Performance index prediction methods, devices, equipment, and media based on industrial process monitoring

Abstract

This application discloses a method, apparatus, equipment, and medium for predicting performance indicators based on industrial process monitoring, relating to the field of industrial process monitoring technology. The method includes: determining the time-series data of the monitored industrial process; using a gating vector to weightedly fuse the time-domain and frequency-domain features of the time-series data to obtain fused features; processing the fused features using a routing function corresponding to a hybrid expert module based on a preset machine learning model, and determining a target routing expert based on the function output value; and processing the fused features through shared experts and the target routing expert to obtain the predicted performance indicators corresponding to the industrial process. Therefore, by fusing time- and frequency-domain features, the periodicity of the data can be captured, reducing prediction errors; and by using hybrid experts to achieve differentiated processing, with shared experts responsible for learning global common patterns and routing experts providing customized processing, the prediction accuracy and robustness of the performance indicators corresponding to the industrial process can be improved.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring technology, and in particular to methods, devices, equipment and media for predicting performance indicators based on industrial process monitoring. Background Technology

[0002] Soft sensor modeling is a key technology in industrial process monitoring and control. Its core is to indirectly obtain key performance indicators that are difficult to measure through observable variables. However, when the operating mechanism of an industrial process is not yet clear, commonly used mechanism-based modeling methods are often difficult to apply effectively. In real-world complex industrial processes, data distribution changes dynamically over time, and commonly used data-driven soft sensor modeling methods struggle to dynamically adapt to different operating conditions and cannot efficiently handle differences and potential correlations between modes. This results in the prediction accuracy and robustness of key performance indicators in complex multimodal industrial processes failing to meet actual requirements.

[0003] Therefore, how to achieve high-precision and robust prediction of performance indicators in complex multimodal industrial processes is a problem that needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for predicting performance indicators based on industrial process monitoring. This method can capture data periodicity through time-frequency domain feature fusion, thereby reducing prediction errors. Furthermore, it achieves differentiated processing through hybrid experts, with shared experts responsible for learning global common patterns and routing experts providing customized processing, thus improving the prediction accuracy and robustness of corresponding performance indicators for industrial processes. The specific solution is as follows:

[0005] Firstly, this application provides a performance index prediction method based on industrial process monitoring, including:

[0006] Determine the time series data of the monitored industrial processes;

[0007] The time-domain and frequency-domain features of the time-series data are weighted and fused using a gating vector to obtain the corresponding fused features.

[0008] The hybrid expert module, based on a preset machine learning model, processes the fused features using a routing function corresponding to the function output value, and determines the target routing expert based on the function output value; the hybrid expert module consists of several shared experts and several routing experts.

[0009] The fused features are processed using the shared expert and the target routing expert to obtain the performance index prediction results corresponding to the industrial process.

[0010] Optionally, the step of using a gating vector to weight and fuse the time-domain and frequency-domain features of the time-series data to obtain the corresponding fused features includes:

[0011] The time series data is processed based on an attention mechanism to obtain the temporal features of the time series data;

[0012] The divide-and-conquer method based on Fast Fourier Transform projects the time-domain sequence of the time series data into the frequency domain to obtain the frequency domain characteristics of the time series data.

[0013] The time-domain features and the frequency-domain features are weighted and fused using pre-trained gating vectors to obtain the corresponding fused features.

[0014] Optionally, the process of processing the time series data based on the attention mechanism to obtain the temporal features of the time series data includes:

[0015] The time series data is processed separately based on a multi-head attention mechanism to obtain the output results of several attention heads;

[0016] The output results are concatenated, and the concatenated features are linearly mapped to obtain the time-domain features of the time series data.

[0017] Optionally, the divide-and-conquer method based on Fast Fourier Transform projects the time-domain sequence of the time-series data into the frequency domain to obtain the frequency domain features of the time-series data, including:

[0018] The divide-and-conquer method based on Fast Fourier Transform projects the time-domain sequence of the time series data into the frequency domain, and extracts the frequency domain features of the time series data through a frequency domain attention mechanism.

[0019] Optionally, the routing function corresponding to the hybrid expert module based on the preset machine learning model processes the fused features and determines the target routing expert based on the function output value, including:

[0020] By using the routing function corresponding to the hybrid expert module of the preset machine learning model, the fused features are subjected to projection clustering to obtain the function output value representing the expert assignment probability.

