An energy consumption optimization control method based on an automated industrial production line

Abstract

The present application relates to a kind of based on automation industrial production line energy consumption optimization control method, it is related to energy consumption optimization control system technical field, the present application introduces one-dimensional convolutional neural network, multiple-source heterogeneous original working condition data is converted into time window matrix, adaptively extracts high robustness deep energy efficiency characteristics, strong filter is excluded random noise interference caused by material batch difference and environment, long short-term memory network is used to deeply mine long-range time sequence law of energy efficiency characteristic sequence, energy efficiency anomaly is converted into early concealed fault vanguard signal, without additional sensor hardware is added, low-cost construction is built high-fidelity multi-physical field coupling benchmark model, high-frequency data is filtered and sinks to edge computing node at the same time, completely eliminate network delay, long-period model is iteratively deployed in cloud, under the premise of confirming no fault safety, global optimal parameter is directly sent to controller, energy consumption adaptive closed-loop optimization considering equipment life and limit cost reduction is realized.

Description

Technical Field

[0001] This invention relates to the field of energy consumption optimization control system technology, and in particular to an energy consumption optimization control method based on automated industrial production lines. Background Technology

[0002] In the field of energy consumption optimization and control of current automated industrial production lines, although the level of digitalization and intelligence is constantly improving, existing control systems still have the following two problems when facing complex industrial environments:

[0003] I. Discrepancy between Real-Time and Virtual Data in Digital Twins: Although digital twins are widely used, the problem of virtual-real synchronization still exists. Many key parameters (such as the wear level of internal components, lubrication status, and hidden heat loss) are difficult to obtain directly through sensors, causing the simulation model to fail to accurately reflect the actual physical losses. Furthermore, there is a delay between sensor data acquisition, network transmission, and algorithm decision-making, causing the operating conditions of the physical line to drift when control commands are issued.

[0004] II. Poor Adaptability Under Abnormal Operating Conditions: While existing systems perform well in ideal environments, they exhibit shortcomings in complex and ever-changing industrial settings. Fixed-parameter control models often fail when raw material batch variations, power grid fluctuations, or minor mechanical faults occur. Furthermore, these systems typically handle energy consumption optimization and fault diagnosis separately. In reality, abnormal energy consumption fluctuations are often precursors to faults, and there is a lack of early warning and fault-tolerant control mechanisms based on energy consumption feature vectors.

[0005] To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to transform multi-source heterogeneous raw operating data into a time window matrix, adaptively extract highly robust deep energy efficiency characteristics, powerfully filter out random noise interference caused by material batch differences and environment, utilize long short-term memory networks to deeply mine the long-range time-series patterns of energy efficiency characteristic sequences, transform energy efficiency anomalies into early warning signals of hidden faults, construct a high-fidelity multi-physics coupled benchmark model at low cost without adding additional sensor hardware, and simultaneously filter high-frequency data down to edge computing nodes to completely eliminate network latency. Long-cycle model iteration is deployed in the cloud, and under the premise of confirmed fault-free safety, the globally optimal parameters are directly sent to the central controller, realizing adaptive closed-loop optimization of energy consumption that takes into account both equipment lifespan and extreme cost reduction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for energy consumption optimization and control based on automated industrial production lines, comprising the following steps:

[0008] S1. By extracting the original attributes of the physical mechanism model inside the digital twin system for dynamic deduction and assignment, and combining them with the explicit state variables that can be measured on site, an energy consumption benchmark model containing implicit parameters and multi-physics field coupling is constructed.

[0009] S2. Deploy edge computing nodes in the physical field close to the automated industrial production line, and send high-frequency real-time environmental perception, data filtering and transient control commands to the edge computing nodes for processing to achieve millisecond-level response. At the same time, deploy long-cycle energy efficiency feature library optimization and deep learning model iteration in the cloud computing center.

[0010] S3. Introduce a one-dimensional convolutional neural network at the edge computing node to convert multi-source heterogeneous working condition data into a time window matrix and input it into the one-dimensional convolutional neural network. Use the convolution kernel to adaptively extract deep energy efficiency feature vectors with noise resistance to filter out random interference caused by material batch differences or external environment, and obtain a deep energy efficiency feature vector sequence.

[0011] S4. Construct a fault prediction model based on long short-term memory network, take the deep energy efficiency feature vector sequence as the input and output the system energy consumption feature. When the system energy consumption feature is found to deviate from the benchmark model in a continuous distortion on the time axis, it is judged as an early hidden fault and the graded early warning mechanism is triggered in advance.

[0012] S5. Under the premise of confirming that there is no early warning fault, substitute the current optimal energy efficiency feature vector into the energy consumption optimization objective function, solve for the globally optimal control parameters, and issue the adjustment command to the main controller in real time to complete the closed-loop adaptive optimization of energy consumption of the automated production line.

