An intelligent manufacturing information decision processing method based on industrial data analysis

Abstract

The application discloses an intelligent manufacturing information decision processing method based on industrial data analysis and relates to the technical field of industrial data analysis, which comprises the following steps: collecting multi-source operation state data of industrial equipment to generate standardized multi-source operation state data; constructing a fractional order Jiles-Atherton hysteresis model to generate a linearized prediction state sequence; segmenting the linearized prediction state sequence to generate a segmented state change function; constructing a manufacturing decision constraint generation module to generate a non-smooth state reachable domain sequence; constructing a manufacturing decision variable set to generate a manufacturing decision candidate set; and evaluating the manufacturing decision candidate set to generate a manufacturing decision result. By introducing the fractional order Jiles-Atherton hysteresis model and the Filippov differential inclusion method, the application realizes operation state dynamic prediction and manufacturing decision adaptive optimization control in a complex industrial manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of industrial data analysis technology, and in particular to an intelligent manufacturing information decision processing method based on industrial data analysis. Background Technology

[0002] With the continuous development of industrial internet, edge computing, and intelligent manufacturing technologies, a large number of industrial devices have acquired real-time data acquisition and networking capabilities. Monitoring operational status, collaborative equipment control, and manufacturing decision optimization in industrial production processes have gradually become important research directions for industrial intelligence. Existing technologies typically analyze the operational status of industrial equipment by collecting equipment operating parameters, load parameters, temperature parameters, and actuator status parameters. This analysis is then combined with predictive models or rule-based control methods to achieve equipment scheduling and process control, thereby improving production efficiency and reducing energy consumption. In complex industrial scenarios, the operational status of industrial equipment often exhibits significant nonlinearity, time-varying characteristics, and coupling. Furthermore, state transitions and operational disturbances exist between different process stages. Therefore, traditional static analysis methods are insufficient to accurately describe the dynamic evolution of complex industrial states.

[0003] Existing technical solutions employ hysteresis models, state-space models, or data-driven prediction models to model the operating state of industrial equipment and generate manufacturing control strategies based on the state prediction results. However, existing basic technologies typically only model continuous and smooth states, lacking effective constraints on non-smooth state boundaries generated during process stage switching, load abrupt changes, and equipment start-up and shutdown. This can easily lead to deviations between the state prediction results and the actual operating state. Furthermore, the lack of dynamic constraints on the reachable range of states during the manufacturing decision-making process results in insufficient stability of the generated decision results.

[0004] Therefore, how to provide a smart manufacturing information decision-making processing method based on industrial data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent manufacturing information decision-making processing method based on industrial data analysis. This invention utilizes fractional-order Jiles-Atherton hysteresis modeling, Koopman high-dimensional state mapping, Filippov non-smooth dynamic constraints, and edge-side industrial data collaborative analysis technology to detail the implementation process of dynamic evolution of industrial equipment operating states, generation of process stage switching constraints, and optimization of manufacturing decisions. It achieves dynamic prediction under complex industrial operating conditions, constraint of non-smooth state reachability domains, and adaptive optimization of manufacturing decisions. Compared to traditional industrial manufacturing decision-making methods, this invention can simultaneously characterize equipment hysteresis characteristics, abrupt changes in operating conditions, and state switching characteristics, possessing advantages such as high state prediction accuracy, strong adaptability to complex operating conditions, high stability of manufacturing decisions, low energy consumption during equipment operation, and high reliability of industrial process control.

[0006] According to an embodiment of the present invention, a smart manufacturing information decision processing method based on industrial data analysis includes: Multi-source operating status data of industrial equipment is collected at edge nodes, and the multi-source operating status data is preprocessed to generate standardized multi-source operating status data. Based on standardized multi-source operating state data, a fractional-order Jiles-Atherton hysteresis model is constructed to calculate magnetization intensity, reversible magnetization component, irreversible magnetization component and hysteresis loss, generate hysteresis state evolution sequence, and map it to the Koopman high-dimensional observable space to construct a linear state evolution matrix and generate a linearized predicted state sequence. Based on standardized multi-source operating status data, process stage identifiers, load and equipment start-up and shutdown status data are read to identify process stage switching points, load mutation points and equipment status switching points. The linearized predicted state sequence is segmented to generate segmented state change functions. Based on the piecewise state change function, a manufacturing decision constraint generation module based on Filippov differential is constructed to calculate the state switching boundary set and solve the state reachable interval, generating a sequence of non-smooth state reachable domains. A set of manufacturing decision variables is constructed at the edge nodes. State prediction is performed by combining the hysteresis state evolution sequence, the linearized predicted state sequence, and the state switching boundary set. Candidate decision variables that exceed the non-smooth state reachability sequence are eliminated to generate a set of manufacturing decision candidates. The system performs multi-objective evaluation on the candidate set of manufacturing decisions, generates manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear and operational stability, and controls industrial equipment to execute manufacturing decisions.

[0007] Optionally, the multi-source operating status data specifically includes excitation current data, magnetic field strength data, magnetic flux density data, temperature data, load data, actuator displacement data, process stage identification data, and equipment start-up and shutdown status data.

[0008] Optionally, the preprocessing of multi-source operational status data specifically includes timestamp alignment, data normalization, outlier removal, missing value imputation, data denoising, data smoothing, and data standardization.

[0009] Optionally, generating the linearized predicted state sequence includes: A fractional-order Jiles-Atherton hysteresis model is constructed based on standardized multi-source operating state data. The fractional-order Jiles-Atherton hysteresis model includes an equivalent magnetization driving module, a magnetization state decomposition module, a fractional-order magnetization evolution module, and a hysteresis state mapping module. The equivalent magnetization drive module calculates the difference between the magnetic flux density and the magnetic field strength to obtain the magnetization intensity state variable. The applied magnetic field and the magnetization intensity state variable are superimposed to construct the equivalent magnetic field. The equivalent magnetic field is then subjected to nonlinear mapping calculation to obtain the hysteresis-free magnetization intensity. The magnetization state decomposition module performs decomposition calculations based on hysteresis-free magnetization intensity and magnetization intensity, decomposing the magnetization intensity into reversible magnetization components and irreversible magnetization components. It also constructs a dynamic proportional relationship based on temperature data, load data, and magnetic field intensity changes, and performs adaptive adjustment processing on the reversible and irreversible magnetization components. Furthermore, it introduces hysteresis loss constraints to perform consistency correction on the decomposition results, obtaining the corrected reversible and irreversible magnetization components. The fractional-order magnetization evolution module performs tensor multi-source coupling processing based on hysteresis-free magnetization intensity, corrected reversible magnetization components and irreversible magnetization components. It constructs a multi-dimensional tensor structure from standardized multi-source operating state data in the time dimension, physical quantity dimension and equipment state dimension, and performs fractional-order dynamic evolution calculation processing on dual time scales. The instantaneous response of magnetization is calculated along a fast timescale path, and the historical cumulative effect of magnetization is calculated along a slow timescale path. The results of the two evolution paths are coupled and fused to obtain the magnetization evolution result. Magnetization intensity is used as nodes and magnetization state transition relationships are used as edges to construct a magnetization state graph. The graph structure is then updated to obtain the hysteresis state evolution sequence. The hysteresis state mapping module performs lap integral processing based on magnetic field strength data and magnetic flux density data to obtain hysteresis loss, and maps the hysteresis state evolution sequence to the Koopman high-dimensional observable space to obtain high-dimensional observable states. It then constructs a linear state evolution matrix and performs state advancement to obtain a linearized predicted state sequence. The fractional-order Jiles-Atherton hysteresis model was trained using the reconstruction error of the magnetization evolution sequence as the optimization objective. The parameters in the equivalent magnetization driving module, magnetization state decomposition module, fractional-order magnetization evolution module, and hysteresis state mapping module were continuously optimized. Training was stopped when the reconstruction error was less than 0.01 and the parameter update amplitude was less than 0.001 within 5 consecutive iterations.

[0010] Optionally, the generation of the piecewise state change function includes: Based on standardized multi-source operating status data, process stage identification data, load data and equipment start-up and shutdown status data are aligned according to a unified time sequence to construct an operating status sequence. Based on the operating condition state sequence, state change detection and processing is performed. The location where the process stage identifier changes is identified as the process stage switching point. The location where the load data difference between adjacent time moments is greater than 0.05 is identified as the load mutation point. The location where the equipment start-up and shutdown status changes is identified as the equipment state switching point. These are then merged to obtain a set of state switching points. The linearized predicted state sequence is segmented based on the set of state switching points. The linearized predicted state sequence is divided into multiple continuous subsequences according to adjacent state switching points, and a corresponding state change function is constructed for each subsequence to obtain a set of segmented state change functions.