[0021] Based on the output value of the function, a second preset number of target routing experts are determined from the routing experts in descending order of expert allocation probability.

[0022] Optionally, the step of processing the fused features using the shared expert and the target routing expert to obtain the performance index prediction result corresponding to the industrial process includes:

[0023] The fused features are processed using the shared expert and the target routing expert to obtain the corresponding processed features;

[0024] The processed features are mapped to a preset output dimension to generate performance index prediction results corresponding to the industrial process.

[0025] Optionally, the training process of the preset machine learning model includes:

[0026] By using two different pre-set data augmentation strategies, the sample features corresponding to the training set are processed respectively to obtain the corresponding first augmented features and second augmented features;

[0027] The first enhanced feature and the second enhanced feature are mapped to the same space, and a sample-level contrastive loss function is constructed by combining feature similarity.

[0028] Projective clustering is performed on the first enhanced feature and the second enhanced feature, and a cluster-level contrastive loss function is constructed by combining the inter-cluster similarity.

[0029] The mean squared error loss function, the sample-level contrast loss function, and the cluster-level contrast loss function are weighted and calculated to obtain the target loss function;

[0030] The target loss function is used to control the model convergence during the training process, so as to train a preset machine learning model.

[0031] Secondly, this application provides a performance index prediction device based on industrial process monitoring, comprising:

[0032] The data determination module is used to determine the time series data of the monitored industrial processes;

[0033] The feature fusion module is used to perform weighted fusion of the time-domain and frequency-domain features of the time-series data using a gating vector to obtain the corresponding fused features.

[0034] The routing expert determination module is used to process the fused features based on the routing function corresponding to the hybrid expert module of the preset machine learning model, and determine the target routing expert according to the function output value; the hybrid expert module consists of several shared experts and several routing experts;

[0035] The prediction module is used to process the fused features through the shared expert and the target routing expert to obtain the performance index prediction results corresponding to the industrial process.

[0036] Thirdly, this application provides an electronic device, comprising:

[0037] Memory, used to store computer programs;

[0038] A processor is used to execute the computer program to implement the performance indicator prediction method based on industrial process monitoring as described above.

[0039] Fourthly, this application provides a method for storing a computer program that, when executed by a processor, implements the above-described method for predicting performance indicators based on industrial process monitoring.

[0040] Therefore, this application first identifies the time-series data of the monitored industrial process; then, it uses a gating vector to weightedly fuse the time-domain and frequency-domain features of the time-series data to obtain the corresponding fused features; next, it processes the fused features based on the routing function corresponding to the hybrid expert module of the preset machine learning model, and determines the target routing expert based on the function output value; the hybrid expert module consists of several shared experts and several routing experts; subsequently, the fused features can be processed by the shared experts and the target routing expert to obtain the performance index prediction results corresponding to the industrial process. In this way, this application, through time-frequency domain feature fusion, captures data periodicity, which can improve the model's ability to model periodicity and global patterns, reducing prediction errors; and through hybrid experts, it achieves differentiated processing of different modal data under different operating conditions of the industrial process, with shared experts responsible for learning global common patterns and routing experts providing customized processing, thereby improving the prediction accuracy and robustness of the performance indicators corresponding to the industrial process. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0042] Figure 1 This is a flowchart of a performance index prediction method based on industrial process monitoring disclosed in this application;

[0043] Figure 2 This is a schematic diagram of a model architecture for predicting industrial process performance indicators disclosed in this application;

[0044] Figure 3 This is a schematic diagram of an attention mechanism disclosed in this application;

[0045] Figure 4 This is a schematic diagram of a comparative clustering model architecture disclosed in this application;

[0046] Figure 5This is a schematic diagram of a performance index prediction device based on industrial process monitoring disclosed in this application.

[0047] Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a performance index prediction method based on industrial process monitoring, including:

[0050] Step S11: Determine the time series data of the monitored industrial process.

[0051] It is understandable that data generated by industrial processes often exhibits high dimensionality and nonlinearity, and in actual operation, multimodal characteristics and distribution drift problems arise due to changes in operating conditions. In predicting the performance indicators of industrial processes, this application first needs to monitor the time-series data of the industrial process, and then mine the time-series evolution characteristics to predict the performance indicators of the industrial process. In a specific embodiment, the monitored data to be processed can be preprocessed and processed according to the time axis to obtain the corresponding time-series data.

