Industrial process monitoring method and system based on multi-scale slow feature auto-encoder
By using a multi-scale slow feature autoencoder, the problem of capturing dynamic characteristics and nonlinear relationships in industrial process monitoring was solved, achieving efficient industrial process monitoring and improving monitoring accuracy and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-03-27
AI Technical Summary
Existing industrial process monitoring methods cannot effectively capture complex dynamic characteristics and nonlinear relationships, and require a large number of fault labeling samples.
A multi-scale slow feature autoencoder is adopted, including a multi-scale convolutional branch module, an attention dynamic weighted fusion module, a residual connection module, and a slowness regularization module. It is combined with a loss function that combines the synergistic effects of slow variation constraints, decorrelation constraints, and reconstruction constraints to train and extract slow features of industrial processes.
It eliminates the need for fault labeling samples, overcomes linear limitations, adapts to dynamic characteristics, improves monitoring accuracy and reliability, and reduces the computational complexity of anomaly detection.
Smart Images
Figure CN121744148A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial data analysis and anomaly detection technology, and more specifically, to an industrial process monitoring method, system, and storage medium based on a multi-scale slow feature autoencoder. Background Technology
[0002] In the process of industrial intelligent upgrading, modern industrial processes are developing towards larger scale, greater complexity, and higher automation. Ensuring the safe, stable, efficient, and high-quality operation of production processes has become a core issue of common concern to both industry and academia. Process monitoring technology, as a key means to achieve this goal, aims to analyze the massive amounts of data collected during the production process to monitor abnormal conditions in real time, thereby avoiding production accidents and reducing downtime losses.
[0003] Traditional physical model-based process monitoring methods heavily rely on a deep understanding of process mechanisms, requiring the establishment of accurate mathematical models. However, modern industrial processes typically possess complex characteristics such as nonlinearity, strong coupling, high variable dimensionality, and varied operating conditions, making accurate mechanism modeling exceptionally difficult. Among data-driven process monitoring methods, multivariate statistical process monitoring is one of the most widely used techniques. This method monitors process operation status by projecting high-dimensional data onto a low-dimensional feature space. However, real-world industrial data is inherently dynamic and nonlinear, and simple linear models often fail to capture the complex dynamic characteristics and nonlinear relationships within the process. While deep learning methods such as neural networks can learn nonlinear features, they require a large number of labeled fault samples for training. However, in industrial scenarios, normal operation data is abundant, but fault samples are scarce, and fault modes are diverse, making it difficult to meet the model's data dependency requirements, resulting in poor generalization ability of such models.
[0004] In summary, when monitoring industrial process data, there is an urgent need for an industrial data analysis technology that can overcome linear limitations, adapt to dynamic characteristics, and does not require a large number of fault labeling samples, in order to solve the pain points of existing technologies and lay the foundation for equipment status characterization and anomaly detection. Summary of the Invention
[0005] The technical problem to be solved by this invention is:
[0006] Existing industrial process monitoring methods have limitations in capturing complex dynamic characteristics and nonlinear relationships within the process, and require a large number of fault labeling samples.
[0007] Therefore, the present invention provides an industrial process monitoring method, system, and storage medium based on a multi-scale slow feature autoencoder.
[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0009] This invention provides an industrial process monitoring method based on a multi-scale slow feature autoencoder, comprising the following steps:
[0010] S1. Utilize sensors to collect diverse time-series data during the normal operation of industrial equipment;
[0011] S2. Construct a multi-scale slow feature autoencoder network, including: a multi-scale convolutional branch module, an attention-based dynamic weighted fusion module, a residual connection module, and a slowness regularization module in the encoder, as well as a decoder; the multi-scale convolutional branch module includes multiple parallel multi-scale branches, used to extract nonlinear features at different time scales in the time series data; the attention-based dynamic weighted fusion module is used to perform attention calculation and attention-weighted fusion on the features of each scale branch to obtain fused features; the slowness regularization module is used to transform the high-dimensional fused features into low-dimensional, smooth slow features;
[0012] S3. Construct a loss function based on the synergistic effect of slow variation constraint, decorrelation constraint, and reconstruction constraint, and use it to train the multi-scale slow feature autoencoder network.
