A ceramic quality anomaly detection method for multi-formulation production lines

Abstract

The present invention discloses a ceramic quality anomaly detection method for a multi-recipe production line, comprising: step S1: collecting detection data from each sensor of the production line, and performing alignment and preprocessing; step S2: constructing a feature extraction and classification model based on a deep convolutional network, performing convolution in the time dimension through convolution kernels of different scales, and pre-training the network through supervised learning to achieve mapping of three-dimensional monitoring data tensors to quality indicators; step S3: perturbing the tensor data through a data enhancement method, performing causal inference based on the distribution of the perturbed feature vectors, and judging the causal relationship between different variables and quality characteristics; the present invention aims to provide a ceramic quality anomaly detection method for a multi-recipe production line, by extracting features of profiling tensor data, screening variables that have a causal relationship with quality indicators, and selecting different important variables, so that when the production line recipe changes, the model can perform quality anomaly detection.

Description

Technical Field

[0001] The present invention relates to the technical field of ceramic quality detection, and in particular to a ceramic quality anomaly detection method for a multi-formula production line. Background Art

[0002] Ceramic production is a multi-stage manufacturing process involving a series of complex physical and chemical reactions, including pulverization, molding, and sintering. The production process parameters at each stage will affect product quality. In addition, due to the different characteristics of product categories and raw material origins in different production batches, the ceramic production formula needs to be adjusted accordingly. Therefore, customized product design and construction lead to long ceramic production processes, frequent production fluctuations, unstable products, and difficulty in producing high-quality products. Although there are many model-based or data-driven multi-stage production process anomaly detection solutions, most of these solutions are aimed at relatively fixed production lines with relatively stable production processes. After changes to the production line, the model needs to be adjusted accordingly, which makes it difficult to adapt to the anomaly detection needs in the complex context of ceramic production. Therefore, in order to better monitor product quality and adjust production lines in a timely manner, how to design a ceramic quality anomaly detection method that can be applied to multi-formula production without any changes is an urgent research problem that needs to be solved.

[0003] In recent years, many researchers have proposed big data anomaly detection methods to address this research problem. Deep learning, a data mining process based on neural network models, can bypass complex modeling processes and discover complex nonlinear mapping relationships within the data. At the time, related research focused on exploring the correlation between production line monitoring data and product quality to improve the accuracy of anomaly detection. However, this approach failed to eliminate the influence of irrelevant variables, resulting in overfitting and hindering the discovery of true causal relationships. Furthermore, manual model adjustments were required after adjustments to the production line recipe to achieve optimal results. Summary of the Invention

[0004] The purpose of the present invention is to propose a ceramic quality anomaly detection method for multi-recipe production lines. The method uses a deep convolutional model to perform multi-scale feature extraction of profiling tensor data, screens variables that have a causal relationship with quality indicators through conditional independence tests, and selects important variables that determine product quality under different recipes. When the production line recipe changes, the model can continue to detect quality anomalies without manual adjustment.

[0005] To achieve this purpose, the present invention adopts the following technical solution: a method for detecting abnormal quality of ceramics for a multi-formulation production line, comprising the following steps:

[0006] Step S1: Collect detection data from each sensor on the production line and perform alignment and preprocessing;

[0007] Step S2: Construct a feature extraction and classification model based on a deep convolutional network, perform convolution in the time dimension through convolution kernels of different scales, and pre-train the network through supervised learning to achieve the mapping of three-dimensional monitoring data tensors to quality indicators;

[0008] Step S3: Perturb the tensor data using data augmentation methods, perform causal inference based on the distribution of the perturbed eigenvectors, and determine the causal relationship between different variables and quality characteristics. Based on the strength of the causal relationship, adjust the weights of different eigenvectors and update the network. Repeatedly train the model on the test data of the production line under several recipes to ultimately obtain a neural network model that reflects the true causal relationship.

