Machine learning-based green production and energy consumption management and control method for equipment manufacturing enterprises

Abstract

The application relates to a kind of green production and energy consumption management and control method of equipment manufacturing enterprise based on machine learning, belong to artificial intelligence and intelligent manufacturing cross technical field.It includes the following steps: collecting equipment manufacturing full chain multi-source heterogeneous data, obtaining standardized feature vector by time stamp unification alignment and standardization processing;Joint representation model of fusion graph neural network and attention mechanism is constructed, and node embedding vector is output;Multi-objective reinforcement learning algorithm based on Pareto front is constructed, node embedding is input as state, and optimization strategy is dynamically generated under production, equipment and process constraints;Model parameters are updated online through sliding window data set, and retraining process is triggered when prediction error exceeds threshold value;Finally, the green energy efficiency index of each production unit is calculated, and a visual evaluation report is generated.The application can improve the adaptability of model working condition, the real-time performance and accuracy of optimization strategy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent manufacturing, specifically relating to a method for green production and energy consumption management in equipment manufacturing enterprises based on machine learning. Background Technology

[0002] With the deepening implementation of the "dual carbon" goals, the demand for coordinated green production and refined energy consumption management among equipment manufacturing enterprises is becoming increasingly prominent. Introducing machine learning technology into manufacturing systems holds promise for achieving dynamic perception, intelligent analysis, and optimization decision-making for highly complex and tightly coupled production processes. Currently, most machine learning-based energy management solutions primarily target scenarios such as buildings, oil fields, or general industrial facilities. Their modeling logic and control strategies are ill-suited to the deeply intertwined nature of process flow, material flow, and energy flow in equipment manufacturing, lacking comprehensive green-energy integrated management capabilities covering the entire chain from order scheduling and equipment operation to resource recycling.

[0003] Existing technologies suffer from the following shortcomings: While graph neural networks are used to construct variable interaction relationships to improve the accuracy of building energy consumption prediction, the feature system focuses on environmental parameters (such as temperature, humidity, and lighting load), failing to incorporate process parameters, equipment energy efficiency curves, and green manufacturing indicators, thus failing to reflect the dynamic energy consumption mechanism during manufacturing task execution. Furthermore, while multimodal long short-term memory neural networks (LSTM) are used to achieve energy consumption prediction and time-sharing optimization for general industrial systems, core green production indicators such as unit product carbon emissions and resource recycling rates are not embedded into the optimization objectives. A collaborative mechanism between production plan adjustments and energy consumption control commands is also lacking, leading to a disconnect between energy-saving measures and green goals. Therefore, there is an urgent need for a machine learning-driven method for equipment manufacturing enterprises that deeply integrates green production and energy consumption management, capable of uniformly modeling multi-dimensional elements of process, equipment, energy, and environment, and achieving synergistic optimization of "energy reduction" and "carbon reduction" throughout the entire production process. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a machine learning-based method for green production and energy consumption management in equipment manufacturing enterprises, comprising the following steps: S1. Collect multi-source heterogeneous data from the entire equipment manufacturing chain; perform unified alignment of timestamps through a high-precision time synchronization module, and then standardize the aligned data to obtain standardized feature vectors; S2. Based on standardized feature vectors, a joint representation model integrating graph neural network and attention mechanism is constructed. Production equipment, process unit and material batch are used as nodes, the coupling strength between process flow, material flow and energy flow is used as edge weight, and the carbon emission per unit product, resource recycling rate and waste generation are used as node attributes. The output is a node embedding vector that integrates green attributes and energy consumption characteristics. S3. With the dual objectives of improving energy efficiency and optimizing environmental performance, a multi-objective reinforcement learning algorithm based on the Pareto front is constructed. The node embedding vector output from step S2 is used as the input to the state space. Under the premise of meeting the production delivery cycle, equipment safety constraints and process quality requirements, optimization strategies are dynamically generated. S4. Collect actual response data during the execution of the optimization strategy, construct a sliding window dataset, update the joint representation model parameters through online gradient descent based on the mean squared error loss function, and trigger the model retraining process when the prediction error exceeds the threshold. S5. Based on the historical data accumulated during the closed-loop control process of steps S1-S4, calculate the green energy efficiency index of each production unit and generate a visual evaluation report to guide long-term process improvement and equipment upgrades.

