is 2×1, the step size is 2×1, the number of kernels is 128, and the activation function is ReLU; the kernel size in the second pooling layer is 2×1, the step size is 2×1, the number of kernels is 128, and the activation function is Max.

[0028] In the polyester esterification stage fault diagnosis method described above, the training process of the multi-scale feature fusion module and the dimensional variational prototype module is as follows:

[0029] (a) During the actual production process, data is collected at the polyester esterification stage of production line A to obtain fault sample A to be diagnosed;

[0030] (b) Experts and technicians perform fault diagnosis on the fault sample A to be diagnosed and obtain the fault type of the fault sample;

[0031] (c) Using VMD technology to decompose the training samples, multi-scale features with H scales are obtained;

[0032] (d) Constructing triple data Are training samples, respectively as anchor points, positive samples, and negative samples, and The fault type is the same as and The fault types are different; in the feature extraction process of the present invention, triplet loss is used to distinguish the most dissimilar samples in the same category (i.e., hard positive samples) and the most similar samples in different categories (i.e., hard negative samples), so as to learn better subtle features;

[0033] (e) The triple data Input into the multi-scale feature fusion module, which outputs the feature vector

[0034] (f) Two nonlinear fully connected layers are introduced to reduce the dimension of the feature vector so that the Euclidean distance can be used directly for measurement. The simplified feature vector Among them, FC2 represents the operation after two fully connected layers;

[0035] (g) Calculated and

[0036] (h) Judge d - ≥d ++γ is established, 0<γ<1, if so, the parameters of the multi-scale feature fusion module are fixed and the training of the multi-scale feature fusion module is completed; otherwise, the next step is entered; the present invention introduces the parameter γ to avoid + =d - When the model cannot determine the category, setting γ can increase the distance between the anchor point and the negative sample, while reducing the distance between the anchor point and the positive sample;

[0037] (i) Gradient back propagation trains the parameters of the multi-scale feature fusion module until the loss function L g convergence, N s Represents the total number of training samples, n represents the number of training samples, fixes the parameters of the multi-scale feature fusion module, and completes the training of the multi-scale feature fusion module;

[0038] (j) Part of the training samples are combined into a support set S, and part of the training samples are combined into a query set Q. There are no identical training samples in S and Q.

[0039] (k) Use S and Q to construct M fault diagnosis tasks. Each fault diagnosis task contains C fault types in S. Each fault type has K training samples in S and L training samples in Q. K<<L. The i-th fault diagnosis task in, Represents the D-dimensional feature vector of the training sample in S, y i,j represents the fault type label of the training sample in S, represents the D-dimensional feature vector of the training samples in Q, represents the fault type label of the training sample in Q, (x i,j ,y i,j ) represents the training sample in S for the i-th fault diagnosis task, represents the training sample in Q for the i-th fault diagnosis task;

[0040] (l) Reparameterize the dimensionality scaling parameter α of the dimensional variational prototype module, α = μ + εσ, where μ represents the expectation of α, σ represents the expectation of α, and ε represents the expectation of ε. i ~N(0,1) random numbers;

[0041] (m) Set the initial values of μ and σ;

[0042] (n) Classify the training samples with the same fault type in S into one category and calculate the fault prototype τ of category c c (i.e. cluster center), c = 1, 2..., C;

[0043]

[0044] Where S C represents the training samples of C fault types in S, |S C |=C×K,f θ (x i,j ) represents the use of the trained multi-scale feature fusion module to i,j The feature vector obtained after feature extraction, h φ (f θ (x i,j )) represents the use of dimensional variational prototype module to f θ (x i,j ) is further processed to obtain the metric features;

[0045] (o) Calculation With τ c Euclidean distance Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of dimensional variational prototype module The metric features obtained by further feature processing;

[0046] (p) calculation The probability that the fault type belongs to category c τ u Represents fault prototypes of other types except type c;

[0047] (q) Calculate the total loss L of M tasks p , the formula is as follows:

[0048]

[0049]

[0050] Where, represents the loss of the jth Q in the i-th task;

[0051] (r) Judge L p Is it converged? If so, fix μ and σ to complete the training of the dimensional variational prototype module; otherwise, proceed to the next step;

[0052] (s) Update μ and σ according to the following rules and return to step (n);

[0053]

[0054]

[0055] Where, l ψ is the update step size; U represents the number of samples of fault prototypes of other categories except category c; μ0 and σ0 are the initial values of the network input.

