A multi-dimensional data monitoring and linkage early warning method for a VIGA powder production process

Abstract

This invention belongs to the field of industrial process monitoring and fault diagnosis technology, specifically relating to a multi-dimensional data monitoring and linkage early warning method for the VIGA pulverizing process. The method includes: acquiring multi-dimensional time-series data and dividing it into multiple data sample matrices; constructing a sparse autoencoder model, inputting the data sample matrices into the trained sparse autoencoder model to obtain a historical reconstruction error sequence, calculating its covariance inverse matrix, and determining an anomaly judgment threshold; calculating the real-time reconstruction error of the real-time data samples, calculating the real-time Mahalanobis distance reconstruction error, and determining an anomaly in the VIGA pulverizing process when the error exceeds the anomaly judgment threshold; calculating partial derivatives, using the absolute value of the partial derivatives as the anomaly contribution; identifying the set of variables whose cumulative contribution meets preset conditions, and issuing an early warning message. This invention solves the problems of insufficient early anomaly detection, difficulty in balancing false alarms and missed alarms, and low anomaly handling efficiency in existing methods.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring and fault diagnosis technology. More specifically, this invention relates to a multi-dimensional data monitoring and linkage early warning method for the VIGA powder production process. Background Technology

[0002] Vacuum induction gas atomization powder making, or VIGA powder making for short, is a core process for preparing high-performance metal powders in high-end manufacturing fields such as metal additive manufacturing and powder metallurgy. The process includes multiple steps such as melting, heat preservation, tilting, and atomization. It is characterized by strong coupling of multiple variables, strong temporal correlation, and sensitivity of powder quality to fluctuations in operating conditions. The process operation status directly determines the core quality indicators of the powder product, such as sphericity, particle size distribution, and oxygen content. At the same time, it is related to the production safety of high-temperature melting and high-pressure atomization. Therefore, real-time monitoring and early warning of abnormalities in the process operation status are the core control links of VIGA powder making production.

[0003] Traditional VIGA milling process monitoring largely relies on statistical process control methods, which monitor the process based on specific statistical distribution assumptions. However, the multidimensional time-series data of the VIGA milling process exhibits strong temporal correlation and inter-variable coupling characteristics, deviating from the core assumptions of these methods. This leads to insufficient sensitivity in identifying early, minor anomalies, resulting in false alarms and missed alarms in practical applications. With the development of deep learning technology, data-driven monitoring methods based on autoencoder models are gradually being applied to VIGA milling process monitoring due to their advantages of unsupervised learning to reconstruct normal operating conditions and using reconstruction errors to determine anomalies.

[0004] However, existing autoencoder models still have the following shortcomings in the practical application of VIGA pulverization process monitoring: On the one hand, their network structures are often not specifically adapted to the characteristics of time-series data, and the error measurement methods do not consider the coupling relationship between variables, making it difficult to accurately capture early minor anomalies and multivariate co-instability problems; on the other hand, the anomaly threshold setting is not adapted to the distribution characteristics of reconstruction error, making it difficult to balance the false alarm rate and the false negative rate, and it is impossible to locate the root cause variable of the anomaly, resulting in low anomaly handling efficiency and making it difficult to give full play to its practical application value in VIGA pulverization process monitoring. Summary of the Invention

[0005] This invention provides a multi-dimensional data monitoring and linkage early warning method for the VIGA powder making process, aiming to solve the technical problems of insufficient ability of existing autoencoder models to capture early minor anomalies in the monitoring of the VIGA powder making process, difficulty in balancing false alarm rate and false negative rate, and low efficiency of anomaly handling.

[0006] This invention provides a multi-dimensional data monitoring and linkage early warning method for the VIGA powder making process, comprising: acquiring multi-dimensional time-series data of the VIGA powder making process, and slidingly dividing it into multiple data sample matrices according to a preset window length and step size; constructing a sparse autoencoder model, wherein the encoder of the sparse autoencoder model includes at least one one-dimensional convolutional layer and a fully connected layer; training the sparse autoencoder model using the data sample matrix under historical normal operating conditions; calculating the covariance matrix and the corresponding inverse matrix of the reconstruction error vector under historical normal operating conditions after training; and inputting the data sample matrix under historical normal operating conditions into the trained sparse autoencoder model to obtain the historical reconstruction error sequence. The algorithm employs an over-threshold sampling method to fit the generalized Pareto distribution to the tail of the historical reconstruction error sequence, and uses its preset upper quantile as the anomaly judgment threshold. The data sample matrix at the current moment is input into the trained sparse autoencoder model to calculate the real-time Mahalanobis distance reconstruction error. When the real-time Mahalanobis distance reconstruction error is greater than the anomaly judgment threshold, an anomaly is determined to have occurred in the VIGA powder making process. The partial derivatives of the Mahalanobis distance reconstruction error with respect to each input process variable are calculated, and the absolute values ​​of the partial derivatives are used as the anomaly contribution. Based on the anomaly contribution of the variables, the set of variables whose cumulative contribution meets the preset conditions is identified, and an early warning message containing the set of variables as the root cause of the anomaly is issued.

