An industrial equipment energy efficiency self-adaptive identification method fusing dynamic incremental learning

Abstract

This invention relates to an adaptive energy efficiency identification method for industrial equipment that integrates dynamic incremental learning. It overcomes the limitations of traditional single electrical quantity features by extracting and fusing components of steady-state energy efficiency features, transient impact features, and operating condition correlation features from multi-source heterogeneous data to generate a standardized multi-dimensional energy efficiency fingerprint vector. By establishing a comprehensive basic energy efficiency fingerprint database, the adaptive capability of the identification model is improved. The construction of a three-layer sparse coding mapping network compresses fingerprint features, and combined with digital twin calibration, eliminates the identification benchmark shift caused by equipment aging. By relying on knowledge graphs and counterfactual reasoning to trace the causes of energy efficiency differences, the energy-saving potential and economic benefits of equipment can be quantified, generating an energy efficiency identification report with logical explanations, thus establishing a technical link from energy efficiency identification to resource collaborative scheduling. This invention has the advantages of integrating multi-source data, strong adaptive capability, and high accuracy and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment energy efficiency identification technology, specifically relating to an adaptive identification method for industrial equipment energy efficiency that integrates dynamic incremental learning. Background Technology

[0002] Currently, against the backdrop of the construction of new power systems and the green transformation of manufacturing, industrial enterprises are increasingly demanding refined energy efficiency management and flexible load control for equipment. Accurate identification of the real-time energy efficiency status of individual equipment and pinpointing inefficient operating conditions are core foundations for achieving industrial energy conservation, demand response, and optimal resource allocation. However, current industrial equipment energy efficiency identification solutions have several shortcomings, specifically: 1. Limited feature dimensions and incomplete equipment status description; accuracy drops significantly when equipment is under variable load or intermittent production conditions; 2. Insufficient generalization and adaptability of identification models; shifts in energy consumption characteristics and model performance occur when new production equipment is added, existing equipment ages, or processing technology is adjusted; 3. Coarse energy efficiency assessment granularity, only capable of overall plant energy consumption statistics, unable to pinpoint short-term inefficiency ranges for individual equipment, and unable to accurately quantify energy-saving potential. Therefore, to address these issues, it is essential to develop an adaptive energy efficiency identification method for industrial equipment that can integrate multi-source data, possesses strong adaptability, and is accurate and efficient. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive identification method for energy efficiency of industrial equipment that can integrate multi-source data, has strong adaptive capabilities, and is accurate and efficient.

[0004] The objective of this invention is achieved as follows: a method for adaptive identification of energy efficiency in industrial equipment that integrates dynamic incremental learning, comprising the following steps: Step 1: Through the interfaces of various types of sensors, PLC controllers and MES systems deployed in industrial equipment, the steady-state electrical parameters, start-stop transient waveform data, workshop environmental sensor data and enterprise production business logs of industrial equipment are collected in real time to form raw multi-source heterogeneous data; Step 2: The extended Kalman filter algorithm is used to perform noise reduction, missing value imputation, outlier detection, time axis alignment and feature normalization on the original multi-source heterogeneous data collected in Step 1. Step 3: Extract three types of feature components from the processed multi-source heterogeneous data: steady-state energy efficiency features, transient impact features, and operating condition correlation features. Then, fuse and splice the extracted feature data to generate a standardized multi-dimensional energy efficiency fingerprint vector. Step 4: Based on the standardized multidimensional energy efficiency fingerprint vector, establish a global basic energy efficiency fingerprint database, and use samples from the global basic energy efficiency fingerprint database to train the multi-branch recognition model offline, construct incremental learning branches, and set memory decay coefficient and feature mapping matching threshold. Step 5: Collect a small amount of labeled operation data of the newly added equipment in a short period of time and generate corresponding multi-dimensional energy efficiency fingerprints. Calculate the cosine similarity between the fingerprint of the newly added equipment and the fingerprints in the fingerprint database. Select similar equipment samples with matching degree higher than the preset threshold as incremental training set. Then, build a three-layer sparse coding mapping network to remove and compress redundant features of the multi-dimensional energy efficiency fingerprints. Introduce the equipment digital twin model to output the theoretical benchmark energy consumption. Use the deviation value between the theoretical energy consumption and the measured energy consumption on site to dynamically calibrate the sparse fingerprint features. Step 6: Input the calibrated sparse fingerprint into the graph neural network, and analyze the energy efficiency correlation between devices based on the graph neural network to analyze the energy consumption coupling and operating condition linkage between multiple devices, thereby outputting the real-time energy efficiency level, inefficient operating period and operating condition matching identification results of a single device. Step 7: Based on knowledge graphs and counterfactual reasoning, trace the causes of energy efficiency differences, quantify the energy-saving potential and economic benefits of equipment, and generate an energy efficiency identification report with logical explanations. At the same time, by collecting actual energy consumption feedback data from the field after energy efficiency optimization and control, dynamically adjust the extended Kalman filter parameters and memory decay coefficient, periodically trigger incremental branch lightweight updates, and continuously optimize the identification model's adaptability.