[0056] In the above-mentioned method for fault diagnosis in the polyester esterification stage, in step (b), the process of experts and technicians performing fault diagnosis on the fault sample A to be diagnosed is as follows: the experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault based on process parameters and fault phenomena, compare them with a known fault database, and assign a fault type based on professional knowledge and specific production conditions.

[0057] In the above-mentioned fault diagnosis method for the polyester esterification stage, in step (c), when the training sample is decomposed using VMD technology, the number of modes is set to 3. When the original signal is decomposed into 4 modes, the center frequency distance between two adjacent modal components is the closest, which can be considered as over-decomposition of the mode, so the number of modes is set to 3.

[0058] The principles of the present invention are as follows:

[0059] In order to solve the fault diagnosis problem in the fully open data scenario, this paper proposes a fault diagnosis method for the polyester esterification stage. First, using fully labeled source domain data, a multi-scale feature fusion module and a hard negative sample recognition module are trained through global supervision to obtain a multi-scale feature fusion module with a parameter θ, which is represented by a function f θ ; then fix f θ , the meta-scenario task constructed using source domain data is trained with a dimensional variational prototype module with parameter φ, which is represented as a function h φ ; Finally, the model is verified on the target domain data; The overall algorithm is based on Figure 1 The dimensional variational prototype network framework for multi-scale feature fusion shown in the figure is shown. The present invention fully utilizes the attribute information of a single sample and the similarity information from the sample group, overcoming the difficulty that existing fault diagnosis technologies are limited by the assumption that there are enough labeled samples for each fault type, which is not met in the real complex process industry with strong noise and strong coupling. The details are as follows:

[0060] Process signals generated by modern industrial processes often exhibit high-dimensional, nonlinear, and multi-scale characteristics, accompanied by strong noise. Current deep learning-based fault diagnosis methods mostly consider only a subset of the features of industrial data, resulting in the loss of some feature information during training, and the strong noise often affects the final diagnostic results. VMD technology decomposes the signal into a series of intrinsic mode functions (EMDs) of varying frequencies and amplitudes to filter out noise disturbances. For each eigenmode function, high-dimensional, nonlinear spatial features of varying scales are extracted, providing effective fault signatures in both the time and frequency domains of the process signal.

[0061] VMD technology is an existing modal decomposition technology and a basic data preprocessing method. It has an existing code package and sets the required number of modes. The present invention sets it to 3. Through experimental analysis, it is found that when decomposed into 4 modes, the center frequency distance between two adjacent modal components is the closest, and it can be considered that VMD is a modal over-decomposition.

[0062] The specific theory is as follows:

[0063] The goal of VMD is to decompose a real-valued input signal into a discrete number of sub-signals (modes) u k , the process mainly includes two stages: constructing the variational problem and solving the variational problem:

[0064] 1) Constructing a variational problem: Assume that the original signal is decomposed into H components, and the decomposition sequence is a modal component with a finite bandwidth and a center frequency. Then, estimate the bandwidth of each modal signal. The resulting constrained variational problem is as follows:

[0065]

[0066] Where H represents the total number of intrinsic mode functions (IMFs); {u h} represents the set of each mode; {ω h} represents the center frequency of each mode; δ(t) represents the impulse function; * represents convolution; m represents the original signal;

[0067] 2) Variational mode solution: In order to find the optimal solution to the constrained variational problem, the Lagrange multiplication operator λ(t) and the quadratic penalty factor are introduced. Convert the constrained problem into an unconstrained problem, namely:

[0068]

[0069] Iterative update calculation using the multiplier alternating direction algorithm and In order to find the "saddle point" of the Lagrangian function above, that is, the optimal solution of the variational problem, the alternating direction method of multipliers (ADMM) is used to iteratively update each mode and its center frequency.

[0070] Through VMD, multi-scale features with H scales are obtained. The multi-scale one-dimensional process signal is used as the network input. H learnable convolution kernels of the same size are used to automatically extract features from the multi-scale feature map to learn comprehensive information from the process signal and achieve effective information fusion. This multi-scale feature fusion module includes an input layer with multiple scales. The first convolution operation uses a larger convolution kernel to obtain a large receptive field. The subsequent convolution layers use smaller convolution kernels to extract more fine-grained feature information. The last convolution layer is not connected to the pooling to retain sufficient information for the subsequent dimensional variational prototype module to perform convolution operations. Figure 2 The first-layer convolutional computing architecture of the multi-scale feature fusion module is shown.

