Industrial process quality abnormity diagnosis method and system based on time sequence convolutional network

Abstract

The invention discloses an industrial process quality abnormity diagnosis method based on a time sequence convolutional network, and the method comprises the steps: collecting original multivariable data of an industrial production process, removing missing values and abnormal values, and decomposing the features of the original multivariable data through a variational mode decomposition method; thirdly, constructing a time sequence convolutional neural network model; then, inputting the structured data into the constructed time sequence convolutional neural network model, and adjusting and optimizing deep learning model parameters by using a cross entropy loss function; and finally, according to the trained time sequence convolutional neural network model, carrying out quality monitoring and diagnosis on the industrial multivariable data generated in real time. According to the method, the accuracy of quality monitoring and diagnosis in the industrial production process can be remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial production process quality anomaly diagnosis, and in particular, relates to an industrial process quality anomaly diagnosis method and system based on a temporal convolutional network. Background Art

[0002] With the advent of a new round of information, data, and industrial revolutions, manufacturing industries such as quantum information, integrated circuits, aerospace technology, high-end equipment, new energy vehicles, and marine equipment are booming. The product processing procedures of emerging manufacturing industries are usually complex and diverse, and the production process standards are more stringent. The product quality level deeply affects the reputation and revenue profits of enterprises. In order to ensure product quality, enterprises install a large number of sensors and wireless radio frequency devices in the production line, such as temperature sensors, quality sensors, pressure sensors, light controllers, etc., to achieve quality monitoring and diagnosis of the manufacturing process; these sensors can collect raw data reflecting the quality status of the production process in real time. However, due to the complexity of production and manufacturing and the different types of sensors, as well as the interference of noise, light, vibration, temperature, humidity, etc. during the production process, the data obtained by the sensor usually has the characteristics of high noise, large volume, time series, and multi-modality. Therefore, how to eliminate interference in data collection, storage, and transmission, and construct an effective real-time quality control model to achieve quality diagnosis of complex multivariable production processes has become the key to ensuring product quality and improving economic benefits.

[0003] At present, the research on the abnormal quality diagnosis system of the production process mainly focuses on the preprocessing of raw data and the identification method of abnormal quality. The original industrial manufacturing data processing methods are mostly concentrated on principal component analysis, wavelet transform, empirical mode decomposition, etc. When REYHANE et al. monitored and analyzed the multivariate production process of complex systems, they first used the expert decision-making method to analyze and preprocess the multivariate data in the complex production process, and then used the principal component analysis model to identify and monitor the two types of data, normal operation and safe operation, respectively, thereby improving the quality monitoring accuracy of data beyond the process limit. In order to solve the problems of multiple correlations between quality characteristics and high-dimensional and high-noise data in the identification of dynamic process quality characteristics, Liu Yumin et al. used the dynamic process LSSVM-BPNN model based on wavelet reconstruction to realize the identification of dynamic process quality characteristics. Shi Shuangping et al. combined the empirical mode decomposition and kurtosis analysis methods to establish a model to suppress the interference of process data so as to more effectively extract the fault characteristic signals of bearings.

[0004] In recent years, in the field of quality anomaly diagnosis, deep learning methods have become a powerful tool for production process quality monitoring because they use deep nonlinear network structures to approximate complex functions, making them powerful in extracting the essential characteristics of input data. Production process quality monitoring technology has transitioned from shallow learning methods such as control chart methods based on statistical techniques and support vector machines (SVM) to deep learning methods such as deep belief networks and convolutional neural networks. For systems with process quality variance offsets due to random failures, Chen Honggen et al. proposed a method of using EWMA control charts to monitor process variances, and combined economic decision-making to build a preventive maintenance and process quality integrated decision model, reducing the complexity of average run length calculations, and obtaining a better economic benefit decision-making method. Wang Ning et al. used partial least squares to improve the adaptive elastic network to solve the problem of identifying key quality characteristics in complex multi-process manufacturing processes, and the monitoring of process quality was more complete and accurate. Aiming at the problem that the quality inspection process of automobile combination instrument assembly process is time-consuming and the production efficiency is low, He Yan et al. proposed a CNN-SVR model based on deep learning convolutional neural network and support vector regression to predict the assembly quality of automobile combination instrument. The model has the characteristics of small prediction error and high prediction accuracy. Saqlain et al. extracted density features, geometric features and radon-based material features from the original wafer image, and compared and applied four machine learning classifiers: Logistic regression, random forest (RF), gradient enhancement machine and artificial neural network (ANN), trained and identified the extracted original data set features, and finally achieved high accuracy and precision.