[0052] Step S12: Use a gating vector to perform weighted fusion of the time-domain and frequency-domain features of the time series data to obtain the corresponding fused features.

[0053] In this embodiment, during the processing of the time series data of the monitored industrial process by the pre-trained machine learning model, it is necessary to mine the time domain features and frequency domain features of the time series data. Then, the time domain features and frequency domain features of the time series data can be weighted and fused through a gating mechanism, so as to fully mine the periodic characteristics of the data and obtain the corresponding fused features.

[0054] In one specific embodiment, the weighted fusion of the time-domain and frequency-domain features of the time-series data using a gating vector to obtain the corresponding fused features may include: processing the time-series data based on an attention mechanism to obtain the time-domain features; projecting the time-domain sequence of the time-series data into the frequency domain based on a divide-and-conquer method using a Fast Fourier Transform (FFT) to obtain the frequency-domain features; and using a pre-trained gating vector to weightedly fuse the time-domain and frequency-domain features to obtain the corresponding fused features. It is understood that after acquiring time-series data of an industrial process, an attention mechanism can be used to process the time-series data and extract time-domain features; frequency information can be extracted using a Discrete Fourier Transform (DFT), and by combining the divide-and-conquer approach of the FFT and leveraging the symmetry of the rotation factor, the FFT can reduce computational complexity, thus projecting the time-domain sequence into the frequency domain and extracting the corresponding frequency-domain features. Furthermore, to achieve the fusion of time-domain and frequency-domain information, after obtaining the time-domain and frequency-domain features corresponding to the time-series data, a pre-trained gating vector can be introduced. This gating vector is used to achieve weighted feature fusion and obtain the corresponding fused features. Here, the gating vector can automatically adjust the weights of different domains to achieve time-frequency complementarity.

[0055] Furthermore, in a specific embodiment, the process of processing the time series data based on the attention mechanism to obtain the temporal features of the time series data may include: processing the time series data separately based on a multi-head attention mechanism to obtain the output results of several attention heads; concatenating the several output results and performing a linear mapping on the concatenated features to obtain the temporal features of the time series data. Specifically, in the process of extracting temporal features, a multi-head attention mechanism can be used to improve feature representation capability; in the temporal domain, the input time series data is first linearly mapped to obtain the corresponding vector sequence, and then a multi-head scaling dot product attention mechanism can be used for feature update; the output results of several attention heads are concatenated and then linearly mapped to obtain the corresponding temporal representation; the encoder finally outputs the temporal features corresponding to the time series data.

[0056] Accordingly, in another specific embodiment, the divide-and-conquer method based on Fast Fourier Transform (FFT) projects the time-domain sequence of the time series data to the frequency domain to obtain the frequency domain features of the time series data. This can include: projecting the time-domain sequence of the time series data to the frequency domain using a FFT-based divide-and-conquer method, and extracting the frequency domain features of the time series data through a frequency domain attention mechanism. Specifically, in the frequency domain, frequency information can be extracted using Discrete Fourier Transform (DFT), while the divide-and-conquer approach based on FFT can reduce the computational complexity of FFT; this transform can be used to project the time-domain sequence to the frequency domain; then, frequency domain representation is extracted through frequency domain attention, ultimately obtaining the frequency domain features corresponding to the time series data.

[0057] Step S13: The routing function corresponding to the hybrid expert module based on the preset machine learning model processes the fused features and determines the target routing expert according to the function output value; the hybrid expert module consists of several shared experts and several routing experts.

[0058] In this embodiment, the time-domain and frequency-domain features of the time-series data can be weighted and fused through the above steps. Then, the weighted and fused features can be processed by a hybrid expert module with a preset machine learning model. It is understood that, in a specific embodiment, the preset machine learning module can directly process the time-series data. After the time-frequency domain features are fused, they directly enter the hybrid expert module, where shared experts and several related routing experts process the fused features. Specifically, the fused features can be processed through the routing function corresponding to the hybrid expert module to determine several routing experts suitable for processing the current time-series data, denoted as target routing experts.

[0059] In a specific embodiment, the routing function corresponding to the hybrid expert module based on the preset machine learning model processes the fused features and determines the target routing experts based on the function output value. This can include: performing projection clustering on the fused features using the routing function corresponding to the hybrid expert module of the preset machine learning model to obtain a function output value representing the expert allocation probability; and determining a second preset number of target routing experts from the plurality of routing experts according to the function output value, in descending order of expert allocation probability. Specifically, in the process of selecting suitable routing experts, projection clustering can be performed on the fused features, and several target routing experts suitable for the current time series data can be determined from several routing experts based on the clustering results. It can be seen that the routing function can combine the feature clustering results to select suitable routing experts.