[0005] Furthermore, in step 1, the steady-state electrical parameters include three-phase voltage, three-phase current, active power, reactive power, harmonic amplitude, and phase; the start-stop transient waveform data includes the peak value of the equipment start-up inrush current, the power rise time, and the characteristics of the shutdown attenuation curve; the workshop environment sensor data includes the workshop temperature, humidity, and cooling system operating status; and the production business log includes the equipment start-up and stop times, processing procedures, load levels, and continuous production duration.

[0006] Furthermore, in step 5, the three-layer sparse coding mapping network consists of the original fingerprint input layer, the L1 regularized sparse compression layer, and the working condition feature fusion layer.

[0007] Furthermore, the noise reduction processing of multi-source heterogeneous data in step 2 specifically involves: establishing the equipment energy consumption time-series state equation and observation equation, using the collected energy consumption data as the observation value, iteratively updating the process noise and observation noise covariance matrix, and smoothing time-series fluctuations.

[0008] Furthermore, the energy efficiency identification report in step 7 includes the real-time energy efficiency value of the equipment, the period of inefficient operation, the energy efficiency difference compared with similar equipment, and the quantitative indicators of potential energy-saving benefits.

[0009] The beneficial effects of the present invention are as follows: The present invention extracts and fuses three types of feature components—steady-state energy efficiency features, transient impact features, and operating condition correlation features—from multi-source heterogeneous data to generate a standardized multi-dimensional energy efficiency fingerprint vector, which breaks through the limitations of traditional single electrical quantity features. By establishing a comprehensive basic energy efficiency fingerprint database and using samples from this database to train a multi-branch recognition model offline, the incremental learning branches are constructed. With this structure, new devices only need a small number of short-term labeled samples to complete model matching, thus improving the adaptive capability of the recognition model. By building a three-layer sparse coding mapping network, fingerprint features can be compressed, and combined with digital twin calibration, the recognition benchmark shift caused by device aging can be eliminated. By relying on knowledge graphs and counterfactual reasoning to design the source of energy efficiency differences, the energy-saving potential and economic benefits of equipment can be quantified, and an energy efficiency identification report with logical explanation can be generated, thereby opening up the technical link from energy efficiency identification to resource collaborative scheduling. In summary, this invention has the advantages of being able to integrate multi-source data, having strong adaptability, and being accurate and efficient. Detailed Implementation

[0010] The present invention will now be further described.

[0011] Example: An adaptive energy efficiency identification method for industrial equipment integrating dynamic incremental learning, comprising the following steps: Step 1: Through the interfaces of various types of sensors, PLC controllers, and MES systems deployed in industrial equipment, real-time data is collected on the steady-state electrical parameters, start-stop transient waveform data, workshop environmental sensor data, and enterprise production logs of the industrial equipment, thereby forming raw multi-source heterogeneous data. The steady-state electrical parameters include three-phase voltage, three-phase current, active power, reactive power, and harmonic amplitude and phase. The start-stop transient waveform data includes the peak value of the equipment's starting inrush current, the power rise time, and the characteristics of the shutdown attenuation curve. The workshop environmental sensor data includes the workshop temperature, humidity, and cooling system operating status. The production logs include the equipment start-stop time, processing steps, load level, and continuous production duration. Step 2: The extended Kalman filter algorithm is used to perform noise reduction, missing value imputation, outlier detection, and time axis alignment and feature normalization on the original multi-source heterogeneous data collected in Step 1. Specifically, the noise reduction of multi-source heterogeneous data is as follows: the equipment energy consumption time series state equation and observation equation are established, the collected energy consumption data are used as the observation values, the process noise and observation noise covariance matrices are iteratively updated, and the time series fluctuations are smoothed. Step 3: Extract three types of feature components from the processed multi-source heterogeneous data: steady-state energy efficiency features, transient impact features, and operating condition correlation features. Then, fuse and splice the extracted feature data to generate a standardized multi-dimensional energy efficiency fingerprint vector. Step 4: Based on the standardized multidimensional energy efficiency fingerprint vector, establish a global basic energy efficiency fingerprint database, and use samples from the global basic energy efficiency fingerprint database to train the multi-branch recognition model offline, construct incremental learning branches, and set memory decay coefficient and feature mapping matching threshold. Step 5: Collect a small amount of short-cycle labeled operating data of the newly added equipment and generate corresponding multi-dimensional energy efficiency fingerprints. Calculate the cosine similarity between the fingerprint of the newly added equipment and the fingerprints in the existing fingerprint database. Select similar equipment samples with a matching degree higher than a preset threshold as incremental training sets. Then, build a three-layer sparse coding mapping network to remove and compress redundant features of the multi-dimensional energy efficiency fingerprints. Introduce the equipment digital twin model to output the theoretical baseline energy consumption. Use the deviation between the theoretical energy consumption and the measured energy consumption on site to dynamically calibrate the sparse fingerprint features. The three-layer sparse coding mapping network consists of an original fingerprint input layer, an L1 regularized sparse compression layer, and a working condition feature fusion layer. Step 6: Input the calibrated sparse fingerprint into the graph neural network, and analyze the energy efficiency correlation between devices based on the graph neural network to analyze the energy consumption coupling and operating condition linkage between multiple devices, thereby outputting the real-time energy efficiency level, inefficient operating period and operating condition matching identification results of a single device. Step 7: Based on knowledge graphs and counterfactual reasoning, trace the causes of energy efficiency differences, quantify the energy-saving potential and economic benefits of equipment, and generate an energy efficiency identification report with logical explanations. The energy efficiency identification report includes the real-time energy efficiency value of the equipment, the period of inefficient operation, the energy efficiency difference compared with similar equipment, and quantitative indicators of potential energy-saving benefits. At the same time, by collecting actual energy consumption feedback data from the field after energy efficiency optimization and control, dynamically adjust the extended Kalman filter parameters and memory decay coefficient, periodically trigger incremental branch lightweight updates, and continuously optimize the identification model's adaptability.