[0005] Although the above methods have been widely used, they still have certain limitations. In terms of raw data processing, statistical analysis methods such as probability partial least squares method and principal component analysis method are limited by the fact that the statistics obey a specific distribution and the random fluctuation of residual variables. Wavelet transform method and empirical mode decomposition method overcome the limitations of statistical methods, but they have the defects of frequency overlap and mode mixing. In terms of using deep learning for quality anomaly diagnosis, machine learning methods such as convolutional neural networks and deep belief networks are limited by the size of the convolution kernel receptive field when extracting data features, and usually do not have long time series memory, so the effect of classifying long-term and large amounts of data is unsatisfactory. In addition, when constructing deep neural networks for machine learning, the theoretical community regards them as black box models, and it is difficult to accurately explain the operating mechanism of internal learning. In summary, quality diagnosis of complex multivariate data generated in the real-time production process of industrial manufacturing enterprises still faces problems such as difficulty in constructing statistical control variables, insufficient long-term memory of neural network learning, and poor interpretability of machine learning black box models. Summary of the invention

[0006] In order to solve the above problems, the purpose of the present invention is to provide an industrial process quality abnormality diagnosis method and system based on a temporal convolutional network.

[0007] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme:

[0008] A method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network comprises the following steps:

[0009] S1. Data preprocessing: Use sensor equipment to collect original multivariate data of industrial production processes; use the 6 sigma principle to eliminate missing values ​​and outliers; and then use the variational mode decomposition method to decompose the characteristics of the original multivariate data;

[0010] S2. Build a time series convolutional neural network model: add a causal dilation convolutional network layer, a skip connection network layer, a normalized neural network layer, and an activated neural network layer to build a deep learning neural network structure that adapts to the data structure;

[0011] S3, model training: input the structured data in step S1 into the temporal convolutional neural network model constructed in step S2, and use the cross entropy loss function to adjust and optimize the deep learning model parameters;

[0012] S4. Quality abnormality diagnosis: Based on the time series convolutional neural network model trained in step S3, quality monitoring and diagnosis are performed on the industrial multivariate data generated in real time.

[0013] Furthermore, the above step S1 includes the following sub-steps:

[0014] S1.1. Use a large number of sensor devices installed in industrial production processes to collect multivariate quality data reflecting production status;

[0015] S1.2. Use the 6 sigma principle to remove missing values ​​and extreme outliers from the original data using the multivariate quality data in step S1.1. The calculation formula is:

[0016]

[0017] Among them, σ represents the standard deviation; L represents the number of data; X i represents each data point; μ represents the mean value of the data;

[0018] After being processed by the 6 sigma principle, the data collected by N sensors within T time is converted into an N × T multivariate production process data matrix X;

[0019]

[0020] Among them, x p,q Represents the data of the p-th sensor at the q-th time point;

[0021] S1.3. Decompose the acquired high-noise production process data stream into K discrete intrinsic mode functions using the variational mode decomposition method And the corresponding center frequency is ω k ,(k=1,2,...,K); where u k (t) = A k (t)cos(Φ k (t)), A k (t) is the amplitude, φ k (t) is the phase, u k (t) is the original data decomposed into the kth eigenmode function; K is the number of eigenmode functions; the operation is:

[0022] First, find K intrinsic mode functions u k (t), (k = 1, 2, ..., K), so that the sum of the estimated bandwidths of all eigenmode functions is minimized and the constraint that the sum of each mode is equal to the data matrix X is satisfied, that is,

[0023]

[0024] To u k (t) Perform Hilbert transform to obtain the analytical expression, that is,

[0025]

[0026] Where δ(·) is the unit impulse function; j is the imaginary unit; * is the convolution operation;