[0005] Further, in step S1, the formula for the standardized feature vector is expressed as follows: , in, Represents standardized data; Represents a vector after structured data has been standardized. This represents a vector after the structured data has been standardized.

[0006] Furthermore, in step S2, the construction process of the joint representation model is as follows: Definition diagram , ,in, Represents a set of nodes. , Indicates the total number of nodes; Denotes the set of edges. , Represents a node and nodes The connection between them; The initial embedding of each node is: ,in, Represents the relationship between the standardized feature vector and the i-th node. Related subvectors; Represents a vector of green manufacturing indicators. , This represents the carbon emissions per unit of product at the i-th node; This represents the resource cycle rate of the i-th node; Indicates the amount of waste generated per unit of product; The graph neural network adopts a three-layer message passing architecture, with each layer including three stages: message construction, attention weighting, message aggregation, and node update. In the message construction phase of layer l, the i-th node Receive from neighboring nodes The message is represented by the following formula: , in, Indicates the l-th layer slave node Passed to node The message; Represents a multilayer perceptron network; Represents the (l-1)th level node The embedding vector; Represents the (l-1)th level node The embedding vector; In the attention-weighted phase, the edge weights are dynamically adjusted using an attention coefficient, as expressed by the following formula: , in, Represents the node at level l For nodes Attention coefficient; This represents the normalized exponential function; Represents a linear rectified function with leakage; Represents the learnable weight matrix; Represents the attention vector; This represents a vector concatenation operation; Message aggregation and node updates are represented by the following formula: , in, Represents the i-th node The set of neighboring nodes; This represents a gated loop unit, used to fuse historical node information with newly aggregated neighborhood information; After three layers of message passing, the final node is embedded. The green attributes and energy consumption characteristics of the local neighborhood are integrated as a joint representation output.

[0007] Furthermore, the carbon emissions per unit product are calculated based on the life cycle assessment method, as expressed by the following formula: , in, Represents the i-th node The corresponding output of products or intermediate components; Represents the i-th node The total number of processes; , These represent the electrical energy and natural gas energy equivalent consumed in the k-th process, respectively. , These represent the carbon emission factors of the corresponding energy sources; This represents the mass of raw materials consumed in the k-th process; Indicates the implicit carbon coefficient of the material; The resource recycling rate It is the ratio of the mass of recyclable and reusable materials to the total mass of input materials; The amount of waste generated per unit product The formula is expressed as: ,in, This represents the total mass of non-recyclable waste generated by the i-th node.

[0008] Furthermore, in step S3, the construction process of the multi-objective reinforcement learning algorithm is as follows: state space : Embedded by the final node of all nodes Composed of splicing, with dimensions of ,in, Indicates the dimension of a single node's embedding; Action space This is a discrete-continuous hybrid space. Discrete actions include equipment operating speed, energy switching commands, and equipment start / stop scheduling commands; continuous actions include coolant flow rate, spindle speed, and cutting feed rate. The reward function is a weighted average of energy efficiency gain and green performance gain, expressed by the following formula: , in, Indicates the state Next action The instant reward received; Indicates the current state; Indicates the action to be performed; Indicates the weighting coefficient; This represents the energy efficiency gain term; This represents the green performance gain item.

[0009] Furthermore, in step S4, actual response data is collected in real time, including actual power consumption of the equipment, actual completion time of the process, measured carbon emissions, actual resource recycling rate, actual waste generation, and product quality inspection data.

[0010] Building a sliding window dataset The window size is set to 100 control cycles, and the step size is set to 1 control cycle. That is, each time data of one control cycle is added, the data of the oldest control cycle is removed. The loss function is the mean squared error loss function, calculated based on the data within the sliding window, and the formula is as follows: , in, Represents the loss function; Represents the parameters of the joint representation model; Indicates the size of the sliding window dataset; This indicates that the joint representation model is related to the input. The predicted value; The values ​​represent the actual observations; the parameters of the joint characterization model are updated using the stochastic gradient descent method.