[0007] This invention divides multidimensional time-series data into a sample matrix using a sliding window. It then utilizes a sparse autoencoder containing one-dimensional convolutional and fully connected layers to perform unsupervised learning on historical normal operating conditions, extracting local correlation features from the time-series data. During the inference phase, Mahalanobis distance is used to measure the deviation between the samples and normal operating conditions, eliminating the coupling effect between variables and effectively capturing early minor anomalies and collaborative instability phenomena among multiple variables. Anomaly judgment thresholds are dynamically determined using over-threshold sampling combined with generalized Pareto distribution fitting, adapting to the long-tailed distribution characteristics of reconstruction errors and balancing false positives and false negatives. Simultaneously, contribution analysis accurately locates the root cause variables of anomalies, guiding rapid on-site handling, thereby significantly improving the targeting and efficiency of anomaly handling.

[0008] Preferably, the step of sliding the data sample matrix according to the preset window length and step size specifically includes: setting the preset window length to a range of 30-120 time steps and setting the step size to 1 time step; and generating N-L+1 data sample matrices by sliding the multidimensional time series data containing M process variables and a total length of N through the sliding window, where L is the preset window length and the dimension of each data sample matrix is ​​L×M.

[0009] Preferably, the construction of the sparse autoencoder model specifically includes: the encoder is composed of a one-dimensional convolutional layer, a max pooling layer, a flattening layer and a fully connected layer connected in sequence; the decoder is composed of a fully connected layer, a reshaping layer and a one-dimensional transposed convolutional layer connected in sequence, used to reconstruct the encoded feature vector back to the dimension of the original data sample matrix.

[0010] By adopting a mirror-symmetric design of encoder and decoder, key local patterns of time-series data can be preserved during compression. At the same time, the integrity of feature representation is verified through reconstruction process, thereby ensuring the accuracy of model representation of normal working conditions and providing a reliable feature extraction basis for subsequent anomaly detection based on reconstruction error.

[0011] Preferably, the step of training the sparse autoencoder model using a data sample matrix under historical normal operating conditions specifically includes: training the sparse autoencoder model using a mean squared error loss function combined with a gradient descent optimizer until the model converges; inputting the data sample matrix under all historical normal operating conditions of VIGA powder production into the trained sparse autoencoder model to obtain the reconstruction error of each variable at each time step and construct a reconstruction error matrix; and calculating the covariance matrix and the corresponding inverse matrix of the reconstruction error matrix.

[0012] Preferably, a smoothing Gini coefficient sparsity penalty term is added during the training process. The calculation process is as follows: obtain the hidden layer activation vector output by the encoder; construct a non-negative vector from the absolute values ​​of each element in the activation vector; calculate the smoothing Gini coefficient of the non-negative vector; take the negative value of the mean of the smoothing Gini coefficients of the activation vectors corresponding to all data samples in a training batch as the smoothing Gini coefficient sparsity penalty term, and combine it with the mean squared error loss function to form a composite loss function for training.

[0013] Preferably, the formula for calculating the Mahalanobis distance reconstruction error is: ; in, The Mahalanobis distance reconstruction error for a single data sample is dimensionless. The length of the sliding window; This is the time step number, with a value ranging from 1 to... ; For the first The reconstruction error vector at each time step has a dimension of M×1, where M is the number of process variables. ,in For the input sample number The original vector at each time step, To reconstruct the sample number A vector at each time step; The covariance matrix for reconstructing the error vector under historical normal operating conditions; is the covariance matrix of the inverse matrix; the superscript is the transpose operation.

[0014] Preferably, the process of calculating the real-time Mahalanobis distance reconstruction error and determining process anomalies specifically includes: collecting multi-dimensional time-series data in real time, performing standardization and filling missing values; taking data with a preset window length before the current moment to form a real-time data sample matrix; inputting the real-time data sample matrix into the trained model to obtain a reconstructed data sample matrix; calculating the real-time Mahalanobis distance reconstruction error based on the pre-calculated inverse covariance matrix, and comparing it with the anomaly judgment threshold to determine whether it is abnormal.

[0015] Preferably, the step of identifying a set of variables whose cumulative contribution degree meets the preset conditions based on the variable's anomaly contribution degree specifically includes: after determining the anomaly, calculating the sum of the absolute values of the partial derivatives of the Mahalanobis distance reconstruction error with respect to each input process variable at all time steps in the current data sample matrix as the anomaly contribution degree of the variable; sorting the anomaly contribution degrees of all variables in descending order; starting from the variable with the largest contribution degree and accumulating until the sum of the cumulative contribution degrees first reaches or exceeds 85%-95% of the sum of all anomaly contribution degrees, and identifying the accumulated variables as the set of anomaly root cause variables.

[0016] By calculating the sum of the absolute values of the partial derivatives of the reconstruction error with respect to each variable as the contribution degree, accumulating in descending order to the preset ratio, and identifying the accumulated variables as the set of anomaly root causes. Its principle is based on gradient sensitivity analysis, which can accurately locate the key variables leading to anomalies, provide clear guidance for on-site troubleshooting, and avoid blind handling.

