Production process determination method, system and electronic device based on machine learning network
By automatically determining production processes through machine learning networks, the problem of low accuracy in identifying processes with large data volumes and complex logic has been solved, thereby optimizing production costs and improving efficiency, and providing precise energy consumption management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUBEI UNIV OF TECH
- Filing Date
- 2022-11-14
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to effectively identify production processes with large data volumes and complex logical relationships, resulting in low accuracy and hindering the optimization of enterprise production costs and efficiency improvements.
A machine learning network-based approach is adopted to obtain the current and previous load power and historical process information. Data processing is performed using encoding and decoding modules, including convolutional layers and classification layers, to achieve automatic determination of production processes.
It improves the accuracy of production process judgment, helps enterprises adjust production schedules, avoid peak electricity consumption periods, reduce costs and improve efficiency, and at the same time achieves accurate energy consumption analysis, detects energy consumption anomalies and deals with them in a timely manner.
Smart Images
Figure CN115936341B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine learning application technology, and relates to a production process determination method, system and electronic equipment, and particularly to a production process determination method, system and electronic equipment based on machine learning networks. Background Technology
[0002] For multi-stage, sequential industrial production processes, determining the production sequence based on relevant production data facilitates real-time monitoring of the production line and allows for the calculation of production time and power consumption for each stage. This enables detailed energy usage analysis in conjunction with an energy management system. Furthermore, by adhering to peak-shaving-valley rules for electricity billing, all production processes can be rationally arranged and planned to avoid peak billing periods while ensuring sufficient output, thereby reducing production costs and improving efficiency. Determining the production sequence also helps managers monitor whether actual production aligns with the plan and take timely appropriate action.
[0003] Currently, process identification for small amounts of data with obvious patterns can be determined through manual analysis. However, for production process identification with large amounts of data and complex logical relationships, manual analysis is difficult to implement, or the accuracy of the identification results may not meet the application requirements. Summary of the Invention
[0004] Based on the technical problems in the background technology, this invention proposes a production process determination method, system and electronic device based on machine learning.
[0005] The technical solution adopted by the method of the present invention is: a production process determination method based on machine learning networks, comprising the following steps:
[0006] Step 1: Obtain the current and previous load power and historical process information; the load power includes historical load power and current load power;
[0007] The power data for the i-th time interval is denoted as power. i The process status corresponding to the j-th power data is denoted as process. j , i∈[1,N], j∈[1,N-1]; N represents the total number of power signals collected at specified time intervals;
[0008] Step 2: Input the dataset composed of the current and previous load power and historical process information into the machine learning network to obtain the current production process information;
[0009] The load power at time t and the previous time is denoted as pow. t pow t-1 The historical process information from the previous moment is recorded as pro.t-287 ,pro t-286 ,...,pro t-1 ;
[0010] The machine learning network includes an encoding module and a decoding module. The input data are the current and previous load power and historical process information, and the output is the process value at the current moment.
[0011] The encoding module comprises an input layer and four cascaded convolutional layers. The first layer processes the input data to obtain the corresponding dataset. The second layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 1, and a stride of 1. The third layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 2, and a stride of 1. The fourth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 4, and a stride of 1. The fifth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 8, and a stride of 1. The first and fifth layers are connected through a residual module to accelerate model convergence and prevent model degradation, avoiding the gradient vanishing problem.
[0012] The decoding module consists of a classification layer and a process conversion layer. The classification layer selects softmax as the activation function and calculates the probability of six categories based on the output layer data. The process conversion layer further processes the classification results to obtain the final corresponding process value.
[0013] The technical solution adopted by the system of the present invention is: a production process determination system based on machine learning, comprising the following modules:
[0014] Module 1 is used to obtain the current and previous load power and historical process information; the load power includes historical load power and current load power;
[0015] The power data for the i-th time interval is denoted as power. i The process status corresponding to the j-th power data is denoted as process. j , i∈[1,N], j∈[1,N-1]; N represents the total number of power signals collected at specified time intervals;
[0016] Module 2 is used to input the dataset composed of the current and previous load power and historical process information into the machine learning network to obtain the current production process information;
[0017] The load power at time t and the previous time is denoted as pow. t pow t-1 The historical process information from the previous moment is recorded as pro. t-287 ,pro t-286 ,...,pro t-1;
[0018] The machine learning network includes an encoding module and a decoding module. The input data are the current and previous load power and historical process information, and the output is the process value at the current moment.