[0011] Furthermore, in step S5, calculating the green energy efficiency index specifically includes: The green energy efficiency index is defined as the geometric mean of the overall energy efficiency score and the green performance score, expressed by the formula: , in, Indicates the green energy efficiency index; This represents the energy efficiency score, calculated based on energy consumption per unit of output. The green performance score is derived from a weighted combination of normalized resource recycling rate and carbon emission intensity.

[0012] The advantages of this invention are: This invention addresses the industry characteristics of deep coupling between process flow, material flow, and energy flow in equipment manufacturing, overcoming the industry pain points of existing technologies such as the disconnect between energy consumption control and green production goals, and insufficient full-chain collaborative optimization capabilities. Based on a joint representation model constructed using graph neural networks and attention mechanisms, core green manufacturing indicators are deeply embedded into node attributes, accurately characterizing the correlation characteristics and coupling mechanisms of each unit in the production network, achieving a deep fusion representation of energy consumption characteristics and green attributes. A multi-objective reinforcement learning algorithm, constructed with energy efficiency improvement and environmental performance optimization as dual objectives, achieves collaborative dynamic optimization of energy consumption reduction and carbon reduction goals while meeting production delivery, equipment safety, and process quality constraints, effectively avoiding the problem of energy-saving measures being disconnected from green production goals. Through a sliding window online parameter update mechanism and an error-triggered model retraining process, the model's adaptability to dynamic changes in production conditions is improved, ensuring the real-time performance and accuracy of the optimization strategy. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0014] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0015] 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.

[0016] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a machine learning-based method for green production and energy consumption management in equipment manufacturing enterprises, with specific steps including: S1. Collect multi-source heterogeneous data from the entire equipment manufacturing chain; perform unified alignment of timestamps through a high-precision time synchronization module, and then standardize the aligned data to obtain standardized feature vectors; Specifically, the multi-source heterogeneous data includes: The order information in the Manufacturing Execution System (MES) includes: order number, product model, planned start time, planned completion time, process route number, process sequence, and planned output. The IoT sensor data for the equipment includes: three-phase current, voltage, active power, reactive power, vibration acceleration, spindle speed, spindle load, tool wear, and coolant temperature; The process database parameters include: cutting speed, feed rate, depth of cut, tool type, coolant pressure, coolant flow rate setting, welding current, and welding voltage; Energy metering instrument data includes: cumulative consumption and instantaneous power / flow data of electricity meters, gas flow meters, and water meters, with a data acquisition granularity of 1 minute; The environmental monitoring system data includes: workshop carbon emission concentration, PM2.5 / PM10 particulate matter emissions, waste classification statistics, and hazardous waste generation. Data sampling frequency requirements: current and voltage signals should be no less than 100Hz, and vibration signals should be no less than 10kHz; all data sources should generate timestamps uniformly through a high-precision time synchronization module, with a timestamp alignment error of less than 50 milliseconds.

[0017] The formula for the standardized feature vector is expressed as follows: , in, Represents standardized data; Represents a vector after structured data has been standardized. This represents a vector after the structured data has been standardized. The structured data is normalized using the min-max normalization method to map the data to the [0,1] interval. For unstructured vibration signals, the frequency domain features are extracted using Fast Fourier Transform (FFT) and then normalized using the min-max normalization method. S2. Based on standardized feature vectors, a joint representation model integrating graph neural network and attention mechanism is constructed. Production equipment, process unit and material batch are used as nodes, the coupling strength between process flow, material flow and energy flow is used as edge weight, and the carbon emission per unit product, resource recycling rate and waste generation are used as node attributes. The output is a node embedding vector that integrates green attributes and energy consumption characteristics. The construction process of the joint representation model is as follows: Definition diagram , ,in, Represents a set of nodes. , Indicates the total number of nodes; Denotes the set of edges. , Represents a node and nodes The connection between them; In one embodiment, the specific formula for calculating the coupling strength between the process flow, material flow, and energy flow is as follows: , in, Represents a node With nodes The three-flow coupling strength between them; Represents a node With nodes The degree of coupling between process flows; Represents a node With nodes The degree of material flow coupling between them; Represents a node With nodes The degree of coupling between energy flows; , , Represents three different weighting coefficients, satisfying In one embodiment, , , .