[0009] Step S3 includes multiple sub-steps:

[0010] Step A1: Perform S data enhancements on each sensor according to the sensor number;

[0011] Step A2: Update all original profiling data after performing causal inference on all sensors using the quality prediction results of the perturbed data using the deep convolutional model;

[0012] Step A3: After updating the profile data, re-optimize the deep convolutional model in step S2 and fine-tune the parameters of the mapping process of the deep convolutional model.

[0013] Preferably, in the step S1, all sensors are numbered, and the time stamp interval when the same tile sample passes through different sensors is determined. Assuming that there are P categories of sensors, each category has Q sensors, and the total number of sensors is recorded as M=PQ, then for the same sample tile, in the order of passing through sensors numbered 1 to M, the time stamp of the sample passing through the first sensor is t1=0, and the time when it passes through the mth sensor is t m , according to the timestamp interval, the data collected when the sample passes each sensor is sampled. Based on the time series data collected by the first sensor, N samples of length T are collected. For the nth sample, it is assumed that the collection time is t n , then the timestamp of the sample segment aligned with the mth sensor should be [τ n +t m , τ n +t m +T],m=1,2,…,M.

[0014] Preferably, in step S2, the network parameters are determined according to the size of the data sample, and the network parameters include the number of network layers and the convolution kernel size, step size, and number of channels of each layer. For the nth sample In the input layer, a convolution with a range of l is performed, and the basic convolution kernel scale is P×Q×l, so that it is along the input tensor Sliding convolution calculation is performed on the time dimension T of

[0015] At the same time, multi-scale convolution kernels are used to extract features, reducing the size of the convolution kernel in the sensor category direction. The convolution kernel needs to slide on the P dimension and then on the T dimension. Let the mapping process of the deep convolution model be F(·). The input sample is expanded into a one-dimensional vector after multiple layers of convolution, and then the corresponding quality index prediction value is calculated through the fully connected classification layer. Compare with the true quality index y(n) corresponding to the sample and substitute it into the formula for calculating the loss function L:

[0016] Where ||·||2 is the 2-norm of the vector. Through the convolutional network, the loss function is minimized to achieve pre-training of the model.

[0017] Preferably, in step A1, specifically, in the q0-th sensor of the p0-th class, data augmentation is performed on the T-dimensional time series, T samples are collected from the standard normal distribution to form a perturbation vector ∈, and the perturbation vector is added to the sensor data for data augmentation, while keeping the scales of the other dimensions unchanged, then:

[0018]

[0019] in Represents the enhanced data, X(p,q,t) represents the coordinates of the data tensor X in the three dimensions of P, Q, and T. After repeated operations, S enhanced data after data enhancement are obtained.

[0020] Preferably, in step A2, the quality prediction result of the data after perturbation by the deep convolution model is Data augmentation through computation And the Hilbert-Schmidt independence of the quality index prediction, quantifying the contribution of the disturbed sensor to the quality index:

[0021]

[0022] Where E represents the expectation of the function, Φ(·), Ψ(·) are Gaussian kernel functions defined on three-dimensional tensor data and mass data, μ y for The expectation of Ψ(y), Represents the tensor product. After performing a data-enhanced causal inference on all M sensors, all the original profiling data are updated.

[0023] Preferably, in step A2, if the Hilbert-Schmidt independence index is higher than the set threshold H0, the contribution value of the q0th sensor of the p0th category to the quality index is small, and the corresponding sensor data is eliminated, that is, the data of the P=p0, Q=q0 dimensions in the original tensor profiling data is set to 0.

[0024] Preferably, in step A3, if the loss function L changes significantly, steps A1 and A2 are repeated to perform a new round of causal reasoning based on data enhancement; if the loss function L does not change significantly, the model training is completed.

[0025] One of the technical solutions of this invention has the beneficial effect of extracting features from production monitoring data across multiple dimensions by constructing a three-dimensional tensor data profile and a multi-scale one-dimensional convolutional neural network. Based on data augmentation methods, multivariate time series causal relationship discovery is established, which allows for screening of effective variables in production monitoring data and enables adjustment-free monitoring of ceramic production anomalies across multiple formulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of a framework of an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation methods.