[0013] Furthermore, the specific process of constructing a multi-physics coupled energy consumption benchmark model is as follows:

[0014] S11. Through the existing PLC control system and peripheral standard sensor network of the automated industrial production line, collect the explicit state variables during equipment operation. The explicit state variables include the three-phase voltage U(t), three-phase current I(t), and real-time speed of the spindle motor. Ambient temperature Te and pneumatic circuit intake pressure Pin;

[0015] S12. The internal physical mechanism model of the digital twin system predefines the material property library and ideal mechanical structure of the equipment. Based on the physical mechanism model, the internal model properties of the equipment are obtained, and the implicit loss parameters are calculated, including:

[0016] By extracting the initial tooth profile geometric parameters and material elastic modulus from the gearbox mechanism model, the dynamic friction torque coefficient during meshing is derived. ;

[0017] The resistivity of the stator winding material, the thermal conductivity of the insulation layer, and the heat dissipation area of ​​the casing are extracted from the three-dimensional digital model of the motor to calculate the equivalent transient thermal resistance inside the equipment. and heat capacity ;

[0018] S13. Deeply couple the explicit state variables obtained in S11 with the implicit loss parameters extracted in S12 to construct the following multiphysics coupled energy consumption equation to calculate the total energy consumption. :

[0019] ;

[0020] Wherein: Input electrical power ,in The phase difference angle between the three-phase voltage and the three-phase current;

[0021] Effective mechanical power =f( , );

[0022] Coupling loss power The core multi-field coupling term includes mechanical friction loss and heat loss, and its calculation formula is as follows:

[0023]

[0024] in, This is an estimated temperature rise value for the core components inside the equipment. Through the implicit parameter heat capacity The solution is obtained by coupling the apparent ambient temperature Te.

[0025] S14. Instantiate the above multiphysics coupling equations in the digital twin engine to generate an energy consumption benchmark model for a specific production process.

[0026] Furthermore, the specific process for deploying edge computing nodes is as follows:

[0027] S21. The edge computing node is deployed at the physical near end of the automated production line and communicates directly with the main controller of the production line via a short-range industrial Ethernet within the automated industrial production line. The edge computing node has a high-performance embedded processing unit.

[0028] S22. The edge computing node is used to undertake high-frequency real-time processing tasks, specifically including:

[0029] Real-time environmental perception: synchronously acquire current, pressure, and position at a sampling frequency higher than the preset frequency;

[0030] Transient data filtering: The moving average filtering algorithm is used to preprocess the original signal and remove glitches caused by electromagnetic interference;

[0031] Transient control issuance: The control parameters solved in S5 are converted locally and transformed into analog voltage signals that can be recognized by the underlying driver;

[0032] S23. The cloud computing center interacts asynchronously with edge computing nodes through an industrial gateway for computationally intensive and non-real-time strategic tasks, as detailed below:

[0033] Long-term energy efficiency feature library optimization: Summarize historical energy efficiency data from multiple production lines and different time periods, and use a global search algorithm to find the minimum energy consumption point in a wider range of dimensions;

[0034] Deep learning model iteration: Offline training and parameter fine-tuning are performed on the one-dimensional convolutional neural network in S3 and the long short-term memory network in S4. When the deep learning model achieves higher fitting accuracy, an updated weight file is generated.

[0035] Furthermore, the specific process for obtaining the deep energy efficiency feature vector sequence is as follows:

[0036] S31. At the edge computing node, perform timestamp alignment and normalization on N different dimensions of runtime data collected at the same time t to form a multi-source data column vector. ;

[0037] S32. Set a sliding time window of length L. As time progresses, concatenate the column vectors of multi-source data from L consecutive sampling points within the sliding time window to construct a two-dimensional time window matrix. As input to the neural network, it is represented as follows:

[0038] , where N is the number of running data;

[0039] The time window matrix integrates the spatial coupling and temporal evolution information of automated industrial production lines over a continuous period of time.

[0040] S33, Constructing the time window matrix Inputting the data into a one-dimensional convolutional neural network along the time axis, and setting K one-dimensional convolutional kernels of size 1×F in the one-dimensional convolutional layer of the one-dimensional convolutional neural network, the weight vector of the k-th convolutional kernel is denoted as... The bias is denoted as One-dimensional convolution is performed by sliding a one-dimensional convolution kernel across the time window matrix to extract local deep features, as shown below:

[0041] ;

[0042] in, Let f be the output of the i-th neuron on the k-th feature map, and let f represent the non-linear activation function.

[0043] S34. Input the local deep features output by the one-dimensional convolutional layer into the max pooling layer, and perform pooling operation through the max pooling layer to sample the local deep features as follows:

[0044] ;

[0045] S35. Flatten the output of the last pooling layer to obtain a one-dimensional vector that comprehensively represents the noise resistance energy efficiency state within the current time window L, i.e., the deep energy efficiency feature vector. As the time window continues to slide throughout the entire operating cycle, the edge computing nodes continuously output a series of discrete feature vectors, which are arranged in chronological order to form a deep energy efficiency feature vector sequence.