[0017] Preferably, the step of sending a warning message including the set of variables as the anomaly root cause variables includes: integrating the anomaly occurrence time, anomaly degree, set of anomaly root cause variables, real-time operating data of the corresponding variables, standard process parameter range, and troubleshooting guidelines into a warning message, and sending a linkage warning through at least one of pop-up windows, audible and visual alarms, email push, and messages on the on-site operation terminal.

[0018] The present invention deeply integrates multi-dimensional status information such as anomaly root cause variables, real-time operating data, and troubleshooting guidelines to construct a panoramic diagnosis report, and triggers a linkage alarm mechanism through multiple channels such as audible and visual signals, push, etc. Its technical effect is that it can not only ensure that on-site personnel learn about the danger in a timely manner, but also directly provide intuitive and clear targeted maintenance guidance, thereby effectively shortening the troubleshooting period of equipment failures and significantly improving the safety control and anomaly handling efficiency of the coal pulverizing production process.

[0019] Preferably, the trained sparse autoencoder model is retrained every 3 months based on newly added normal working condition data during the production process, and the covariance matrix, inverse matrix, and anomaly judgment threshold are recalculated.

[0020] This solution retrains the model periodically with new normal operating data and updates the inverse of the covariance matrix and the anomaly detection threshold simultaneously. This allows the model to adapt to changes in operating conditions during production, ensuring that the model always maintains its monitoring capabilities that match the current production status and guaranteeing that the anomaly identification and early warning results are stable and reliable in the long term.

[0021] The technical solution of the present invention has the following beneficial technical effects: By acquiring multidimensional time-series data of the VIGA powder-making process and dividing it into a sample matrix using a sliding window, a sparse autoencoder model containing one-dimensional convolutional layers and fully connected layers is constructed. This model is trained using historical normal operating data. After training, the reconstruction error covariance matrix and inverse matrix are calculated. Anomaly detection thresholds are determined by combining over-threshold sampling with generalized Pareto distribution fitting. Real-time data is input into the model to calculate the real-time Mahalanobis distance reconstruction error to determine anomalies. The partial derivatives of the reconstruction error with respect to each process variable are then calculated as contributions. Root cause variables are identified and warnings are issued based on the cumulative contribution condition. This invention effectively extracts local correlation features from time-series data using a one-dimensional convolutional structure, enabling sensitive capture of early minor anomalies and signs of multivariate coordinated instability. Threshold settings based on extreme value theory overcome the shortcomings of fixed thresholds in balancing false positives and false negatives. Simultaneously, contribution analysis accurately locates the root cause variables of anomalies, guiding rapid on-site handling and solving the problems of insufficient early anomaly capture, difficulty in balancing false positives and false negatives, and low anomaly handling efficiency in existing methods. Attached Figure Description

[0022] Figure 1 A flowchart of a multi-dimensional data monitoring and linkage early warning method for the VIGA powder production process; Figure 2 A comparative chart of the overall performance of the VIGA powder production process monitoring model. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] This invention discloses a multi-dimensional data monitoring and linkage early warning method for the VIGA powder production process, referring to... Figure 1 This includes the following steps: S1. Obtain multidimensional time-series data during the VIGA powder production process, and divide it into multiple data sample matrices according to the preset window length and step size; construct a sparse autoencoder model, in which the encoder of the sparse autoencoder model contains at least one one-dimensional convolutional layer and a fully connected layer.

[0026] A communication connection is established with the PLC control system of the VIGA powder-making equipment through an industrial standard communication protocol to collect multi-dimensional time-series data of the VIGA powder-making process in real time at a frequency of 1Hz. Historical operating data can be read from the production history SQL database using the read_sql function in the Pandas library of Python, or from historical data files using the read_csv function. Optional industrial communication protocols include OPC-UA, ModbusTCP, and Profinet, all of which are common industrial communication protocols for VIGA powder-making equipment. Those skilled in the art can select the appropriate protocol based on the actual interface configuration of the equipment. The collected process variables are the core process parameters of the entire VIGA powder-making process, including 12 process variables: melting temperature, vacuum chamber pressure, atomizing gas pressure, tundish temperature, crucible tilt angle, rod speed, atomizer vibration frequency, melting power, cooling water inlet temperature, cooling water outlet temperature, molten metal superheat, and atomizing gas temperature, i.e., the number of variables M=12.

[0027] The collected multidimensional time-series data is preprocessed, specifically including: Outlier handling: using The criteria identify outliers in the original data, including those exceeding... Replace the range of data with the historical average of the same variable under normal operating conditions; Missing value handling: For missing data caused by data acquisition interruption or signal loss, the mean of the same variable under normal operating conditions is used to fill the missing data to ensure data continuity; Standardization: The fitting transformation method of the standard scaler class in the scikit-learn library is called to perform Z-Score standardization on all data based on the mean and standard deviation of the historical normal operating data of VIGA milling, so as to eliminate the difference in the units of different variables. The mean of the standardized data is 0 and the standard deviation is 1.