[0013] S4. Using the collected time-series data of the normal operation process as input data, train a multi-scale slow feature autoencoder network to extract slow feature values; then calculate the monitoring statistics and the monitoring control limit threshold.
[0014] S5. Input the multivariate time-series data, which serves as test data, into the trained multi-scale slow feature autoencoder network, extract the slow feature values, and then calculate the monitoring statistics to monitor the process data.
[0015] Furthermore, the functional implementation process of the multi-scale convolutional branch module is as follows:
[0016] The size of the convolution kernel in each branch of the module is m. k k=1,2,…,K; Convolution is performed on each variable individually, and the extracted features are:
[0017]
[0018] Where h k j Variable x extracted for convolution kernels of different sizes j Features, j=1,2,…,M; b k For bias;
[0019] Standardize the convolution output:
[0020]
[0021] Where μ kj σ is the mean. k j Let γ be the variance. k j Scaling factor, offset factor β k j For hyperparameters, To prevent the use of infinitesimal values where the denominator is 0, let the characteristic of the M-dimensional variable be denoted as H. k BN By introducing nonlinearity using activation functions, the output features H at different time scales are obtained. k :
[0022]
[0023] Where f(·) is the activation function.
[0024] Furthermore, the functional implementation process of the attention dynamic weighted fusion module is as follows:
[0025] First, global average pooling is used to compress the multi-scale branch features into a single-vector multi-scale descriptor V. k :
[0026]
[0027] An attention evaluation module is constructed by stacking two fully connected layers, which maps the feature descriptors of multi-scale branches to importance scores at each scale. :
[0028]
[0029] Among them W fc1 b fc1 W fc2 b fc2 These are the weights and bias parameters of the two fully connected layers, respectively.
[0030] The importance scores for each scale are normalized to obtain the weighting coefficients. :
[0031]
[0032] in ;
[0033] We obtain the fusion feature H by weighted summation of the scales of each branch:
[0034] .
[0035] Furthermore, the functional implementation process of the slow-rate regularization module is as follows:
[0036] High-dimensional fusion features after residual concatenation Transformed into a low-dimensional, smooth, slow feature S:
[0037]
[0038] Among them W smooth and b smooth For smooth convolution kernels and biases, H smooth The output after smooth convolution; W reduce and b reduce The kernel and bias of the dimensionality-reducing convolution.
[0039] Furthermore, the functional implementation process of the decoder is as follows:
[0040] First, perform a 1D convolution on each variable of the slow feature S individually to capture the evolution of the slow feature over time:
[0041]
[0042] Among them W c depth Let b be the depthwise convolution kernel for the c-th variable. depth For bias, H depth Features output by depthwise convolution;
[0043] By introducing nonlinearity using an activation function and then using point convolution to map D-dimensional slow features back to the original M-dimensional data, the reconstructed data is output. :
[0044]
[0045] Where f is the activation function, W point For point convolution kernels, b point For bias.
[0046] Furthermore, the loss function based on the synergistic effect of slow-varying constraints, decorrelation constraints, and reconstruction constraints is:
[0047]
[0048] in, For slow-degree loss, To eliminate the related sub-loss, For the reconstructor loss, α, β, and γ are hyperparameters, where,
[0049]
[0050] Where S(t, d) is the value of the slow feature S at the t-th sample point and in the d-th dimension, and N is the number of samples;
[0051]
[0052] Where C is the covariance matrix of the slow feature S, and C(d, d') is the covariance of the d-th and d'-th dimensions of the feature, where d ≠ d';
[0053]
[0054] Where X(t, m) and Let M be the value of the original input data and the m-th variable at the t-th sample point, respectively, and M be the number of variables.
