Industrial process key performance index prediction method, device and system, and storage medium
By constructing a training data set for a single target domain of multi-source domain in the industrial process, and using a multi-source domain model clustering algorithm with federated aggregation and dynamic weights, the problem of inconsistent data distribution between source domain and target domain is solved, the generalization and adaptability of the model is improved, and efficient prediction of key performance indicators is achieved.
Patent Information
- Application Number
- CN202411558084.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-02-25
AI Technical Summary
In industrial processes, the data distribution of the source and target domains is inconsistent, resulting in the performance of traditional federated learning methods in predictions, and data privacy protection and data silos make it difficult to construct high-precision prediction models.
The training data set of the single target domain of multi-source domain is adopted to build a multi-source domain local model, and a transfer learning model suitable for the target domain is generated through federated aggregation. Combined with the multi-source domain model clustering algorithm with dynamic weights, the performance of the model in the target domain is optimized.
The generalization and adaptability of the model are improved, ensuring effectiveness and reliability in practical applications, and being able to effectively predict key performance indicators in the target domain.
Smart Images
Figure CN119513531A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of industrial process soft measurement, and in particular relates to an industrial process key performance indicator prediction method and device, system, and storage medium. Background Art
[0002] Industrial processes are the core of modern manufacturing and production, and their efficient operation directly affects production efficiency, product quality and corporate competitiveness. The key performance indicators of the operation process can reflect the health status and operating efficiency of the equipment. Real-time monitoring of key performance indicators can effectively ensure the stability of the industrial process, avoid downtime and capacity loss caused by equipment failure, and provide reliable data support for quality management, help optimize and adjust the production process, thereby improving the overall quality and consistency of the product. Therefore, with the help of advanced monitoring technology and data analysis tools, comprehensive and accurate real-time monitoring of key performance indicators has become a key strategy that cannot be ignored in modern industrial process management, and is also an important foundation for promoting enterprises to achieve intelligent manufacturing and sustainable development.
[0003] In actual production, the collection of key performance indicators of industrial processes faces the problem of data sparsity, especially the difficulty in obtaining labeled data. Since key performance indicators often require manual labeling, which is time-consuming, labor-intensive and costly, the labeled data that can be obtained is very limited. Such a small amount of data is difficult to support the independent construction of a high-precision prediction model, resulting in limited accuracy and generalization of the model. In addition, data between different factories, production equipment and enterprises are usually managed in a closed manner. Due to the restrictions of privacy protection, commercial competition and laws and regulations, data cannot be freely shared between different entities. This data island phenomenon further exacerbates the difficulty of building an efficient prediction model. Federated learning provides an innovative solution for this. It allows all parties to jointly train the model while protecting data privacy by sharing model parameters rather than raw data between multiple users. This can not only break data silos and achieve multi-party collaboration, but also avoid data leakage or illegal use, and ultimately provide the possibility of building a more accurate prediction model.
[0004] Traditional federated learning methods assume that the data distribution of all users is consistent, but due to differences in equipment, process conditions, operators, etc., the data distribution of different source domains may be inconsistent. Directly using the source domain model to predict the target domain usually leads to performance degradation due to distribution differences. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a method and device, system, and storage medium for predicting key performance indicators of industrial processes, which can achieve privacy-preserving prediction and improve the generalization and adaptability of the model when the data distribution in the source domain and the target domain are inconsistent.
[0006] To achieve the above object, the present invention adopts the following technical solution:
[0007] A method for predicting a key performance indicator of an industrial process, comprising:
[0008] Step 1: Construct a training dataset of multiple source domains and a single target domain;
[0009] Step 2: Build a multi-source domain local model based on the training data set;
[0010] Step 3: Federate and aggregate the local models of multiple source domains;
[0011] Step 4: Migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain;
[0012] Step 5: Distribute the target domain model to multiple source domains and update the model based on the local data sets of the multiple source domains;
[0013] Step 6: Repeat the above steps 3-5 until the total loss function converges, the training is completed, and the trained COD prediction model is obtained;
[0014] Step 7: Use the trained COD prediction model to predict COD in the target domain.