[0027] Secondly, the squared norm of the gradient is used to estimate the eigenfunction u k The bandwidth of (t) is expressed as:

[0028]

[0029] in, is the gradient operation, is the imaginary transformation;

[0030] Obtain the variational mode decomposition constrained model:

[0031]

[0032] Among them, {u k}={u1,...,u K} and {ω k}={ω1,...,ω K} are the abbreviations of each mode and its center frequency set; is understood as the set of all modes; u k With ω k are the center frequencies of the kth modal component and its corresponding modal component, for a total of K modes;

[0033] Finally, the penalty factor α and the Lagrange multiplier λ are introduced, and the resulting augmented Lagrange expression is:

[0034]

[0035] The saddle point of equation (7) is solved by iteratively updating the alternating direction multiplier algorithm to obtain the optimal u k ,ω k And λ, the optimal solution of variational mode decomposition is obtained as:

[0036]

[0037] in, is the Fourier transform of the modal component; is the Fourier transform of the input signal; ω is the modal frequency.

[0038] Furthermore, in the above step S2, the temporal convolutional network includes an input layer, a dilated causal convolutional layer, a weight normalization layer, an activation function Relu and Dropout layer, and an output layer;

[0039] (1) In the dilated causal convolutional layer, for the batch input model data U∈R n , the convolution kernel f:{0,…,h-1}→R, the specific calculation of the dilated causal convolution operation is:

[0040]

[0041] Where d represents the dilation factor; h represents the size of the convolution kernel; U s-d.j Indicates the element in the input sequence that corresponds to the convolution kernel during dilated convolution;

[0042] (2) In the weight normalization layer, the weight matrix of the initialized network model is set to W, and the network training is optimized through the loss function and gradient descent method; the operation is:

[0043] First, the weight matrix W is decoupled in length g and direction v. The calculation formula is:

[0044]

[0045] Secondly, using the loss function based on cross entropy, when the loss function is minimized, the network model reaches the optimal value. The expression of the loss function L is as follows:

[0046]

[0047] Among them, S i is the probability distribution of the real data; F i is the probability distribution of the model predicting data; i is the number of data;

[0048] The gradient of the loss function L with respect to g and v is expressed as:

[0049]

[0050] Finally, through the gradient descent method, new gradients are continuously obtained to reversely determine the length g and direction v, which are passed to the neural network weight matrix W;

[0051] (3) In the activation network layer, the input X and the dilated causal convolution layer F(s) are summed and passed through the activation function to obtain the output Y. That is, the activation relationship between the input X and the output Y is expressed as:

[0052] Y = Relu(X + F(s)) (14)

[0053] Among them, Relu(·) is the activation function;

[0054] (4) The regularized network layer is constructed through the Dropout network layer. The N monitoring variables are diagnosed simultaneously by inputting the multivariate data matrix X, and the output is Y, Y = (y1, y2, ..., y T ), diagnose the quality status of industrial production process products within T time.

[0055] Furthermore, the above step S3 includes the following sub-steps:

[0056] S3.1, the multivariate data of the training samples are processed sequentially through the dilated causal convolution layer, the weight normalization layer, the random dropout layer and the activation layer to obtain the data quality status;

[0057] S3.2, calculating the cross entropy loss function value between the actual quality state and the multivariate data quality state identified by the training model in step S3.1, and reversely adjusting and optimizing the weight matrix of the temporal convolutional neural network model through the loss function value;

[0058] S3.3, re-input the training sample multivariate data into the optimized model in step S3.2 and continue iterating;

[0059] S3.4. Determine the parameter range of the temporal convolutional neural network model through multiple experiments.

[0060] Furthermore, in the above step S4, the operation is as follows: the real-time multivariate data stream is input into the time series convolutional network model in batches through the sliding window value extraction technology, and the quality of the products corresponding to the data is monitored and diagnosed; when the multivariate data meets the normal quality characteristics, it slides to the data of the next monitoring window in turn; when the value in the sliding window is abnormal, the time series convolutional network model identifies the abnormal quality characteristics, issues an alarm warning for the products produced, and production is stopped immediately.