[0012] In use, this invention first collects steady-state electrical parameters, start-stop transient waveform data, workshop environmental sensor data, and enterprise production logs in real time through various types of sensors, PLC controllers, and MES system interfaces deployed in industrial equipment, thus forming raw multi-source heterogeneous data. Then, an extended Kalman filter algorithm is used to perform noise reduction, missing value imputation, outlier detection, and multi-source data time axis alignment and feature normalization on the collected raw multi-source heterogeneous data. Next, three types of feature components—steady-state energy efficiency features, transient impact features, and operating condition correlation features—are extracted from the processed multi-source heterogeneous data. The extracted feature data are then fused and stitched together to generate standardized data. A multi-dimensional energy efficiency fingerprint vector is generated. Then, based on the standardized multi-dimensional energy efficiency fingerprint vector, a global basic energy efficiency fingerprint database is established. The multi-branch recognition model is trained offline using samples from this database to construct incremental learning branches. Finally, a small amount of labeled operational data from newly added devices is collected over a short period to generate corresponding multi-dimensional energy efficiency fingerprints. The cosine similarity between the fingerprints of the newly added devices and the existing fingerprints in the database is calculated. Samples of similar devices with a matching degree higher than a preset threshold are selected as the incremental training set. Subsequently, a three-layer sparse coding mapping network is constructed to remove and compress redundant features from the multi-dimensional energy efficiency fingerprints. A digital twin model of the device is introduced to output the theoretical baseline energy consumption. The deviation between the theoretical energy consumption and the measured energy consumption is used to dynamically calibrate the sparse fingerprint features. After completing the above operations, the calibrated sparse fingerprint is input into a graph neural network. Based on the graph neural network, energy efficiency correlation identification between devices is performed, analyzing the energy consumption coupling and operating condition linkage relationships between multiple devices. This outputs the real-time energy efficiency level, inefficient operating periods, and operating condition matching identification results for each device. Finally, relying on knowledge graphs and counterfactual reasoning, the causes of energy efficiency differences are traced, the energy-saving potential and economic benefits of devices are quantified, and an energy efficiency identification report with logical explanations is generated. At the same time, by collecting actual energy consumption feedback data from the field after energy efficiency optimization and control, the parameters of the extended Kalman filter and the memory decay coefficient are dynamically adjusted, and incremental branch lightweight updates are periodically triggered to continuously optimize the adaptability of the identification model. This invention breaks through the limitations of traditional single electrical quantity features by extracting and fusing three types of feature components—steady-state energy efficiency features, transient impact features, and operating condition correlation features—from multi-source heterogeneous data to generate standardized multi-dimensional energy efficiency fingerprint vectors. It establishes a global basic energy efficiency fingerprint database and uses samples from this database to train a multi-branch recognition model offline, thus constructing incremental learning branches. Using this structure, new devices only require a small number of short-term labeled samples to complete model matching, improving the adaptive capability of other models. Furthermore, by constructing a three-layer sparse coding mapping network, fingerprint features can be compressed, and combined with digital twin calibration, the identification benchmark shift caused by equipment aging can be eliminated.By relying on knowledge graphs and counterfactual reasoning to trace the causes of energy efficiency differences, this invention quantifies the energy-saving potential and economic benefits of equipment, generating an energy efficiency identification report with logical explanations. This establishes a technical link from energy efficiency identification to resource collaborative scheduling. In summary, this invention has the advantages of being able to integrate multi-source data, possessing strong adaptability, and being accurate and efficient.