[0015] As a preferred method, four process flows with activated sludge as the core are selected, three of which are source domains and one is the target domain. The data set is constructed and recorded as {S1, S2, S3}. The training data set of each source domain is in, It represents the input variables that can be collected. Inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, total nitrogen concentration and other indicators are selected as input process variables. represents the corresponding key performance indicator value, and the chemical oxygen demand (COD) of the effluent is selected; the sewage treatment process is the target domain, denoted as T, in, The samples representing the target domain are scarce, and all source domains and target domains are assumed to have different distributions, that is, when i≠j,
[0016] As a preferred method, samples in the target domain are collected online, denoted as The process data Input COD prediction model F T* , get the predicted value for:
[0017]
[0018] The present invention also provides an industrial process key performance indicator prediction device, comprising:
[0019] The first processing module is used to construct a training data set of multiple source domains and a single target domain;
[0020] The second processing module is used to construct a multi-source domain local model based on the training data set;
[0021] The third processing module is used to perform federated aggregation on the local models of multiple source domains;
[0022] The fourth processing module is used to migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain;
[0023] a fifth processing module, configured to distribute the target domain model to multiple source domains and update the model based on the local data sets of the multiple source domains;
[0024] The sixth processing module is used to complete the training until the total loss function converges and obtain the trained COD prediction model;
[0025] The seventh processing module is used to use the trained COD prediction model to predict COD in the target domain.
[0026] As a preferred method, four process flows with activated sludge as the core are selected, three of which are source domains and one is the target domain. The data set is constructed and recorded as {S1, S2, S3}. The training data set of each source domain is in, It represents the input variables that can be collected. Inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, total nitrogen concentration and other indicators are selected as input process variables. represents the corresponding key performance indicator value, and the chemical oxygen demand (COD) of the effluent is selected; the sewage treatment process is the target domain, denoted as T, in, The samples representing the target domain are scarce, and all source domains and target domains are assumed to have different distributions, that is, when i≠j,
[0027] As a preferred embodiment, the seventh module is used to collect samples in the target domain online, denoted as The process data Input COD prediction model F T* , get the predicted value for:
[0028]
[0029] An embodiment of the present invention further provides an industrial process key performance indicator prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes an industrial process key performance indicator prediction method when executed by the processor.
[0030] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the method for predicting key performance indicators of industrial processes when running.
[0031] In the current context of increasing attention to data privacy protection, the present invention proposes an effective domain difference measurement method that can evaluate the contribution of each data source to the target task. Based on the analysis of this domain difference, a multi-source domain model clustering algorithm based on dynamic weights is designed to effectively integrate the characteristics of each source domain model to generate a model that is more in line with the characteristics of the target domain. The design of this dynamic weight not only takes into account the importance of different source domains, but also can be automatically adjusted during the training process, making the aggregated model more adaptable and robust. Finally, the step of fine-tuning the target domain model further optimizes the performance of the model in the target domain. By combining a small amount of labeled data in the target domain, the prediction accuracy and generalization ability of the model are improved, ensuring its effectiveness and reliability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0033] Figure 1 is a flow chart of a method for predicting key performance indicators of an industrial process according to an embodiment of the present invention;
[0034] Figure 2 This is the network architecture diagram of this instance;
[0035] Figure 3 This is the sewage treatment process flow chart used in this example, including 3 source domains and 1 target domain;
[0036] Figure 4 These are the 48 input process variables and 1 output variable selected in this example;
[0037] Figure 5 It is the prediction result diagram of the present invention in this example. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0040] Embodiment 1:
[0041] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting key performance indicators of an industrial process, comprising:
[0042] Step 1: Construct a training dataset of multiple source domains and a single target domain;
[0043] In step 1, taking the sewage treatment process as the background, four process flows with activated sludge method as the core are selected, three of which are source domains and one is the target domain, such as Figure 3 As shown. Construct a data set, denoted as {S1, S2, S3}, and the training data set of each source domain is in, Indicates the input variables that can be collected. In this example, 48 process variables are selected. This example selects indicators such as inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, and total nitrogen concentration as input process variables. For details, see Figure 4 . Indicates the corresponding key performance indicator value. In this example, the chemical oxygen demand (COD) of the effluent is selected. The input process variables and output key performance indicators are selected based on the actual collection requirements and mechanism experience of the sewage treatment process. The sewage treatment process to be modeled is the target domain, denoted by T, in The samples representing the target domain are scarce and it is impossible to build a prediction model independently. Assume that all source domains and target domains have different distributions, that is, when i≠j,
[0044] Step 2: Given the model structure, build a multi-source domain local model;
[0045] In step 2, it is assumed that a general neural network model structure F = {F en ,F de ,F pr}, where F en Indicates the encoder, Fde Denotes the decoder, F pr Represents the predictor, see Figure 2 Each wastewater treatment process (i.e. each source domain) has a local model The model is initialized and trained on the local dataset of each source domain. For the i-th source domain, based on its local dataset The training loss function can be defined as:
[0046]
[0047] in, Represents 48 input process variables such as influent flow, effluent flow, etc. in the sewage treatment process Pass through the encoder Decoder The resulting reconstructed input, represents the reconstruction error, Input variables Pass through the encoder Predictor The predicted value corresponding to the key indicator COD is obtained. Represents the prediction error. l(·) represents the loss function, which is usually calculated using the mean square error function.