[0011] Optionally, generating the sequence of non-smooth reachable domains includes: A manufacturing decision constraint generation module based on Filippov differential is constructed. The module includes a state switching topology representation unit, a state encoding unit, a Filippov reachable domain calculation unit, and a state domain generation unit. The state switching topology representation unit is based on the segmented state change function. Each segmented state change function is represented as a node, the switching relationship between adjacent segments is represented as a directed edge, and the state switching position is determined at the position corresponding to the directed edge to construct a segmented switching graph. The state coding unit performs path-dependent coding based on the linearized predicted state sequence, encodes the evolution information of the linearized predicted state on the historical path, obtains the extended state sequence, and embeds it into the segmented switching graph to obtain the state switching boundary set. The Filippov reachability computation unit performs set-valued dynamics computation based on the state switching boundary set. It represents the derivative of the piecewise state change function as multiple vector fields, performs weighted convex combination operation on the vector fields at the state switching boundary to obtain the state change rate belonging to the convex hull of the vector fields, introduces a hysteresis operator kernel to perform weighted convolution on the historical linearized predicted state, and combines the extended state sequence to perform time-progression calculation on the linearized predicted state to obtain the state reachability interval. The state domain generation unit constructs the state domain based on the state reachable interval. It divides the state reachable interval into multiple sub-intervals according to the piecewise state change function, and performs boundary reconstruction processing on each sub-interval to form a continuous boundary surface. It combines the hysteresis loss to constrain and filter the state domain, and performs set expansion and boundary integration processing on each sub-interval to generate a non-smooth state reachable domain sequence.

[0012] Optionally, generating the candidate set for manufacturing decisions includes: A set of manufacturing decision variables is constructed at the edge nodes. The manufacturing decision variables are represented as a multi-dimensional variable sequence consisting of actuator displacement adjustment, load adjustment and equipment operating status adjustment. The time-aligned data is then processed with standardized multi-source operating status data to obtain a set of candidate decision variables. State prediction is performed based on the hysteresis state evolution sequence, the linearized predicted state sequence, and the set of state switching boundaries. The set of candidate decision variables is substituted into the linear state evolution matrix for state advancement calculation. The predicted state is then corrected by combining the hysteresis state evolution sequence to obtain the corresponding set of candidate state trajectories. The candidate state trajectory set is constrained and filtered based on the non-smooth state reachability domain sequence. Candidate decision variables whose predicted states are within the state range corresponding to the non-smooth state reachability domain sequence are retained, while candidate decision variables whose predicted states are outside the state range corresponding to the non-smooth state reachability domain sequence are removed, thus obtaining the manufacturing decision candidate set.

[0013] Optionally, the step of generating manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear, and operational stability includes: A multi-objective evaluation process is performed on the candidate set of manufacturing decisions. The production efficiency, energy consumption, equipment loss and operational stability corresponding to each candidate decision are calculated and normalized to obtain the evaluation index sequence. The evaluation index sequence is weighted and fused. Production efficiency, energy consumption, equipment loss and operational stability are weighted and summed according to the set weights to obtain the comprehensive evaluation value of the corresponding candidate decision. The comprehensive evaluation values ​​are then sorted to obtain the priority sequence of candidate decisions. The manufacturing decision is selected based on the priority sequence of candidate decisions, and the candidate decision with the best comprehensive evaluation value is selected as the manufacturing decision result. Based on the manufacturing decision result, the displacement data, load data and equipment start-up and shutdown status data of the actuator are controlled and adjusted to complete the manufacturing decision execution.

[0014] The beneficial effects of this invention are: This invention proposes an intelligent manufacturing information decision-making processing method based on industrial data analysis. By uniformly collecting, preprocessing, and dynamically modeling multi-source operating status data of industrial equipment at edge nodes, it deeply couples the fractional-order Jiles-Atherton hysteresis model, Koopman high-dimensional state mapping, and Filippov differential inclusion constraint mechanism. This achieves dynamic prediction of operating status and collaborative optimization of manufacturing decisions in complex industrial manufacturing processes. Compared to traditional industrial manufacturing decision-making methods that rely solely on static rules or single prediction models, this invention can simultaneously characterize the hysteresis characteristics, historical memory characteristics, and non-smooth state change characteristics of industrial equipment during process switching, load abrupt changes, and equipment start-up and shutdown. This improves the consistency between state prediction results and actual operating states under complex industrial conditions and reduces prediction bias caused by sudden state changes.

[0015] This invention constructs a manufacturing decision constraint generation module based on Filippov differentials, dynamically calculates state switching boundaries and state reachable intervals, and combines non-smooth state reachable domain sequences to screen candidate manufacturing decision variables, so that the manufacturing decision results are always within the stable operating range of the equipment, thereby reducing the risk of abnormal fluctuations during equipment operation.

[0016] Meanwhile, this invention utilizes the hysteresis state evolution sequence to dynamically correct the linearized predicted state, thereby improving the state evolution stability of complex nonlinear industrial systems under long-term operating conditions. Furthermore, through a joint evaluation mechanism of production efficiency, energy consumption, equipment wear and tear, and operational stability, it achieves comprehensive optimization of manufacturing decision-making results, possessing advantages such as high manufacturing decision-making stability, strong adaptability to complex operating conditions, high equipment operational reliability, and good industrial energy consumption control. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent manufacturing information decision processing method based on industrial data analysis proposed in this invention; Figure 2This is a schematic diagram of the fractional-order Jiles-Atherton hysteresis model of an intelligent manufacturing information decision processing method based on industrial data analysis proposed in this invention. Figure 3 This is a schematic diagram of the manufacturing decision constraint generation module of an intelligent manufacturing information decision processing method based on industrial data analysis proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A smart manufacturing information decision processing method based on industrial data analysis includes: Multi-source operating status data of industrial equipment is collected at edge nodes, and the multi-source operating status data is preprocessed to generate standardized multi-source operating status data. Based on standardized multi-source operating state data, a fractional-order Jiles-Atherton hysteresis model is constructed to calculate magnetization intensity, reversible magnetization component, irreversible magnetization component and hysteresis loss, generate hysteresis state evolution sequence, and map it to the Koopman high-dimensional observable space to construct a linear state evolution matrix and generate a linearized predicted state sequence. Based on standardized multi-source operating status data, process stage identifiers, load and equipment start-up and shutdown status data are read to identify process stage switching points, load mutation points and equipment status switching points. The linearized predicted state sequence is segmented to generate segmented state change functions. Based on the piecewise state change function, a manufacturing decision constraint generation module based on Filippov differential is constructed to calculate the state switching boundary set and solve the state reachable interval, generating a sequence of non-smooth state reachable domains. A set of manufacturing decision variables is constructed at the edge nodes. State prediction is performed by combining the hysteresis state evolution sequence, the linearized predicted state sequence, and the state switching boundary set. Candidate decision variables that exceed the non-smooth state reachability sequence are eliminated to generate a set of manufacturing decision candidates. The system performs multi-objective evaluation on the candidate set of manufacturing decisions, generates manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear and operational stability, and controls industrial equipment to execute manufacturing decisions.

[0020] In this embodiment, the multi-source operating status data specifically includes excitation current data, magnetic field strength data, magnetic flux density data, temperature data, load data, actuator displacement data, process stage identification data, and equipment start-up and shutdown status data.

[0021] In this embodiment, the preprocessing of multi-source operational status data specifically includes timestamp alignment, data normalization, outlier removal, missing value imputation, data denoising, data smoothing, and data standardization.