[0061] An industrial process quality anomaly diagnosis system based on a temporal convolutional network includes a data acquisition module, a data preprocessing module, a model building module, a model training module, and a quality anomaly diagnosis module, wherein:

[0062] The data acquisition module is used to collect original multivariate data of the industrial production process;

[0063] A data preprocessing module is used to remove missing values ​​and abnormal values ​​from the original multivariate data acquired by the data acquisition module using the 6 sigma principle; and then decompose the characteristics of the original multivariate data using the variational mode decomposition method;

[0064] The model building module is used to add an extended causal convolution layer to keep the sequence relationship of the data unchanged while expanding the receptive field and extracting more features; construct a weight normalization layer so that the neural network can update the weight parameters in reverse in a timely manner; and use the data output by the activation layer neural network to analyze the quality status of the current process;

[0065] The model training module is used to input the structured data of the data preprocessing module into the temporal convolutional neural network model, and adjust and optimize the deep learning model parameters using the cross entropy loss function;

[0066] The quality anomaly diagnosis module is used to perform quality monitoring and diagnosis on industrial multivariate data generated in real time based on the trained temporal convolutional neural network model.

[0067] Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages:

[0068] The invention discloses an industrial process quality abnormality diagnosis method and system based on a time series convolutional network. The method collects and acquires data through intelligent sensors, unifies the data form and structure, facilitates subsequent data processing, and improves data reliability. The time series convolutional network model based on machine learning is used to monitor multivariate process data, which greatly improves automation and intelligence. By constructing a multi-layer network structure, deeper and more abstract data features can be extracted, the monitoring accuracy is higher, and the diagnosis effect is more in line with actual production. When constructing the network structure, dilated convolution is added, which greatly enhances the receptive field of a single time series convolution kernel, improves the global feature extraction capability of industrial big data, and avoids the deficiency of focusing only on local features. The use of time series causal convolution keeps the input and output synchronized, maintains the time attribute of the data stream, avoids data confusion and leakage, improves the accuracy of model diagnosis, and increases the interpretability of the model, further solving the black box problem faced in machine learning. The invention effectively solves the complex monitoring problem of multivariate real-time data in the industrial production process, and compared with the statistical quality control in the prior art and other recurrent neural network quality diagnosis, the accuracy of industrial production process quality monitoring and diagnosis is significantly improved, and the production efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the industrial process quality abnormality diagnosis method based on the temporal convolutional network of the present invention;

[0070] Figure 2 It is a variational mode decomposition flow chart in the industrial process quality abnormality diagnosis method based on the temporal convolutional network of the present invention;

[0071] Figure 3 It is a structural diagram of an embodiment of a temporal convolutional neural network model in an industrial process quality abnormality diagnosis method based on a temporal convolutional network of the present invention;

[0072] Figure 4 It is a structural diagram of the industrial process quality anomaly diagnosis system based on the temporal convolutional network of the present invention. DETAILED DESCRIPTION

[0073] The technical solution of the present invention is further described in detail below through the drawings and embodiments.

[0074] Example 1

[0075] like Figure 1 As shown, a method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network comprises the following steps:

[0076] S1. Data preprocessing: Use a large number of sensors installed in the production line to collect the original multivariate data of the industrial production process; use the 6 sigma principle to eliminate missing values ​​and outliers; then use the variational mode decomposition method to decompose the characteristics of the original multivariate data, unify the data structure, visualize the normal and abnormal characteristics of quality, and improve the reliability of multivariate data; including the following sub-steps:

[0077] S1.1. Use a large number of sensor devices installed in industrial production processes to collect multivariate quality data reflecting production status;

[0078] S1.2. Use the 6 sigma principle to remove missing values ​​and extreme outliers from the original data using the multivariate quality data in step S1.1. The calculation formula is:

[0079]

[0080] Among them, σ represents the standard deviation; L represents the number of data; X i represents each data point; μ represents the mean value of the data;

[0081] When the target value is consistent with the expectation and the original abnormal data is removed, most of the product values ​​fall within the range of [T-6σ, T+6σ) of the target value T, and the defective rate per million reaches 3.4, which has extremely high accuracy.