[0071] For f θ In the training, we learn better subtle features to distinguish the most dissimilar samples in the same category (i.e. hard positive samples) from the most similar samples in different categories (i.e. hard negative samples). This paper builds a hard negative sample recognition module. This module uses the source domain fault data to First, select a sample from any fault type as an anchor point Then select another sample from this type of fault as a positive sample Select a sample from other fault types as a negative sample to form a triplet data By f θ Complete feature extraction and get Since the features obtained at this time are high-dimensional, the Euclidean distance cannot be measured. Therefore, two nonlinear fully connected layers are added to reduce the dimension of the feature vector. The simplified feature vector is represented as follows: Among them, FC2 represents the calculation after two fully connected layers. and Then require d + The smaller the better, d - The bigger the better, when d - ≥d + +γ, then Loss = 0, no optimization is required, otherwise it means that the model cannot distinguish between the positive and negative samples in this group, then Loss = d + +γ-d - , then the loss of global supervised training is: The relative distance parameter γ is set mainly to avoid d + =d -When the model cannot determine which category it belongs to, the model can be forced to make the distance between the anchor point and the negative sample larger, and the distance between the anchor point and the positive sample smaller. By optimizing the difficult negative sample recognition module, the hard positive samples and hard negative samples can be effectively identified to learn better subtle features.

[0072] Use the supervised information of a single sample in the source domain to train f θ ,These layers are then fixed and shared to the meta-scenario training to transform the raw data of the meta-task into the basic feature space.

[0073] In global supervised training, we get f θ Finally, the present invention can use meta-scenario training to solve the fault diagnosis problem in the fully open set scenario with limited data. In order to train the dimensional variational prototype module, the present invention considers using a prototype network with a scaling parameter to learn prototype vectors for representing different categories of data. However, the scaling parameter in scalar form can only simply change the scale of the embedding space and cannot adjust the relative position between the query set and the support set. Therefore, the present invention generalizes the scaling parameter in scalar form to a dimensional scaling vector, which can transform the embedding space into a space suitable for the data.

[0074] Choosing an appropriate dimension is crucial to ensure that data points can be projected into a linearly separable space. Too low a dimension can lead to information loss, while too high a dimension can introduce redundancy. The optimal dimension is related to the correlation of the data and is difficult to determine as a hyperparameter before training. Therefore, this paper adopts a Bayesian perspective to redefine the prototypical network based on the dimensionality scaling metric and automatically learns the scaling parameter α of the dimensionality metric through the statistical method of variational inference.

[0075] Assume that the posterior distribution of α is p φ (α|T s ), then the prediction distribution can be represented as:

[0076]

[0077] The conditional distribution is a discriminative classifier represented by the parameter φ.

[0078] The goal of variational inference is to find an approximate posterior probability distribution So that it is consistent with the true posterior probability distribution p φ (α|T s ) is minimized, which is represented by:

[0079]

[0080] The second term of the above formula is logp(T s )and Is irrelevant, then minimizing the KL divergence is transformed into maximizing the variational lower bound ELBO, which is represented by:

[0081]

[0082] Dimensional variational prototype module h φ The final objective function is represented as:

[0083]

[0084] Since sampling directly from the variational parameters is a nonlinear operation, the gradient cannot be calculated directly.

[0085] To solve this problem, we reparameterize α as ε i ~N(0,1). Assume Obey N(μ,σ 2 ), the variational parameter is represented by μ=(μ 1 ,μ 2 ,...,μ W ) Τ ,σ=(σ 1 ,σ 2 ,...,σ W ) Τ , where W is the dimension of the variational parameter, and the reparameterization is represented by:

[0086] α=g μ,σ (ε)=μ+εσ;

[0087] Assume q μ,σ (α)~N(μ,σ 2 ), prior distribution p(α)~N(μ0,σ0), the objective function is represented as:

[0088]

[0089] Taking partial derivatives of μ and σ respectively, we have:

[0090]

[0091]

[0092] Choose Euclidean distance as the metric, then With class prototype c c The distance between them is represented by:

[0093]

[0094] Then the update rules for the variational parameters μ and σ are:

[0095]

[0096]

[0097] The total loss of M tasks is L p , the formula is as follows:

[0098]

[0099] Where, Represents the loss of the jth Q in the i-th task.