[0022] In this embodiment, generating the linearized predicted state sequence includes: A fractional-order Jiles-Atherton hysteresis model is constructed based on standardized multi-source operating state data. This model includes an equivalent magnetization driving module, a magnetization state decomposition module, a fractional-order magnetization evolution module, and a hysteresis state mapping module. The construction of the fractional-order Jiles-Atherton hysteresis model is as follows: Based on the equivalent magnetic field calculation process in the traditional Jiles-Atherton hysteresis model, this paper introduces magnetic flux density data, temperature data, load data, and equipment status data to improve the equivalent magnetization drive module. Based on the reversible and irreversible magnetization decomposition process in the traditional Jiles-Atherton hysteresis model, this paper introduces dynamic proportional adjustment relationships and hysteresis loss constraints to improve the magnetization state decomposition module. Based on the magnetization state evolution process in the traditional Jiles-Atherton hysteresis model, this paper introduces fractional-order historical memory evolution and dual-timescale dynamic evolution to improve the fractional-order magnetization evolution module. Furthermore, a Koopman high-dimensional state mapping process is added to the output of the traditional Jiles-Atherton hysteresis model to perform high-dimensional state mapping and linear state advancement on the hysteresis state evolution results, forming a hysteresis state mapping module. This results in the fractional-order Jiles-Atherton hysteresis model, where: The equivalent magnetization drive module includes: Difference register: stores the magnetic flux density, magnetic field strength, and the magnetization state variable obtained by the difference between the two; Superposition register: stores the external magnetic field, magnetization intensity state variables, and the superimposed equivalent magnetic field; Mapping register: stores the equivalent magnetic field, hyperbolic tangent mapping result, and hysteresis-free magnetization; Operating status cache: stores temperature data, load data, and device status data; The equivalent magnetization drive module calculates the difference between the magnetic flux density and the magnetic field strength to obtain the magnetization state variable. The applied magnetic field and the magnetization state variable are then superimposed to construct the equivalent magnetic field. Nonlinear mapping calculations are performed on the equivalent magnetic field to obtain the hysteresis-free magnetization, where: The construction of the equivalent magnetic field is as follows: Read the magnetic flux density data and magnetic field strength data under the same time series, and perform point-by-point calculation for each time sampling point. First, subtract the corresponding magnetic field strength value from the magnetic flux density value at the current time to obtain the magnetization intensity state variable. Then, add the magnetization intensity state variable directly to the magnetic field strength value at the same time to obtain the equivalent magnetic field at the current time. Repeat the same process from the first sampling point to the last sampling point according to the time series to obtain the complete equivalent magnetic field sequence. The equivalent magnetic field is processed by nonlinear mapping calculation, specifically as follows: For each equivalent magnetic field value in the equivalent magnetic field sequence, point-by-point calculations are performed. The equivalent magnetic field value is multiplied by a coefficient of 0.8 to obtain the equivalent magnetic field scaling value. Then, the exponent value of the equivalent magnetic field scaling value is calculated using the natural exponential function. The natural constant 2.71828 is used as the base, and the scaling value is used as the exponent for exponent operation to obtain the first exponent value. The opposite of the equivalent magnetic field scaling value is used as the exponent for exponent operation to obtain the second exponent value. The difference is obtained by subtracting the second exponent value from the first exponent value. The sum is obtained by adding the second exponent value to the first exponent value. The difference is divided by the sum to obtain the hyperbolic tangent function value. Finally, the hyperbolic tangent function value is multiplied by 1.6 to obtain the hysteresis-free magnetization intensity at the corresponding time point. The same operation is repeated point-by-point for the entire equivalent magnetic field sequence to obtain the hysteresis-free magnetization intensity sequence. The magnetization state decomposition module includes: Proportional buffer: stores temperature data, load data, changes in magnetic field strength, and dynamic proportional coefficients; Reversible component register: stores the initial reversible magnetization component, the adjusted reversible magnetization component, and the corrected reversible magnetization component; Irreversible component register: stores the initial irreversible magnetization component, the adjusted irreversible magnetization component, and the corrected irreversible magnetization component; Loss correction register: stores hysteresis loss, cumulative absolute value of irreversible magnetization component, and correction coefficient; The magnetization state decomposition module performs decomposition calculations based on hysteresis-free magnetization intensity and magnetization strength, decomposing the magnetization intensity into reversible and irreversible magnetization components. A dynamic proportional relationship is constructed based on temperature data, load data, and magnetic field strength changes to adaptively adjust the reversible and irreversible magnetization components. Furthermore, a hysteresis loss constraint is introduced to correct the consistency of the decomposition results, yielding the corrected reversible and irreversible magnetization components, where: The magnetization intensity is decomposed into reversible magnetization components and irreversible magnetization components, specifically: For each time sampling point, the magnetization intensity value and the hysteresis-free magnetization intensity value are read. First, the difference between the hysteresis-free magnetization intensity and the magnetization intensity is calculated. Then, the difference is multiplied by a coefficient of 0.5 to obtain the initial irreversible magnetization component. The initial irreversible magnetization component is obtained by subtracting the magnetization intensity value from the initial irreversible magnetization component. The same operation is performed point by point on the entire time series to obtain the initial reversible magnetization component sequence and the initial irreversible magnetization component sequence. A dynamic proportional relationship is constructed based on temperature data, load data, and changes in magnetic field strength, specifically as follows: For each time sampling point, temperature data, load data, and the difference between the current magnetic field strength and the magnetic field strength at the previous moment are read. The temperature data is divided by 300 to obtain the temperature normalization value, the load data is divided by 100 to obtain the load normalization value, and the difference in magnetic field strength is divided by the absolute value of the magnetic field strength at the previous moment plus 0.01 to obtain the change normalization value. Then, the three normalization values ​​are multiplied by 0.3, 0.4, and 0.3 respectively and added together to obtain the scaling factor. The same operation is repeated point by point for the entire time series to obtain the dynamic scaling factor sequence. Adaptive adjustment processing is performed on both the reversible and irreversible magnetization components, specifically as follows: For each time sampling point, the initial reversible magnetization component, the initial irreversible magnetization component, and the corresponding scaling factor are read. The initial irreversible magnetization component is multiplied by the scaling factor to obtain the adjusted irreversible magnetization component. The adjusted reversible magnetization component is obtained by subtracting the adjusted irreversible magnetization component from the magnetization intensity. Multiplication and subtraction operations are performed point by point on the entire time series to obtain the adjusted reversible magnetization component sequence and the adjusted irreversible magnetization component sequence. Hysteresis loss constraint refers to aligning the cumulative amount of the adjusted irreversible magnetization component with the calculated hysteresis loss amount in the corresponding time window within each time window. By calculating the proportional relationship between the two, the proportional relationship is used as a constraint coefficient to apply to the irreversible magnetization component. The cumulative change of the irreversible magnetization component is kept in a consistent relationship with the hysteresis loss amount, thus limiting the range of values ​​of the irreversible magnetization component. A consistency correction is performed on the decomposition results by introducing a hysteresis loss constraint, specifically as follows: For each time window, read the adjusted irreversible magnetization component sequence and calculate the sum of the cumulative absolute values. Then read the hysteresis loss value in the same time window and divide the two to obtain the correction coefficient. Multiply the correction coefficient by the adjusted irreversible magnetization component to obtain the corrected irreversible magnetization component. Subtract the corrected irreversible magnetization component from the magnetization intensity to obtain the corrected reversible magnetization component. Repeat the same operation for the entire time series in a window sliding manner to obtain the corrected reversible magnetization component sequence and the corrected irreversible magnetization component sequence. The fractional-order magnetization evolution module includes: Tensor buffer: Stores three-dimensional tensors constructed according to the time dimension, physical quantity dimension, and device state dimension; Fast Evolution Register: Stores instantaneous magnetization response values ​​and corresponding weights on fast time scales; Slow Evolution Register: Stores historical cumulative magnetization values ​​and corresponding weights over a slow timescale; Fusion register: stores the magnetization evolution results after coupling fast timescale results with slow timescale results; State graph buffer: stores magnetization nodes, state transition edges between nodes, and edge weight changes; The fractional-order magnetization evolution module performs tensor-based multi-source coupling processing based on hysteresis-free magnetization intensity, corrected reversible magnetization components, and irreversible magnetization components. It constructs a multi-dimensional tensor structure from standardized multi-source operating state data across time, physical quantity, and device state dimensions, and performs dual-time-scale fractional-order dynamic evolution calculations. The process of Zhang quantitative multi-source coupling processing is as follows: At each time sampling point, the hysteresis-free magnetization intensity, the corrected reversible magnetization component, the corrected irreversible magnetization component, as well as temperature data, load data, actuator displacement data, and equipment start-stop status data are read simultaneously. All types of data are arranged in order into a numerical vector of length 7. The arrangement process is repeated for all time sampling points and stacked in time order to obtain a data sequence arranged by time index. The construction of multidimensional tensor structures is as follows: Based on the data sequence arranged by time index, 1000 numerical vectors of length 7 arranged in time order are formed into a matrix of 1000 rows and 7 columns. The matrix is ​​divided into layers according to the device start and stop status values ​​of 0 and 1. The row data corresponding to the device status of 0 and the row data corresponding to the device status of 1 are filled into two channels respectively. The channels are stacked to form a three-dimensional tensor of 1000 by 7 by 2. The instantaneous response of magnetization is calculated along a fast timescale path, and the historical cumulative effect of magnetization is calculated along a slow timescale path using fractional-order evolution. The results from the two evolution paths are then coupled and fused to obtain the magnetization evolution result, where: A fast timescale path refers to a magnetization change sequence with the time interval between two adjacent time sampling points as the calculation step size, and the time interval is taken as 0.001s; The instantaneous response of magnetization is calculated along a fast timescale path, specifically as follows: At each time sampling point, the corresponding 7-dimensional vector in the current time 3D tensor is read, and three data items are taken: hysteresis-free magnetization intensity, reversible magnetization component, and irreversible magnetization component. The magnetization intensity at the current time is subtracted from the magnetization intensity at the previous time to obtain the difference. The difference is then multiplied by 0.7 and added to the magnetization intensity at the previous time to obtain the fast response magnetization intensity at the current time. The same operation is performed point by point on all time sampling points to obtain the fast time-scale magnetization intensity sequence. The slow timescale path refers to the magnetization change sequence with a sliding window of 10 time sampling points as the calculation interval, corresponding to a time length of 0.01s; The historical cumulative effect of magnetization is calculated using fractional-order evolution along a slow timescale path, specifically as follows: At each time sampling point, the magnetization intensity values ​​of the current time and the previous 9 time sampling points are read. The magnetization intensity related data in the 7-dimensional vector of the corresponding time index is taken from the 3D tensor. The magnetization intensity of the 10 time points is multiplied by the weight coefficients 0.4, 0.2, 0.1, 0.08, 0.06, 0.05, 0.04, 0.03, 0.02 and 0.02 respectively and then added together to obtain the slow evolution magnetization intensity of the time point. The same operation is repeated for all time sampling points in the sliding window manner to obtain the slow time scale magnetization intensity sequence. The evolution of magnetization intensity is obtained as follows: At each time sampling point, the fast response magnetization is multiplied by 0.6 and the slow evolution magnetization is multiplied by 0.4 and then added together to obtain the magnetization evolution result at the time point. The same operation is performed on all time sampling points to obtain the magnetization evolution sequence. Magnetization intensity is used as nodes, and magnetization state transition relationships are used as edges to construct a magnetization state graph. After graph structure update processing, a hysteresis state evolution sequence is obtained, where: Constructing a magnetization state diagram, specifically: The magnetization intensity evolution result corresponding to each time sampling point is taken as a node, and the magnetization intensity difference between adjacent time sampling points is used as the weight of the directed edge to establish a connection relationship. The nodes and directed edges between adjacent nodes are established sequentially for all time sampling points to form a complete magnetization state diagram. Perform graph structure update processing, specifically: For each node, calculate the weight of all outgoing edges, retain the edges with weights greater than 0.05, delete the edges with weights less than or equal to 0.05, renumber the retained edges in chronological order, and update the connection relationships between nodes. Perform comparison, deletion, and numbering operations on all nodes one by one to obtain the hysteresis state evolution sequence. The hysteresis state mapping module includes: Integrator register: stores the average magnetic flux density, the change in magnetic field strength, and the integral value over the interval; Loss register: stores the cumulative hysteresis loss of the window; Koopman embedded register: stores a five-dimensional high-dimensional vector formed by the original magnetization intensity, square value, cubic value, sine value and cosine value; Matrix buffer: stores the linear state evolution matrix and the corresponding high-dimensional observable state sequence; Linear advance register: stores the current high-dimensional state vector and the linearized predicted state for the next time step; The hysteresis state mapping module performs lap integral processing on magnetic field strength data and magnetic flux density data to obtain hysteresis loss. It then maps the hysteresis state evolution sequence to the Koopman high-dimensional observable space to obtain high-dimensional observable states. A linear state evolution matrix is ​​constructed and state progression is performed to obtain a linearized predicted state sequence, where: The lap integral processing is performed as follows: The magnetic field strength and magnetic flux density data are read in chronological order. An integration interval is formed for two adjacent time sampling points. Within the integration interval, the magnetic flux density values ​​of the previous and next time moments are added together and then divided by 2 to obtain the average magnetic flux density. The magnetic field strength value of the previous moment is then subtracted from the magnetic field strength value of the next moment to obtain the change in magnetic field strength. The average magnetic flux density is multiplied by the change in magnetic field strength to obtain the integral value of the integration interval. The same operation is repeated for all time intervals, and the integral value of each segment is accumulated sequentially to obtain the hysteresis loss value. Mapping the hysteresis state evolution sequence to the Koopman high-dimensional observable space is as follows: At each time sampling point, the magnetization intensity value in the hysteresis state evolution sequence is read. The magnetization intensity value is used as the original observation. The square and cube values ​​of the magnetization intensity value are calculated respectively. The corresponding sine and cosine values ​​are calculated respectively. The original magnetization intensity value, square value, cube value, sine and cosine values ​​are combined in a fixed order to form a 5-dimensional vector. The 5-dimensional vector is used as the high-dimensional observable state representation in the Koopman high-dimensional observable space. The same calculation is performed point by point for all time sampling points. All 5-dimensional vectors are arranged in time order to obtain the high-dimensional observable state sequence. Constructing a linear state evolution matrix and performing state progression are as follows: The high-dimensional observable state sequence is arranged in chronological order. For each time sampling point, the 5-dimensional vector at the current time is paired with the 5-dimensional vector at the next time. All the high-dimensional vectors at the current time are arranged in columns to form the first set of data, and the corresponding high-dimensional vectors at the next time are arranged in columns to form the second set of data. The first set of data is transposed, and the transposed result is multiplied by the original first set of data to obtain a square matrix. The square matrix is ​​inverted, and the second set of data is multiplied by the transposed result of the first set of data. The result of the matrix multiplication is multiplied by the matrix obtained by inverting the square matrix to obtain the linear state evolution matrix. At each time sampling point, the 5-dimensional vector at the current time is multiplied by the linear state evolution matrix to obtain the prediction vector at the next time. The matrix multiplication operation is performed point by point for all time sampling points to obtain the linearized prediction state sequence. A fractional-order Jiles-Atherton hysteresis model was trained, using the reconstruction error of the magnetization evolution sequence as the optimization objective. The parameters in the equivalent magnetization driving module, magnetization state decomposition module, fractional-order magnetization evolution module, and hysteresis state mapping module were continuously optimized. Training was stopped when the reconstruction error was below 0.01 for five consecutive iterations and the parameter update magnitude was less than 0.001. The fractional-order Jiles-Atherton hysteresis model is trained as follows: The magnetization intensity sequence obtained from actual observations is read in chronological order as the target sequence. At the same time, the standardized multi-source operating state data under the same time index is input into the fractional-order Jiles-Atherton hysteresis model to obtain the magnetization intensity evolution sequence. The difference between the magnetization intensity evolution sequence and the target sequence is calculated at each time sampling point. The difference is multiplied by the magnetization intensity evolution sequence to obtain the squared error. The squared errors of all time sampling points are summed and divided by the total number of time sampling points to obtain the mean squared reconstruction error. Using the mean squared reconstruction error as the optimization objective, iterative updates are performed on the parameters in the equivalent magnetization driving module, magnetization state decomposition module, fractional magnetization evolution module, and hysteresis state mapping module of the fractional-order Jiles-Atherton hysteresis model. The parameter updates are achieved by gradually adjusting the values ​​of each parameter along the direction of error reduction, with each adjustment increment being 0.001. The changes in mean squared reconstruction error and parameter values ​​between two adjacent iterations are recorded. When the mean squared reconstruction error is less than 0.01 in five consecutive iterations and the absolute value of the change of each parameter between two adjacent iterations is less than 0.001, the iteration is stopped, and the trained fractional-order Jiles-Atherton hysteresis model is obtained.

