An intelligent production scheduling method and system for an automobile parts production workshop

Abstract

This invention belongs to the field of intelligent scheduling technology, specifically relating to an intelligent scheduling method and system for automotive parts production workshops. The method includes: constructing a state transition model for the production workshop, integrating equipment occupancy, work-in-process turnover, and predictive resource availability matrices; establishing a multi-dimensional transfer cost vector; extending the transition to a probabilistic multi-successor transition including processing qualification, rework, and equipment downtime; iteratively subdividing state blocks based on a progressive composite criterion of resource availability matrix differences, weighted Chebyshev distance, and Bach distance to achieve reduction; dynamically adjusting the weight vector according to the availability probability, and dynamically adjusting the cost threshold and probability threshold inversely based on the weight vector; constructing an aggregated state model based on the state reduction set; and generating scheduling instruction strategies using an improved A-heuristic search algorithm. This invention avoids the omission of optimal strategies due to oversimplification of the model, achieving a balance between scheduling efficiency and robustness in complex environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent production scheduling technology. More specifically, this invention relates to an intelligent production scheduling method and system for an automotive parts manufacturing workshop. Background Technology

[0002] Automotive parts manufacturing is the cornerstone of the automotive industry. Its production workshops are typically complex systems comprised of multiple CNC machining centers, automated conveyor systems, and diverse work-in-process inventory. To ensure timely and high-quality order delivery, companies need to scientifically plan the production elements within the workshop. For example, when faced with processing hundreds or thousands of parts, the system needs to determine which machine each workpiece should be on at a specific time, and the processing logic between these machines. Intelligent scheduling addresses such complex scenarios by constructing mathematical models to simulate the dynamic migration of production elements and using optimization algorithms to automatically generate production instruction strategies. However, in actual automotive parts production, the workshop environment is often accompanied by uncertainties such as random equipment failures, tool performance degradation, and fluctuations in processing yield. This necessitates that the scheduling system accurately characterize the micro-level changes in production status and their resulting multi-dimensional cost impacts.

[0003] Currently, to address the dynamic nature of the workshop, existing technical solutions typically tend to construct a production workshop state transition model. This approach establishes logical transition relationships between states by defining system states that include equipment occupancy information and the location of work-in-process inventory, and pre-sets certain judgment criteria to merge and reduce similar states. When performing production scheduling searches, this approach uses pre-set cost indicators such as processing time to evaluate the search path, and then attempts to find an optimal or near-optimal path that can complete all production tasks.

[0004] However, the existing solutions exhibit significant limitations when faced with the high complexity and dynamic demands of modern automotive parts manufacturing. First, as production scale expands and random events increase, the number of states in the shop floor state transition model often grows exponentially, easily leading to state space explosion. This makes it difficult for traditional search and solution algorithms to obtain optimal or near-optimal production scheduling solutions within a reasonable timeframe. Second, existing state reduction methods typically focus on assessing the current deterministic production state, lacking predictions of future resource availability over multiple time periods, resulting in insufficient stability of scheduling strategies when resource supply changes abruptly. Furthermore, when determining whether different states can be merged, traditional techniques often rely on fixed single criteria and static thresholds, making it difficult to integrate multi-dimensional transition costs, probabilistic multi-branch successor transitions, and resource change trends. Finally, this rigid and fragmented reduction mechanism cannot flexibly adjust the reduction scale according to future resource changes, easily leading to oversimplification of the model and omission of optimal conditional branch scheduling strategies, or continued constraints due to insufficient reduction. Therefore, how to improve the efficiency of solving complex shop floor scheduling problems while maintaining model accuracy has become a pressing technical challenge for the industry. Summary of the Invention

[0005] To address the technical problems of low scheduling efficiency due to excessively high state space dimensions in existing technologies, and insufficient stability of scheduling instruction strategies under fluctuating resource supply environments, this invention provides an intelligent scheduling method and system for automotive parts production workshops.

[0006] In a first aspect, the present invention provides an intelligent scheduling method for an automotive parts production workshop, comprising: S1: constructing a state transition model for the production workshop, wherein the state is composed of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability in multiple future time periods; constructing a multi-dimensional transition cost vector for state transitions, including processing time, energy consumption, and tool wear; and extending deterministic transitions to probabilistic multi-successor transitions representing processing qualification, rework, and equipment downtime; S2: performing state reduction on the state transition model for the production workshop, iteratively subdividing the state blocks until convergence to obtain a state reduction set, the subdivision being based on a progressive composite criterion, the progressive composite criterion including: when the predictive resources of two states... When the difference in the availability matrix exceeds a preset difference threshold, the two states are separated; the weighted Chebyshev distance between the multidimensional transition cost vectors when the two states perform the same operation is calculated, and the two states are separated when the weighted Chebyshev distance exceeds the cost threshold; the Barthel distance between the probability distribution vectors of the two states transitioning to each subsequent state block is calculated, and the two states are separated when the Barthel distance exceeds the probability threshold; the weight vector corresponding to the weighted Chebyshev distance is generated by the predictive resource availability matrix. When the predicted future availability probability of a resource decreases, the corresponding cost dimension weight is increased, and the cost threshold and probability threshold are dynamically adjusted inversely according to the weight vector; S3: An aggregated state model is constructed based on the state reduction set, using an improved A Heuristic search algorithms are used to generate production scheduling instruction strategies.