[0028] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0029] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0030] See Figure 1 As shown, a method for detecting abnormal ceramic quality in a multi-recipe production line includes the following steps:

[0031] Step S1: Collect detection data from each sensor on the production line and perform alignment and preprocessing;

[0032] Step S2: Construct a feature extraction and classification model based on a deep convolutional network, perform convolution in the time dimension through convolution kernels of different scales, and pre-train the network through supervised learning to achieve the mapping of three-dimensional monitoring data tensors to quality indicators;

[0033] Step S3: Perturb the tensor data using data augmentation methods, perform causal inference based on the distribution of the perturbed eigenvectors, and determine the causal relationship between different variables and quality characteristics. Based on the strength of the causal relationship, adjust the weights of different eigenvectors and update the network. Repeatedly train the model on the test data of the production line under several recipes to ultimately obtain a neural network model that reflects the true causal relationship.

[0034] Step S3 includes multiple sub-steps:

[0035] Step A1: Perform S data enhancements on each sensor according to the sensor number;

[0036] Step A2: Update all original profiling data after performing causal inference on all sensors using the quality prediction results of the perturbed data using the deep convolutional model;

[0037] Step A3: After updating the profile data, re-optimize the deep convolutional model in step S2 and fine-tune the parameters of the mapping process of the deep convolutional model.

[0038] Inspection data is collected from various sensors along the production line and aligned and preprocessed. The timestamp interval between each tile sample passing through different sensors is determined. Based on this timestamp interval, the data collected from each sensor is sampled, with the same sampling length. A profile data tensor for each sample is constructed based on sensor type, sensor number, and sampling timestamp. Batch normalization is performed, and the corresponding quality data is used as a label. A feature extraction and classification model based on a deep convolutional network is constructed. Convolution kernels of varying scales are used in the time dimension. The network is pre-trained using supervised learning, initially achieving a mapping from the three-dimensional monitoring data tensor to quality indicators. Data augmentation is used to perturb the tensor data. Causal inference is performed based on the distribution of the perturbed feature vectors to determine whether different variables have a causal relationship with the quality characteristics. The weights of the different feature vectors are adjusted based on the strength of the causal relationship, and the network is updated. This model is repeatedly trained using inspection data from the production line under several different recipes, ultimately resulting in a neural network model that reflects true causal relationships.

[0039] Through multi-scale feature extraction of profiling tensor data using a deep convolutional model, and through conditional independence tests, variables that have a causal relationship with quality indicators are screened. Compared with traditional models, this can better grasp the true causal relationship and select important variables that determine product quality under different formulas. When the production line formula changes, the model can continue to detect quality anomalies without manual adjustment.

[0040] By constructing a three-dimensional tensor data profile and a multi-scale one-dimensional convolutional neural network, we extract features from production monitoring data across multiple dimensions. Based on data augmentation methods, we establish a multivariate time series causal relationship discovery model, screen effective variables in production monitoring data, and implement adjustment-free monitoring of multi-formula ceramic production anomalies.

[0041] Specifically, in the step S1, all sensors are numbered, and the time stamp interval when the same tile sample passes through different sensors is determined. Assuming that there are P categories of sensors, each category has Q sensors, and the total number of sensors is recorded as M=PQ, then for the same sample tile, in the order of passing through sensors numbered 1 to M, the time stamp of the sample passing through the first sensor is t1=0, and the time when it passes through the mth sensor is t m , according to the timestamp interval, the data collected when the sample passes each sensor is sampled. Based on the time series data collected by the first sensor, N samples of length T are collected. For the nth sample, it is assumed that the collection time is t n , then the timestamp of the sample segment aligned with the mth sensor should be [τ n +t m , τ n +t m +T],m=1,2,…,M.