[0060] Furthermore, in a specific embodiment, the training process of the preset machine learning model may include: processing the sample features corresponding to the training set using two different pre-set data augmentation strategies to obtain corresponding first augmented features and second augmented features; mapping the first augmented features and the second augmented features to the same space and constructing a sample-level contrastive loss function by combining feature similarity; performing projection clustering on the first augmented features and the second augmented features and constructing a cluster-level contrastive loss function by combining inter-cluster similarity; performing weighted calculations on the mean squared error loss function, the sample-level contrastive loss function, and the cluster-level contrastive loss function to obtain a target loss function; and using the target loss function to control the model convergence during the training process to train the preset machine learning model. Specifically, in the training process of expert routing, two different data augmentation strategies can be used to augment the samples. For any augmented sample, except for its corresponding other augmented sample, all other samples are considered negative samples; that is, the first augmented feature alone and its corresponding second augmented feature are considered positive samples, and all other augmented features are considered negative samples. Subsequently, the samples can be mapped to a one-dimensional latent space using a feature transformer. Further, the projected values ​​are compared to the latent space using a fully connected layer, yielding two types of samples. Combining feature similarity and traversing all enhanced samples, a sample-level contrastive loss function can be constructed in this space. Building upon this, a cluster-level projection head can project the samples to a clustering probability space, with different clusters corresponding to different routing experts. Further combining inter-cluster similarity, contrastive loss calculations can be performed on all clusters under the two enhancement strategies, constructing a cluster-level contrastive loss function. During the training of the machine learning model, mean squared error can be used as the regression loss function and as a convergence metric to control the model training process. To reflect the expert routing effect, the sample-level and cluster-level contrastive loss functions corresponding to the expert routing process can be combined for model training. Specifically, a hyperparameter adjusting the weight of the routing loss term can be introduced to balance the optimization objectives between the main task and the routing task. Then, these three loss functions can be weighted and calculated, and the corresponding target loss function can be used to control model convergence, thereby training the preset machine learning model.

[0061] In a specific embodiment, the process of constructing the cluster-level contrastive loss function may include: constructing the clustering loss corresponding to each cluster based on inter-cluster similarity; and constructing the cluster-level contrastive loss function by combining the corresponding clusters and their respective clustering losses with the entropy of the cluster distribution, based on the clusters' inter-cluster similarity. It is understood that the cluster-level contrastive loss can be constructed based on inter-cluster similarity, involving temperature parameters related to the inter-cluster contrastive loss. During the contrastive loss calculation for all clusters, the uniformity of the clustering results can be constrained by the entropy of the cluster distribution. This prevents the model from falling into trivial solutions where all samples are assigned to the same cluster, thereby improving the diversity and discriminative power of expert routing.

[0062] Step S14: The fused features are processed by the shared expert and the target routing expert to obtain the performance index prediction results corresponding to the industrial process.

[0063] In this embodiment, the above steps can be used to determine the target routing expert in the hybrid expert module that can process the current time series data; then, by sharing the experts and selecting the target routing expert, the fused features can be processed to finally obtain the performance index prediction results corresponding to the industrial process.

[0064] In a specific embodiment, processing the fused features using the shared expert and the target routing expert to obtain the performance index prediction result corresponding to the industrial process may include: processing the fused features using the shared expert and the target routing expert to obtain corresponding processed features; mapping the processed features to a preset output dimension to generate the performance index prediction result corresponding to the industrial process. Specifically, to reduce the computational complexity of the model, the decoding part can be omitted, and after encoding, a projection layer can be directly introduced to map the features to the target output dimension to generate the prediction result of the key performance indicators; that is, after processing the relevant fused features through the hybrid expert module, the processed features can be mapped to a preset output dimension, thus generating the final performance index prediction result corresponding to the industrial process.

[0065] Therefore, this application can improve the model's ability to model periodic and global patterns and reduce prediction errors by capturing data periodicity through time-frequency domain feature fusion. Furthermore, by using hybrid experts to achieve differentiated processing of different modal data under different operating conditions of industrial processes, with shared experts responsible for learning global common patterns and routing experts providing customized processing, the prediction accuracy and robustness of corresponding performance indicators of industrial processes can be improved.