[0100] Beneficial effects

[0101] (1) The present invention designs a variational prototype network based on multi-scale feature fusion, which fully utilizes the attribute information from a single sample and the similarity information from a sample group to solve the fault diagnosis problem in a fully open set scenario with limited data.

[0102] (2) The present invention redefines the prototype network based on dimensionality scaling metrics from a Bayesian perspective, transforms the embedding space into a space suitable for the data, and adaptively measures the feature similarity of sample pairs.

[0103] (3) This paper analyzes and constructs a polyester polymerization and esterification stage fault dataset and conducts extensive experiments on the proposed method. The results show that MFF-VPNet (the algorithm of this paper) can solve the fault diagnosis problem in the scenario of limited data and fully open datasets. By comparing with existing methods, the effectiveness and superiority of this method are verified, providing a feasible solution for fault diagnosis in complex industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 Schematic diagram of the dimensional variational prototype network framework for multi-scale feature fusion in the present invention;

[0105] Figure 2 This is a schematic diagram of the structure of the multi-channel feature fusion module in the present invention; only three channels are shown in the figure, and the processing methods of the remaining channels are the same;

[0106] Figure 3 It is a schematic diagram of the industrial process of the polyester polymerization esterification stage in the present invention. DETAILED DESCRIPTION

[0107] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0108] A method for diagnosing a fault in a polyester esterification stage comprises the following steps:

[0109] (1) Collect data during the polyester esterification stage of production line B during the actual production process to obtain fault sample B to be diagnosed;

[0110] (2) Experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault according to the process parameters and fault phenomena, and compare them with the known fault database. According to professional knowledge and specific production conditions, the fault type is given, and the fault samples B to be diagnosed are divided into the diagnosed fault samples B1 with known fault types and the fault samples B2 with unknown fault types. All B1 are combined into the support set S * , S * There are C* fault types in it;

[0111] (3) Using VMD technology to decompose B1 and B2, a multi-scale feature with a scale number of H is obtained; when decomposing B1, the number of modes is set to 3;

[0112] (4) S * The samples with the same fault type are classified into one category, and the fault prototype τ of the c*th category is calculated. c* , c*=1,2...,C*;

[0113]

[0114] Where, Stands for S * There are C* samples of fault types in Stands for S * The number of samples of C* fault types in Stands for S * The D-dimensional feature vector of the sample in, Stands for S * The fault type labels of the samples in Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing;

[0115] (5) Calculation With τ c* Euclidean distance represents the D-dimensional feature vector of B2, Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing;

[0116] (6) Calculate the probability that the fault type of B2 belongs to the c* class Represents the model's prediction Label, α represents the dimensionality scaling parameter of the dimensional variation prototype module, τ u* Represents fault prototypes of other categories except category c*;

[0117] (7) The fault type corresponding to the maximum probability is taken as the fault type of B2;

[0118] The multi-scale feature fusion module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, the second pooling layer, the third one-dimensional convolution layer, and the fourth one-dimensional convolution layer connected in sequence; the first one-dimensional convolution layer uses H convolution kernels of the same size to fuse and superimpose multi-scale features of H scales into a single-scale feature map;

[0119] In the first one-dimensional convolutional layer of the multi-scale feature fusion module, the kernel size is 64×1×3, the stride is 2×1, the number of kernels is 16, and the activation function is ReLU; in the first pooling layer, the kernel size is 2×1, the stride is 2×1, the number of kernels is 16, and the activation function is Max; in the second one-dimensional convolutional layer, the kernel size is 3×1, the stride is 2×1, the number of kernels is 32, and the activation function is ReLU; in the second pooling layer, the kernel size is 2×1, the stride is 2×1, the number of kernels is 32, and the activation function is Max; in the third one-dimensional convolutional layer, the kernel size is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU; in the fourth one-dimensional convolutional layer, the kernel size is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU;

[0120] The dimensional variation prototype module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, and the second pooling layer connected in sequence;

[0121] The kernel size of the first one-dimensional convolutional layer of the dimensional variational prototype module is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU; the kernel size of the first pooling layer is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is Max; the kernel size of the second one-dimensional convolutional layer is 2×1, the stride is 2×1, the number of kernels is 128, and the activation function is ReLU; the kernel size of the second pooling layer is 2×1, the stride is 2×1, the number of kernels is 128, and the activation function is Max;