[0055] Furthermore, the monitoring statistics described in S4 include:
[0056]
[0057]
[0058] Where, X(m) and These are the values of the m-th variable in the input multivariate time-series data and the decoder-reconstructed data, respectively.
[0059] This invention provides an industrial process monitoring system based on a multi-scale slow feature autoencoder. The system has a program module corresponding to the steps of any of the above-described technical solutions, and executes the steps in the above-described industrial process monitoring method based on a multi-scale slow feature autoencoder when running.
[0060] The present invention provides a computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of the industrial process monitoring method based on a multi-scale slow feature autoencoder as described in any of the above technical solutions.
[0061] Compared with the prior art, the beneficial effects of the present invention are:
[0062] This invention is based on a multi-scale slow feature autoencoder, which can effectively extract nonlinear slow-varying features at different time scales in industrial time series data, covering short-term local fluctuations and long-term global trends. Parallel attention computation ensures that features at different time scales are fully and independently expressed and weighted, and adaptively and dynamically evaluates the importance of features at each scale, thereby improving the global representation capability of features.
[0063] This invention eliminates the need for labeled fault samples and overcomes linear limitations while adapting to dynamic characteristics. It achieves dimensionality compression while retaining key slowly changing information, reducing the computational complexity of subsequent anomaly detection and thus effectively improving the accuracy and reliability of industrial process monitoring. Attached Figure Description
[0064] Figure 1 This is a flowchart of an industrial process monitoring method based on a multi-scale slow feature autoencoder according to an embodiment of the present invention.
[0065] Figure 2 This is a network model diagram of the multi-scale slow feature autoencoder and multi-constraint loss function constructed in the embodiments of the present invention;
[0066] Figure 3 This is a graph showing the T2 monitoring results of fault 2 in the TE test set in this embodiment of the invention.
[0067] Figure 4 This is a graph showing the MSE monitoring results of fault 2 in the TE test set in this embodiment of the invention;
[0068] Figure 5 This is a graph showing the T2 monitoring results of fault 17 in the TE test set in this embodiment of the invention.
[0069] Figure 6 This is a graph showing the MSE monitoring results of fault 17 in the TE test set in this embodiment of the invention. Detailed Implementation
[0070] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.
[0071] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0072] Specific Implementation Plan 1: Combining Figures 1 to 2 As shown, this invention provides an industrial process monitoring method based on a multi-scale slow feature autoencoder, comprising the following steps:
[0073] S1. Use sensors to collect process data X during normal operation of industrial equipment as the training set. The data is in the form of multivariate time series, including N samples and M variables.
[0074] S2. Construct a multi-scale slow feature autoencoder network, including: a multi-scale convolutional branch module, an attention-based dynamic weighted fusion module, a residual connection module, and a slowness regularization module in the encoder, as well as a decoder; the multi-scale convolutional branch module includes multiple parallel multi-scale branches, used to extract nonlinear features at different time scales in the time series data; the attention-based dynamic weighted fusion module is used to perform attention calculation and attention-weighted fusion on the features of each scale branch to obtain fused features; the slowness regularization module is used to transform the high-dimensional fused features into low-dimensional, smooth slow features;
[0075] S3. Construct a loss function based on the synergistic effect of slow variation constraint, decorrelation constraint, and reconstruction constraint, and use it to train the multi-scale slow feature autoencoder network.
[0076] S4. Using the time-series data of the normal operation process collected in S1 as input data, train a multi-scale slow feature autoencoder network to extract slow feature values; then calculate the monitoring statistics and the monitoring control limit threshold.
[0077] S5. Input the multivariate time-series data, which serves as test data, into the trained multi-scale slow feature autoencoder network, extract the slow feature values, and then calculate the monitoring statistics to monitor the process data.