[0028] The standardized multidimensional time-series data is segmented using a sliding window. The preset window length ranges from 30 to 120 time steps, with a step size of 1 time step. The window length is dynamically adjusted according to the VIGA powder-making process: the melting and holding stage has slower changes in operating conditions, so a time step of 60-120 is used; the atomization stage has faster changes in operating conditions, so a time step of 30-60 is used. In this embodiment, for a general monitoring scenario covering the entire VIGA powder-making process, a window length of 60 time steps is preferred, corresponding to 60 seconds of process data. This effectively covers the operating condition fluctuation cycle of the VIGA powder-making process and characterizes the short-term operating characteristics of the process. The step size is set to 1 time step, corresponding to 1 second, to ensure the continuity of data utilization and rapid response to abnormal changes.

[0029] For a multidimensional time-series data set of the VIGA pulverizing process containing M process variables and a total length of N, N-L+1 data sample matrices are generated by partitioning the data using a sliding window. In this embodiment, the window length L is 60, thus generating N-59 data sample matrices, each with a dimension of 60×M. For example, if a 36000-byte segment of normal operating condition time-series data is collected, this sliding window method will generate 35941 data sample matrices, where each data sample matrix X(t,m) corresponds to the observation value of the m-th variable at the t-th time step within the window.

[0030] A sparse autoencoder model is built based on a deep learning framework. Optional deep learning frameworks include TensorFlow, PyTorch, and PaddlePaddle, all of which are commonly used in industrial scenarios. Those skilled in the art can choose according to their actual development environment. This embodiment builds the model based on the Keras API of the TensorFlow framework. The model includes an encoder and a decoder, with the specific structure as follows: The encoder consists of a one-dimensional convolutional layer, a max-pooling layer, a flattening layer, and a fully connected layer connected sequentially. It maps the input data to a low-dimensional feature space, extracting local temporal features from the VIGA powder production time-series data. The one-dimensional convolutional layer uses 8-32 kernels of size 3-5, with a stride of 1, and employs same padding to maintain the temporal dimension. The activation function is a linear rectified unit. The max-pooling layer has a pooling window size of 2 and a stride of 2. The fully connected layer outputs the hidden layer activation vectors, i.e., the encoded features.

[0031] Decoder: Consists of a fully connected layer, a reshaping layer and a one-dimensional transposed convolutional layer connected sequentially. The network structure is mirror-symmetric to the encoder. It is used to reconstruct the original input data from the low-dimensional feature space and outputs a reconstructed data sample matrix with the same dimensions as the input data sample matrix.

[0032] In this embodiment, the preferred model structure for 12-dimensional VIGA powder preparation data is as follows: The encoder's one-dimensional convolutional layer has 16 convolutional kernels of size 3, a stride of 1, same padding, and ReLU activation function, outputting a feature map dimension of 60×16; the max pooling layer has a pooling window size of 2, a stride of 2, and an output dimension of 30×16; the flattening layer converts the feature map into a one-dimensional vector of length 480; the fully connected layer has 128 neurons, outputting a 128-dimensional hidden layer activation vector. The decoder's fully connected layer expands the 128-dimensional encoded features to 480 dimensions, and the reconstructing layer converts it into a 30×16 feature matrix; the one-dimensional transposed convolutional layer has 12 convolutional kernels of size 2, a stride of 2, same padding, and outputs a reconstructed data sample matrix of dimension 60×12, completely consistent with the input data dimension.

[0033] S2. Train the sparse autoencoder model using the data sample matrix under historical normal operating conditions. After training, calculate the covariance matrix and corresponding inverse matrix of the reconstruction error vector under historical normal operating conditions. Input the data sample matrix under historical normal operating conditions into the trained sparse autoencoder model to obtain the historical reconstruction error sequence. Use the over-threshold sampling algorithm to fit the distribution tail of the historical reconstruction error sequence to a generalized Pareto distribution and take its preset upper quantile as the anomaly judgment threshold.

[0034] The sparse autoencoder model was trained using N-59 data sample matrices from the historical normal operating conditions of VIGA milling: the mean squared error loss function was used, the number of iterations was 100, the batch size was 128, the Adam optimizer was used, and the learning rate was 0.0001; the historical normal operating condition dataset was divided into training set and validation set in a 9:1 ratio, and training was terminated early when the validation set loss did not decrease for 10 consecutive iterations.

[0035] Covariance matrix calculation: Input the matrix of N-59 data samples from all VIGA powder production history under normal operating conditions into the trained model to obtain the reconstruction error of each variable at each time step, forming a reconstruction error matrix; use the covariance function in the NumPy library to calculate the covariance matrix of this reconstruction error matrix. The dimension is M×M, and the inverse of the covariance matrix is ​​calculated using the matrix inverse function. If the covariance matrix exhibits a singularity (determinant is 0, cannot be inverted), add a 1×10 matrix element to the diagonal. -6 The inverse is then calculated after the regularization term to ensure the stability of the calculation process.