[0018] Process flow coupling reflects whether there is a direct upstream and downstream process relationship between two nodes. When, it indicates a node It is a node Directly following process; when When, it indicates a node It is a node Indirect subsequent processes; when When there is no process association, it indicates that there is no process association.

[0019] Material flow coupling degree reflects the material flow from the node Flow to Node The traffic share and frequency are expressed by the formula: , in, Indicates the time from node t Flow to Node Material quality or quantity; Represents the node at time t The total mass or quantity of materials output; T represents the statistical period.

[0020] Energy flow coupling reflects the degree to which two nodes share an energy loop and their energy flows influence each other. The formula is as follows: , in, , Representing nodes respectively ,node Real-time power / energy consumption sequence; Represents covariance; Indicates variance.

[0021] The initial embedding of each node is: ,in, Represents the relationship between the standardized feature vector and the i-th node. Related subvectors; Represents a vector of green manufacturing indicators. , This represents the carbon emissions per unit of product at the i-th node; This represents the resource cycle rate of the i-th node; This indicates the amount of waste generated per unit of product, expressed in kg / piece. The carbon emissions per unit product are calculated based on the life cycle assessment method, and the formula is as follows: , in, Represents the i-th node The corresponding output of products or intermediate components; Represents the i-th node The total number of processes; , These represent the electrical energy and natural gas energy equivalent consumed in the k-th process, respectively. , These represent the carbon emission factors of the corresponding energy sources, preferably... The value is taken as 0.785 kg CO2 / kWh. The value is taken as 2.03 kg CO2 / m³; This represents the mass of raw materials consumed in the k-th process; This indicates the implicit carbon coefficient of the material; for steel, it is 2.1 kg CO2 / kg, and for aluminum, it is 11.0 kg CO2 / kg. The resource recycling rate The ratio of the mass of recyclable and reusable materials to the total mass of input materials is expressed by the following formula: , in, This represents the mass of recyclable and reusable material at the i-th node; This represents the total mass of input materials at the i-th node; The amount of waste generated per unit product The formula is expressed as: ,in, This represents the total mass of non-recyclable waste generated by the i-th node.

[0022] The graph neural network adopts a three-layer message passing architecture, with each layer including three stages: message construction, attention weighting, message aggregation, and node update.

[0023] In the message construction phase of layer l, the i-th node Receive from neighboring nodes The message is represented by the following formula: , in, Indicates the l-th layer slave node Passed to node The message; This represents a multilayer perceptron network containing two hidden layers, with ReLU activation function. Represents the (l-1)th level node The embedding vector; Represents the (l-1)th level node The embedding vector.

[0024] In the attention-weighted phase, the edge weights are dynamically adjusted using an attention coefficient, as expressed by the following formula: , in, Represents the node at level l For nodes Attention coefficient; This represents the normalized exponential function; This represents a linear rectified function with leakage, and the negative slope is 0.2. Represents the learnable weight matrix; Represents the attention vector; This indicates a vector concatenation operation.

[0025] Message aggregation and node updates are represented by the following formula: , in, Represents the i-th node The set of neighboring nodes; This represents a gated loop unit used to merge historical node information with newly aggregated neighborhood information.

[0026] After three layers of message passing, the final node is embedded. The green attributes and energy consumption characteristics of the local neighborhood are integrated as a joint representation output.