[0023] In this embodiment, the generation of the piecewise state change function includes: Based on standardized multi-source operating status data, process stage identification data, load data and equipment start-up and shutdown status data are aligned according to a unified time sequence to construct an operating status sequence. Based on the operating condition state sequence, state change detection and processing is performed. The location where the process stage identifier changes is identified as the process stage switching point. The location where the load data difference between adjacent time moments is greater than 0.05 is identified as the load mutation point. The location where the equipment start-up and shutdown status changes is identified as the equipment state switching point. These are then merged to obtain a set of state switching points. The linearized predicted state sequence is segmented based on the set of state transition points. The linearized predicted state sequence is divided into multiple continuous subsequences according to adjacent state transition points, and a corresponding state change function is constructed for each subsequence, resulting in a set of segmented state change functions, where: For each subsequence, a corresponding state change function is constructed, specifically as follows: For each subsequence, read the linearized predicted state values ​​in the subsequence in chronological order, number the time sampling points as a continuous integer sequence starting from 1, sum all the linearized predicted state values ​​in the subsequence and divide by the subsequence length to obtain the state average, sum all the time numbers and divide by the subsequence length to obtain the time average, for each sampling point, subtract the time average from the sampling point's time number, subtract the state average from the corresponding linearized predicted state value, multiply the two and sum over all sampling points to obtain the numerator; The time value of each sampling point is subtracted from the time average, multiplied by the time value, and summed over all sampling points to obtain the denominator. The slope coefficient is obtained by dividing the numerator by the denominator. The intercept is obtained by subtracting the slope coefficient from the state average and multiplying by the time average. The slope coefficient and the intercept are combined to form the state change function of the subsequence. The state value at any time value is equal to the slope coefficient multiplied by the time value and the intercept. The same operation is repeated for all subsequences to obtain a set of piecewise state change functions.