[0082] After being processed by the 6 sigma principle, the data collected by N sensors within T time is converted into an N × T multivariate production process data matrix X;

[0083]

[0084] Among them, x p,q Represents the data of the p-th sensor at the q-th time point;

[0085] S1.3. Using the variational modal decomposition (VMD) method, the acquired high-noise production process data stream is decomposed into K discrete intrinsic mode functions (k=1, 2, ..., K), and the corresponding center frequency is ω k ,(k=1,2,...,K); where u k (t) = A k (t)cos(Φ k (t)), A k (t) is the amplitude, Φ k (t) is the phase, u k (t) is the original data decomposed into the kth eigenmode function; K is the number of eigenmode functions; the operation is:

[0086] First, the bandwidth of the intrinsic mode function is assumed to be limited, and the quality information of the high-noise production process is recorded as f(t). The variational problem of the quality mode of the high-noise production process is: Find K intrinsic mode functions u k (t), (k = 1, 2, ..., K), so that the sum of the estimated bandwidths of all eigenmode functions is minimized and the constraint that the sum of each mode is equal to the data matrix X is satisfied, that is,

[0087]

[0088] To estimate the data spectrum bandwidth and determine the corresponding center frequency, the intrinsic mode function u is obtained. k (t), for u k (t) Perform Hilbert transform to obtain the analytical expression, that is,

[0089]

[0090] Where δ(·) is the unit impulse function; j is the imaginary unit; * is the convolution operation;

[0091] Secondly, the squared norm of the gradient is used to estimate the eigenfunction u k The bandwidth of (t) is expressed as:

[0092]

[0093] in, is the gradient operation, is the imaginary transformation;

[0094] In the VMD method, the mode u k Closely around the center frequency ω k And choose the bandwidth of the spectrum as each mode u k The sparse characteristics of the center frequency ω k Determined together with the decomposition, in order to evaluate the bandwidth of each mode, the relevant analytical signal is calculated by Hilbert transform, and the Gaussian smoothness of the analytical signal, that is, the square norm of the gradient, is used to estimate the bandwidth, thereby obtaining the variational mode decomposition constraint model:

[0095]

[0096] Among them, {u k}={u1,...,u K} and {ω k}={ω1,...,ω K} are the abbreviations of each mode and its center frequency set; is understood as the set of all modes; u k With ωk are the center frequencies of the kth modal component and its corresponding modal component, for a total of K modes;

[0097] Finally, in order to obtain the optimal solution of the variational mode decomposition model of formula (6), the penalty factor α and the Lagrange multiplier λ are introduced, and the obtained augmented Lagrange expression is:

[0098]

[0099] The saddle point of the above equation is solved by iterative update using the Alternate Direction Method of Multiplers (ADMM) algorithm to obtain the optimal u k ,ω k And λ, so the optimal solution of variational mode decomposition is:

[0100]

[0101] in, is the Fourier transform of the modal component; is the Fourier transform of the input signal; ω is the modal frequency, such as Figure 2 As shown;

[0102] S2. Construct a temporal convolutional neural network model: add a causal dilated convolutional network layer, a skip connection network layer, a normalized neural network layer, and an activated neural network layer to construct a deep learning neural network structure that adapts to the data structure; the temporal convolutional network includes an input layer, a dilated causal convolutional layer, a weight normalization layer (WeightNorm), an activation function Relu and a Dropout layer, and an output layer;

[0103] (1) In the dilated causal convolutional layer, for the batch input model data U∈R n , the convolution kernel f:{0,…,h-1}→R, the specific calculation of the dilated causal convolution operation is:

[0104]

[0105] Where d represents the dilation factor; h represents the size of the convolution kernel; U s-d.j Indicates the element in the input sequence that corresponds to the convolution kernel during dilated convolution;

[0106] (2) In the weight normalization layer, the weight matrix of the initialized network model is assumed to be W. The process of optimizing the network training is performed through the loss function and the gradient descent method, that is, the process of solving the optimal W. The operation is: first, the weight matrix W is decoupled in length g and direction v. The calculation formula is:

[0107]

[0108] The temporal convolutional neural network uses a loss function based on cross entropy. When the loss function is minimized, the network model reaches the optimal value. The expression of the loss function L is as follows:

[0109]