[0046] Furthermore, the specific process for determining early-stage, latent faults is as follows:

[0047] S41. Obtain the deep energy efficiency feature vector sequence S V =V1, V2, ..., As input, it is fed into a Long Short-Term Memory (LSTM) network, which includes an input gate, a forget gate, and an output gate. At time t, the hidden layer state ht and the cell state ct are updated using the state from the previous time step and the current feature vector.

[0048] ;

[0049] By using a long short-term memory network for forward prediction, the expected energy efficiency feature vector corresponding to normal, trouble-free operating conditions is output. ;

[0050] S42. Calculate the deep energy efficiency feature vector actually extracted at the current time t. With the benchmark expected vector Residual distance in feature space:

[0051] ;

[0052] Within a diagnostic assessment sliding window of length W, a time-weighted integral function with an exponentially decaying forgetting factor is introduced, and the distortion index is calculated according to the following formula. Used to characterize the energy consumption characteristics of the system:

[0053]

[0054] in, This is a preset time decay coefficient, and its value is greater than 0;

[0055] When the eigenvalue deviation exhibits continuity, cumulativeity, and irreversibility over time, the distortion index... The value increased significantly, effectively filtering out transient high-frequency noise that had been weakened but still remained under certain extreme conditions;

[0056] S43. Obtain the preset dynamic diagnostic threshold. When the distortion index is greater than the dynamic diagnostic threshold and the duration exceeds the set step size, it is determined that the equipment has an early hidden fault.

[0057] S44. Based on the absolute value of the distortion index and its first derivative, trigger dynamic hierarchical early warning:

[0058] Attention-level warning: If the distortion accumulates slowly but at a moderate rate, it will not disrupt the current production loop, but will push flexible maintenance suggestions to the cloud and MES system;

[0059] Intervention-level early warning: If the distortion exceeds the median threshold, the processing cycle time and peak acceleration are reduced to sacrifice some efficiency for equipment life and prevent the failure from escalating.

[0060] Circuit breaker warning: If the first derivative suddenly increases, the safety interrupt logic of the underlying PLC will be triggered immediately to cut off the power source and avoid catastrophic shutdown.

[0061] Furthermore, the specific process of solving for the globally optimal control parameters is as follows:

[0062] S51. Obtain the fault diagnosis status output by S4. When the diagnostic distortion index is less than the preset dynamic diagnosis threshold, obtain the current deep energy efficiency feature vector V output by S3. t As the current optimal energy efficiency feature vector, and to obtain the current batch production task requirements, initialize the solution boundary conditions;

[0063] S52. Based on energy consumption minimization, control smoothness, and production cycle constraints, and based on the logic of model predictive control, the following objective function J(U) is constructed in the future prediction time domain H:

[0064] ;

[0065] Among them: U=[u T+1 u T+2 ,…,u T+H ] represents the sequence of future control parameters to be solved;

[0066] The predicted energy consumption term utilizes a deep energy efficiency feature vector V that contains implicit parameter information. t This was obtained through high-fidelity deduction.

[0067] To control incremental penalty terms, which are used to limit the drastic acceleration and deceleration of the actuator and ensure the stability and lifespan of the underlying kinematic pairs;

[0068] This is a penalty function used to rigidly constrain the actual production cycle time. Not exceeding the scheduled delivery time;

[0069] α, β, and γ are dynamic weight coefficients issued based on long-term optimization in the cloud.

[0070] S53. Obtain the internal model attributes of the equipment in S1 and the deep energy efficiency feature vector in S3. Perform Taylor first-order expansion and linearization approximation on the nonlinear objective function at the current operating point, transform it into a standard quadratic programming problem, and quickly calculate the optimal control parameter sequence that minimizes J(U) within a limited number of iterations through edge computing nodes. Only the first control vector in the sequence is taken as the global optimal control parameter at the current moment.

[0071] S54. The globally optimal control parameters are mapped to specific adjustment commands for the underlying physical devices. The mapped adjustment commands are then sent directly to the production line's main controller via real-time industrial Ethernet, replacing the original fixed conservative parameters.

[0072] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0073] This energy consumption optimization and control method based on automated industrial production lines introduces a one-dimensional convolutional neural network to transform multi-source heterogeneous raw operating data into a time window matrix. It adaptively extracts highly robust deep energy efficiency features, powerfully filters out random noise interference caused by material batch differences and the environment, and uses a long short-term memory network to deeply mine the long-term time-series patterns of energy efficiency feature sequences. It transforms energy efficiency anomalies into early warning signals of hidden faults. Without adding additional sensor hardware, it constructs a high-fidelity multi-physics coupled benchmark model at low cost. At the same time, it filters high-frequency data and sinks it to edge computing nodes to completely eliminate network latency. Long-cycle model iteration is deployed in the cloud. Under the premise of confirming no faults and safety, the globally optimal parameters are directly sent to the central controller, realizing adaptive closed-loop optimization of energy consumption that takes into account both equipment life and extreme cost reduction. Attached Figure Description

[0074] Figure 1 A schematic diagram of the overall method flow of the present invention is shown. Detailed Implementation

[0075] 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.