[0019] The encoding module comprises an input layer and four cascaded convolutional layers. The first layer processes the input data to obtain the corresponding dataset. The second layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 1, and a stride of 1. The third layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 2, and a stride of 1. The fourth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 4, and a stride of 1. The fifth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 8, and a stride of 1. The first and fifth layers are connected through a residual module to accelerate model convergence and prevent model degradation, avoiding the gradient vanishing problem.
[0020] The decoding module consists of a classification layer and a process conversion layer. The classification layer selects softmax as the activation function and calculates the probability of six categories based on the output layer data. The process conversion layer further processes the classification results to obtain the final corresponding process value.
[0021] The technical solution adopted by the electronic device of the present invention is: an electronic device comprising:
[0022] One or more processors;
[0023] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the machine learning-based production process determination method.
[0024] This invention uses machine learning methods to determine production processes. Compared with traditional manual analysis methods, it simplifies the data processing and analysis process, making the production process determination system easier to implement and achieving a higher accuracy rate.
[0025] This invention, through the judgment of production processes, can help enterprises achieve the following energy-saving and efficiency-enhancing objectives: (1) Adjust production arrangements in a timely manner in conjunction with the production plan to avoid peak electricity metering periods, thereby reducing enterprise production costs and improving production efficiency. (2) Combine with the energy consumption management system to accurately analyze the energy consumption of each process involved in the production process, promptly identify processes with abnormal energy consumption, and remind managers to take appropriate action.
[0026] The electronic device of the present invention has a flexible installation method, which can be used in newly designed equipment or in the modification of old equipment, and has a wide range of applications, and can be used for various types of production process judgment. Attached Figure Description
[0027] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;
[0028] Figure 2 This is a diagram of the machine learning network structure according to an embodiment of the present invention;
[0029] Figure 3 This is a flowchart of the machine learning network training process according to an embodiment of the present invention;
[0030] Figure 4 This is a diagram showing the actual test results of this invention, combined with data from a titanium metal smelting production process. Detailed Implementation
[0031] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0032] Please see Figure 1 The present invention provides a production process determination method based on machine learning networks, comprising the following steps:
[0033] Step 1: Obtain the current and previous load power and historical process information; load power includes historical load power and current load power;
[0034] The power data for the i-th time interval is denoted as power. i The process status corresponding to the j-th power data is denoted as process. j , i∈[1,N], j∈[1,N-1]; N represents the total number of power signals collected at specified time intervals;
[0035] Step 2: Input the dataset composed of the current and previous load power and historical process information into the machine learning network to obtain the current production process information;
[0036] The load power at time t and the previous time is denoted as pow. t pow t-1 The historical process information from the previous moment is recorded as pro. t-287 ,pro t-286 ,...,pro t-1 ;
[0037] Please see Figure 2The machine learning network in this embodiment includes an encoding module and a decoding module. The input data are the load power and historical process information at the current and previous moments, and the output is the process value at the current moment. The encoding module is composed of a multi-layer network in sequence, and the first layer and the bottom five layers are connected by residual links. Each layer is convolved with a certain dilation factor. The decoding module consists of a classification layer and a transformation process layer.
[0038] The encoding module in this embodiment includes an input layer and four cascaded convolutional layers. The first layer processes the input data to obtain the corresponding dataset. The second layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 1, and a stride of 1. The third layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 2, and a stride of 1. The fourth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 4, and a stride of 1. The fifth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 8, and a stride of 1. The first and fifth layers are connected through a residual module to accelerate model convergence and prevent model degradation, avoiding the gradient vanishing problem.
[0039] The decoding module in this embodiment consists of a classification layer and a process conversion layer. The classification layer selects softmax as the activation function and calculates the probability of six categories based on the output layer data. The process conversion layer further processes the classification results to obtain the final corresponding process value.