[0066] like Figure 2As shown in the embodiments of this application, the architecture of the machine learning model corresponding to the performance index prediction method based on industrial process monitoring is disclosed. It can be seen that the machine learning model adopts a modular design, involving a frequency domain enhanced attention mechanism and a contrastive routing hybrid expert system. Specifically, it includes:

[0067] In this embodiment, the main body of the model is based on the Transformer encoder structure, and a time-frequency fusion attention mechanism is further introduced to enhance the model's ability to model periodic and global patterns. This mechanism integrates time-domain multi-head self-attention, frequency-domain attention, and gating fusion modules, which can effectively improve the model's global temporal modeling capability. Furthermore, a hybrid expert module driven by contrastive clustering is set up to replace the fully connected layer in the Transformer, realizing dynamic matching between data and experts to improve the model's adaptability in multimodal data scenarios. Further, to reduce the model's computational complexity, the decoder part of the standard Transformer model is discarded, and a projection layer is directly introduced after the encoder to map features to the target output dimension to generate prediction results for key performance indicators.

[0068] To compensate for the shortcomings of standard attention mechanisms in modeling sequence position information, fixed-position encoding can be introduced and added to the embedding results to provide explicit time-series awareness. In a specific embodiment, the raw data needs to be preprocessed. Specifically, the raw data to be processed can be converted into CSV or NumPy array format (one row per sample or file), with each sample corresponding to a time window (e.g., 50 seconds). The rows are arranged in order of sensor channels, and the last column is the label (0 = normal, 1-6 = fault type). After the data is read, it is first normalized by min-max (Min-Max Normalization, deviation standardization) (by sensor channel), and then truncated by window (e.g., every 50 seconds). The normalized sequence is flattened along the time axis, and the fault label is appended at the end.

[0069] like Figure 3 As shown, the original input sequence After being mapped to the D-dimensional hidden representation space through the embedding layer, we obtain In the time domain, the input time series is first mapped linearly to obtain query, key, and value vectors, denoted as Q, K, and V, respectively. Then, a scaled dot product attention mechanism is used for feature updates, the calculation formula of which is as follows:

[0070] ;

[0071] Where d kLet be the dimension of the key vector. To enhance the model's expressive power, a multi-head attention mechanism can be introduced. The outputs of h attention heads are concatenated and linearly mapped to obtain the temporal representation:

[0072] ;

[0073] ;

[0074] in, , , These are the projection matrices of the i-th attention head, respectively, ultimately yielding the temporal output representation of the encoder. .

[0075] In the frequency domain, the model first extracts frequency information through the Discrete Fourier Transform (DFT). For a one-dimensional time series of length L, x=[x1, x2, ..., x...] L-1 Its DFT expression is as follows:

[0076] ;

[0077] Among them, X k Let be the complex coefficient of the k-th frequency component, and j represent the imaginary unit. Based on the divide-and-conquer approach of the Fast Fourier Transform (FFT), the above equation can be further decomposed into:

[0078] ;

[0079] By leveraging the symmetry of the twiddle factor, FFT reduces the computational complexity from O(L...) 2 The time domain can be reduced to O(L log L). Using this transformation, time-domain sequences can be projected into the frequency domain:

[0080] ;

[0081] in This represents the query, key, and value in the frequency domain. In a single-sided spectrum, the first element is typically retained. Each frequency component. Frequency domain attention can be calculated as follows:

[0082] ;

[0083] in This indicates the conjugate transpose. To prevent the stability constant from being divided by zero, the frequency domain representation is then restored to the time domain using an inverse Fourier transform, yielding... At the same time, it preserves the global information of frequency domain enhancement.

[0084] To achieve the fusion of time-domain and frequency-domain information, a trainable gating vector is introduced. To achieve weighted fusion of features:

[0085] ;

[0086] This mechanism can automatically adjust the weights of features from different domains during training to achieve time-frequency complementary information modeling.