[0076] Example:

[0077] like Figure 1 As shown, an energy consumption optimization and control method based on an automated industrial production line includes the following steps:

[0078] S1. By extracting the original attributes of the physical mechanism model inside the digital twin system for dynamic deduction and assignment, and combining them with the explicit state variables that can be measured on site, an energy consumption benchmark model containing implicit parameters and multi-physics field coupling is constructed.

[0079] The specific process of constructing a multiphysics coupled energy consumption benchmark model is as follows:

[0080] S11. Through the existing PLC control system and peripheral standard sensor network of the automated industrial production line, collect the explicit state variables during equipment operation. The explicit state variables include the three-phase voltage U(t), three-phase current I(t), and real-time speed of the spindle motor. Ambient temperature Te and pneumatic circuit intake pressure Pin;

[0081] S12. The internal physical mechanism model of the digital twin system predefines the material property library and ideal mechanical structure of the equipment. Based on the physical mechanism model, the internal model properties of the equipment are obtained, and the implicit loss parameters are calculated, including:

[0082] By extracting the initial tooth profile geometric parameters and material elastic modulus from the gearbox mechanism model, the dynamic friction torque coefficient during meshing is derived. ;

[0083] The resistivity of the stator winding material, the thermal conductivity of the insulation layer, and the heat dissipation area of ​​the casing are extracted from the three-dimensional digital model of the motor to calculate the equivalent transient thermal resistance inside the equipment. and heat capacity ;

[0084] S13. Deeply couple the explicit state variables obtained in S11 with the implicit loss parameters extracted in S12 to construct the following multiphysics coupled energy consumption equation to calculate the total energy consumption. :

[0085] ;

[0086] Wherein: Input electrical power ,in The phase difference angle between the three-phase voltage and the three-phase current;

[0087] Effective mechanical power =f( , );

[0088] Coupling loss power The core multi-field coupling term includes mechanical friction loss and heat loss, and its calculation formula is as follows:

[0089]

[0090] in, This is an estimated temperature rise value for the core components inside the equipment. Through the implicit parameter heat capacity The solution is obtained by coupling the apparent ambient temperature Te.

[0091] S14. Instantiate the above multiphysics coupling equations in the digital twin engine to generate an energy consumption benchmark model for a specific production process. The energy consumption benchmark model operates under ideal conditions without external interference and outputs a benchmark energy consumption curve for each standard production cycle.

[0092] S2. Deploy edge computing nodes in the physical field close to the automated industrial production line, and send high-frequency real-time environmental perception, data filtering and transient control commands to the edge computing nodes for processing to achieve millisecond-level response. At the same time, deploy long-cycle energy efficiency feature library optimization and deep learning model iteration in the cloud computing center.

[0093] The specific process for deploying edge computing nodes is as follows:

[0094] S21. The edge computing node is deployed at the physical near end of the automated production line and communicates directly with the main controller of the production line via a short-range industrial Ethernet within the automated industrial production line. The edge computing node has a high-performance embedded processing unit (such as an FPGA or a high-performance ARM processor). The edge computing node can directly subscribe to high-frequency raw data streams from the sensor layer, avoiding the wide area network latency caused by uploading data to a remote server.

[0095] S22. The edge computing node is used to undertake high-frequency real-time processing tasks, specifically including:

[0096] Real-time environmental perception: synchronously acquire current, pressure, and position at a sampling frequency higher than the preset frequency;

[0097] Transient data filtering: The moving average filtering algorithm is used to preprocess the original signal and remove glitches caused by electromagnetic interference;

[0098] Transient control issuance: The control parameters solved in S5 are converted locally and transformed into analog voltage signals that can be recognized by the underlying driver. Since the decision-making process is completed at the edge, the system's response loop is restricted to the local area network, thereby achieving a millisecond-level real-time response of 1ms-10ms, ensuring dynamic capture and rapid compensation of instantaneous energy consumption fluctuations in the production line.

[0099] S23. The cloud computing center interacts asynchronously with edge computing nodes through an industrial gateway for computationally intensive and non-real-time strategic tasks, as detailed below:

[0100] Long-cycle energy efficiency feature library optimization: Summarize historical energy efficiency data from multiple production lines and different time periods, and use global search algorithms (such as genetic algorithms or particle swarm algorithms) to find the minimum energy consumption point in a wider range of dimensions;

[0101] Deep learning model iteration: Offline training and parameter fine-tuning are performed on the one-dimensional convolutional neural network in S3 and the long short-term memory network in S4. When the deep learning model achieves higher fitting accuracy, an updated weight file is generated.

[0102] The cloud computing center regularly distributes the iterative deep learning model weights and the global optimal energy efficiency library to the edge computing nodes. The edge computing nodes then upload the feature data, which has been anonymized and dimensionality reduced, to the cloud computing center. This not only solves the stringent latency requirements of industrial sites, but also makes up for the shortcomings of insufficient computing power of individual edge nodes and the difficulty in self-evolution of deep learning, thus supporting the closed-loop adaptive optimization of the system under complex working conditions.