[0036] Optionally, a sparse penalty term for the smoothing Gini coefficient can be added during training, and the total loss function is expressed as follows: ; in, The total loss function value is dimensionless. The mean squared error loss term is dimensionless. The weight coefficient for the sparsity penalty term is dimensionless and ranges from 0.01 to 1. In this embodiment, the preferred value is 0.1. The Gini coefficient sparsity penalty term is dimensionless and a classic indicator of distribution inequality; a larger value indicates greater differences among elements in the vector. This formula is based on the classic design logic of a composite loss function for neural network training, combining a mean squared error term and a regularization penalty term. While ensuring the model's accuracy in reconstructing VIGA pulverizing normal operating condition data, it constrains the sparsity of hidden layer activations, improving the model's learning ability and generalization performance of core features, thus adapting to the needs of VIGA pulverizing process monitoring. A smaller total loss function value indicates higher model accuracy in reconstructing VIGA pulverizing normal operating condition data, better feature sparsity, and stronger model representation ability for normal operating conditions.

[0037] In the subsequent real-time monitoring and anomaly detection phase, the real-time Mahalanobis distance reconstruction error term The relationship is as follows: ; In the formula, The length of the sliding window is specified in this embodiment. Take 60; This is the time step number, ranging from 1 to W, corresponding to each time step within the sliding window; For the first The reconstruction error vector at each time step has a dimension of M×1. ,in For the input sample number The original vector at each time step, To reconstruct the sample number A vector at each time step; The covariance matrix of the reconstructed error vector under normal operating conditions in the VIGA powder production history is M×M and dimensionless. Covariance matrix The inverse matrix of the matrix, dimensionless; superscript This is the transpose operation for a vector. The higher the value, the greater the deviation of the sample from the normal operating conditions of VIGA powder production, and the higher the probability of process abnormalities; during real-time monitoring, When the anomaly threshold is exceeded, an anomaly is determined to have occurred in the milling process. This relationship is based on the classic Mahalanobis distance metric in statistics. Mahalanobis distance effectively measures the distance between a sample and a distribution and can eliminate the influence of correlation between variables. This invention, tailored to the characteristics of VIGA milling time-series data, accumulates the Mahalanobis distances of the reconstruction error vectors at all time steps within a sliding window, comprehensively measuring the reconstruction deviation throughout the entire time-series window. Adapting to the time-series data characteristics of the VIGA milling process, it can simultaneously capture univariate amplitude anomalies and multivariate correlation structure anomalies.

[0038] Sparse penalty term for smoothing Gini coefficient The process is as follows: Obtain the hidden layer activation vectors output by the encoder. ,in In this embodiment, the number of neurons in the hidden layer is... Set the value to 128; calculate the absolute value of each element in the activation vector to obtain the non-negative vector. ; Calculate the smoothing Gini coefficient of a nonnegative vector. The relationship is as follows: ; In the formula, The smoothing Gini coefficient is dimensionless and ranges from 0 to 1. The larger the value, the more uneven the distribution of the activation vector and the stronger the sparsity. , Non-negative vectors The first in , One element, dimensionless; The smoothing coefficient is a very small positive number, dimensionless, and in this embodiment, it is taken as 1×10⁻⁶. -8 This is used to ensure that the function is smooth and differentiable near the zero point; The number of neurons in the hidden layer is dimensionless. , The neuron number, with a value ranging from 1 to... Smooth Gini coefficient A larger value indicates stronger sparsity of the hidden layer activation vectors, and a better ability of the model to compress the core features under normal VIGA milling conditions. This relationship is based on the classic Gini coefficient calculation formula, which effectively measures the degree of unevenness in distribution. This invention further enhances this by adding a smoothing term. This solves the problem that the original Gini coefficient is not differentiable at zero, making it compatible with backpropagation training of neural networks, achieving differentiable sparsity constraints, encouraging the model to learn the core features of normal VIGA pulverization operation, and suppressing interference from irrelevant features.

[0039] If a training batch contains Data samples were obtained. smooth Gini coefficient , , , Then the sparse penalty term of the smoothing Gini coefficient The calculation formula is as follows: ; In the formula, The number of data samples contained in a single training batch, dimensionless; For the first The smoothing Gini coefficients for each sample are dimensionless. During the minimization of the total loss function... The smaller the value of , the larger the mean of the smoothing Gini coefficient, and the stronger the sparsity of the hidden layer activation vector. This relationship is established based on the regularization constraint logic of the sparse autoencoder. By taking a negative value for the smoothing Gini coefficient, the model tends to improve the sparsity of the hidden layer activation vector in the process of minimizing the total loss function, thereby achieving effective learning of the core features of VIGA powder making under normal operating conditions and improving the model's generalization performance.

[0040] To adapt to the characteristics of VIGA milling time-series data and the convergence requirements of the composite loss function, while ensuring the model's generalization ability and long-term monitoring stability, this embodiment has made targeted settings for the hyperparameters, training rules, and model maintenance mechanisms in the formal training process: the number of iterations is set to 150, the batch size is 128, the optimizer is Adam, and the learning rate is set to 0.001. This parameter combination can balance the convergence speed and stability of model training, avoiding oscillations or insufficient convergence during the training process. The historical normal operating condition dataset is divided into a training set and a validation set in a 9:1 ratio. The training set is used for updating model parameters, and the validation set is used to monitor the model's generalization ability. When the validation set loss does not decrease for 10 consecutive iterations, training is terminated early to avoid overfitting caused by the model overlearning the training set data and to reduce the false alarm rate of warnings under normal operating conditions. After training, the model is retrained every 3 months based on the new normal operating condition data added during the production process to avoid the decrease in monitoring accuracy caused by operating condition drift during the production process and to ensure the reliability of the method in long-term continuous production.