[0078] Specific Implementation Scheme Two: The functional implementation process of the multi-scale convolution branch module is as follows:
[0079] The multi-scale convolutional branching module comprises K parallel branches to extract nonlinear features at different time scales from industrial time-series data, covering short-term local fluctuations and long-term global trends. The kernel size of each branch is m. k k=1,2,…,K; Convolution is performed on each variable individually, and the extracted features are:
[0080]
[0081] Where h k j Variable x extracted for convolution kernels of different sizes j Features, j=1,2,…,M; b k For bias;
[0082] Standardizing the convolutional output eliminates distributional differences between different data sets, accelerates network training, and avoids gradient shift.
[0083]
[0084] Where μ k j σ is the mean. kj Let γ be the variance. k j Scaling factor, offset factor β k j For hyperparameters, To prevent the use of infinitesimal values where the denominator is 0, let the characteristic of the M-dimensional variable be denoted as H. k BN By introducing nonlinearity using activation functions, the output features H at different time scales are obtained. k :
[0085]
[0086] Where f(·) is the activation function; the output features H at different time scales k Maintain the same dimensions as the input data.
[0087] This implementation plan is otherwise the same as Specific Implementation Plan 1.
[0088] Specific Implementation Scheme 3: The functional implementation process of the attention dynamic weighted fusion module is as follows:
[0089] The attention-based dynamic weighted fusion module dynamically evaluates the feature importance of each multi-scale branch. Through weighted fusion, it highlights scale features with high information content, avoids single-scale bias, and improves the global representation ability of features.
[0090] First, global average pooling is used to compress the multi-scale branch features into a single-vector multi-scale descriptor V. k :
[0091]
[0092] An attention evaluation module is constructed by stacking two fully connected layers, which maps the feature descriptors of multi-scale branches to importance scores at each scale. Quantify the importance of the k-th branch:
[0093]
[0094] Among them W fc1 b fc1 W fc2 b fc2 These are the weights and bias parameters of the two fully connected layers, respectively.
[0095] The importance scores for each scale are normalized to obtain the weighting coefficients. :
[0096]
[0097] in ;
[0098] By weighting and summing the data at each branch scale and dynamically balancing the multi-scale information, the fused feature H is obtained.
[0099] .
[0100] Residual connections are used to solve the vanishing gradient problem in deep networks, preserving the original data information:
[0101]
[0102] Residual connections preserve key information from the original data through direct superposition, avoiding gradient vanishing caused by increased network depth, stabilizing the training process, and enhancing the robustness of features. This implementation scheme is otherwise identical to specific implementation scheme two.
[0103] Specific Implementation Scheme Four: The functional implementation process of the slowness regularization module is as follows:
[0104] Slow regularization and dimensionality compression are achieved through convolution operations, resulting in high-dimensional fused features after residual concatenation. This is transformed into a low-dimensional, smooth, slow feature S to match the gradual evolution characteristics of industrial processes.
[0105]
[0106] Among them W smooth and b smooth For smooth convolution kernels and biases, H smooth The output is after smooth convolution, with the same dimension as the input data; W reduce and b reduce The convolution kernel and bias are used for dimensionality reduction convolution, resulting in a slow feature dimension of N×D (D≤M). The final output slow feature S dimension is pre-specified, which preserves key slow-varying information while significantly reducing the computational complexity of subsequent anomaly detection, thus meeting the needs of real-time industrial monitoring. Other aspects of this implementation scheme are the same as in specific implementation scheme three.
[0107] Specific Implementation Scheme Five: The functional implementation process of the decoder is as follows:
[0108] The original input X is reconstructed from the slow features S to verify whether the slow features retain sufficient information, balancing the slow variation characteristics with information integrity, and assisting the encoder in optimizing feature extraction. First, a 1D convolution is performed on each variable of the slow features S individually to capture the evolution of the slow features over time.