[0122] The training method for the multi-scale feature fusion module and the dimensional variational prototype module is as follows: For fault samples from production line A, the source domain data's fault type is first identified through global supervised training. Then, the fixed multi-scale feature fusion module is used for task meta-learning training to learn a feature representation that makes similar faults closer to different faults in the embedding space. Finally, the dimensional variational prototype module is used to adaptively measure the feature similarity of sample pairs. The statistical method of variational inference is used to automatically learn the metric scaling parameter to transform the embedding space. Production lines A and B are two production lines with the same process flow that operate simultaneously. The specific training steps are as follows:

[0123] (a) During the actual production process, data is collected at the polyester esterification stage of production line A to obtain fault sample A to be diagnosed;

[0124] (b) Experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault based on process parameters and fault phenomena, compare them with the known fault database, and assign a fault type based on professional knowledge and specific production conditions;

[0125] (c) Using VMD technology to decompose the training samples, multi-scale features with a scale number of H are obtained. When decomposing the training samples, the number of modes is set to 3.

[0126] (d) Constructing triple data are all training samples, and The fault type is the same as and The fault types are different;

[0127] (e) The triple data Input into the multi-scale feature fusion module, which outputs the feature vector

[0128] (f) Two nonlinear fully connected layers are introduced to reduce the dimension of the feature vector. The simplified feature vector Among them, FC2 represents the operation after two fully connected layers;

[0129] (g) Calculated and

[0130] (h) Determine d-≥d + +γ is established, 0<γ<1, if so, the parameters of the multi-scale feature fusion module are fixed and the training of the multi-scale feature fusion module is completed; otherwise, proceed to the next step;

[0131] (i) Gradient back propagation trains the parameters of the multi-scale feature fusion module until the loss function Lg convergence, N s Represents the total number of training samples, n represents the number of training samples, fixes the parameters of the multi-scale feature fusion module, and completes the training of the multi-scale feature fusion module;

[0132] (j) Part of the training samples are combined into a support set S, and part of the training samples are combined into a query set Q. There are no identical training samples in S and Q.

[0133] (k) Use S and Q to construct M fault diagnosis tasks. Each fault diagnosis task contains C fault types in S. Each fault type has K training samples in S and L training samples in Q. K<<L. The i-th fault diagnosis task in, Represents the D-dimensional feature vector of the training sample in S, y i,j represents the fault type label of the training sample in S, represents the D-dimensional feature vector of the training samples in Q, represents the fault type label of the training sample in Q, (x i,j ,y i,j ) represents the training sample in S for the i-th fault diagnosis task, represents the training sample in Q for the i-th fault diagnosis task;

[0134] (l) Reparameterize the dimensionality scaling parameter α of the dimensional variational prototype module, α = μ + εσ, where μ represents the expectation of α, σ represents the expectation of α, and ε represents the expectation of ε. i ~N(0,1) random numbers;

[0135] (m) Set the initial values of μ and σ;

[0136] (n) Classify the training samples with the same fault type in S into one category and calculate the fault prototype τ of category c c , c=1,2...,C;

[0137]

[0138] Where S C represents the training samples of C fault types in S, |S C |=C×K,f θ (x i,j ) represents the use of the trained multi-scale feature fusion module to i,j The feature vector obtained after feature extraction, h φ (f θ (x i,j )) represents the use of dimensional variational prototype module to f θ (xi,j ) is further processed to obtain the metric features;

[0139] (o) Calculation With τ c Euclidean distance Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of dimensional variational prototype module The metric features obtained by further feature processing;

[0140] (p) calculation The probability that the fault type belongs to category c τ u Represents fault prototypes of other types except type c;

[0141] (q) Calculate the total loss L of M tasks p , the formula is as follows:

[0142]

[0143]

[0144] Where, represents the loss of the jth Q in the i-th task;

[0145] (r) Judge L p Is it converged? If so, fix μ and σ to complete the training of the dimensional variational prototype module; otherwise, proceed to the next step;

[0146] (s) Update μ and σ according to the following rules and return to step (n);

[0147]

[0148]

[0149] Where, l ψ is the update step size; U represents the number of samples of fault prototypes of other categories except category c; μ0 and σ0 are the initial values of the network input.