[0048] Step 3: Federate and aggregate local models from multiple source domains;
[0049] In step 3, when the local models of multiple source domains are aggregated into the target domain model, in order to make the aggregated model more suitable for the target domain, a dynamic weight clustering method based on inter-domain similarity is adopted, that is, the more similar the source domain and the target domain are, the greater the contribution of the source domain model to the target domain model. Therefore, the key to the dynamic weight is to measure the similarity between each source domain and the target domain. However, due to the protection of data privacy, the original data cannot be directly obtained, and the previous method of directly calculating the difference between the source domain and the target domain based on data or features is not feasible. Considering that only the parameters of the source domain model can be obtained, the present invention designs a reconstruction error matching method to indirectly evaluate the similarity between the source domain and the target domain. The basic idea is that if the data distribution between the source domain and the target domain is similar, the encoder and decoder of the source domain trained on the target domain should be able to effectively reconstruct the data in the target domain. Specifically, the encoder and decoder trained well in the i-th sewage treatment process (source domain) are transmitted to the sewage treatment process to be predicted (target domain), and on this basis, the reconstructed input corresponding to the 48 process variables in the target domain is obtained. The reconstruction error It can be calculated as:
[0050]
[0051] The dynamic weight α is derived using the reconstruction error of the autoencoder i As the similarity measure between the i-th source domain and the target domain, that is:
[0052]
[0053] Where softmax(·) represents a normalization function to ensure Each source domain model The parameters are recorded as Then, the aggregated target domain model parameters are:
[0054]
[0055] Step 4: Migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain;
[0056] In step 4, the aggregated target domain model F T Fine-tune in the target domain to better adapt to the target domain. First, F T As the initial model, a small amount of labeled data in the target domain is then used for supervised learning and fine-tuning. The supervised loss function is:
[0057]
[0058] in They represent the input process variables (influent flow, effluent flow, etc.) and the output process variables (effluent COD) in the sewage treatment process to be predicted. The model is fine-tuned using the standard gradient descent method or its variants (such as the Adam optimizer). The update formula is:
[0059]
[0060] Where η represents the learning rate, t represents the number of rounds of fine-tuning, represents the gradient. Since there are few labeled data in the target domain, the fine-tuning process may be prone to overfitting. In order to avoid overfitting, an early stopping mechanism can be introduced. When the validation set loss no longer decreases significantly, the training is stopped. After the above supervised fine-tuning, the transfer learning model F suitable for the target domain is obtained. T* .
[0061] Step 5: Distribute the target domain model to multiple source domains and update the model based on the local datasets of the multiple source domains;
[0062] In step 5, the fine-tuned target domain model parameter φ T* Distribute back to each source domain, each source domain uses a local dataset The model is further updated. This process updates the model parameters of each user through local training:
[0063]
[0064] Step 6: Repeat the above steps 3-5 until the total loss function converges and the training is completed.
[0065] In step 6, through multiple rounds of federated aggregation and fine-tuning, in each iteration, the updated local model is federated again and migrated to the target domain again. This process is repeated until the global loss function converges. The global loss function is defined as:
[0066]
[0067] when When the preset threshold is reached or the change is small enough, the training ends. The trained COD prediction model F suitable for the target domain is obtained. T* .
[0068] Step 7: Use the offline trained F T* Used for prediction of COD in the target domain.
[0069] In step 7, samples in the target domain are first collected online. The collected process variables should be consistent with the 48 process variables in step 1, which are recorded as Then the process data Input COD prediction model F T* , get the predicted value for:
[0070]
[0071] Then, the root mean square error and R are introduced. 2 As a quantitative evaluation indicator of prediction accuracy. It is defined as:
[0072]
[0073] Among them, y i , They represent the true value, predicted value and average value of COD in the test data set. The experimental results are RMSE=72.35, R 2 =0.8067, the visual prediction results are as follows Figure 5 As shown, the effectiveness of the present invention is demonstrated.
[0074] Embodiment 2:
[0075] The embodiment of the present invention further provides an industrial process key performance indicator prediction device, comprising:
[0076] The first processing module is used to construct a training data set of multiple source domains and a single target domain;
[0077] The second processing module is used to construct a multi-source domain local model based on the training data set;
[0078] The third processing module is used to perform federated aggregation on the local models of multiple source domains;
[0079] The fourth processing module is used to migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain;
[0080] a fifth processing module, configured to distribute the target domain model to multiple source domains and update the model based on the local data sets of the multiple source domains;
[0081] The sixth processing module is used to complete the training until the total loss function converges and obtain the trained COD prediction model;
[0082] The seventh processing module is used to use the trained COD prediction model to predict COD in the target domain.