[0103] S3. Introduce a one-dimensional convolutional neural network at the edge computing node to convert multi-source heterogeneous working condition data into a time window matrix and input it into the one-dimensional convolutional neural network. Use the convolution kernel to adaptively extract deep energy efficiency feature vectors with noise resistance to filter out random interference caused by material batch differences or external environment, and obtain a deep energy efficiency feature vector sequence.

[0104] The specific process for obtaining the deep energy efficiency feature vector sequence is as follows:

[0105] S31. At the edge computing node, perform timestamp alignment and normalization on N different dimensions of operational data collected at the same time t (such as the instantaneous current It of the servo motor, the air pressure fluctuation Pt of the pneumatic system, the ambient temperature Tt, and the implicit loss parameters derived in S1) to form a multi-source data column vector. ;

[0106] S32. Set a sliding time window of length L. As time progresses, concatenate the column vectors of multi-source data from L consecutive sampling points within the sliding time window to construct a two-dimensional time window matrix. As input to the neural network, it is represented as follows:

[0107] , where N is the number of running data;

[0108] The time window matrix integrates the spatial coupling and temporal evolution information of automated industrial production lines over a continuous period of time.

[0109] S33, Constructing the time window matrix Inputting the data into a one-dimensional convolutional neural network along the time axis (i.e., the direction of dimension L), and setting K one-dimensional convolutional kernels of size 1×F in the one-dimensional convolutional layer of the one-dimensional convolutional neural network, the weight vector of the k-th convolutional kernel is denoted as... The bias is denoted as One-dimensional convolution is performed by sliding a one-dimensional convolution kernel across the time window matrix to extract local deep features, as shown below:

[0110] ;

[0111] in, Let f be the output of the i-th neuron on the k-th feature map, and let f represent the non-linear activation function.

[0112] S34. Input the local deep features output by the one-dimensional convolutional layer into the max pooling layer, and perform pooling operation through the max pooling layer to sample the local deep features as follows:

[0113] ;

[0114] Max pooling can adaptively preserve significant abrupt changes that characterize the true energy efficiency of a system, while smoothing out minor fluctuations caused by external environment or random disturbances, thereby greatly enhancing the noise resistance of extracted features and the robustness of the system.

[0115] S35. Flatten the output of the last pooling layer to obtain a one-dimensional vector that comprehensively represents the noise resistance energy efficiency state within the current time window L, i.e., the deep energy efficiency feature vector. As the time window continues to slide throughout the entire operating cycle, the edge computing nodes continuously output a series of discrete feature vectors, arranged in chronological order, forming a deep energy efficiency feature vector sequence. This deep energy efficiency feature vector sequence not only eliminates transient noise but also highly condenses the coupling characteristics of multiphysics fields, directly serving as a high-quality input source for the long short-term memory network in S4 to predict early fault evolution trends.

[0116] S4. Construct a fault prediction model based on long short-term memory network, take the deep energy efficiency feature vector sequence as the input and output the system energy consumption feature. When the system energy consumption feature is found to deviate from the benchmark model in a continuous distortion on the time axis, it is judged as an early hidden fault and the graded early warning mechanism is triggered in advance.

[0117] The specific process for identifying early-stage, hidden faults is as follows:

[0118] S41. Obtain the deep energy efficiency feature vector sequence S V =V1, V2, ..., As input, it is fed into a Long Short-Term Memory (LSTM) network, which includes an input gate, a forget gate, and an output gate. At time t, the hidden layer state ht and the cell state ct are updated using the state from the previous time step and the current feature vector.

[0119] ;

[0120] By using a long short-term memory network for forward prediction, the expected energy efficiency feature vector corresponding to normal, trouble-free operating conditions is output. ;

[0121] S42. Calculate the deep energy efficiency feature vector actually extracted at the current time t. With the benchmark expected vector Residual distance in feature space:

[0122] ;

[0123] Within a diagnostic assessment sliding window of length W, a time-weighted integral function with an exponentially decaying forgetting factor is introduced, and the distortion index is calculated according to the following formula. Used to characterize the energy consumption characteristics of the system:

[0124]

[0125] in, This is a preset time decay coefficient, and its value is greater than 0;

[0126] When the eigenvalue deviation exhibits continuity, cumulativeity, and irreversibility over time, the distortion index... The value increased significantly, effectively filtering out transient high-frequency noise that had been weakened but still remained under certain extreme conditions;

[0127] S43. Obtain the preset dynamic diagnostic threshold. When the distortion index is greater than the dynamic diagnostic threshold and the duration exceeds the set step size, it is determined that the equipment has an early hidden fault.