[0041] S3. Input the data sample matrix at the current moment into the trained sparse autoencoder model to calculate the real-time Mahalanobis distance reconstruction error. When the real-time Mahalanobis distance reconstruction error is greater than the anomaly judgment threshold, it is determined that an anomaly has occurred in the VIGA powder making process.

[0042] The 95th-99th percentile of the historical reconstruction error sequence was used as the initial threshold. In this embodiment, the 98th percentile is preferably used as the initial threshold; all errors exceeding the initial threshold are selected from the historical reconstruction error sequence. The error value is used to obtain the over-threshold sample set, and the number of samples is denoted as . Subtract the initial threshold from each sample in the set of samples exceeding the threshold to obtain the set of samples with excess threshold. ,in The first in the historical reconstruction error sequence One sample; the maximum likelihood estimation method is used for the over-threshold margin sample set. By fitting a generalized Pareto distribution, the scale parameters are obtained. and shape parameters A preset confidence level of 99%-99.99% is set; in this embodiment, a preferred confidence level is 99.9%, corresponding to a false alarm rate of [missing information]. The anomaly detection threshold is calculated using the following formula. : ; In the formula, The total number of samples in the historical reconstruction error sequence; The number of samples in the sample set that exceeds the threshold; This is the preset false alarm rate. The anomaly detection threshold is... As a benchmark for judging anomalies in the VIGA powder production process, during real-time monitoring, if the real-time reconstruction error of the sample is greater than... If the process is deemed abnormal, then the process operation status is determined to be abnormal. In this embodiment, the total number of samples in the historical reconstruction error sequence... The value is 35941 (meaning it was obtained by sliding the original data of a total length of 36000 through a window length of 60). ), initial threshold Take 25.6 as the scale parameter. Take 5.2, shape parameters Taking 0.1, we substitute it into the formula to calculate the anomaly detection threshold. This serves as the benchmark for anomaly detection in real-time monitoring. The above relationship is established based on the over-threshold sampling method in extreme value theory. The generalized Pareto distribution is a classic probability model in extreme value theory used to fit the over-threshold residual distribution of random variables. It can effectively adapt to the long-tailed distribution characteristics of the reconstruction error in the VIGA milling process. This invention applies it to the determination of the anomaly threshold for monitoring the VIGA milling process, making the threshold setting statistically reasonable and ensuring a high anomaly detection rate with a low false alarm rate.

[0043] Multidimensional time-series data of the VIGA powder making process are collected in real time. The data is standardized using the same standardized parameters as the training data. Missing data are filled with the mean of the historical normal operating conditions of the same variable. Data from 60 time steps before the current moment are taken to form a real-time data sample matrix with a dimension of 60×12. The matrix is ​​then input into the trained sparse autoencoder model and reconstructed through forward propagation to obtain the data sample matrix.

[0044] Using the pre-calculated inverse of the covariance matrix Calculate the Mahalanobis distance reconstruction error for the real-time sample, and then compare the real-time reconstruction error with the anomaly detection threshold. Comparison: If the real-time reconstruction error is greater than... If the current VIGA powder-making process is deemed abnormal, the anomaly tracing and early warning process will be triggered; if the real-time reconstruction error is less than or equal to... If the process is deemed to be running normally, continuous real-time monitoring will be conducted.

[0045] S4. Calculate the partial derivatives of the Mahalanobis distance reconstruction error with respect to each input process variable, and use the absolute value of the partial derivatives as the abnormal contribution. Based on the abnormal contribution of the variables, identify the set of variables whose cumulative contribution meets the preset conditions, and issue an early warning message containing the set of variables as the root cause of the abnormality.

[0046] When an anomaly is detected in the VIGA powder production process, the anomaly tracing process is initiated, and the specific steps are as follows: Within the TensorFlow framework, the gradient band recording tool is used to calculate the partial derivative of the Mahalanobis distance reconstruction error scalar value with respect to each element in the real-time data sample matrix, resulting in a gradient matrix with the same dimension as the input samples. For each process variable, sum the absolute values ​​of its gradients across all time steps within the sliding window to obtain the variable's outlier contribution. The relationship is as follows: ; in, For the first The outlier contribution of each process variable, dimensionless; This is the ordinal number of the process variable, ranging from 1 to M, where M is the total number of process variables; The length of the sliding window is specified in this embodiment. Take 60; This is the time step number, with a value ranging from 1 to... ; The gradient matrix is ​​the first... The time step, the first The partial derivatives of each variable are dimensionless. (Abnormal contribution) The larger the value, the more significant the number. The greater the influence of a variable on the current anomaly, the greater the probability that it is the root cause of the anomaly. This relationship is constructed based on the gradient sensitivity analysis method. By calculating the partial derivative of the loss function with respect to the input variables, it measures the contribution of the input variables to the anomaly loss. This method is a recognized interpretability analysis method in the field of industrial process anomaly tracing. This invention adapts it to the time-series window data of VIGA milling to achieve accurate location of the root cause variable of the anomaly, providing clear troubleshooting guidance for on-site operators.