[0024] In this embodiment, generating the sequence of non-smooth reachable domains includes: A manufacturing decision constraint generation module based on Filippov differential is constructed. This module includes a state switching topology representation unit, a state encoding unit, a Filippov reachable domain calculation unit, and a state domain generation unit, wherein: A module for generating manufacturing decision constraints based on Filippov differentials is constructed, specifically as follows: The state switching topology representation unit, state encoding unit, Filippov reachable domain calculation unit, and state domain generation unit are connected in series to obtain the manufacturing decision constraint generation module, wherein: The state-switching topology representation unit includes: Node register: stores the node number, segment slope, and corresponding time interval for each segment state change function; Edge register: Stores the directed connection relationships between adjacent nodes and the corresponding switching directions; Switching position register: stores the time position corresponding to process stage switching points, load change points, and equipment status switching points; Topology table cache: stores the segmented switching graph formed by node registers and edge registers, as well as the node connection relationships; The state transition topology representation unit is based on piecewise state change functions. Each piecewise state change function is mapped to a node, the transition relationship between adjacent pieces is represented as a directed edge, and the state transition position is determined at the corresponding position of the directed edge, thus constructing a piecewise transition graph, where: Construct a segmented switching graph, specifically as follows: The set of segmented state change functions is numbered in chronological order. Each segmented state change function is assigned to a node. The time boundary between two adjacent segmented functions is taken as the switching position. A directed connection relationship from the previous segment to the next segment is established between nodes, and the value of the time boundary is recorded at the corresponding directed connection position. The node numbering, adjacent pairing and connection establishment process is executed sequentially for all segmented functions to obtain a segmented switching graph composed of nodes, directional connections and switching positions. The status coding unit includes: Path Cache: Stores the historical node sequence and node switching order corresponding to the linearized prediction state in chronological order; Extended feature register: stores the extended state feature sequence consisting of the current linearized predicted state value, first-order change, and second-order change; Boundary Index Register: Stores the node number, edge number, and switching time position corresponding to the state switching boundary; Hysteresis Convolution Register: Stores the 10 most recent linearized predicted state values ​​and their corresponding hysteresis kernel weights; The state coding unit performs path-dependent coding based on the linearized predicted state sequence, encoding the evolution information of the linearized predicted state along the historical path to obtain an extended state sequence, which is then embedded into the segmented switching graph to obtain a set of state switching boundaries, where: The path-dependent encoding process is performed as follows: Read the linearized predicted state sequence in chronological order. For each time sampling point, the difference between the current linearized predicted state value and the linearized predicted state value at the previous time step is used to obtain the first-order change. The difference between the current linearized predicted state value and the linearized predicted state values ​​at the previous two time steps is used to obtain the second-order change. The current linearized predicted state value, the first-order change, and the second-order change are combined in a fixed order to form a state feature group. The same processing is performed on all time sampling points to obtain the extended state sequence. The set of state transition boundaries is obtained as follows: In the segmented switching graph, read the switching position time point corresponding to each directed connection one by one, locate the state feature group corresponding to the time point in the extended state sequence, pair the state feature group with the corresponding switching position one by one to form boundary description data, repeat the same pairing operation for all directed connections, and summarize all pairing results to obtain the state switching boundary set. The Filippov reachability calculation unit includes: Vector field register: stores the set of state change rates corresponding to the derivatives of piecewise state change functions; Convex combination register: stores the rate of change of states on both sides of the state switching boundary and the weighted convex combination result; Hysteresis state register: stores historical linearized prediction states, weighted convolution results, and corresponding time indices; Time advance register: stores boundary change rate, historical weighted state and state value at the next time step; Reachable range buffer: stores the upper and lower boundary sequences corresponding to the reachable range of states; The Filippov reachability computation unit performs set-valued dynamics computation based on the state transition boundary set. It represents the derivative of the piecewise state change function as multiple vector fields, and performs a weighted convex combination operation on these vector fields at the state transition boundaries to obtain the state change rate belonging to the convex hull of the vector fields. Furthermore, it introduces a hysteresis operator kernel to perform weighted convolution on the historical linearized predicted states, and combines this with an extended state sequence to perform time-progression computation on the linearized predicted states, thus obtaining the state reachability interval, where: The ensemble-valued dynamics computation process is performed as follows: For each segment of the state change function, the slope value is read according to the time interval, and the slope is used as the corresponding state change rate. If there are two adjacent segments in the same time interval, the slope values ​​of the two segments are read at the same time, and the two slopes are used as two candidate state change rate sets. The slope reading and set combination operation is performed segment by segment for all time intervals to obtain the state change rate set distributed by time. At the state transition boundary, a weighted convex combination operation is performed on the vector field, specifically as follows: At each state transition boundary position, read the two slope values ​​corresponding to the adjacent segments on both sides of the state transition boundary, multiply the first slope by a weight of 0.5, multiply the second slope by a weight of 0.5, and then add the two products to obtain the state change rate at the boundary. Perform the same process on all state transition boundaries one by one to obtain the combined state change rate at the boundary. The hysteresis operator kernel is a sequence of weight coefficients arranged in chronological order. The closer the historical state is to the current time, the greater its weight, and the farther away it is, the smaller its weight. A hysteresis operator kernel is introduced to perform a weighted convolution on the historical linearized prediction state, specifically as follows: At each time sampling point, read the linearized predicted state values ​​of the current time and the previous 9 time sampling points. Multiply these 10 values ​​by weighting coefficients of 0.3, 0.2, 0.15, 0.1, 0.08, 0.06, 0.04, 0.03, 0.02, and 0.02, respectively. Then add all the product results to obtain the weighted result of the time point. Repeat the same operation for all time sampling points in a sliding window manner to obtain the weighted linearized predicted state sequence. The linearized predicted state is time-progressed by combining the extended state sequence, specifically as follows: At each time sampling point, the current linearized predicted state value, the corresponding weighted linearized predicted state value, and the state feature group in the extended state sequence are read. The current state change rate is multiplied by the time step size of 0.001 to obtain the change amount. The change amount is then added to the weighted linearized predicted state value and multiplied by 0.5 as the correction amount. The correction amount is added to the current linearized predicted state value to obtain the linearized predicted state value at the next time step. The same operation is performed point by point on all time sampling points to obtain the upper and lower boundary sequences of the state reachable interval. The state domain generation unit includes: Sub-interval index table: stores the reachable interval range of the state corresponding to the state change function of each segment; Boundary Smoothing Register: Stores the upper boundary sequence, the lower boundary sequence, and the smoothed boundary result after moving average; Loss constraint register: stores hysteresis loss, cumulative absolute value of irreversible magnetization component and constraint coefficient; Boundary integration register: stores the unified boundary value and extended boundary value corresponding to the intersection of adjacent sub-intervals; Non-smooth domain buffer: stores the sequence of non-smooth reachable domains after boundary integration; The state domain generation unit constructs the state domain based on the reachable interval. It divides the reachable interval into multiple sub-intervals according to a piecewise state change function, and reconstructs the boundaries of each sub-interval to form a continuous boundary surface. The state domain is then constrained and filtered using hysteresis loss. Finally, the sub-intervals undergo set expansion and boundary integration to generate a sequence of non-smooth reachable domains, where: The reachable interval is divided into multiple sub-intervals according to the piecewise state change function, specifically: Read the upper and lower boundary sequences of the reachable state interval in chronological order, and at the same time read the start and end times of the segment corresponding to the segmented state change function. Extract the upper and lower boundary values ​​of the corresponding time index in the reachable state interval within each segmented time range. Continuously extract the upper and lower boundary values ​​belonging to the same segmented time range to form a sub-interval. Perform time matching and interval extraction operations on all segments one by one to obtain multiple sub-intervals that correspond one-to-one with the segmented state change function. The boundaries of each sub-interval are reconstructed to form a continuous boundary surface, specifically as follows: For each sub-interval, read the upper boundary sequence and the lower boundary sequence respectively. Perform a three-point moving average calculation on the upper boundary sequence according to the time index. Add the values ​​of the current point with the previous point and the next point and divide by 3 to obtain the smoothed value. Take the average of the two adjacent points at both ends of the upper boundary sequence to obtain the endpoint smoothed value. Perform the same three-point moving average calculation on the lower boundary sequence. Arrange the smoothed upper boundary sequence and the lower boundary sequence in time order to form a continuous boundary data set. Repeat the same operation on all sub-intervals to obtain the continuously changing upper boundary and lower boundary curves. The state domain is constrained and filtered based on the hysteresis loss, specifically as follows: Read the hysteresis loss value of the corresponding time window in each sub-interval, divide the hysteresis loss value by the average of the absolute values ​​of the upper boundary sequence of the sub-interval to obtain the constraint coefficient. When the constraint coefficient is greater than 1, it is truncated to one. Multiply the constraint coefficient by the lower boundary sequence and the upper boundary sequence respectively to obtain the constrained lower boundary value and upper boundary value. Perform the same operation on all sub-intervals segment by segment to obtain the state interval constrained by the hysteresis loss. The set expansion and boundary integration processes for each sub-interval are performed as follows: Subtract 0.02 from the lower boundary value of each sub-interval to obtain the extended lower boundary, and add 0.02 to the upper boundary value to obtain the extended upper boundary. Then, unify the boundary values ​​of two adjacent sub-intervals at the intersection time point. Add the end boundary value of the previous sub-interval to the start boundary value of the next sub-interval and divide by 2 to obtain the unified boundary value. Perform the same operation on all sub-intervals one by one and splice the extended intervals in time order to obtain the sequence of non-smooth reachable domains.