[0040] After obtaining the current production process information, when the production process changes, the change in the production process status is announced via voice broadcast. At the same time, a detailed analysis of the electrical energy consumed in this production process is performed. After obtaining the current production process, in conjunction with the production planning management system, a judgment and analysis are made on whether the current production process is consistent with the planned production process.
[0041] The machine learning network in this embodiment performs convolution operations strictly in chronological order to ensure no leakage of future information; that is, the output at any time t is only related to the input at time t and before. Its mathematical expression is:
[0042]
[0043] Among them, f i Here are the parameters corresponding to the filter, K is the kernel size, and x is the value of x. t-K+i The input is given at time t-k+i, where i represents the number of convolution operations.
[0044] The machine learning network in this embodiment uses convolution operations with dilation factors. Each layer selects data at a certain distance from the previous layer, and its mathematical expression is as follows:
[0045]
[0046] Among them, f i Here are the parameters corresponding to the filters, M is the number of filters, d is the expansion coefficient, and x is the expansion coefficient. t-d×i The input is at time td×i, where i represents the number of filtering operations.
[0047] The residual connection mathematical expression of the machine learning network in this embodiment is:
[0048] F(x l W l )=H(x l )-x l ;
[0049] In the formula, x l W is the actual target value. l Let H(x) be the weight parameter matrix of the layer. l ) represents the actual target mapping value.
[0050] Please see Figure 3 The machine learning network in this embodiment is a pre-trained machine learning network; its training process includes the following sub-steps:
[0051] Step 2.1: Collect production-related information, including production-related voltage and current information and historical process information;
[0052] Step 2.2: Extract the load power and historical process information related to the judgment process from the production-related information; the load power includes historical load power and current load power;
[0053] The power data for the i-th time interval is denoted as power. i The process status corresponding to the j-th power data is denoted as process. j , i∈[1,N], j∈[1,N-1]; N represents the total number of power signals collected at specified time intervals;
[0054] Step 2.3: Construct the training dataset;
[0055] Based on the load power and historical process information extracted in step 2, in process1, process2,..., process N-1 and power1, power2, ..., power N-1 The original process and power data were resampled sequentially according to a window size of 288, resulting in M power datasets of length 288, pow. i,1 pow i,2 ,...,pow i,288 ,i∈[1,M] and the corresponding process status data proi,1 ,pro i,2 ,...,pro i,288 For each power data set, i∈[1,M], the first 287 data values of each process state are selected from the power data set and the corresponding process state data, i.e., pro. i,1 ~pro i,287 The last two power values of i∈[1,M] and the corresponding power value array, i.e., pow i,287 pow i,288 The elements i ∈ [1, M] are combined into a new array, which serves as the training and testing dataset for the machine learning network. i,289 ,i∈[1,M], select the corresponding process state, i.e., pro i,288 The data in the range i∈[1,M] are the label data for this dataset. i ,i∈[1,M], finally obtaining the training and testing datasets data after data preprocessing. i,289 ,i∈[1,M] and label set i ,i∈[1,M];
[0056] Step 2.4: Input the training and testing datasets into the machine learning network to learn and extract features and patterns of load power and historical process information;
[0057] The training process uses the Adam optimizer and categorical cross entropy loss as the loss function. After a certain number of iterations (80 in this example), the trained machine learning network is obtained.
[0058] The machine learning network in this embodiment is built using Python and based on the Keras framework. The relevant dataset is input, and the machine learning network is trained and tested according to the specified machine learning network optimizer and loss function and other parameters. The trained machine learning network is obtained and saved for use by the data storage and processing device.
[0059] This embodiment uses Java, JavaScript, and other languages to write the corresponding front-end and back-end software and deploys it on a data storage and processing device, i.e., a local Windows 10 desktop computer. A MySQL database is used for data storage to complete the system setup, which is responsible for receiving, processing, storing, and displaying data.
[0060] In this embodiment, real-time power information is collected by a data acquisition device and uploaded to a data storage and processing device via a wired or wireless network; historical process information is entered by the operator or imported into the data storage, processing and display device via a table;
[0061] The data storage, processing and display device uses a pre-trained machine learning network to determine the current process information by combining real-time power data and historical process information. It calculates the current process status every 10 minutes and stores and displays it as a reference for subsequent process status identification.