[0128] S44. Based on the absolute value of the distortion index and its first derivative (i.e., the rate of deterioration), trigger a dynamic, graded early warning:

[0129] Attention-level warning: If the distortion accumulates slowly but at a moderate rate, it will not disrupt the current production loop, but will push flexible maintenance suggestions to the cloud and MES system (such as "It is recommended to check the lubricating oil pressure during the next batch changeover").

[0130] Intervention-level early warning: If the distortion exceeds the median threshold, the processing cycle time and peak acceleration are reduced to sacrifice some efficiency for equipment life and prevent the failure from escalating.

[0131] Circuit breaker warning: If the first derivative suddenly increases (indicating a step deterioration in energy consumption), the safety interrupt logic of the underlying PLC will be triggered immediately to cut off the power source and avoid catastrophic shutdown.

[0132] S5. Under the premise of confirming that there is no early warning fault, substitute the current optimal energy efficiency feature vector into the energy consumption optimization objective function, solve for the globally optimal control parameters, and issue the adjustment command to the main controller in real time to complete the closed-loop adaptive optimization of energy consumption of the automated production line.

[0133] The specific process of solving for the globally optimal control parameters is as follows:

[0134] S51. Obtain the fault diagnosis status output by S4. When the diagnostic distortion index is less than the preset dynamic diagnosis threshold (i.e., it is confirmed as a healthy operating condition without early hidden faults), obtain the current deep energy efficiency feature vector V output by S3. t As the current optimal energy efficiency feature vector, and to obtain the current batch production task requirements, initialize and solve the boundary conditions (such as the target production cycle time Ttarget, processing accuracy tolerance, etc.).

[0135] S52. Based on energy consumption minimization, control smoothness, and production cycle constraints, and based on the logic of model predictive control, the following objective function J(U) is constructed in the future prediction time domain H:

[0136] ;

[0137] Among them: U=[u T+1 u T+2 ,…,u T+H ] represents the sequence of future control parameters to be solved;

[0138] The predicted energy consumption term utilizes a deep energy efficiency feature vector V that contains implicit parameter information. t This was obtained through high-fidelity deduction.

[0139] To control incremental penalty terms, which are used to limit the drastic acceleration and deceleration of the actuator and ensure the stability and lifespan of the underlying kinematic pairs;

[0140] This is a penalty function used to rigidly constrain the actual production cycle time. Not exceeding the scheduled delivery time;

[0141] α, β, and γ are dynamic weight coefficients issued based on long-term optimization in the cloud.

[0142] S53. Obtain the internal model attributes of the equipment in S1 and the deep energy efficiency feature vector in S3. Perform Taylor first-order expansion and linearization approximation on the nonlinear objective function at the current operating point, transform it into a standard quadratic programming problem, and quickly calculate the optimal control parameter sequence that minimizes J(U) within a limited number of iterations through edge computing nodes. Only the first control vector in the sequence is taken as the global optimal control parameter at the current moment.

[0143] S54. The globally optimal control parameters are mapped to specific adjustment commands of the underlying physical devices. The mapped adjustment commands are directly sent to the production line's main controller via real-time industrial Ethernet, replacing the original fixed conservative parameters. In the next sampling cycle, the system enters a new state and continues to repeat the process from S1 to S5, thereby realizing the energy consumption closed-loop adaptive optimization of the automated production line under varying operating conditions.

[0144] For example, for complex production line collaborative operating conditions, the control vector can be decoupled as follows:

[0145] Torque feedforward coefficient and maximum speed limit of servo motor;

[0146] The opening advance and throttling buffer time of pneumatic actuator valves;

[0147] The timing logic for waiting processes between multiple devices.

[0148] This invention introduces a one-dimensional convolutional neural network to transform multi-source heterogeneous raw operating data into a time window matrix, adaptively extracting highly robust deep energy efficiency features, powerfully filtering out random noise interference caused by material batch differences and the environment, and using a long short-term memory network to deeply mine the long-range time-series patterns of energy efficiency feature sequences, transforming energy efficiency anomalies into early warning signals of hidden faults. Without adding additional sensor hardware, a high-fidelity multi-physics coupled benchmark model is constructed at low cost. At the same time, high-frequency data is filtered and pushed down to edge computing nodes to completely eliminate network latency. Long-cycle model iterations are deployed in the cloud, and under the premise of confirmed fault-free safety, the globally optimal parameters are directly sent to the central controller, realizing adaptive closed-loop optimization of energy consumption that balances equipment lifespan and extreme cost reduction.