[0109]
[0110] Among them W c depth Let b be the depthwise convolution kernel for the c-th variable. depth For bias, Hdepth For features output by deep convolution, fine-grained temporal patterns are preserved;
[0111] By introducing nonlinearity through activation functions, the decoder's ability to reconstruct complex data is enhanced. Point convolution is then used to map D-dimensional slow features back to the original M-dimensional data, outputting the reconstructed data. :
[0112]
[0113] Where f is the activation function, W point For point convolution kernels, b point For bias.
[0114] This implementation plan is otherwise the same as Specific Implementation Plan Four.
[0115] Specific Implementation Plan Six: The loss function serves as a multi-objective optimization criterion during network training. With the synergistic effect of "slow-varying constraints - decorrelation constraints - reconstruction constraints", it guides the network to extract slow-varying, feature-independent, and fidelity-preserving slow features.
[0116] 1) Slowness constraint
[0117] Slowness loss forces features to conform to slow variation patterns by penalizing temporal abrupt changes in features:
[0118]
[0119] Where S(t, d) is the value of the slow feature S at sample point t and in dimension d. The more smoothly the feature changes over time, the lower the loss value. This constraint allows the feature to focus on long-term trends and filters out short-term noise.
[0120] 2) Remove relevant constraints
[0121] High correlation between feature dimensions can lead to information redundancy. Decorrelation loss, by penalizing the statistical correlation between dimensions, forces each feature dimension to be independent and complementary, thereby increasing information density.
[0122]
[0123] Where C is the covariance matrix of the slow feature S, and C(d, d') is the covariance of the d-th and d'-th features (d≠d'). The smaller the sum of squared covariances, the weaker the correlation between dimensions, and the lower the loss. This constraint ensures that each slow feature has unique information and avoids redundancy.
[0124] 3) Reconfigure constraints
[0125] Slow features extract slowly changing information from the original data without losing key details. Reconstruction constraints ensure that the features retain sufficiently important information by comparing the original data with the data reconstructed from the slow features:
[0126]
[0127] Where X(t, m) and X̂(t, m) are the values of the original input data and the decoder reconstructed data at the t-th sample point and the m-th variable, respectively. The smaller the error, the more accurate the reconstruction, and the more complete the original information preserved by the slow features. This constraint balances the slow variation constraint with detail preservation, preventing the loss of key information.
[0128] 4) Joint loss function
[0129] The three sub-losses constrain the features from different perspectives and guide the network parameter updates by balancing their contributions through weights:
[0130]
[0131] α, β, and γ are hyperparameters that control the influence of each sub-loss.
[0132] A multi-scale slow-feature autoencoder is trained using training data. Data is transmitted via forward propagation to achieve signal transfer, and then backpropagation minimizes the total loss, simultaneously optimizing the parameters of each module in the network model. Ultimately, the extracted slow features simultaneously meet the requirements of slow variation, low redundancy, and complete information, providing high-quality features for industrial process monitoring. The constructed multi-scale slow-feature autoencoder and multi-constraint loss function network model are as follows: Figure 2 As shown. This implementation plan is otherwise identical to specific implementation plan five.
[0133] Specific Implementation Plan Seven: After the multi-scale slow feature autoencoder training is completed, the required slow features can be obtained. Two monitoring statistics are constructed using these features:
[0134]
[0135]
[0136] Where, X(m) and These are the values of the m-th variable in the input multivariate time-series data and the decoder-reconstructed data, respectively.
[0137] For the two monitoring statistics mentioned above, the control limit thresholds are calculated using kernel density estimation.
[0138] For the test samples collected online and to be monitored, they are input into the trained multi-scale slow feature autoencoder network to obtain slow feature values. Monitoring statistics are then calculated and compared with control limit thresholds. If the corresponding control limit is exceeded, it indicates an anomaly in the process data, and an alarm is triggered. Other aspects of this implementation scheme are the same as in specific implementation scheme six.
[0139] The industrial process monitoring method (algorithm) based on a multi-scale slow feature autoencoder proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.