[0087] Furthermore, the machine learning model includes a conditional computation module based on MoE (Mixture of Experts), which helps the model better model complex multimodal industrial processes. For a given sample, an appropriate expert can be selected for computation based on the operating mode. Figure 1 As shown, MoE is composed of N s A shared expert and N r The system consists of several routing experts. Shared experts are responsible for learning global common patterns, while routing experts provide customized modeling based on the modal features of each sample. For each sample's feature vector z, the output can be represented as:

[0088] ;

[0089] in An expert model is used to transform input features. The value of the j-th element of the routing function represents the j-th expert E. j The weight, This represents the output of the hybrid expert layer. To improve the sparsity and specificity of the selection, the KeepTopK strategy can be used to retain only the K largest routing experts by weight:

[0090] ;

[0091] ;

[0092] That is, after selecting TopK experts for each sample based on the output value of the routing function, the Softmax function is used to normalize their weights.

[0093] Furthermore, in the MoE structure, a routing network is typically required to dynamically select appropriate experts for feature extraction based on the input data. Considering contrastive clustering, a contrastive routing strategy can be formulated based on contrastive learning, such as... Figure 4 As shown, each sample has a feature representation z. First, two different data augmentation strategies T are applied to the feature space. ɑ and T b Generate the enhanced representation respectively and For B samples in a batch, 2B augmented representations can be obtained. For any augmented sample In addition to its corresponding augmented sample Aside from the positive samples, the remaining 2B-2 samples are treated as negative samples. Subsequently, the samples are mapped to a one-dimensional latent space representation h by a feature transformer G(∙), i.e. Adding a nonlinear projection head before the contrast loss can mitigate the information loss problem. Therefore, it can be further projected into the contrast space through a fully connected layer to obtain... A sample-level contrastive loss function is constructed within this space. The similarity between samples is defined as follows:

[0094] ;

[0095] in, ,and To enhance the consistency of representation of the same sample under different enhancement methods, for the sample... Its contrast loss is defined as:

[0096] ;

[0097] in, The temperature coefficient of the sample-level contrastive loss is used to adjust the smoothness of the distribution. Finally, by traversing all augmented samples, the overall sample-level contrastive loss is defined as:

[0098] ;

[0099] Based on this, further clustering-level projection head P c (∙) Project the samples onto the clustering probability space to obtain .in, Indicates sample n in augmentation mode The probability of a sample being assigned to the m-th cluster is the probability that the sample will be processed by the m-th expert. This allows each expert to focus on processing samples of a specific modality, thus forming a functional division of labor among experts.

[0100] Ideally, each sample should be assigned to only one cluster, i.e. and Each row should approximate a one-hot vector. Let... express The representation vector of the i-th cluster under the augmentation method has a corresponding positive cluster under another augmentation method b, and the rest are 2N. r - Two clusters are considered negative samples. The similarity between clusters is defined as follows:

[0101] ;

[0102] in, ,and For clusters The clustering contrast loss is defined as follows:

[0103] ;

[0104] in, The temperature parameter represents the inter-cluster contrast loss. By calculating the contrast loss for all clusters under both enhancement methods, the overall expression for the cluster contrast loss is:

[0105] ;

[0106] Where H(Y) is the entropy of the cluster distribution in the enhanced samples, used to constrain the uniformity of the clustering results, and is defined as follows:

[0107] ;

[0108] in, This entropy regularization term can effectively prevent the model from getting stuck in trivial solutions where all samples are assigned to the same cluster, thereby improving the diversity and discriminativeness of expert routes.

[0109] During model training, mean squared error is generally used as the regression loss function, defined as follows:

[0110] ;

[0111] in, and These represent the predicted and actual values ​​of the j-th variable at time i, respectively. To achieve an effective expert routing mechanism, the aforementioned sample-level contrastive loss can be used. Comparative loss with clustering level This is incorporated into the overall objective function. The complete training loss function is defined as follows:

[0112] ;

[0113] in, To adjust the hyperparameters of the routing loss term weights, and to balance the optimization objectives between the main task and the routing task.

[0114] Therefore, this solution introduces structural optimization and information fusion mechanisms, enabling the model to more effectively extract time-series dependencies and potential modal distributions from industrial process data, thereby improving the prediction accuracy and robustness of soft sensing. Combined with a comparative clustering-driven hybrid expert module, it can automatically identify data of different modalities and assign appropriate experts to each data point, better capturing the changing characteristics of data under different operating conditions in industrial processes.