[0149] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0150] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0151] In the two embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices or modules, and may be electrical, mechanical or other forms.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing energy consumption control in automated industrial production lines, characterized in that, Includes the following steps: S1. By extracting the original attributes of the physical mechanism model inside the digital twin system for dynamic deduction and assignment, and combining them with the explicit state variables that can be measured on site, an energy consumption benchmark model containing implicit parameters and multi-physics field coupling is constructed. S2. Deploy edge computing nodes in the physical field close to the automated industrial production line, and send high-frequency real-time environmental perception, data filtering and transient control commands to the edge computing nodes for processing to achieve millisecond-level response. At the same time, deploy long-cycle energy efficiency feature library optimization and deep learning model iteration in the cloud computing center. S3. Introduce a one-dimensional convolutional neural network at the edge computing node to convert multi-source heterogeneous working condition data into a time window matrix and input it into the one-dimensional convolutional neural network. Use the convolution kernel to adaptively extract deep energy efficiency feature vectors with noise resistance to filter out random interference caused by material batch differences or external environment, and obtain a deep energy efficiency feature vector sequence. S4. Construct a fault prediction model based on long short-term memory network, take the deep energy efficiency feature vector sequence as the input and output the system energy consumption feature. When the system energy consumption feature is found to deviate from the benchmark model in a continuous distortion on the time axis, it is judged as an early hidden fault and the graded early warning mechanism is triggered in advance. S5. Under the premise of confirming that there is no early warning fault, substitute the current optimal energy efficiency feature vector into the energy consumption optimization objective function, solve for the globally optimal control parameters, and issue the adjustment command to the main controller in real time to complete the closed-loop adaptive optimization of energy consumption of the automated production line.

2. The energy consumption optimization and control method based on an automated industrial production line according to claim 1, characterized in that, The specific process of constructing a multiphysics coupled energy consumption benchmark model is as follows: S11. Through the existing PLC control system and peripheral standard sensor network of the automated industrial production line, collect the explicit state variables during equipment operation. The explicit state variables include the three-phase voltage U(t), three-phase current I(t), and real-time speed of the spindle motor. Ambient temperature Te and pneumatic circuit intake pressure Pin; S12. The internal physical mechanism model of the digital twin system predefines the material property library and ideal mechanical structure of the equipment. Based on the physical mechanism model, the internal model properties of the equipment are obtained, and the implicit loss parameters are calculated, including: By extracting the initial tooth profile geometric parameters and material elastic modulus from the gearbox mechanism model, the dynamic friction torque coefficient during meshing is derived. ; The resistivity of the stator winding material, the thermal conductivity of the insulation layer, and the heat dissipation area of ​​the casing are extracted from the three-dimensional digital model of the motor to calculate the equivalent transient thermal resistance inside the equipment. and heat capacity ; S13. Deeply couple the explicit state variables obtained in S11 with the implicit loss parameters extracted in S12 to construct the following multiphysics coupled energy consumption equation to calculate the total energy consumption. : ； Wherein: Input electrical power ,in The phase difference angle between the three-phase voltage and the three-phase current; Effective mechanical power =f( , ); Coupling loss power The core multi-field coupling term includes mechanical friction loss and heat loss, and its calculation formula is as follows: in, This is an estimated temperature rise value for the core components inside the equipment. Through the implicit parameter heat capacity The solution is obtained by coupling the apparent ambient temperature Te. S14. Instantiate the above multiphysics coupling equations in the digital twin engine to generate an energy consumption benchmark model for a specific production process.

3. The energy consumption optimization control method based on an automated industrial production line according to claim 1, characterized in that, The specific process for deploying edge computing nodes is as follows: S21. The edge computing node is deployed at the physical near end of the automated production line and communicates directly with the main controller of the production line via a short-range industrial Ethernet within the automated industrial production line. The edge computing node has a high-performance embedded processing unit. S22. The edge computing node is used to undertake high-frequency real-time processing tasks, specifically including: Real-time environmental perception: synchronously acquire current, pressure, and position at a sampling frequency higher than the preset frequency; Transient data filtering: The moving average filtering algorithm is used to preprocess the original signal and remove glitches caused by electromagnetic interference; Transient control issuance: The control parameters solved in S5 are converted locally and transformed into analog voltage signals that can be recognized by the underlying driver; S23. The cloud computing center interacts asynchronously with edge computing nodes through an industrial gateway for computationally intensive and non-real-time strategic tasks, as detailed below: Long-term energy efficiency feature library optimization: Summarize historical energy efficiency data from multiple production lines and different time periods, and use a global search algorithm to find the minimum energy consumption point in a wider range of dimensions; Deep learning model iteration: Offline training and parameter fine-tuning are performed on the one-dimensional convolutional neural network in S3 and the long short-term memory network in S4. When the deep learning model achieves higher fitting accuracy, an updated weight file is generated.