[0062] Data storage, processing, and display devices provide editable historical process information, further improving the accuracy of subsequent process identification;
[0063] The data acquisition device used in this implementation adopts the ZH-4041-14N1 power acquisition module provided by Zhongchuang Zhihe. This module can measure the effective value data of various sine waves and non-standard waveforms. It has high measurement accuracy, fast sampling speed, and allows users to set the data update speed. The frequency response range is 30-1000Hz, and it is suitable for detection of various power frequency and intermediate frequency signals. The communication protocol is compatible with Modbus-RTU and Modbus-TCP communication protocols.
[0064] Data storage, processing, and display devices: A Windows 10 desktop computer equipped with an Intel i7 9700K processor, 16GB of RAM, and a 512GB hard drive. The use of the widely adopted Windows system facilitates the setup and application of the process identification system software, while also allowing users to quickly familiarize themselves with and operate it.
[0065] Please see Figure 4 This is a diagram showing the actual test results of a titanium metal smelting production line, based on data from the present invention. After compilation and statistics, a total of 327 sets of test data for this production line process were identified, of which 317 sets of data corresponded to model process identification results consistent with the actual process. Figure 4 It can be seen that the actual process and the process identified by the model almost overlap, with a corresponding recognition accuracy of 96.94%. This further proves the accuracy and practicality of the proposed process identification method, indicating that the method can be used to identify and judge the process of the enterprise production line, helping enterprises to achieve energy saving and efficiency improvement.
[0066] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.
Claims
1. A method for determining production processes based on machine learning networks, characterized in that, Includes the following steps: Step 1: Obtain the current and previous load power and historical process information; the load power includes historical load power and current load power; No. The power data for each time interval is denoted as , No. The process status corresponding to each power data point is denoted as: , , N represents the total number of power signals collected at specified time intervals; Step 2: Input the dataset composed of the current and previous load power and historical process information into the machine learning network to obtain the current production process information; The load power at the current time t and the previous time is denoted as: The historical process information from the previous moment is recorded as follows: ; The machine learning network includes an encoding module and a decoding module. The input data are the current and previous load power and historical process information, and the output is the process value at the current moment. The encoding module comprises an input layer and four cascaded convolutional layers. The first layer processes the input data to obtain the corresponding dataset. The second layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 1, and a stride of 1. The third layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 2, and a stride of 1. The fourth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 4, and a stride of 1. The fifth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 8, and a stride of 1. The first and fifth layers are connected through a residual module to accelerate model convergence and prevent model degradation, avoiding the gradient vanishing problem. The decoding module consists of a classification layer and a process conversion layer. The classification layer selects softmax as the activation function and calculates the probability of six categories based on the output layer data. The process conversion layer further processes the classification results to obtain the final corresponding process value. The machine learning network mentioned is a pre-trained machine learning network; Its training process includes the following sub-steps: Step 2.1: Collect production-related information, including production-related voltage and current information and historical process information; Step 2.2: Extract the load power and historical process information related to the judgment process from the production-related information; the load power includes historical load power and current load power; Among them, the The power data for each time interval is denoted as , No. The process status corresponding to each power data point is denoted as: , , N represents the total number of power signals collected at specified time intervals; Step 2.3: Construct the training dataset; Based on the load power and historical process information extracted in step 2.2, in and The original process and power data were resampled sequentially using a window size of 288, resulting in M power datasets of length 288. and corresponding process status data From each set of power data and the corresponding process status data, the first 287 data values of each process status are selected. And the last two power data of the corresponding power value array, i.e. Combined into a new array, which serves as the training and testing dataset for the machine learning network. Select the corresponding process status. The label data for this dataset is Finally, we obtain the training and testing datasets after data preprocessing. and tag set ; Step 2.4: Input the training and test datasets into the machine learning network to learn and extract features and patterns of load power and historical process information; The training process uses the Adam optimizer and employs the classification cross-entropy loss as the loss function. After a certain number of iterations, a well-trained machine learning network is obtained. Step 3: After obtaining the current production process information, when the production process changes, the change in the production process status is announced via voice broadcast. At the same time, a detailed analysis of the power consumption of this production process is performed. After obtaining the current production process, in conjunction with the production planning management system, a judgment and analysis are made on whether the current production process is consistent with the planned production process.