[0110] Among them, S i is the probability distribution of the real data; F i is the probability distribution of the model predicting data; i is the number of data;

[0111] The gradient of the loss function L with respect to g and v is expressed as:

[0112]

[0113] Through the gradient descent method, new gradients are continuously obtained to reversely determine the length g and direction v, which are passed to the neural network weight matrix W, making the temporal convolutional neural network model achieve better performance;

[0114] (3) In the activation network layer, the input X and the dilated causal convolution layer F(s) are summed and passed through the activation function to obtain the output Y. That is, the activation relationship between the input X and the output Y is expressed as:

[0115] Y = Relu(X + F(s)) (14)

[0116] Among them, Relu(·) is the activation function;

[0117] (4) By adding the Dropout network layer, a regularized network layer is constructed, and the multivariate data matrix X is input.

[0118]

[0119] Diagnose N monitoring variables simultaneously, and the output is Y, Y=(y1,y2,...,y T ), thereby diagnosing the quality status of the products in the industrial production process within time T, such as Figure 3 As shown;

[0120] S3, model training: input the multivariate data matrix in step S1 into the time series convolutional neural network model constructed in step S2 for iterative learning and training, adjust the deep learning model parameters, extract the normal and abnormal quality features in the industrial multivariate data, so that the model can efficiently and accurately process the multivariate industrial big data; the operation is:

[0121] S3.1, the multivariate data of the training samples are processed sequentially through the dilated causal convolution layer, the weight normalization layer, the random dropout layer and the activation layer to obtain the data quality status;

[0122] S3.2, calculating the cross entropy loss function value between the actual quality state and the multivariate data quality state identified by the training model in step S3.1, and reversely adjusting and optimizing the weight matrix of the temporal convolutional neural network model through the loss function value;

[0123] S3.3, re-input the training sample multivariate data into the optimized model in step S3.2 and continue iterating;

[0124] S3.4. Determine the parameter range of the temporal convolutional neural network model through multiple experiments;

[0125] Due to the different specific collected data, the feature complexity contained in the data is different. By giving a series of parameter ranges of the model, the robustness of the model is increased, and the optimal parameters are determined by specific debugging of the collected data to enhance the matching degree between the data and the model; preferably, the number of convolution kernels is in the range of [16 512], that is, 16, 32, 64, 128, 256, 512; the range of causal convolution kernels is 1×1, 1×3, 1×5 or 1×7; the expansion factor is in the range of [2 16], that is, 2, 4, 8, 16; the number of iterations is 30 to 10,000 times; the droupout regularization technique is used in the training process to randomly mask and inactivate a certain proportion of neurons to reduce the dependence on specific neurons, and the random inactivation rate is 0 to 0.3; the training and optimal parameter range of the temporal convolutional network model make the model robust, can achieve better monitoring effects and effectively control the complexity of the model;

[0126] Through the model parameter range given in step 3, the temporal convolutional network model parameters that meet specific data tasks can be debugged faster and more accurately for specific data;

[0127] S4. Quality anomaly diagnosis: A large number of sensors installed in the production line collect and generate multivariate data streams representing the quality status of the product in real time, and use the sliding window value acquisition technology to obtain quantitative data in sequence, and input the real-time multivariate data streams into the time series convolutional network model in batches; when the multivariate data meets the normal quality characteristics, it slides to the data of the next monitoring window in sequence; when the value in the sliding window is abnormal, the time series convolutional network model identifies the quality anomaly characteristics, issues an alarm warning for the products produced, and production is stopped immediately; the staff inspects the equipment, parts, raw materials, etc., finds out the cause of the failure, eliminates the factors that lead to quality anomalies, continues to produce products that meet quality standards, realizes automatic inspection of the product production process, and achieves the goal of safe production and improving the company's operating efficiency.