4. The energy consumption optimization and control method based on an automated industrial production line according to claim 1, characterized in that, The specific process for obtaining the deep energy efficiency feature vector sequence is as follows: S31. At the edge computing node, perform timestamp alignment and normalization on N different dimensions of runtime data collected at the same time t to form a multi-source data column vector. ; S32. Set a sliding time window of length L. As time progresses, concatenate the column vectors of multi-source data from L consecutive sampling points within the sliding time window to construct a two-dimensional time window matrix. As input to the neural network, it is represented as follows: , where N is the number of running data; The time window matrix integrates the spatial coupling and temporal evolution information of automated industrial production lines over a continuous period of time. S33, Constructing the time window matrix Inputting the data into a one-dimensional convolutional neural network along the time axis, and setting K one-dimensional convolutional kernels of size 1×F in the one-dimensional convolutional layer of the one-dimensional convolutional neural network, the weight vector of the k-th convolutional kernel is denoted as... The bias is denoted as One-dimensional convolution is performed by sliding a one-dimensional convolution kernel across the time window matrix to extract local deep features, as shown below: ； in, Let f be the output of the i-th neuron on the k-th feature map, and let f represent the non-linear activation function. S34. Input the local deep features output by the one-dimensional convolutional layer into the max pooling layer, and perform pooling operation through the max pooling layer to sample the local deep features as follows: ； S35. Flatten the output of the last pooling layer to obtain a one-dimensional vector that comprehensively represents the noise resistance energy efficiency state within the current time window L, i.e., the deep energy efficiency feature vector. As the time window continues to slide throughout the entire operating cycle, the edge computing nodes continuously output a series of discrete feature vectors, which are arranged in chronological order to form a deep energy efficiency feature vector sequence.

5. The energy consumption optimization and control method based on an automated industrial production line according to claim 1, characterized in that, The specific process for identifying early-stage, hidden faults is as follows: S41. Obtain the deep energy efficiency feature vector sequence S V =V1, V2, ..., As input, it is fed into a Long Short-Term Memory (LSTM) network, which includes an input gate, a forget gate, and an output gate. At time t, the hidden layer state ht and the cell state ct are updated using the state from the previous time step and the current feature vector. ； By using a long short-term memory network for forward prediction, the expected energy efficiency feature vector corresponding to normal, trouble-free operating conditions is output. ; S42. Calculate the deep energy efficiency feature vector actually extracted at the current time t. With the benchmark expected vector Residual distance in feature space: ； Within a diagnostic assessment sliding window of length W, a time-weighted integral function with an exponentially decaying forgetting factor is introduced, and the distortion index is calculated according to the following formula. Used to characterize the energy consumption characteristics of the system: in, This is a preset time decay coefficient, and its value is greater than 0; When the eigenvalue deviation exhibits continuity, cumulativeity, and irreversibility over time, the distortion index... The value increased significantly, effectively filtering out transient high-frequency noise that had been weakened but still remained under certain extreme conditions; S43. Obtain the preset dynamic diagnostic threshold. When the distortion index is greater than the dynamic diagnostic threshold and the duration exceeds the set step size, it is determined that the equipment has an early hidden fault. S44. Based on the absolute value of the distortion index and its first derivative, trigger dynamic hierarchical early warning: Attention-level warning: If the distortion accumulates slowly but at a moderate rate, it will not disrupt the current production loop, but will push flexible maintenance suggestions to the cloud and MES system; Intervention-level early warning: If the distortion exceeds the median threshold, the processing cycle time and peak acceleration are reduced to sacrifice some efficiency for equipment life and prevent the failure from escalating. Circuit breaker warning: If the first derivative suddenly increases, the safety interrupt logic of the underlying PLC will be triggered immediately to cut off the power source and avoid catastrophic shutdown.

6. The energy consumption optimization control method based on an automated industrial production line according to claim 1, characterized in that, The specific process of solving for the globally optimal control parameters is as follows: S51. Obtain the fault diagnosis status output by S4. When the diagnostic distortion index is less than the preset dynamic diagnosis threshold, obtain the current deep energy efficiency feature vector V output by S3. t As the current optimal energy efficiency feature vector, and to obtain the current batch production task requirements, initialize the solution boundary conditions; S52. Based on energy consumption minimization, control smoothness, and production cycle constraints, and based on the logic of model predictive control, the following objective function J(U) is constructed in the future prediction time domain H: ； Among them: U=[u T+1 u T+2 ,…,u T+H ] represents the sequence of future control parameters to be solved; The predicted energy consumption term utilizes a deep energy efficiency feature vector V that contains implicit parameter information. t This was obtained through high-fidelity deduction. To control incremental penalty terms, this is used to limit the drastic acceleration and deceleration of the actuator and ensure the stability and lifespan of the underlying kinematic pairs; This is a penalty function used to hard-constrain the actual production cycle time. Not exceeding the scheduled delivery time; α, β, and γ are dynamic weight coefficients issued based on long-term optimization in the cloud. S53. Obtain the internal model attributes of the equipment in S1 and the deep energy efficiency feature vector in S3. Perform Taylor first-order expansion and linearization approximation on the nonlinear objective function at the current operating point, transform it into a standard quadratic programming problem, and quickly calculate the optimal control parameter sequence that minimizes J(U) within a limited number of iterations through edge computing nodes. Only the first control vector in the sequence is taken as the global optimal control parameter at the current moment. S54. The globally optimal control parameters are mapped to specific adjustment commands for the underlying physical devices. The mapped adjustment commands are then sent directly to the production line's main controller via real-time industrial Ethernet, replacing the original fixed conservative parameters.

Images