[0150] The effectiveness of the model of the method of the present invention is verified by specific experimental data below:

[0151] The esterification stage is the first stage of the polyester polymerization process. This stage is a complex industrial process with complex chemical processes and diverse production equipment. As for product viscosity, as long as the esterification reaction rate is within the required range, the product viscosity will be controlled within the required range. Real-time status monitoring, diagnosis and fault data analysis of the system process should start with the esterification reaction.

[0152] According to the process principle of Jima polyester esterification production, PTA and EG are mixed in proportion and then enter the esterification kettle from the bottom through the conveying pipe, and the flow of materials is controlled by the material pump; the esterification kettle is a double-chamber, vertical reactor with no stirring device in the kettle. The esterification reaction is uniform by natural circulation formed by excess EG; under high temperature conditions, the evaporated ethylene glycol and water vapor are separated in the esterification separation tower; the water vapor escapes from the top of the separation tower and is recycled after condensation; the distilled ethylene glycol comes out from the bottom of the separation tower and can be reused; the process flow of the esterification stage Figure 3 shown.

[0153] The esterification reaction occurs at high temperature and positive pressure, and is influenced by multiple reaction parameters. Temperature, liquid level, pressure, n(EG) / n(PTA), and reflux EG all influence the esterification reaction. Based on these characteristics and field experience, 20 monitoring variables (Table 1) were selected for each production line (Line A and Line B). These variables include 11 control variables and 9 process variables. These variables comprehensively characterize the production conditions during the esterification stage.

[0154] Table 1 Monitoring variables in polyester esterification stage

[0155]

[0156] The experimental data are valid sensor data collected by the DSC equipment of the two production lines in the esterification stage mentioned above;

[0157] Based on the existing literature and reaction mechanism, the collected sensor data were analyzed for variables, and the description of the fault type and sample size in the polyester esterification stage were finally obtained, as shown in Table 2:

[0158] Table 2 Description of failure types in polyester esterification stage

[0159]

[0160] manual

[0161]

[0162] As shown in Table 2, both production lines A and B contain 8 types of faults, and there are 6 types of faults common to both production lines;

[0163] The meta-learning of the present invention has a support set S and a query set Q in the training phase, which are completed on production line A; after the network training is completed, the fault sample types (such as labeled samples in the three categories of II1, II5, and II7) analyzed by experts and technicians on production line B are used to calculate the fault prototype, and then the network is tested with (other faults in the three categories of II1, II5, and II7). Specifically, the present invention uses the fault samples of production line A (I1, I2, I3, I4, and I6) to train the network, and tests it on the fault samples of production line B (II1, II5, and II7), respectively, verifying the accuracy of the model in two open set diagnostic scenarios with limited data: 3-way-10-shot and 3-way-20-shot; although the sensors of the two production lines are arranged in the same way, the working conditions of the two production lines are very different; the number of training fault class samples is uniformly selected as 750, and the number of test fault class samples is uniformly selected as 350, and each sample contains 32 consecutive time step information; the Z-Score normalization method is used to normalize the samples to obtain high-quality training and test results;

[0164] The fault samples are decomposed using VMD technology, with the number of scales set to 3; the triplet loss hyperparameter is set to γ = 0.3; the prior distribution of the dimension scaling parameter is set to p(α) = N(1, 1), and the variational parameter is initialized to μ init =100, σ = 0.2, and the regularization parameter is set to η = 0.7; to avoid overfitting, dorpout is set to 0.2, and the overall model is optimized by the Adam optimizer with an initial learning rate of 1e-4 and a learning rate decay rate of 0.8. The learning rate decreases with the decay rate every 20 iterations;

[0165] In order to evaluate the effectiveness and advancement of the algorithm model (MFF-VPNet) of the present invention, several classic insufficient data fault diagnosis machine learning algorithms, including FDDPN, MRN, CFDM, FSM3, and TRNet, were used to identify the predicted samples using the fault samples. The recognition results are shown in Table 3:

[0166] Table 3 Recognition accuracy of production line target test faults under different sample sizes

[0167]

[0168] As can be seen from Table 3, the algorithm model of the present invention can more accurately identify fault types than the classic machine learning algorithm, and has achieved considerable results in the actual polyester esterification reaction process, further demonstrating the effectiveness and advancement of the proposed model, which can solve the fault diagnosis problem in the fully open data set scenario with limited data.