[0115] like Figure 5 As shown in the figure, this application discloses a performance index prediction device based on industrial process monitoring, comprising:

[0116] Data determination module 11 is used to determine the time series data of the monitored industrial process;

[0117] Feature fusion module 12 is used to perform weighted fusion of the time-domain features and frequency-domain features of the time series data using a gate vector to obtain the corresponding fused features;

[0118] The routing expert determination module 13 is used to process the fused features based on the routing function corresponding to the hybrid expert module of the preset machine learning model, and determine the target routing expert according to the function output value; the hybrid expert module consists of several shared experts and several routing experts;

[0119] The prediction module 14 is used to process the fused features through the shared expert and the target routing expert to obtain the performance index prediction results corresponding to the industrial process.

[0120] Therefore, this application can improve the model's ability to model periodic and global patterns and reduce prediction errors by capturing data periodicity through time-frequency domain feature fusion. Furthermore, by using hybrid experts to achieve differentiated processing of different modal data under different operating conditions of industrial processes, with shared experts responsible for learning global common patterns and routing experts providing customized processing, the prediction accuracy and robustness of corresponding performance indicators of industrial processes can be improved.

[0121] In one specific embodiment, the feature fusion module 12 may include:

[0122] The time-domain feature processing submodule is used to process the time-series data based on an attention mechanism to obtain the time-domain features of the time-series data;

[0123] The frequency domain feature processing submodule is used to project the time-domain sequence of the time series data into the frequency domain based on the divide-and-conquer method of fast Fourier transform to obtain the frequency domain features of the time series data.

[0124] The feature fusion unit is used to perform weighted fusion of the time-domain features and the frequency-domain features using pre-trained gating vectors to obtain the corresponding fused features.

[0125] In another specific embodiment, the time-domain feature processing submodule may include:

[0126] The temporal attention processing unit is used to process the time series data separately based on a multi-head attention mechanism to obtain the output results of several attention heads;

[0127] The time-domain feature processing unit is used to concatenate several of the output results and perform linear mapping on the concatenated features to obtain the time-domain features of the time series data.

[0128] In yet another specific embodiment, the frequency domain feature processing submodule may include:

[0129] The frequency domain feature processing unit is used to project the time-domain sequence of the time series data into the frequency domain based on the divide-and-conquer method of fast Fourier transform, and extract the frequency domain features of the time series data through the frequency domain attention mechanism.

[0130] In one specific embodiment, the routing expert determination module 13 may include:

[0131] The projection clustering unit is used to perform projection clustering on the fused features through the routing function corresponding to the hybrid expert module of the preset machine learning model, so as to obtain the function output value representing the expert assignment probability.

[0132] The routing expert determination unit is used to determine a second preset number of target routing experts from a plurality of routing experts according to the output value of the function and in a manner that the expert allocation probability is arranged from large to small.

[0133] In one specific embodiment, the prediction module 14 may include:

[0134] The feature processing unit is used to process the fused features using the shared expert and the target routing expert to obtain the corresponding processed features;

[0135] The prediction result generation unit is used to map the processed features to a preset output dimension to generate the performance index prediction results corresponding to the industrial process.

[0136] In one specific embodiment, the device may include:

[0137] The data augmentation module is used to process the sample features corresponding to the training set through two different pre-set data augmentation strategies to obtain the corresponding first augmented features and second augmented features.

[0138] The first loss function construction module is used to map the first enhanced feature and the second enhanced feature to the same space, and combine feature similarity to construct a sample-level contrast loss function;

[0139] The second loss function construction module is used to perform projection clustering on the first enhanced features and the second enhanced features, and combine the inter-cluster similarity to construct a cluster-level contrastive loss function;

[0140] The third loss function construction module is used to perform weighted calculations on the mean squared error loss function, the sample-level contrast loss function, and the cluster-level contrast loss function to obtain the target loss function;

[0141] The model training module is used to control the model convergence during the training process using the target loss function, so as to train a preset machine learning model.

[0142] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0143] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the performance index prediction method based on industrial process monitoring disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0144] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0145] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0146] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the performance index prediction method based on industrial process monitoring executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0147] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned performance index prediction method based on industrial process monitoring. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0148] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0149] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0150] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0151] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0152] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A performance indicator prediction method based on industrial process monitoring, characterized in that, The method comprises the steps of: determining time series data of a monitored industrial process; performing weighted fusion on time domain features and frequency domain features of the time series data by using a gating vector to obtain corresponding fused features; processing the fused features by a routing function corresponding to a mixed expert module of a preset machine learning model, and determining a target routing expert according to a function output value; processing the fused features by the shared expert and the target routing expert to obtain a performance index prediction result corresponding to the industrial process; wherein the training process of the preset machine learning model comprises: processing sample features corresponding to a training set by two different data enhancement strategies set in advance to obtain corresponding first enhanced features and second enhanced features; mapping the first enhanced features and the second enhanced features to the same space, and combining feature similarity to construct a sample-level contrast loss function; projecting and clustering the first enhanced features and the second enhanced features, and combining inter-cluster similarity to construct a clustering-level contrast loss function; performing weighted calculation on a mean square error loss function, the sample-level contrast loss function and the clustering-level contrast loss function to obtain a target loss function; controlling model convergence of the training process by using the target loss function to train the preset machine learning model.