[0025] In this embodiment, generating the manufacturing decision candidate set includes: A set of manufacturing decision variables is constructed at the edge nodes. These variables are represented as a multidimensional sequence of variables consisting of actuator displacement adjustment, load adjustment, and equipment operating status adjustment. This sequence is then time-aligned with standardized multi-source operating status data to obtain a set of candidate decision variables, where: Construct a set of manufacturing decision variables at the edge nodes, specifically as follows: At each time sampling point, the actuator displacement data, load data, and equipment start-stop status data are read. The actuator displacement data is discretized by step size to obtain a set of displacement adjustment values, and the load data is discretized by step size to obtain a set of load adjustment values. The equipment start-stop status data is set to 0 and 1 as the equipment operation status adjustment values. For the same time sampling point, the displacement adjustment values, load adjustment values, and equipment operation status adjustment values ​​are fully combined and arranged to form a multidimensional variable vector. The multidimensional variable vectors corresponding to all time sampling points are aligned by time index to obtain a set of candidate decision variables. State prediction is performed based on the hysteresis state evolution sequence, the linearized predicted state sequence, and the set of state switching boundaries. The candidate decision variable set is substituted into the linear state evolution matrix for state advancement calculation. The predicted state is then corrected using the hysteresis state evolution sequence, resulting in the corresponding set of candidate state trajectories. The set of candidate decision variables is substituted into the linear state evolution matrix to perform state advancement calculations, specifically: At each time sampling point, the current high-dimensional observable state vector and the corresponding candidate decision variable vector are read. Each component in the candidate decision variable vector is multiplied by a weight coefficient of 0.5, 0.3 and 0.2 respectively and summed to obtain the decision influence. The decision influence is added to each component of the current high-dimensional state vector to form a corrected state vector. Then, the corrected state vector is multiplied by the linear state evolution matrix to obtain the predicted state vector for the next time step. The same operation is performed point by point for all time sampling points to obtain the candidate state trajectory set. The predicted state is corrected by combining the hysteresis state evolution sequence, specifically as follows: At each time sampling point, the predicted state vector and the corresponding hysteresis state evolution sequence value are read. The hysteresis state value is multiplied by 0.01 to obtain the correction amount. The correction amount is added to each component of the predicted state vector to obtain the corrected predicted state vector. The same operation is performed point by point on all time sampling points to obtain the corrected candidate state trajectory set. Based on the non-smooth state reachability domain sequence, a constraint screening process is performed on the candidate state trajectory set. Candidate decision variables whose predicted states fall within the corresponding state range of the non-smooth state reachability domain sequence are retained, while candidate decision variables whose predicted states fall outside the corresponding state range of the non-smooth state reachability domain sequence are removed, resulting in a manufacturing decision candidate set, where: The candidate state trajectory set is constrained and filtered based on the sequence of non-smooth reachable domains, specifically as follows: At each time sampling point, the corresponding state vector in the candidate state trajectory is read, and the upper and lower boundary values ​​of the non-smooth state reachable domain sequence at the corresponding time point are read. A comparison operation is performed on each component of the state vector. When the component value is greater than the lower boundary and less than the upper boundary, the component is marked as valid. When any component does not meet the range condition, the corresponding candidate decision variable is removed. The comparison and filtering operation is performed one by one for all time sampling points. All candidate decision variables that meet the range condition are retained to obtain the manufacturing decision candidate set.

[0026] In this embodiment, the step of generating manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear, and operational stability includes: A multi-objective evaluation process is performed on the candidate set of manufacturing decisions. The production efficiency, energy consumption, equipment wear and tear, and operational stability corresponding to each candidate decision are calculated and normalized to obtain a sequence of evaluation indicators, where: The production efficiency, energy consumption, equipment wear and tear, and operational stability corresponding to each candidate decision are calculated as follows: At each time sampling point, the candidate decision variable vector is read, the displacement adjustment of the actuator is summed and divided by the time window length to obtain the displacement change per unit time, which is used as the production efficiency. The load adjustment is squared point by point, summed and then divided by the time window length to obtain the average power consumption, which is used as the energy consumption. The hysteresis loss value in the corresponding time window is read and added to the sum of the absolute values ​​of the load change to obtain the equipment loss. The absolute value of the state difference between adjacent time points in the linearized predicted state sequence is calculated point by point and averaged to obtain the fluctuation amplitude. The fluctuation amplitude is subtracted from 1 to obtain the operating stability. Repeat the same operation for all time sampling points to obtain 4 evaluation index sequences. Then, perform the operation of subtracting the minimum value and dividing by the difference between the maximum value and the minimum value for each evaluation index sequence to obtain the normalized evaluation index sequence. The evaluation index sequence is weighted and fused, and production efficiency, energy consumption, equipment wear and tear, and operational stability are weighted and summed according to set weights to obtain the comprehensive evaluation value of the corresponding candidate decisions. The comprehensive evaluation values ​​are then sorted to obtain the priority sequence of candidate decisions, where: The production efficiency, energy consumption, equipment wear and tear, and operational stability are weighted and summed according to set weights, specifically as follows: At each time sampling point, normalized values ​​of production efficiency, energy consumption, equipment wear and tear, and operational stability are read. Production efficiency is multiplied by 0.4, energy consumption by 0.2, equipment wear and tear by 0.2, and operational stability by 0.2. The four products are then added together to obtain a comprehensive evaluation value. Multiplication and addition operations are performed point by point for all time sampling points, and the resulting comprehensive evaluation values ​​are sorted from largest to smallest to obtain a candidate decision priority sequence. The manufacturing decision is selected based on the priority sequence of candidate decisions, and the candidate decision with the best comprehensive evaluation value is selected as the manufacturing decision result. Based on the manufacturing decision result, the displacement data, load data and equipment start-up and shutdown status data of the actuator are controlled and adjusted to complete the manufacturing decision execution.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to an intelligent manufacturing production line that coordinates electromagnetic heating and servo pressing. This production line needs to complete heating, heat preservation, transfer, pressing, and unloading operations during continuous production. The equipment's operating status is affected by excitation current, magnetic field strength, magnetic flux density, temperature, load, actuator displacement, process stage switching, and equipment start / stop status. Traditional methods use fixed threshold control, triggering adjustments only when the temperature exceeds 315K, the load exceeds 90, or the actuator displacement exceeds 6.5mm. In one continuous production segment, it was found that the pressing unit exhibited a response lag even before the load reached the traditional alarm threshold. The load increased from 68.4 to 78.6, the actuator displacement increased from 4.12mm to 5.03mm, and the magnetic flux density increased from 0.63 to 0.78. However, the traditional control system still determined the equipment was normal and did not adjust the load and displacement in advance, leading to the risk of displacement overshoot in the next production operation.