[0042] The monitoring data of this sample, with P sensor categories, Q sensors in each category, and sample timestamp T as indicators, constitutes a profiling data tensor with dimensions of P×Q×T. After collecting all samples, each sensor category is standardized in the Q×T dimension to have a mean of 0 and a variance of 1.

[0043] Preferably, in step S2, the network parameters are determined according to the size of the data sample, and the network parameters include the number of network layers and the convolution kernel size, step size, and number of channels of each layer. For the nth sample In the input layer, a convolution with a range of l is performed, and the basic convolution kernel scale is P×Q×l, so that it is along the input tensor Sliding convolution calculation is performed on the time dimension T of

[0044] At the same time, multi-scale convolution kernels are used to extract features, reducing the size of the convolution kernel in the sensor category direction. The convolution kernel needs to slide on the P dimension and then on the T dimension. Let the mapping process of the deep convolution model be F(·). The input sample is expanded into a one-dimensional vector after multiple layers of convolution, and then the corresponding quality index prediction value is calculated through the fully connected classification layer. Compare with the true quality index y(n) corresponding to the sample and substitute it into the formula for calculating the loss function L:

[0045] Where ‖·‖2 is the 2-norm of the vector. Through the convolutional network, the loss function is minimized to achieve pre-training of the model.

[0046] Specifically, in step A1, data augmentation is performed on the T-dimensional time series in the q0-th sensor of the p0-th class. T samples are collected from the standard normal distribution to form a perturbation vector ∈. The perturbation vector is added to the sensor data for data augmentation, while the scales of the other dimensions remain unchanged. Then:

[0047]

[0048] in Represents the enhanced data, X(p,q,t) represents the coordinates of the data tensor X in the three dimensions of P, Q, and T. After repeated operations, S enhanced data after data enhancement are obtained.

[0049] Preferably, in step A2, the quality prediction result of the data after perturbation by the deep convolution model is Data augmentation through computation And the Hilbert-Schmidt independence of the quality index prediction, quantifying the contribution of the disturbed sensor to the quality index:

[0050]

[0051] Where E represents the expectation of the function, Φ(·), Ψ(·) are Gaussian kernel functions defined on three-dimensional tensor data and mass data, μ y for The expectation of Ψ(y), Represents the tensor product. After performing a data-enhanced causal inference on all M sensors, all the original profiling data are updated.

[0052] Specifically, in step A2, if the Hilbert-Schmidt independence index is higher than the set threshold H0, the contribution value of the q0th sensor of the p0th category to the quality index is small, and the corresponding sensor data is eliminated, that is, the data of the P=p0, Q=q0 dimension is set to 0 in the original tensor profiling data.

[0053] Preferably, in step A3, if the loss function L changes significantly, steps A1 and A2 are repeated to perform a new round of causal reasoning based on data enhancement; if the loss function L does not change significantly, the model training is completed.

[0054] Throughout this specification, reference to terms such as "embodiment" or "example" indicates that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0055] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are intended solely to illustrate the principles of the present invention and are not to be construed in any way as limiting the scope of protection of the present invention. Based on the explanations herein, those skilled in the art will readily conceive of other specific embodiments of the present invention without inventive effort, and such embodiments will fall within the scope of protection of the present invention.

Claims

1. A method for detecting abnormal ceramic quality in a multi-formulation production line, characterized in that: The steps include: Step S1: Collect detection data from each sensor on the production line and perform alignment and preprocessing; Step S2: Construct a feature extraction and classification model based on a deep convolutional network, perform convolution in the time dimension through convolution kernels of different scales, and pre-train the network through supervised learning to achieve the mapping of three-dimensional monitoring data tensors to quality indicators; Step S3: Perturb the tensor data using data augmentation methods, perform causal inference based on the distribution of the perturbed eigenvectors, and determine the causal relationship between different variables and quality characteristics. Based on the strength of the causal relationship, adjust the weights of different eigenvectors and update the network. Repeatedly train the model on the test data of the production line under several recipes to ultimately obtain a neural network model that reflects the true causal relationship. Step S3 includes multiple sub-steps: Step A1: Perform S data enhancements on each sensor according to the sensor number; Step A2: Update all original profiling data after performing causal inference on all sensors using the quality prediction results of the perturbed data using the deep convolutional model; Step A3: After updating the profile data, re-optimize the deep convolutional model in step S2 and fine-tune the parameters of the mapping process of the deep convolutional model.