[0128] Example 2

[0129] The present invention discloses an industrial process quality abnormality diagnosis system based on a time series convolutional network, which is used to implement the industrial process quality abnormality diagnosis method based on a time series convolutional network described in Example 1, such as Figure 4 As shown, the system includes the following functional modules:

[0130] The data acquisition module is used to collect original multivariate data of the industrial production process;

[0131] The data preprocessing module is used to remove missing values ​​and abnormal values ​​from the original multivariate data acquired by the data acquisition module using the 6 sigma principle; and then decompose the characteristics of the original multivariate data through the variational mode decomposition method, unify the data structure, and visualize the normal and abnormal quality characteristics;

[0132] The model building module is used to add an extended causal convolution layer to keep the sequence relationship of the data unchanged while expanding the receptive field (RF) to extract more features; construct a weight normalization layer to enable the neural network to update the weight parameters in reverse in a timely manner; and use the data output by the activation layer neural network to analyze the quality status of the current process;

[0133] The model training module is used to input the structured data of the data preprocessing module into the time series convolutional neural network model, adjust the deep learning model parameters, and extract normal and abnormal quality features in the industrial multivariate data;

[0134] The quality anomaly diagnosis module is used to monitor and diagnose the quality of industrial multivariate data generated in real time based on the trained time series convolutional neural network model, and identify abnormal products in the production process.

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network, characterized by: It includes the following steps: S1. Data preprocessing: using sensor equipment to collect raw multivariate data of industrial production processes; Use the 6 sigma principle to eliminate missing values ​​and outliers; Then, the characteristics of the original multivariate data are decomposed by using the variational mode decomposition method; S2. Build a time series convolutional neural network model: add a causal dilation convolutional network layer, a skip connection network layer, a normalized neural network layer, and an activated neural network layer to build a deep learning neural network structure that adapts to the data structure; S3, model training: input the structured data in step S1 into the temporal convolutional neural network model constructed in step S2, and use the cross entropy loss function to adjust and optimize the deep learning model parameters; S4. Quality abnormality diagnosis: Based on the time series convolutional neural network model trained in step S3, quality monitoring and diagnosis are performed on the industrial multivariate data generated in real time.

2. The method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network according to claim 1 is characterized in that: The step S1 comprises the following sub-steps: S1.

1. Use a large number of sensor devices installed in industrial production processes to collect multivariate quality data reflecting production status; S1.

2. Use the 6 sigma principle to remove missing values ​​and extreme outliers from the original data using the multivariate quality data in step S1.

1. The calculation formula is: Among them, σ represents the standard deviation; L represents the number of data; X i represents each data point; μ represents the mean value of the data; After being processed by the 6 sigma principle, the data collected by N sensors within T time is converted into an N × T multivariate production process data matrix X; Among them, x p,q Represents the data of the p-th sensor at the q-th time point; S1.

3. Decompose the acquired high-noise production process data stream into K discrete intrinsic mode functions using the variational mode decomposition method And the corresponding center frequency is ω k ,(k=1,2,...,K); where u k (t) = A k (t)cos(Φ k (t)), A k (t) is the amplitude, Φ k (t) is the phase, u k (t) is the original data decomposed into the kth eigenmode function; K is the number of eigenmode functions; the operation is: First, find K intrinsic mode functions u k (t), (k = 1, 2, ..., K), so that the sum of the estimated bandwidths of all eigenmode functions is minimized and the constraint that the sum of each mode is equal to the data matrix X is satisfied, that is, To u k (t) Perform Hilbert transform to obtain the analytical expression, that is, Where δ(·) is the unit impulse function; j is the imaginary unit; * is the convolution operation; Secondly, the squared norm of the gradient is used to estimate the eigenfunction u k The bandwidth of (t) is expressed as: in, is the gradient operation, is the imaginary transformation; Obtain the variational mode decomposition constrained model: Among them, {u k }={u1,...,u K } and {ω k }={ω1,...,ω K } are the abbreviations of each mode and its center frequency set; is understood as the set of all modes; u k With ω k are the center frequencies of the kth modal component and its corresponding modal component, for a total of K modes; Finally, the penalty factor α and the Lagrange multiplier λ are introduced, and the resulting augmented Lagrange expression is: The saddle point of equation (7) is solved by iteratively updating the alternating direction multiplier algorithm to obtain the optimal u k ,ω k And λ, the optimal solution of variational mode decomposition is obtained as: in, is the Fourier transform of the modal component; is the Fourier transform of the input signal; ω is the modal frequency.