[0028] Within this production segment, edge nodes continuously collected 1000 sampling points at a sampling interval of 0.001 seconds. In a certain analysis window, the excitation current increased from 0.46A to 0.71A, the magnetic field strength increased from 0.58 to 0.74, the magnetic flux density increased from 0.61 to 0.83, the temperature increased from 302.8K to 314.6K, the load increased from 66.2 to 91.4, the actuator displacement increased from 3.86mm to 6.18mm, and the process stage indicator switched from heat preservation and transfer to press-fitting adjustment, while the equipment remained operational. Traditional methods only trigger control when the load reaches 91.4, while this invention identifies a normalized load difference of 0.057 when the load increases from 72.1 to 76.5, exceeding 0.05, and records this location as a state switching point in conjunction with the change in the process stage indicator.

[0029] Near this state transition point, the present invention reads the magnetic field strength (0.74) and magnetic flux density (0.83), calculates the magnetization state variable (0.09), and adds it to the magnetic field strength to obtain the equivalent magnetic field (0.83). After nonlinear mapping, the hysteresis-free magnetization is obtained as 1.05. The system continues to read the magnetization (0.09) and hysteresis-free magnetization (1.05), calculates the difference (0.96), multiplies it by 0.5 to obtain the initial irreversible magnetization component (0.48), and then subtracts the initial irreversible magnetization component from the magnetization to obtain the initial reversible magnetization component (-0.39). At this time, the temperature is 314.6K, the load is 91.4, and the change in magnetic field strength is 0.04. The system calculates the dynamic proportional coefficient as 0.53, obtaining the adjusted irreversible magnetization component (0.254) and the adjusted reversible magnetization component (-0.164). Therefore, it can be seen that traditional methods only identify overload, while the present invention can further identify the hysteresis accumulation state behind the load increase.

[0030] This invention performs lap-loop integration on the magnetic field strength and magnetic flux density of the window. Taking two consecutive sampling points as an example, the magnetic field strength changes from 0.70 to 0.74, and the magnetic flux density changes from 0.79 to 0.83, with an average magnetic flux density of 0.81. The change in magnetic field strength is 0.04, resulting in a single-interval integral value of 0.0324. After summing over all sampling intervals, the hysteresis loss is obtained as 0.068. Traditional methods calculate equipment loss based solely on the average squared load, resulting in a loss score of 0.41. This invention incorporates the hysteresis loss of 0.068 along with the absolute value of the load change (17.2) into the equipment loss calculation, yielding an equipment loss score of 0.76, which more accurately reflects the potential equipment loss risk.

[0031] In the state prediction process, this invention maps the hysteresis state evolution sequence to the Koopman high-dimensional observable space. Taking the magnetization evolution result of 0.72 as an example, the system calculates the square value of 0.5184, the cube value of 0.3732, the sine value of 0.6594, and the cosine value of 0.7518, and combines them with 0.72 to form a high-dimensional observable state, and then constructs a linear state evolution matrix for prediction. This invention predicts the next window state value to be 0.824, the actual observed state value is 0.839, and the error is 0.015. The traditional linear prediction method predicts a value of 0.752 with an error of 0.087, indicating that this invention has higher prediction accuracy when hysteresis and operating condition switching coexist.

[0032] After the process stage switch, the system reads the linearized predicted state before and after the switch point. Within the five sampling points before the switch point, the predicted state increases from 0.701 to 0.739; within the five sampling points after the switch point, the predicted state increases from 0.741 to 0.812. The system calculates the slope of the state change function before the switch to 0.0095 and the slope after the switch to 0.0355, and determines the boundary between the two as the state switch boundary. Traditional methods still use a single smoothing function, yielding a slope of 0.017, which is insufficient to reflect the rapid state changes after the start of the pressing stage. This invention weights and combines the slopes of 0.0095 and 0.0355 on both sides of the boundary to obtain a boundary state change rate of 0.0225. It then reads the state values ​​of the current and the previous nine sampling points (0.812, 0.798, 0.783, 0.769, 0.752, 0.741, 0.733, 0.726, 0.718, and 0.701), and weights them according to the hysteresis operator kernel to obtain a historical weighted state value of 0.782. Finally, it generates the reachable state range of 0.742 to 0.866 for this window. Traditional methods fix the safety range at 0.70 to 0.90, which is too wide and cannot exclude some high-risk candidate decisions.

[0033] In the manufacturing decision screening, the system generates 240 candidate decisions based on the current actuator displacement of 5.84 mm, load of 88.6, and equipment operating state 1. Traditional methods directly select the scheme with the highest production efficiency: increasing displacement by 0.42 mm, keeping the load constant, and continuing equipment operation. This scheme has a production efficiency score of 0.96, but its predicted state reaches a maximum of 0.891, exceeding the upper limit of 0.866 calculated by this invention. If this scheme is executed, the peak load in the next window will rise to 94.8, and the temperature will rise to 318.2 K. In contrast, this invention predicts the state trajectory of each candidate decision group. The scheme with an increase in displacement of 0.42 mm is eliminated. The scheme with an increase in displacement of 0.18 mm, a decrease in load of 0.10, and continued equipment operation has a predicted state range of 0.758 to 0.843, which is within the reachable range and is therefore retained. After the system completes the screening, 163 of the 240 candidate decisions are retained, and 77 are eliminated.

[0034] In the multi-objective evaluation, the normalized values ​​for production efficiency, energy consumption, equipment loss, and operational stability of the retained scheme were 0.89, 0.54, 0.49, and 0.94, respectively. The system calculated the comprehensive evaluation value with weights of 0.4, 0.2, 0.2, and 0.2, determining an increase of 0.18 mm in actuator displacement, a decrease of 0.10 in load, and continued equipment operation as the manufacturing decision outcome. After execution, the traditional method corresponded to a peak load of 94.8, a peak actuator displacement of 6.47 mm, a peak temperature of 318.2 K, a hysteresis loss of 0.073, and a state fluctuation amplitude of 0.142. The present invention corresponded to a peak load of 87.9, a peak actuator displacement of 6.02 mm, a peak temperature of 313.8 K, a hysteresis loss of 0.052, and a state fluctuation amplitude of 0.061.

[0035] To further verify the effectiveness, the traditional threshold control method, the traditional linear predictive control method, and the method of this invention were run in 120 consecutive analysis windows. The average state prediction error of the traditional threshold control method was 0.086, that of the traditional linear predictive control method was 0.061, and that of this invention was 0.024. The average load peak value of the traditional threshold control method was 93.6, that of the traditional linear predictive control method was 90.8, and that of this invention was 86.9. The average hysteresis loss of the traditional threshold control method was 0.070, that of the traditional linear predictive control method was 0.061, and that of this invention was 0.049. The average actuator displacement fluctuation of the traditional threshold control method was 0.58 mm, that of the traditional linear predictive control method was 0.43 mm, and that of this invention was 0.26 mm. The manufacturing decision efficiency of the traditional threshold control method was 81.5%, that of the traditional linear predictive control method was 88.3%, and that of this invention was 96.7%. The results show that the present invention can identify the risk of state change in advance under load changes, process switching and equipment start-up and shutdown disturbances, eliminate candidate decisions that are beyond the reach of the state, reduce prediction bias, equipment loss and operation fluctuations, and improve the stability and effectiveness of manufacturing decisions.