[0047] Obtain the set of abnormal contributions of all variables. Then, calculate the total contribution. The abnormal contributions of all variables are sorted in descending order. Starting with the variable with the largest contribution, the contribution is accumulated one by one until the sum of the accumulated contributions first reaches or exceeds 85%-95% of the total contribution, which is then identified as the minimum set of abnormal root cause variables. The preset percentage is adjusted according to the severity of the abnormality: 85%-88% for severe abnormalities to quickly locate the core root cause; and 92%-95% for minor abnormalities to comprehensively cover related and influencing variables. In this embodiment, a preset percentage of 90% is preferred. The system integrates the time of anomaly occurrence, severity of anomaly, set of root cause variables, real-time operational data of the corresponding variables, standard process parameter ranges, and troubleshooting guidelines into early warning information. This information is then disseminated via at least one of the following methods from the VIGA pulverizing production monitoring system: pop-up windows, audible and visual alarms, email push notifications, and messages from on-site operation terminals. This guides on-site operators to quickly handle the anomaly. The aforementioned troubleshooting guidelines can be pre-defined using a mapping table of troubleshooting steps for the root cause variables. For example, when the root cause variable is atomizing gas pressure, the corresponding troubleshooting guidelines would be to check the high-pressure gas cylinder's remaining capacity, the status of the pressure regulator valve, and pipeline leaks.

[0048] For example, if the contribution after sorting is The total contribution score is 110, the cumulative contribution threshold is 99, and the cumulative contribution score is... ;Accumulation ;Accumulation Continue adding subsequent variables until the cumulative contribution reaches 99. All the variables that have been added are the set of variables that are the root cause of the anomaly. The warning message will clearly indicate that the set of variables is the core root cause of the anomaly and push the corresponding investigation guidance.

[0049] To further verify the effectiveness of the present invention, the following comparative experiment was conducted.

[0050] The dataset consists of 12-dimensional process variable data collected from a VIGA device running continuously for 120 hours. The first 100 hours of data, representing normal operating conditions, were used for training, while the last 20 hours contained data on five typical process anomalies for testing. All models were implemented on the same hardware platform (NVIDIA RTX3080 GPU and Intel i7-12700 CPU) using Python and the TensorFlow framework. Evaluation metrics included precision, recall, F1 score, and false positive rate. The control group model used a standard autoencoder with fully connected layers, employed the mean squared error loss function, and set the anomaly threshold at the 99.9 percentile.

[0051] The control group model achieved a precision of 87.5%, a recall of 84.2%, an overall F1 score of 85.8%, and a false positive rate of 1.5% on the test set. The method proposed in this invention achieved a precision of 97.8%, a recall of 96.5%, an overall F1 score of 97.1%, and a false positive rate reduced to 0.2% on the same test set. Regarding anomaly root cause diagnosis, the control group method, lacking a source tracing mechanism, could not pinpoint specific variables. The source tracing method of this invention achieved an accuracy of 92.3% in identifying the top three root cause variables among all successfully detected anomaly samples.

[0052] Figure 2 This paper visually demonstrates the core performance differences between the sparse autoencoder model used in this invention and the standard autoencoder (control group) in the multidimensional data monitoring task of the VIGA powder making process. The comparison clearly shows that the sparse autoencoder model of this invention outperforms the standard autoencoder in all core performance aspects of the VIGA powder making process monitoring, providing more accurate and comprehensive anomaly monitoring and early warning support for the powder production process, thus fully verifying the effectiveness of the technical solution of this invention.

[0053] The experimental data above demonstrate that this invention exhibits significant performance advantages over traditional autoencoder methods, with the F1 score increasing from 85.8% to 97.1%, a rise of 11.3 percentage points. This is attributed to the fact that the one-dimensional convolutional network structure can effectively extract the temporal dependencies of process data, making it more sensitive to local pattern changes compared to fully connected layers; the Mahalanobis distance reconstruction error, combined with the correlation between variables, can detect situations where a single variable is within limits but the joint distribution of multiple variables is abnormal, thus improving recall; the sparse penalty term of the smoothing Gini coefficient enhances the model's generalization ability, and together with the over-threshold sampling threshold setting method based on extreme value theory, the false alarm rate is reduced from 1.5% to 0.2%, achieving a balance between false alarms and false negatives.

[0054] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-dimensional data monitoring and linkage early warning method for the VIGA powder production process, characterized in that, include: Acquire multidimensional time-series data during the VIGA powder production process, and divide it into multiple data sample matrices according to preset window length and step size; Construct a sparse autoencoder model, wherein the encoder of the sparse autoencoder model contains at least one one-dimensional convolutional layer and a fully connected layer; The sparse autoencoder model is trained using a data sample matrix under historical normal operating conditions; After training, the covariance matrix and the corresponding inverse matrix of the reconstruction error vector under historical normal working conditions are calculated. The data sample matrix under historical normal working conditions is input into the trained sparse autoencoder model to obtain the historical reconstruction error sequence. The distribution tail of the historical reconstruction error sequence is fitted with a generalized Pareto distribution using the over-threshold sampling algorithm, and its preset upper quantile is taken as the anomaly judgment threshold. Input the data sample matrix at the current moment into the trained sparse autoencoder model to calculate the real-time Mahalanobis distance reconstruction error. When the real-time Mahalanobis distance reconstruction error is greater than the anomaly judgment threshold, it is determined that an anomaly has occurred in the VIGA powder making process. Calculate the partial derivatives of the Mahalanobis distance reconstruction error with respect to each input process variable, and use the absolute value of the partial derivatives as the anomaly contribution. Based on the abnormal contribution of variables, the system identifies the set of variables whose cumulative contribution meets preset conditions and issues an early warning message that includes the set of variables as the root cause of the abnormality.

2. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, The step of sliding the data into multiple data sample matrices according to a preset window length and step size specifically includes: setting the preset window length to a range of 30-120 time steps and setting the step size to 1 time step; and generating N-L+1 data sample matrices by sliding the multidimensional time series data containing M process variables and a total length of N through a sliding window, where L is the preset window length and the dimension of each data sample matrix is ​​L×M.

3. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, The construction of the sparse autoencoder model specifically includes: the encoder is composed of a one-dimensional convolutional layer, a max pooling layer, a flattening layer and a fully connected layer connected in sequence; the decoder is composed of a fully connected layer, a reshaping layer and a one-dimensional transposed convolutional layer connected in sequence, used to reconstruct the encoded feature vector back to the dimension of the original data sample matrix.

4. The multi-dimensional data monitoring and linkage early warning method for the VIGA powder making process according to claim 1, characterized in that, The method of training the sparse autoencoder model using a data sample matrix under historical normal operating conditions specifically includes: training the sparse autoencoder model using a mean squared error loss function combined with a gradient descent optimizer until the model converges. Input the data sample matrix of all VIGA powder production history under normal operating conditions into the trained sparse autoencoder model to obtain the reconstruction error of each variable at each time step and form a reconstruction error matrix. Calculate the covariance matrix and the corresponding inverse matrix of the reconstruction error matrix.

5. The multi-dimensional data monitoring and linkage early warning method for the VIGA powder making process according to claim 4, characterized in that, During training, a smoothing Gini coefficient sparsity penalty term is added. The calculation process is as follows: obtain the hidden layer activation vector output by the encoder; construct a non-negative vector from the absolute values ​​of each element in the activation vector; calculate the smoothing Gini coefficient of the non-negative vector; take the negative value of the mean of the smoothing Gini coefficients of the activation vectors corresponding to all data samples in a training batch as the smoothing Gini coefficient sparsity penalty term, and combine it with the mean squared error loss function to form a composite loss function for training.

6. The multi-dimensional data monitoring and linkage early warning method for the VIGA powder making process according to claim 1, characterized in that, The formula for calculating the Mahalanobis distance reconstruction error is as follows: ； in, The Mahalanobis distance reconstruction error for a single data sample is dimensionless. The length of the sliding window; This is the time step number, with a value ranging from 1 to... ; For the first The reconstruction error vector at each time step has a dimension of M×1, where M is the number of process variables. ,in For the input sample number The original vector at each time step, To reconstruct the sample number A vector at each time step; The covariance matrix for reconstructing the error vector under historical normal operating conditions; Covariance matrix The inverse matrix; superscript This is for the transpose operation.

7. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, The calculation of the real-time Mahalanobis distance reconstruction error and the determination of process anomalies specifically include: real-time acquisition of multi-dimensional time series data, and standardization and missing value imputation. A real-time data sample matrix is ​​constructed by taking data of a preset window length prior to the current time. Input the real-time data sample matrix into the trained model to obtain the reconstructed data sample matrix; Based on the pre-calculated inverse of the covariance matrix, the real-time Mahalanobis distance reconstruction error is calculated and compared with the anomaly judgment threshold to determine whether it is an anomaly.

8. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, The variable-based abnormal contribution degree identifies a set of variables whose cumulative contribution degree meets preset conditions. Specifically, after determining an anomaly, the sum of the absolute values ​​of the partial derivatives of the Mahalanobis distance reconstruction error with respect to each input process variable at all time steps in the current data sample matrix is ​​calculated as the abnormal contribution degree of that variable. Sort all variables in descending order of their outlier contributions; Start by accumulating the variables with the largest contribution until the sum of the cumulative contributions first reaches or exceeds 85%-95% of the sum of all abnormal contributions. Then, identify the accumulated variables as the set of abnormal root cause variables.

9. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, Issue early warning information containing a set of variables as the root cause of the anomaly, including: integrating the time of the anomaly occurrence, the degree of the anomaly, the set of root cause variables, the real-time operating data of the corresponding variables, the standard process parameter range and the fault troubleshooting guide into early warning information, and issuing a linked early warning through at least one of the following methods: pop-up window, audible and visual alarm, email push, and on-site operation terminal message.

10. The method for multi-dimensional data monitoring and linkage early warning in the VIGA powder making process according to claim 1, characterized in that, The trained sparse autoencoder model is retrained every 3 months based on new normal operating data added during the production process, and the covariance matrix, inverse matrix, and anomaly judgment threshold are recalculated.

Images