[0013] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for adaptive energy efficiency recognition of industrial equipment integrating dynamic incremental learning, characterized in that, Includes the following steps: Step 1: Through the interfaces of various types of sensors, PLC controllers and MES systems deployed in industrial equipment, the steady-state electrical parameters, start-stop transient waveform data, workshop environmental sensor data and enterprise production business logs of industrial equipment are collected in real time to form raw multi-source heterogeneous data; Step 2: The extended Kalman filter algorithm is used to perform noise reduction, missing value imputation, outlier detection, time axis alignment and feature normalization on the original multi-source heterogeneous data collected in Step 1. Step 3: Extract three types of feature components from the processed multi-source heterogeneous data: steady-state energy efficiency features, transient impact features, and operating condition correlation features. Then, fuse and splice the extracted feature data to generate a standardized multi-dimensional energy efficiency fingerprint vector. Step 4: Based on the standardized multidimensional energy efficiency fingerprint vector, establish a global basic energy efficiency fingerprint database, and use samples from the global basic energy efficiency fingerprint database to train the multi-branch recognition model offline, construct incremental learning branches, and set memory decay coefficient and feature mapping matching threshold. Step 5: Collect a small amount of labeled operation data of the newly added equipment in a short period of time and generate corresponding multi-dimensional energy efficiency fingerprints. Calculate the cosine similarity between the fingerprint of the newly added equipment and the fingerprints in the fingerprint database. Select similar equipment samples with matching degree higher than the preset threshold as incremental training set. Then, build a three-layer sparse coding mapping network to remove and compress redundant features of the multi-dimensional energy efficiency fingerprints. Introduce the equipment digital twin model to output the theoretical benchmark energy consumption. Use the deviation value between the theoretical energy consumption and the measured energy consumption on site to dynamically calibrate the sparse fingerprint features. Step 6: Input the calibrated sparse fingerprint into the graph neural network, and analyze the energy efficiency correlation between devices based on the graph neural network to analyze the energy consumption coupling and operating condition linkage between multiple devices, thereby outputting the real-time energy efficiency level, inefficient operating period and operating condition matching identification results of a single device. Step 7: Based on knowledge graphs and counterfactual reasoning, trace the causes of energy efficiency differences, quantify the energy-saving potential and economic benefits of equipment, and generate an energy efficiency identification report with logical explanations. At the same time, by collecting actual energy consumption feedback data from the field after energy efficiency optimization and control, dynamically adjust the extended Kalman filter parameters and memory decay coefficient, periodically trigger incremental branch lightweight updates, and continuously optimize the identification model's adaptability.

2. The industrial equipment energy efficiency adaptive identification method integrating dynamic incremental learning as described in claim 1, characterized in that: In step 1, the steady-state electrical parameters include three-phase voltage, three-phase current, active power, reactive power, harmonic amplitude, and phase; the start-stop transient waveform data includes the peak value of the equipment start-up inrush current, the power rise time, and the characteristics of the shutdown attenuation curve; the workshop environmental sensor data includes the workshop temperature, humidity, and cooling system operating status; and the production business log includes the equipment start-up and stop times, processing procedures, load levels, and continuous production duration.

3. The industrial equipment energy efficiency adaptive identification method integrating dynamic incremental learning as described in claim 1, characterized in that: In step 5, the three-layer sparse coding mapping network consists of the original fingerprint input layer, the L1 regularized sparse compression layer, and the working condition feature fusion layer.

4. The industrial equipment energy efficiency adaptive identification method integrating dynamic incremental learning as described in claim 1, characterized in that: The noise reduction process for multi-source heterogeneous data in step 2 specifically involves: establishing the equipment energy consumption time-series state equation and observation equation, using the collected energy consumption data as the observation value, iteratively updating the process noise and observation noise covariance matrix, and smoothing time-series fluctuations.

5. The industrial equipment energy efficiency adaptive identification method integrating dynamic incremental learning as described in claim 1, characterized in that: The energy efficiency identification report in step 7 includes the real-time energy efficiency value of the equipment, the period of inefficient operation, the energy efficiency difference compared with similar equipment, and the quantitative indicators of potential energy-saving benefits.