[0036] 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 smart manufacturing information decision processing method based on industrial data analysis, characterized in that, include: Multi-source operating status data of industrial equipment is collected at edge nodes, and the multi-source operating status data is preprocessed to generate standardized multi-source operating status data. Based on standardized multi-source operating state data, a fractional-order Jiles-Atherton hysteresis model is constructed to calculate magnetization intensity, reversible magnetization component, irreversible magnetization component and hysteresis loss, generate hysteresis state evolution sequence, and map it to the Koopman high-dimensional observable space to construct a linear state evolution matrix and generate a linearized predicted state sequence. Based on standardized multi-source operating status data, process stage identifiers, load and equipment start-up and shutdown status data are read to identify process stage switching points, load mutation points and equipment status switching points. The linearized predicted state sequence is segmented to generate segmented state change functions. Based on the piecewise state change function, a manufacturing decision constraint generation module based on Filippov differential is constructed to calculate the state switching boundary set and solve the state reachable interval, generating a sequence of non-smooth state reachable domains. A set of manufacturing decision variables is constructed at the edge nodes. State prediction is performed by combining the hysteresis state evolution sequence, the linearized predicted state sequence, and the state switching boundary set. Candidate decision variables that exceed the non-smooth state reachability sequence are eliminated to generate a set of manufacturing decision candidates. The system performs multi-objective evaluation on the candidate set of manufacturing decisions, generates manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear and operational stability, and controls industrial equipment to execute manufacturing decisions. The generation of the linearized predicted state sequence includes: A fractional-order Jiles-Atherton hysteresis model is constructed based on standardized multi-source operating state data. The fractional-order Jiles-Atherton hysteresis model includes an equivalent magnetization driving module, a magnetization state decomposition module, a fractional-order magnetization evolution module, and a hysteresis state mapping module. The equivalent magnetization drive module calculates the difference between the magnetic flux density and the magnetic field strength to obtain the magnetization intensity state variable. The applied magnetic field and the magnetization intensity state variable are superimposed to construct the equivalent magnetic field. The equivalent magnetic field is then subjected to nonlinear mapping calculation to obtain the hysteresis-free magnetization intensity. The magnetization state decomposition module performs decomposition calculations based on hysteresis-free magnetization intensity and magnetization intensity, decomposing the magnetization intensity into reversible magnetization components and irreversible magnetization components. Based on temperature data, load data, and changes in magnetic field intensity, a dynamic proportional relationship is constructed to adaptively adjust the reversible and irreversible magnetization components. Furthermore, a hysteresis loss constraint is introduced to perform consistency correction on the decomposition results, resulting in corrected reversible and irreversible magnetization components. The fractional-order magnetization evolution module performs tensor multi-source coupling processing based on hysteresis-free magnetization intensity, corrected reversible magnetization components and irreversible magnetization components. It constructs a multi-dimensional tensor structure from standardized multi-source operating state data in the time dimension, physical quantity dimension and equipment state dimension, and performs fractional-order dynamic evolution calculation processing on dual time scales. The instantaneous response of magnetization is calculated along a fast timescale path, and the historical cumulative effect of magnetization is calculated along a slow timescale path. The results of the two evolution paths are coupled and fused to obtain the magnetization evolution result. Magnetization intensity is used as nodes and magnetization state transition relationships are used as edges to construct a magnetization state graph. The graph structure is then updated to obtain the hysteresis state evolution sequence. The hysteresis state mapping module performs lap integral processing based on magnetic field strength data and magnetic flux density data to obtain hysteresis loss, and maps the hysteresis state evolution sequence to the Koopman high-dimensional observable space to obtain high-dimensional observable states. It then constructs a linear state evolution matrix and performs state advancement to obtain a linearized predicted state sequence. The fractional-order Jiles-Atherton hysteresis model was trained using the reconstruction error of the magnetization evolution sequence as the optimization objective. The parameters in the equivalent magnetization driving module, magnetization state decomposition module, fractional-order magnetization evolution module, and hysteresis state mapping module were continuously optimized. Training was stopped when the reconstruction error was less than 0.01 and the parameter update amplitude was less than 0.001 within 5 consecutive iterations.

2. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The multi-source operating status data specifically includes excitation current data, magnetic field strength data, magnetic flux density data, temperature data, load data, actuator displacement data, process stage identification data, and equipment start-up and shutdown status data.

3. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The preprocessing of multi-source operational status data specifically includes timestamp alignment, data normalization, outlier removal, missing value imputation, data denoising, data smoothing, and data standardization.

4. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The generation of the piecewise state change function includes: Based on standardized multi-source operating status data, process stage identification data, load data and equipment start-up and shutdown status data are aligned according to a unified time sequence to construct an operating status sequence. Based on the operating condition state sequence, state change detection and processing is performed. The location where the process stage identifier changes is identified as the process stage switching point. The location where the load data difference between adjacent time moments is greater than 0.05 is identified as the load mutation point. The location where the equipment start-up and shutdown status changes is identified as the equipment state switching point. These are then merged to obtain a set of state switching points. The linearized predicted state sequence is segmented based on the set of state switching points. The linearized predicted state sequence is divided into multiple continuous subsequences according to adjacent state switching points, and a corresponding state change function is constructed for each subsequence to obtain a set of segmented state change functions.

5. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The generation of the non-smooth state reachable domain sequence includes: A manufacturing decision constraint generation module based on Filippov differential is constructed. The module includes a state switching topology representation unit, a state encoding unit, a Filippov reachable domain calculation unit, and a state domain generation unit. The state switching topology representation unit is based on the segmented state change function. Each segmented state change function is represented as a node, the switching relationship between adjacent segments is represented as a directed edge, and the state switching position is determined at the position corresponding to the directed edge to construct a segmented switching graph. The state coding unit performs path-dependent coding based on the linearized predicted state sequence, encodes the evolution information of the linearized predicted state on the historical path, obtains the extended state sequence, and embeds it into the segmented switching graph to obtain the state switching boundary set. The Filippov reachability computation unit performs set-valued dynamics computation based on the state switching boundary set. It represents the derivative of the piecewise state change function as multiple vector fields, performs weighted convex combination operation on the vector fields at the state switching boundary to obtain the state change rate belonging to the convex hull of the vector fields, introduces a hysteresis operator kernel to perform weighted convolution on the historical linearized predicted state, and combines the extended state sequence to perform time-progression calculation on the linearized predicted state to obtain the state reachability interval. The state domain generation unit constructs the state domain based on the state reachable interval. It divides the state reachable interval into multiple sub-intervals according to the piecewise state change function, and performs boundary reconstruction processing on each sub-interval to form a continuous boundary surface. It combines the hysteresis loss to constrain and filter the state domain, and performs set expansion and boundary integration processing on each sub-interval to generate a non-smooth state reachable domain sequence.

6. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The generated manufacturing decision candidate set includes: A set of manufacturing decision variables is constructed at the edge nodes. The manufacturing decision variables are represented as a multi-dimensional variable sequence consisting of actuator displacement adjustment, load adjustment and equipment operating status adjustment. The time-aligned data is then processed with standardized multi-source operating status data to obtain a set of candidate decision variables. State prediction is performed based on the hysteresis state evolution sequence, the linearized predicted state sequence, and the set of state switching boundaries. The set of candidate decision variables is substituted into the linear state evolution matrix for state advancement calculation. The predicted state is then corrected by combining the hysteresis state evolution sequence to obtain the corresponding set of candidate state trajectories. The candidate state trajectory set is constrained and filtered based on the non-smooth state reachability domain sequence. Candidate decision variables whose predicted states are within the state range corresponding to the non-smooth state reachability domain sequence are retained, while candidate decision variables whose predicted states are outside the state range corresponding to the non-smooth state reachability domain sequence are removed, thus obtaining the manufacturing decision candidate set.

7. The intelligent manufacturing information decision processing method based on industrial data analysis according to claim 1, characterized in that, The process of generating manufacturing decision results based on production efficiency, energy consumption, equipment wear and tear, and operational stability includes: A multi-objective evaluation process is performed on the candidate set of manufacturing decisions. The production efficiency, energy consumption, equipment loss and operational stability corresponding to each candidate decision are calculated and normalized to obtain the evaluation index sequence. The evaluation index sequence is weighted and fused. Production efficiency, energy consumption, equipment loss and operational stability are weighted and summed according to the set weights to obtain the comprehensive evaluation value of the corresponding candidate decision. The comprehensive evaluation values ​​are then sorted to obtain the priority sequence of candidate decisions. The manufacturing decision is selected based on the priority sequence of candidate decisions, and the candidate decision with the best comprehensive evaluation value is selected as the manufacturing decision result. Based on the manufacturing decision result, the displacement data, load data and equipment start-up and shutdown status data of the actuator are controlled and adjusted to complete the manufacturing decision execution.

Images