3. The method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network according to claim 1 is characterized in that: The temporal convolutional network in step S2 includes an input layer, an expanded causal convolutional layer, a weight normalization layer, an activation function Relu and Dropout layer, and an output layer; (1) In the dilated causal convolutional layer, for the batch input model data U∈R n , the convolution kernel f:{0,...,h-1}→R, the specific calculation of the dilated causal convolution operation is: Where d represents the dilation factor; h represents the size of the convolution kernel; U s-d.j Indicates the element in the input sequence that corresponds to the convolution kernel during dilated convolution; (2) In the weight normalization layer, the weight matrix of the initialized network model is set to W, and the network training is optimized through the loss function and gradient descent method; the operation is: First, the weight matrix W is decoupled in length g and direction v. The calculation formula is: Secondly, using the loss function based on cross entropy, when the loss function is minimized, the network model reaches the optimal value. The expression of the loss function L is as follows: Among them, S i is the probability distribution of the real data; F i is the probability distribution of the model predicting data; i is the number of data; The gradient of the loss function L with respect to g and v is expressed as: Finally, through the gradient descent method, new gradients are continuously obtained to reversely determine the length g and direction v, which are passed to the neural network weight matrix W; (3) In the activation network layer, the input X and the dilated causal convolution layer F(s) are summed and passed through the activation function to obtain the output Y. That is, the activation relationship between the input X and the output Y is expressed as: Y = Relu(X + F(s)) (14) Among them, Relu(·) is the activation function; (4) The regularized network layer is constructed through the Dropout network layer. The N monitoring variables are diagnosed simultaneously by inputting the multivariate data matrix X, and the output is Y, Y = (y1, y2, ..., y T ), diagnose the quality status of industrial production process products within T time.

4. The method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network according to claim 1 is characterized in that: The step S3 comprises the following sub-steps: S3.1, the multivariate data of the training samples are processed sequentially through the dilated causal convolution layer, the weight normalization layer, the random dropout layer and the activation layer to obtain the data quality status; S3.2, calculating the cross entropy loss function value between the actual quality state and the multivariate data quality state identified by the training model in step S3.1, and reversely adjusting and optimizing the weight matrix of the temporal convolutional neural network model through the loss function value; S3.3, re-input the training sample multivariate data into the optimized model in step S3.2 and continue iterating; S3.

4. Determine the parameter range of the temporal convolutional neural network model through multiple experiments.

5. The method for diagnosing abnormal quality of industrial processes based on a temporal convolutional network according to claim 1 is characterized in that: In step S4, the operation is as follows: the real-time multivariate data stream is input into the time series convolutional network model in batches through the sliding window value extraction technology, and the quality of the products corresponding to the data is monitored and diagnosed; when the multivariate data meets the normal quality characteristics, it slides to the data of the next monitoring window in sequence; when the value in the sliding window is abnormal, the time series convolutional network model identifies the quality abnormality characteristics, issues an alarm warning for the produced products, and production is stopped immediately.

6. An industrial process quality anomaly diagnosis system based on a temporal convolutional network, characterized by: The method for diagnosing abnormal quality of an industrial process based on a temporal convolutional network as described in any one of claims 1 to 5 comprises a data acquisition module, a data preprocessing module, a model building module, a model training module, and a quality abnormality diagnosis module, wherein: The data acquisition module is used to collect original multivariate data of the industrial production process; The data preprocessing module is used to remove missing values ​​and abnormal values ​​from the original multivariate data acquired by the data acquisition module using the 6 sigma principle; and then decompose the characteristics of the original multivariate data using the variational mode decomposition method; The model building module is used to add an extended causal convolution layer to keep the sequence relationship of the data unchanged while expanding the receptive field and extracting more features; construct a weight normalization layer so that the neural network can update the weight parameters in reverse in a timely manner; and use the data output by the activation layer neural network to analyze the quality status of the current process; The model training module is used to input the structured data of the data preprocessing module into the temporal convolutional neural network model, and adjust and optimize the deep learning model parameters using the cross entropy loss function; The quality anomaly diagnosis module is used to perform quality monitoring and diagnosis on industrial multivariate data generated in real time based on the trained temporal convolutional neural network model.

Images