[0169] The data used in this paper is constructed by analyzing real complex process industry data. The process industry is complex and the working environment and operating conditions are harsh. The data has the characteristics of strong noise and strong coupling. The above-mentioned comparison methods (FDDPN, MRN, CFDM, FSM3, TRNet) basically complete the diagnostic tasks on simulation data sets. The simulation data cannot well reflect the actual performance of the industrial process.

Claims

1. A method for fault diagnosis in the polyester esterification stage, characterized in that: The following steps are involved: (1) Collect data during the polyester esterification stage of production line B during the actual production process to obtain fault sample B to be diagnosed; (2) Experts and technicians perform fault diagnosis on the fault sample B to be diagnosed, and divide the fault sample B to be diagnosed into the diagnosed fault sample B1 with known fault type and the fault sample B2 with unknown fault type. All B1 are combined into the support set S * , S * There are C* fault types in it; (3) Using VMD technology to decompose B1 and B2, a multi-scale feature with a scale number of H is obtained; (4) S * The samples with the same fault type are classified into one category, and the fault prototype of the c*th category is calculated. c*=1,2...,C*; Where, Stands for S * There are C* samples of fault types in Stands for S * The number of samples of C* fault types in Stands for S * The D-dimensional feature vector of the sample in, Stands for S * The fault type labels of the samples in Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing; (5) Calculation and Euclidean distance represents the D-dimensional feature vector of B2, Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of trained dimensional variation prototype modules to The metric features obtained by further feature processing; (6) Calculate the probability that the fault type of B2 belongs to the c* class Represents the model's prediction Label, α represents the dimensionality scaling parameter of the dimensional variation prototype module, Represents fault prototypes of other categories except category c*; (7) The fault type corresponding to the maximum probability is taken as the fault type of B2; The multi-scale feature fusion module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, the second pooling layer, the third one-dimensional convolution layer, and the fourth one-dimensional convolution layer connected in sequence; the first one-dimensional convolution layer uses H convolution kernels of the same size to fuse and superimpose multi-scale features of H scales into a single-scale feature map; The dimensional variation prototype module is composed of the first one-dimensional convolution layer, the first pooling layer, the second one-dimensional convolution layer, and the second pooling layer connected in sequence; The training method for the multi-scale feature fusion module and the dimensional variational prototype module is as follows: for the fault samples of production line A, the fault type of the source domain data is first identified through global supervised training. Then, the multi-scale feature fusion module is fixed for task meta-learning training to learn a feature representation so that the same type of faults are closer to different types of faults in the embedding space. Finally, the dimensional variational prototype module is used to adaptively measure the feature similarity of sample pairs, and the metric scaling parameters are automatically learned through the statistical method of variational inference to transform the embedding space. Production lines A and B are two production lines with the same process flow and are operated simultaneously.

2. The method for fault diagnosis in the polyester esterification stage according to claim 1, characterized in that: In step (2), the process of experts and technicians performing fault diagnosis on the fault sample B to be diagnosed is as follows: experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault based on process parameters and fault phenomena, compare them with the known fault database, and give the fault type based on professional knowledge and specific production conditions.

3. The method for fault diagnosis in the polyester esterification stage according to claim 1, characterized in that: In step (3), when using VMD technology to decompose B1, the number of modes is set to 3.

4. The method for fault diagnosis in the polyester esterification stage according to claim 1, characterized in that: The kernel size of the first one-dimensional convolutional layer of the multi-scale feature fusion module is 64×1×3, the stride is 2×1, the number of kernels is 16, and the activation function is ReLU; the kernel size of the first pooling layer is 2×1, the stride is 2×1, the number of kernels is 16, and the activation function is Max; The kernel size in the second one-dimensional convolution layer is 3×1, the stride is 2×1, the number of kernels is 32, and the activation function is ReLU; the kernel size in the second pooling layer is 2×1, the stride is 2×1, the number of kernels is 32, and the activation function is Max; The kernel size in the third one-dimensional convolutional layer is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU; The kernel size in the fourth one-dimensional convolutional layer is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU.