[0083] As an implementation method of the present invention, four process flows with activated sludge method as the core are selected, three of which are source domains and one is the target domain, and a data set is constructed, which is recorded as {S1, S2, S3}. The training data set of each source domain is in, It represents the input variables that can be collected. Inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, total nitrogen concentration and other indicators are selected as input process variables. represents the corresponding key performance indicator value, and the chemical oxygen demand (COD) of the effluent is selected; the sewage treatment process is the target domain, denoted as T, in, The samples representing the target domain are scarce, and all source domains and target domains are assumed to have different distributions, that is, when i≠j,
[0084] As an implementation method of the embodiment of the present invention, the seventh module is used to collect samples in the target domain online, denoted as The process data Input COD prediction model F T* , get the predicted value for:
[0085]
[0086] Embodiment 3:
[0087] An embodiment of the present invention further provides an industrial process key performance indicator prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes an industrial process key performance indicator prediction method when executed by the processor.
[0088] Embodiment 4:
[0089] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the method for predicting key performance indicators of industrial processes when running.
[0090] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for predicting key performance indicators of industrial processes, characterized in that: include: Step 1: Construct a training dataset of multiple source domains and a single target domain; Step 2: Build a multi-source domain local model based on the training data set; Step 3: Federate and aggregate the local models of multiple source domains; Step 4: Migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain; Step 5: Distribute the target domain model to multiple source domains and update the model based on the local data sets of the multiple source domains; Step 6: Repeat the above steps 3-5 until the total loss function converges, the training is completed, and the trained COD prediction model is obtained; Step 7: Use the trained COD prediction model to predict COD in the target domain.
2. The method for predicting key performance indicators of an industrial process according to claim 1, characterized in that: Four process flows with activated sludge as the core are selected, three of which are source domains and one is the target domain. The dataset is constructed and recorded as {S1, S2, S3}. The training dataset of each source domain is in, It represents the input variables that can be collected. Inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, total nitrogen concentration and other indicators are selected as input process variables. represents the corresponding key performance indicator value, and the chemical oxygen demand (COD) of the effluent is selected; the sewage treatment process is the target domain, denoted as T, in, The samples representing the target domain are scarce, and all source domains and target domains are assumed to have different distributions, that is, when i≠j, 3. The method for predicting key performance indicators of an industrial process according to claim 2, characterized in that: Collect samples in the target domain online, denoted as The process data Input COD prediction model F T* , get the predicted value for:
4. A device for predicting key performance indicators of industrial processes, characterized in that: include: The first processing module is used to construct a training data set of multiple source domains and a single target domain; The second processing module is used to construct a multi-source domain local model based on the training data set; The third processing module is used to perform federated aggregation on the local models of multiple source domains; The fourth processing module is used to migrate the aggregated federated model to the target domain for model fine-tuning to obtain a transfer learning model suitable for the target domain; a fifth processing module, configured to distribute the target domain model to multiple source domains and update the model based on the local data sets of the multiple source domains; The sixth processing module is used to complete the training until the total loss function converges and obtain the trained COD prediction model; The seventh processing module is used to use the trained COD prediction model to predict COD in the target domain.
5. The industrial process key performance indicator prediction device according to claim 4, characterized in that: Four process flows with activated sludge as the core are selected, three of which are source domains and one is the target domain. The dataset is constructed and recorded as {S1, S2, S3}. The training dataset of each source domain is in, It represents the input variables that can be collected. Inlet flow, outlet flow, inorganic suspended solids, ammonia nitrogen concentration, dissolved oxygen capacity, total nitrogen concentration and other indicators are selected as input process variables. represents the corresponding key performance indicator value, and the chemical oxygen demand (COD) of the effluent is selected; the sewage treatment process is the target domain, denoted as T, in, The samples representing the target domain are scarce, and all source domains and target domains are assumed to have different distributions, that is, when i≠j, 6. The industrial process key performance indicator prediction device according to claim 5, characterized in that: The seventh module is used to collect samples in the target domain online, denoted as The process data Input COD prediction model F T* , get the predicted value for:
7. An industrial process key performance indicator prediction system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the method for predicting key performance indicators of an industrial process as described in any one of claims 1 to 3 is executed.
8. A storage medium, characterized in that: The storage medium stores a computer program, which executes the industrial process key performance indicator prediction method according to any one of claims 1 to 3 when running.