[0140] Based on the method proposed in this invention, an industrial process monitoring system based on a multi-scale slow feature autoencoder is developed using a programming language. This system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned industrial process monitoring method based on a multi-scale slow feature autoencoder when running.
[0141] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-described industrial process monitoring method based on a multi-scale slow feature autoencoder when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.
[0142] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0143] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.
[0144] The beneficial effects of the present invention will be described below with reference to specific embodiments.
[0145] Example 1
[0146] The Tennessee-Eastman (TE) process is used as an example. This process is an industrial-grade chemical process simulation platform that provides a standard test object closely resembling real-world industrial scenarios for process control, fault diagnosis, and other fields. Its complexity, nonlinearity, and dynamic characteristics highly match those of real chemical production processes. The TE process data includes 22 continuous process variables such as temperature, pressure, flow rate, and liquid level, 19 component variables, and 11 operational variables. The data sampling time is 3 minutes, including process data under one normal operating condition and various fault data.
[0147] The continuous and operational variables during normal operation in the TE process training set were used as training data, denoted as X. The data is in the form of a multivariate time series, including 500 samples and 33 variables, as shown in Table 1.
[0148] Table 1
[0149]
[0150] The multi-scale convolution branch module has three branches at different scales, with kernel sizes m1=3, m2=5, and m3=7 for each branch. Each variable is convolved individually to extract multi-scale features. The convolution output is standardized to eliminate distributional differences between different data points, accelerate network training, and avoid gradient shift. Let the features of the 33-dimensional variable be denoted as H. k BN Introducing nonlinearity using the ReLU function:
[0151] .
[0152] The slow regularization module uses convolution operations to achieve slow regularization and dimensionality compression. The dimensionality of the slow features after dimensionality reduction is 500×10, which not only preserves the key slow variation information, but also greatly reduces the computational complexity of subsequent anomaly detection, thus meeting the needs of real-time industrial monitoring.
[0153] In the decoder, the ReLU function is used to introduce nonlinearity, which enhances the decoder's ability to reconstruct complex data, and point convolution is used to map the 10-dimensional slow features back to the original 33-dimensional features.
[0154] The hyperparameters of the loss function based on the synergistic effect of slow-varying constraints, decorrelation constraints, and reconstruction constraints are set to α=1, β=0.2, and γ=0.2, respectively, to control the influence of each sub-loss.
[0155] For T 2 Using kernel density estimation, the control limit thresholds for the two monitoring statistics, MSE, were calculated at a 95% confidence level for each statistic.
[0156] The online monitoring TE process test set contains test samples of 21 types of faults, each with a dimension of 960×33. The fault occurs at the 161st sample point. These samples are then input into a trained multi-scale slow feature autoencoder network to obtain slow feature values, and then T is calculated. 2 The MSE monitoring statistics are compared with control limit thresholds. If the corresponding control limit is exceeded, it indicates that the process data is abnormal, and an alarm is triggered. The T value for fault 2 in the test set is... 2 The monitoring results of MSE monitoring statistics are attached. Figure 3 and 4 As shown, the T in test set fault 17 2 The monitoring results of MSE monitoring statistics are attached as follows. Figure 5 and Figure 6 As shown.
[0157] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.
Claims
1. An industrial process monitoring method based on a multi-scale slow feature autoencoder, characterized in that, Includes the following steps: S1. Utilize sensors to collect diverse time-series data during the normal operation of industrial equipment; S2. Construct a multi-scale slow feature autoencoder network, including: a multi-scale convolutional branch module, an attention-based dynamic weighted fusion module, a residual connection module, and a slowness regularization module in the encoder, as well as a decoder; the multi-scale convolutional branch module includes multiple parallel multi-scale branches, used to extract nonlinear features at different time scales in the time series data; the attention-based dynamic weighted fusion module is used to perform attention calculation and attention-weighted fusion on the features of each scale branch to obtain fused features; the slowness regularization module is used to transform the high-dimensional fused features into low-dimensional, smooth slow features; S3. Construct a loss function based on the synergistic effect of slow variation constraint, decorrelation constraint, and reconstruction constraint, and use it to train the multi-scale slow feature autoencoder network. S4. Using the collected time-series data of the normal operation process as input data, train a multi-scale slow feature autoencoder network to extract slow feature values; then calculate the monitoring statistics and the monitoring control limit threshold. S5. Input the multivariate time-series data, which serves as test data, into the trained multi-scale slow feature autoencoder network, extract the slow feature values, and then calculate the monitoring statistics to monitor the process data.
2. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 1, characterized in that, The functional implementation process of the multi-scale convolutional branch module is as follows: The size of the convolutional kernel in each branch of the module is m. k k=1,2,…,K; Convolution is performed on each variable individually, and the extracted features are: Where h k j Variable x extracted for convolution kernels of different sizes j Features, j=1,2,…,M; b k For bias; Standardize the convolution output: Where μ k j σ is the mean. k j Let γ be the variance. k j Scaling factor, offset factor β k j For hyperparameters, To prevent tiny values with a denominator of 0; Let the characteristic of the M-dimensional variable be denoted as H. k BN By introducing nonlinearity using activation functions, the output features H at different time scales are obtained. k : Where f(·) is the activation function.
3. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 2, characterized in that, The functional implementation process of the attention dynamic weighted fusion module is as follows: First, global average pooling is used to compress the multi-scale branch features into a single-vector multi-scale descriptor V. k : An attention evaluation module is constructed by stacking two fully connected layers, which maps the feature descriptors of multi-scale branches to importance scores at each scale. : Among them W fc1 b fc1 W fc2 b fc2 These are the weights and bias parameters of the two fully connected layers, respectively. The importance scores for each scale are normalized to obtain the weighting coefficients. : in ; We obtain the fusion feature H by weighted summation of the scales of each branch: 。 4. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 3, characterized in that, The function implementation process of the slow regularization module is as follows: High-dimensional fusion features after residual concatenation Transformed into a low-dimensional, smooth, slow feature S: Among them W smooth and b smooth For smooth convolution kernels and biases, H smooth The output after smooth convolution; W reduce and b reduce The kernel and bias of the dimensionality-reducing convolution.
5. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 4, characterized in that, The functional implementation process of the decoder is as follows: First, perform a 1D convolution on each variable of the slow feature S individually to capture the evolution of the slow feature over time: Among them W c depth Let b be the depthwise convolution kernel for the c-th variable. depth For bias, H depth Features output by depthwise convolution; By introducing nonlinearity using an activation function and then using point convolution to map D-dimensional slow features back to the original M-dimensional data, the reconstructed data is output. : Where f is the activation function, W point The kernel of the point convolution, b point For bias.
6. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 5, characterized in that, The loss function based on the synergistic effect of slow-varying constraints, decorrelation constraints, and reconstruction constraints is: in, For slow-degree loss, To eliminate the related sub-loss, For the reconstructor loss, α, β, and γ are hyperparameters, where, Where S(t, d) is the value of the slow feature S at the t-th sample point and in the d-th dimension, and N is the number of samples; Where C is the covariance matrix of the slow feature S, and C(d, d') is the covariance of the d-th and d'-th dimensions of the feature, where d ≠ d'; Where X(t, m) and Let M be the value of the original input data and the m-th variable at the t-th sample point, respectively, and M be the number of variables.
7. The industrial process monitoring method based on a multi-scale slow feature autoencoder according to claim 6, characterized in that, The monitoring statistics described in S4 include: Where, X(m) and These are the values of the m-th variable in the input multivariate time-series data and the decoder-reconstructed data, respectively.
8. An industrial process monitoring system based on a multi-scale slow feature autoencoder, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 7, and executes the steps in the above-described industrial process monitoring method based on multi-scale slow feature autoencoders when running.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the industrial process monitoring method based on a multi-scale slow feature autoencoder as described in any one of claims 1 to 7.