[0027] S3. With the dual objectives of improving energy efficiency and optimizing environmental performance, a multi-objective reinforcement learning algorithm based on the Pareto front is constructed. The node embedding vector output from step S2 is used as the input to the state space. Under the premise of meeting the production delivery cycle, equipment safety constraints and process quality requirements, optimization strategies are dynamically generated. The optimization strategies include equipment start-up and shutdown scheduling, process parameter adjustment and energy allocation instructions. Specifically, the construction process of the multi-objective reinforcement learning algorithm is as follows: state space : Embedded by the final node of all nodes Composed of splicing, with dimensions of ,in, Indicates the dimension of a single node's embedding; Action space The system is a mixed discrete-continuous space. Discrete actions include equipment operating speed (stop, low speed, medium speed, high speed), energy switching commands (switching to photovoltaic power supply, activating energy storage discharge, etc.), and equipment start-up and shutdown scheduling commands. Continuous actions include coolant flow rate (0–50L / min), spindle speed (500–8000rpm), and cutting feed rate (0.1–2.0mm / r). The reward function is a weighted average of energy efficiency gain and green performance gain, expressed by the following formula: , in, Indicates the state Next action The instant reward received; Indicates the current state; Indicates the action to be performed; This represents the weighting coefficient, which is dynamically adjusted based on the company's phased "dual carbon" targets, during the critical period of emission reduction. The value is 0.3, representing the energy efficiency improvement period. The value is 0.7; This represents the energy efficiency gain term; Indicates green performance gain items; The energy efficiency gain term is defined as the percentage reduction in energy consumption per unit of output, expressed by the formula: , in, Historical baseline energy consumption is the average energy consumption per unit of output for this production unit over the past 30 days. This indicates the current predicted energy consumption, i.e., the action to be performed. The predicted energy consumption per unit of output; The formula for the green performance gain term is expressed as follows: ,in, , , These represent the increase in resource recycling rate, the reduction in waste per unit of product, and the reduction in carbon emissions per unit of product, respectively. The formula for calculating the increase in resource recycling rate is: , Indicates the execution of an action The predicted resource recycling rate after that, Indicates the baseline resource recycling rate; The formula for calculating the reduction in waste per unit product is as follows: , Indicates the execution of an action The predicted waste generation per unit of product is as follows. This indicates the amount of waste generated per unit of product. The formula for calculating the reduction in carbon emissions per unit product is as follows: , Indicates the execution of an action The predicted carbon emissions per unit of product after that, Indicates carbon emissions per unit of product; The policy network adopts an Actor-Critic architecture. The Actor network outputs the action probability distribution, and the Critic network evaluates the state value. Both share the underlying feature extractor. The policy gradient method based on the Pareto front is used for training. In each iteration, a solution is randomly selected from the Pareto optimal solution set to update the network parameters.

[0028] In one embodiment, constraints are hard-coded into the action selection process, including equipment safety constraints, process quality constraints, and delivery cycle constraints. Equipment safety constraints: If the combination of spindle speed and coolant flow rate causes the predicted power to exceed 90% of the equipment's rated power, the spindle speed will be automatically reduced to below the safety threshold; if the equipment vibration acceleration exceeds the threshold, the equipment will be forced to stop for inspection. Process quality constraints: If the predicted product dimensional deviation exceeds the process tolerance zone, a negative bonus of -10 will be applied; if the predicted surface roughness exceeds the requirements, a negative bonus of -5 will be applied. Delivery cycle constraint: Calculate the remaining buffer time for critical path processes. If the remaining buffer time is less than 5% of the total project duration, it is prohibited to extend the processing time of critical path processes and delivery should be prioritized.

[0029] S4. Collect actual response data during the execution of the optimization strategy, construct a sliding window dataset, update the joint representation model parameters through online gradient descent based on the mean squared error loss function, and trigger the model retraining process when the prediction error exceeds the threshold. Specifically, real-time collection of actual response data includes actual equipment power consumption, actual process completion time, measured carbon emissions, actual resource recycling rate, actual waste generation, and product quality inspection data. Building a sliding window dataset The window size is set to 100 control cycles, and the step size is set to 1 control cycle. That is, each time data of one control cycle is added, the data of the oldest control cycle is removed. The loss function is the mean squared error loss function, calculated based on the data within the sliding window, and the formula is as follows: , in, Represents the loss function; Represents the parameters of the joint representation model; Indicates the size of the sliding window dataset; This indicates that the joint representation model is related to the input. The predicted values ​​include predicted energy consumption per unit of output, predicted carbon emissions per unit of product, predicted resource recycling rate, and predicted waste generation per unit of product. This represents the actual observed values, which are the real data obtained from energy meters, sensors, environmental monitoring systems, and MES systems, and are completely aligned with the dimensions of the predicted values. The parameters of the joint representation model are updated using stochastic gradient descent, as shown in the following formula: , in, This represents the model parameters at time t; This represents the learning rate, initially set to 0.001, which decays exponentially with the number of training iterations. If the relative error between predicted energy consumption and actual energy consumption exceeds 8% or the relative error between predicted carbon emissions and actual carbon emissions exceeds 10% within three consecutive control cycles, the model retraining process is triggered. The retraining process uses full historical data to fine-tune the parameters of the last two layers of the joint representation model, while the remaining layers are frozen to maintain model stability.