5. The method for fault diagnosis in the polyester esterification stage according to claim 1, characterized in that: The kernel size of the first one-dimensional convolutional layer of the dimensional variational prototype module is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is ReLU; the kernel size of the first pooling layer is 2×1, the stride is 2×1, the number of kernels is 64, and the activation function is Max; The kernel size in the second one-dimensional convolution layer is 2×1, the stride is 2×1, the number of kernels is 128, and the activation function is ReLU; the kernel size in the second pooling layer is 2×1, the stride is 2×1, the number of kernels is 128, and the activation function is Max.

6. The method for fault diagnosis in the polyester esterification stage according to claim 1, characterized in that: The training process of the multi-scale feature fusion module and the dimensional variation prototype module is as follows: (a) During the actual production process, data is collected at the polyester esterification stage of production line A to obtain fault sample A to be diagnosed; (b) Experts and technicians perform fault diagnosis on the fault sample A to be diagnosed and obtain the fault type of the fault sample; (c) Using VMD technology to decompose the training samples, multi-scale features with H scales are obtained; (d) Constructing triple data are all training samples, and The fault type is the same as and The fault types are different; (e) The triple data Input into the multi-scale feature fusion module, which outputs the feature vector (f) Two nonlinear fully connected layers are introduced to reduce the dimension of the feature vector. The simplified feature vector Among them, FC2 represents the operation after two fully connected layers; (g) Calculated (h) Judge d - ≥d + +γ is established, 0<γ<1, if so, the parameters of the multi-scale feature fusion module are fixed and the training of the multi-scale feature fusion module is completed; otherwise, proceed to the next step; (i) Gradient back propagation trains the parameters of the multi-scale feature fusion module until the loss function L g convergence, N s Represents the total number of training samples, n represents the number of training samples, fixes the parameters of the multi-scale feature fusion module, and completes the training of the multi-scale feature fusion module; (j) Part of the training samples are combined into a support set S, and part of the training samples are combined into a query set Q. There are no identical training samples in S and Q. (k) Use S and Q to construct M fault diagnosis tasks. Each fault diagnosis task contains C fault types in S. Each fault type has K training samples in S and L training samples in Q. K<<L. The i-th fault diagnosis task in, Represents the D-dimensional feature vector of the training sample in S, y i,j represents the fault type label of the training sample in S, represents the D-dimensional feature vector of the training samples in Q, represents the fault type label of the training sample in Q, (x i,j ,y i,j ) represents the training sample in S for the i-th fault diagnosis task, represents the training sample in Q for the i-th fault diagnosis task; (l) Reparameterize the dimensionality scaling parameter α of the dimensional variational prototype module, α = μ + εσ, where μ represents the expectation of α, σ represents the expectation of α, and ε represents the expectation of ε. i ~N(0,1) random numbers; (m) Set the initial values of μ and σ; (n) Classify the training samples with the same fault type in S into one category and calculate the fault prototype τ of category c c , c=1,2...,C; Where S C represents the training samples of C fault types in S, |S C |=C×K,f θ (x i,j ) represents the use of the trained multi-scale feature fusion module to i,j The feature vector obtained after feature extraction, h φ (f θ (x i,j )) represents the use of dimensional variational prototype module to f θ (x i,j ) is further processed to obtain the metric features; (o) Calculation With τ c Euclidean distance Represents the use of trained multi-scale feature fusion module to The feature vector obtained after feature extraction is Represents the use of dimensional variational prototype module The metric features obtained by further feature processing; (p) calculation The probability that the fault type belongs to category c τ u Represents fault prototypes of other types except type c; (q) Calculate the total loss L of M tasks p , the formula is as follows: Where, represents the loss of the jth Q in the i-th task; (r) Judge L p Is it converged? If so, fix μ and σ to complete the training of the dimensional variational prototype module; otherwise, proceed to the next step; (s) Update μ and σ according to the following rules and return to step (n); Where, l ψ is the update step size; U represents the number of samples of fault prototypes of other categories except category c; μ0 and σ0 are the initial values of the network input.

7. A method for fault diagnosis in the polyester esterification stage according to claim 6, characterized in that: In step (b), the process of experts and technicians performing fault diagnosis on the fault sample A to be diagnosed is as follows: experts and technicians collect and record key data and information, analyze the characteristics and manifestations of the fault based on process parameters and fault phenomena, compare them with the known fault database, and give the fault type based on professional knowledge and specific production conditions.

8. The method for fault diagnosis in the polyester esterification stage according to claim 6, characterized in that: In step (c), when the VMD technique is used to decompose the training samples, the number of modes is set to 3.

Images