2. The production process determination method based on machine learning according to claim 1, characterized in that: The machine learning network described in step 2 performs convolution operations strictly according to the time sequence to ensure that no future information is leaked. That is, the output at any time t is only related to the input at time t and before. Its mathematical expression is: ； in, These are the parameters corresponding to the filter. The kernel size is [size]. for Input at any time This indicates the number of convolution operations.
3. The production process determination method based on machine learning according to claim 1, characterized in that: The machine learning network described in step 2 employs convolution operations with a dilation factor, and each layer selects data at a certain distance from the previous layer. Its mathematical expression is: ； in, These are the parameters corresponding to the filter. The number of filters. The coefficient of thermal expansion is 1 / 3. for Input at any time Indicates the number of filtering operations.
4. The production process determination method based on machine learning according to claim 1, characterized in that: The residual connection mathematical expression of the machine learning network described in step 2 is: ； In the formula, For the actual target value, The weight parameter matrix of the layer, This is the actual target mapping value.
5. A production process determination system based on machine learning, characterized in that, Includes the following modules: Module 1 is used to obtain the current and previous load power and historical process information; the load power includes historical load power and current load power; No. The power data for each time interval is denoted as , No. The process status corresponding to each power data point is denoted as: , , N represents the total number of power signals collected at specified time intervals; Module 2 is used to input the dataset composed of the current and previous load power and historical process information into the machine learning network to obtain the current production process information; The load power at the current time t and the previous time is denoted as: The historical process information from the previous moment is recorded as follows: ; The machine learning network includes an encoding module and a decoding module. The input data are the current and previous load power and historical process information, and the output is the process value at the current moment. The encoding module comprises an input layer and four cascaded convolutional layers. The first layer processes the input data to obtain the corresponding dataset. The second layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 1, and a stride of 1. The third layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 2, and a stride of 1. The fourth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 4, and a stride of 1. The fifth layer is a convolutional operation layer with a kernel size of 3, a dilation factor of 8, and a stride of 1. The first and fifth layers are connected through a residual module to accelerate model convergence and prevent model degradation, avoiding the gradient vanishing problem. The decoding module consists of a classification layer and a process conversion layer. The classification layer selects softmax as the activation function and calculates the probability of six categories based on the output layer data. The process conversion layer further processes the classification results to obtain the final corresponding process value. The machine learning network mentioned is a pre-trained machine learning network; Its training process includes the following sub-steps: Step 2.1: Collect production-related information, including production-related voltage and current information and historical process information; Step 2.2: Extract the load power and historical process information related to the judgment process from the production-related information; the load power includes historical load power and current load power; Among them, the The power data for each time interval is denoted as , No. The process status corresponding to each power data point is denoted as: , , N represents the total number of power signals collected at specified time intervals; Step 2.3: Construct the training dataset; Based on the load power and historical process information extracted in step 2.2, in and The original process and power data were resampled sequentially using a window size of 288, resulting in M power datasets of length 288. and corresponding process status data From each set of power data and the corresponding process status data, the first 287 data values of each process status are selected. And the last two power data of the corresponding power value array, i.e. Combined into a new array, which serves as the training and testing dataset for the machine learning network. Select the corresponding process status. The label data for this dataset is Finally, we obtain the training and testing datasets after data preprocessing. and tag set ; Step 2.4: Input the training and test datasets into the machine learning network to learn and extract features and patterns of load power and historical process information; The training process uses the Adam optimizer and employs the classification cross-entropy loss as the loss function. After a certain number of iterations, a well-trained machine learning network is obtained. Module 3 is used to obtain the current production process information and, when the production process changes, announce the change in the production process status via voice. At the same time, it performs a detailed analysis of the power consumption of the production process. After obtaining the current production process, it combines with the production planning management system to judge and analyze whether the current production process is consistent with the planned production process.
6. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the machine learning-based production process determination method as described in any one of claims 1 to 4.