[0030] S5. Based on the historical data accumulated during the closed-loop control process of steps S1-S4, calculate the green energy efficiency index of each production unit and generate a visual evaluation report to guide long-term process improvement and equipment upgrades.

[0031] Specifically, calculating the green energy efficiency index includes: The historical data accumulated during the closed-loop control process includes: original multi-source heterogeneous data, generated node embedding data, generated optimization strategy data, collected actual response data, and model parameter update records; The green energy efficiency index is defined as the geometric mean of the overall energy efficiency score and the green performance score, expressed by the formula: , in, This represents the green energy efficiency index, with a value range of [value range missing]. A higher value indicates a better level of green energy efficiency. Indicates the energy efficiency score; Indicates green performance score; The energy efficiency score is calculated based on energy consumption per unit of output. ,in, This indicates the benchmark energy consumption per unit of output in the same industry or the company's historical best energy consumption per unit of output. This represents the actual energy consumption per unit of output value in the current period, expressed in kWh / 10,000 yuan. This indicates taking the minimum value; The green performance score It is obtained based on a weighted combination of the normalized resource recycling rate and carbon emission intensity, and the formula is expressed as: ,in, , These represent the normalized resource recycling rate and the normalized carbon emission intensity, respectively. Normalized resource recycling rate The calculation formula is expressed as follows: , in, , These represent the maximum and minimum values ​​of the resource recycling rate within the industry, respectively. The formula for calculating normalized carbon emission intensity is as follows: , in, , These represent the maximum and minimum carbon emissions per unit of product within the industry, respectively. , These represent two different weighting coefficients, each with a value of 0.5; In one embodiment, the assessment report is automatically generated at three time granularities: daily, weekly, and monthly. The report includes GEI rankings for each workshop, energy efficiency trend charts for production lines, and carbon footprint heatmaps for product families. It also supports drill-down analysis by workshop, production line, or product family.

[0032] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for green production and energy consumption management in equipment manufacturing enterprises, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data from the entire equipment manufacturing chain, perform unified timestamp alignment, and standardize the aligned data to obtain standardized feature vectors. S2. Construct a joint representation model based on standardized feature vectors that integrates graph neural networks and attention mechanisms. Take production equipment, process units, and material batches as nodes, the coupling strength between process flow, material flow, and energy flow as edge weights, and take unit product carbon emissions, resource recycling rate, and waste generation as node attributes, and output node embedding vectors. S3. Construct a multi-objective reinforcement learning algorithm based on the Pareto front. Use the node embedding vector output in step S2 as the input to the state space. Under the premise of meeting the production delivery cycle, equipment safety constraints and process quality requirements, dynamically generate optimization strategies. S4. Collect actual response data during the execution of the optimization strategy, construct a sliding window dataset, update the joint representation model parameters through online gradient descent based on the mean squared error loss function, and trigger the model retraining process when the prediction error exceeds the threshold. S5. Based on the historical data accumulated during the closed-loop control process of steps S1-S4, calculate the green energy efficiency index of each production unit and generate a visual evaluation report.

2. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning as described in claim 1, characterized in that, In step S1, the formula for the standardized feature vector is expressed as follows: ， in, Represents standardized data; Represents a vector after structured data has been standardized. This represents a vector after the structured data has been standardized.

3. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning as described in claim 2, characterized in that, In step S2, the construction process of the joint representation model is as follows: Definition diagram , , where represents the set of nodes, , represents the total number of nodes; represents the set of edges, , represents the connection between node and node . The initial embedding of each node is: ,in, Represents the relationship between the standardized feature vector and the i-th node. Related subvectors; Represents a vector of green manufacturing indicators. , This represents the carbon emissions per unit of product at the i-th node; This represents the resource cycle rate of the i-th node; Indicates the amount of waste generated per unit of product; The graph neural network adopts a three-layer message passing architecture, with each layer including three stages: message construction, attention weighting, message aggregation, and node update. In the message construction phase of layer l, the i-th node Receive from neighboring nodes The message is represented by the following formula: ， in, Indicates the l-th layer slave node Passed to node The message; Represents a multilayer perceptron network; Represents the (l-1)th level node The embedding vector; Represents the (l-1)th level node The embedding vector; In the attention-weighted phase, the edge weights are dynamically adjusted using an attention coefficient, as expressed by the following formula: ， in, Represents the node at level l For nodes Attention coefficient; This represents the normalized exponential function; Represents a linear rectified function with leakage; Represents the learnable weight matrix; Represents the attention vector; This represents a vector concatenation operation; Message aggregation and node updates are represented by the following formula: ， in, Represents the i-th node The set of neighboring nodes; This represents a gated loop unit used to merge historical node information with newly aggregated neighborhood information; After three layers of message passing, the final node embedding is obtained. .

4. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning as described in claim 3, characterized in that, The carbon emissions per unit product are calculated based on the life cycle assessment method, and the formula is as follows: ， in, Represents the i-th node The corresponding output of products or intermediate components; Represents the i-th node The total number of processes; , These represent the electrical energy and natural gas energy equivalent consumed in the k-th process, respectively. , These represent the carbon emission factors of the corresponding energy sources; This represents the mass of raw materials consumed in the k-th process; Indicates the implicit carbon coefficient of the material; The resource recycling rate It is the ratio of the mass of recyclable and reusable materials to the total mass of input materials; The amount of waste generated per unit product The formula is expressed as: ,in, This represents the total mass of non-recyclable waste generated by the i-th node.

5. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning as described in claim 3, characterized in that, In step S3, the construction process of the multi-objective reinforcement learning algorithm is as follows: state space : Embedded by the final node of all nodes Composed of splicing, with dimensions of ,in, Indicates the dimension of a single node's embedding; Action space This is a discrete-continuous hybrid space. Discrete actions include equipment operating speed, energy switching commands, and equipment start / stop scheduling commands; continuous actions include coolant flow rate, spindle speed, and cutting feed rate. The reward function is a weighted average of energy efficiency gain and green performance gain, expressed by the following formula: ， in, Indicates the state Next action The instant reward received; Indicates the current state; Indicates the action to be performed; Indicates the weighting coefficient; This represents the energy efficiency gain term; This represents the green performance gain item.

6. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning according to claim 5, characterized in that, In step S4, a sliding window dataset is constructed. The window size is set to 100 control cycles, and the step size is set to 1 control cycle. That is, each time data of one control cycle is added, the data of the oldest control cycle is removed. The loss function is the mean squared error loss function, calculated based on the data within the sliding window, and the formula is as follows: ， in, Represents the loss function; Represents the parameters of the joint representation model; Indicates the size of the sliding window dataset; This indicates that the joint representation model is related to the input. The predicted value; The values ​​represent the actual observations; the parameters of the joint characterization model are updated using the stochastic gradient descent method.

7. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning as described in claim 1, characterized in that, Step S5, calculating the green energy efficiency index specifically includes: The green energy efficiency index is defined as the geometric mean of the overall energy efficiency score and the green performance score, expressed by the formula: ， in, Indicates the green energy efficiency index; This represents the energy efficiency score, calculated based on energy consumption per unit of output. The green performance score is derived from a weighted combination of normalized resource recycling rate and carbon emission intensity.

8. The method for green production and energy consumption control of equipment manufacturing enterprises based on machine learning according to claim 1, characterized in that, Real-time collection of actual response data, including actual equipment power consumption, actual process completion time, measured carbon emissions, actual resource recycling rate, actual waste generation, and product quality inspection data.

Images