2. The performance indicator prediction method based on industrial process monitoring according to claim 1, characterized in that, The method comprises the steps of: processing the time series data based on an attention mechanism to obtain time domain features of the time series data; projecting time domain sequences of the time series data to a frequency domain based on a divide-and-conquer method of fast Fourier transform to obtain frequency domain features of the time series data; performing weighted fusion on the time domain features and the frequency domain features by using a pre-trained gating vector to obtain corresponding fused features.

3. The performance indicator prediction method based on industrial process monitoring according to claim 2, characterized in that, The method comprises the steps of: processing the time series data based on a multi-head attention mechanism to obtain output results of a plurality of attention heads; splicing a plurality of the output results, and performing linear mapping on the spliced features to obtain time domain features of the time series data.

4. The performance indicator prediction method based on industrial process monitoring according to claim 2, characterized in that, The method comprises the steps of: projecting time domain sequences of the time series data to a frequency domain based on a divide-and-conquer method of fast Fourier transform, and extracting frequency domain features of the time series data by a frequency domain attention mechanism.

5. The performance indicator prediction method based on industrial process monitoring according to claim 1, characterized in that, The method comprises the steps of: projecting time domain sequences of the time series data to a frequency domain based on a divide-and-conquer method of fast Fourier transform, and extracting frequency domain features of the time series data by a frequency domain attention mechanism. The method comprises the steps of: projecting and clustering the fused features by a routing function corresponding to a mixed expert module of a preset machine learning model to obtain a function output value representing expert allocation probability; According to the function output value, a second preset number of target routing experts are determined from the routing experts in a descending order of expert allocation probability.

6. The performance indicator prediction method based on industrial process monitoring according to claim 1, characterized in that, The processing of the fused features by the shared experts and the target routing experts to obtain the performance index prediction result corresponding to the industrial process comprises: The processing of the fused features by the shared experts and the target routing experts to obtain the corresponding processed features; Mapping the processed features to a preset output dimension to generate the performance index prediction result corresponding to the industrial process.

7. An apparatus for predicting a performance indicator based on monitoring of an industrial process, characterized in that Comprise: The data determination module is configured to determine time series data of a monitored industrial process; The feature fusion module is configured to perform weighted fusion on time domain features and frequency domain features of the time series data using a gating vector to obtain corresponding fused features; The routing expert determination module is configured to process the fused features based on a routing function of a mixed expert module of a preset machine learning model, and determine a target routing expert according to a function output value; the mixed expert module comprises a plurality of shared experts and a plurality of routing experts; The prediction module is configured to process the fused features by the shared experts and the target routing experts to obtain a performance index prediction result corresponding to the industrial process; The device comprises: The data enhancement module is configured to process sample features corresponding to a training set by two different data enhancement strategies to obtain corresponding first enhanced features and second enhanced features; The first loss function construction module is configured to map the first enhanced features and the second enhanced features to the same space, and construct a sample-level contrastive loss function in combination with feature similarity; The second loss function construction module is configured to project and cluster the first enhanced features and the second enhanced features, and construct a cluster-level contrastive loss function in combination with inter-cluster similarity; The third loss function construction module is configured to perform weighted calculation on a mean square error loss function, the sample-level contrastive loss function and the cluster-level contrastive loss function to obtain a target loss function; The model training module is configured to control model convergence of a training process using the target loss function to train a preset machine learning model.

8. An electronic device, comprising: Comprise: A memory for saving a computer program; A processor for executing the computer program to implement the performance index prediction method based on industrial process monitoring according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A memory for saving a computer program; the computer program is executed by a processor to implement the performance index prediction method based on industrial process monitoring according to any one of claims 1 to 6. A memory for saving a computer program; the computer program is executed by a processor to implement the performance index prediction method based on industrial process monitoring according to any one of claims 1 to 6.

Images