2. The ceramic quality abnormality detection method for a multi-formulation production line according to claim 1, characterized in that: In the step S1, all sensors are numbered and the time stamp intervals when the same tile sample passes through different sensors are determined. Assuming that there are P categories of sensors, each category has Q sensors, and the total number of sensors is recorded as M=PQ, then for the same sample tile, in the order of passing through sensors numbered 1 to M, the time stamp of the sample passing through the first sensor is =0, the time when it passes the mth sensor is , according to the timestamp interval, the data collected when the sample passes each sensor is sampled. Based on the time series data collected by the first sensor, N samples of length T are collected. For the nth sample, assuming that the collection time is , then the timestamp of the sample segment aligned with the mth sensor should be ,m=1,2,…,M.

3. The method for detecting abnormal ceramic quality in a multi-formulation production line according to claim 1, characterized in that: In the step S2, the network parameters are determined according to the size of the data sample. The network parameters include the number of network layers and the convolution kernel size, step size, and number of channels of each layer. For the nth sample , in the input layer, a convolution with a range of l is performed, then the basic convolution kernel scale is P×Q×l, so that it is along the input tensor Sliding convolution calculation is performed on the time dimension T of At the same time, multi-scale convolution kernels are used to extract features, reducing the size of the convolution kernel in the sensor category direction. The convolution kernel needs to slide on the P dimension and then on the T dimension. Let the mapping process of the deep convolution model be F(·). The input sample is expanded into a one-dimensional vector after multiple layers of convolution, and then the corresponding quality index prediction value is calculated through the fully connected classification layer. , compared with the true quality index y(n) corresponding to the sample, substitute it into the formula for calculating the loss function L: ; in is the 2-norm of the vector, and the loss function is minimized through the convolutional network to achieve pre-training of the model.

4. The method for detecting abnormal ceramic quality in a multi-formulation production line according to claim 1, characterized in that: In the step A1, specifically, Class For each sensor, perform data augmentation on the T-dimensional time series. Collect T samples from the standard normal distribution to form a perturbation vector ϵ. Add the perturbation vector to the sensor data for data augmentation, while keeping the scales of the other dimensions unchanged. Then: ； in Represents the enhanced data, X(p,q,t) represents the coordinates of the data tensor X in the three dimensions of P, Q, and T. After repeated operations, S enhanced data are obtained.

5. The ceramic quality abnormality detection method for a multi-formulation production line according to claim 4, characterized in that: In step A2, the quality prediction result of the data after perturbation by the deep convolution model is ,S=1,2,…,S, by computing enhanced data And the Hilbert-Schmidt independence of the quality index prediction, quantifying the contribution of the disturbed sensor to the quality index: ； Where E represents the expectation of the function, Φ(·), Ψ(·) are Gaussian kernel functions defined on three-dimensional tensor data and mass data, , for , , ⊗ represents the tensor product. After performing a data-enhanced causal inference on all M sensors, all the original profiling data are updated.

6. The method for detecting abnormal ceramic quality in a multi-formulation production line according to claim 5, characterized in that: In step A2, if the Hilbert-Schmidt independence index is higher than the set threshold , then Class The contribution of each sensor to the quality index is small, and the corresponding sensor data is removed, that is, the original tensor profiling data is , The data of the dimension is set to 0.

7. The method for detecting abnormal ceramic quality in a multi-formulation production line according to claim 1, characterized in that: In step A3, if the loss function L changes significantly, steps A1 and A2 are repeated to perform a new round of causal reasoning based on data enhancement; if the loss function L does not change significantly, the model training is completed.

Images