[0007] By adopting the above technical solution, this invention solves the technical problems of state space explosion caused by the expansion of production task scale in modern workshops and poor stability of production scheduling schemes caused by random equipment failures and fluctuations in processing qualification rates. It constructs an integrated predictive resource availability matrix and a probabilistic multi-successor transition state transition model, coupled with a progressive composite reduction criterion. The predictive matrix introduces forward-looking risk perception into production scheduling, and, combined with a three-layer progressive filtering mechanism of resource difference, weighted cost, and probability distribution similarity, mathematically ensures that state merging only occurs when functional expectations and risk trends are highly consistent. Simultaneously, the inverse linkage logic of weights and thresholds automatically tightens the state separation scale when the risk of resource degradation increases. This significantly reduces computational overhead while avoiding the omission of optimal strategies caused by oversimplification in traditional models, thus improving the efficiency and stability of production scheduling solutions in complex environments.

[0008] Preferably, the status consists of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability in multiple future time periods, including: acquiring occupancy status data streams of equipment at the bottom level of the workshop to represent equipment occupancy; acquiring turnover location status data of work-in-process to represent work-in-process turnover; statistically analyzing the continuous normal operation duration and historical failure time points of the bottom-level equipment, and using the Weibull distribution reliability model to calculate the expected fault-free operation probability of the bottom-level equipment in multiple future time periods; and combining the multiple future time periods with the corresponding expected fault-free operation probabilities to construct the predictive resource availability matrix.

[0009] By adopting the above technical solution, this invention utilizes the Weibull distribution reliability model to transform the physical operating patterns of equipment into predicted fault-free operation probabilities, realizing the transformation of workshop resources from static occupancy descriptions to dynamic availability predictions. This model accurately captures the wear and tear and failure trends of equipment over time, providing statistically confident data support for the predictive resource availability matrix, enabling subsequent production scheduling decisions to avoid potential resource bottlenecks in advance based on actual physical degradation patterns.

[0010] Preferably, the construction of a multidimensional transfer cost vector for state transition, including processing time, energy consumption, and tool wear, includes: calculating the processing time based on the start and end times of a single processing action; collecting the power consumption data of the processing equipment during processing and deducting the standby basic energy consumption to obtain the net energy consumption value as the energy consumption; acquiring the characteristic value of the spindle vibration signal and calculating the tool wear amount through a preset tool wear evaluation model as the tool wear; and combining the processing time, the energy consumption, and the tool wear to generate the multidimensional transfer cost vector.

[0011] Preferably, the step of extending the deterministic transition to represent probabilistic multiple successor transitions representing processing qualification, rework, and equipment downtime includes: calculating the equipment downtime branch probability based on the unit-time failure rate of the equipment during the duration of a single processing action; calculating the conditional qualification probability and conditional rework probability based on the number of qualified parts and the number of reworked parts in the same historical process, provided that no equipment downtime branch occurs; and calculating the processing qualification branch probability and rework branch probability based on the equipment downtime branch probability, the conditional qualification probability, and the conditional rework probability to generate the probabilistic multiple successor transitions.

[0012] By adopting the above technical solution, this invention establishes a probabilistic multi-successor transition based on the unit-time failure rate and historical processing statistics, transforming uncontrollable disturbances in the real workshop into predictable branches in a mathematical model. This multi-branch structure, through the Markov chain principle, treats qualified processing, rework, and downtime as a coexisting potential path space, enabling the system to pre-configure response actions based on probability distribution when generating instructions, significantly reducing the impact of random production fluctuations on scheduling execution rates.

[0013] Preferably, separating the two states when the difference between the predictive resource availability matrices of the two states is greater than a preset difference threshold includes: extracting the predictive resource availability matrices corresponding to the first state and the second state to be compared; traversing the same row and column coordinate positions of the two predictive resource availability matrices and calculating the absolute value of the difference between the values ​​of the corresponding position elements; summing the absolute values ​​of all coordinate positions to obtain the sum of absolute deviations; when the sum of absolute deviations is greater than the difference threshold, determining that the first state and the second state do not satisfy the isomorphism condition, and separating the first state and the second state.

[0014] Preferably, the weight vector is generated from a predictive resource availability matrix, including: obtaining the average value of the expected availability probability of each resource in the predictive resource availability matrix over multiple future time periods; when the average value is lower than a preset normal availability probability benchmark value, calculating the probability decrease deviation between the average value and the normal availability probability benchmark value; calculating the corresponding weight increment based on the probability decrease deviation and a preset compensation coefficient; and superimposing the weight increment onto the corresponding cost dimension weight according to a preset mapping relationship and performing normalization processing to generate the weight vector.

[0015] By adopting the above technical solution, this invention dynamically generates a weight vector based on the deviation of the resource availability probability from the benchmark value, establishing a cost-sensitive adjustment mechanism driven by resource health. When a failure risk is predicted for a certain piece of equipment or material, the cost weight of the relevant dimension automatically increases, and the forced reduction algorithm focuses on monitoring this risk dimension, guiding the production scheduling logic to achieve more targeted and refined status classification during periods of resource shortage.

[0016] Preferably, the step of calculating the weighted Chebyshev distance between the multidimensional transfer cost vectors when two states perform the same operation includes: obtaining the first multidimensional transfer cost vector and the second multidimensional transfer cost vector corresponding to the two states when performing the same operation; calculating the absolute values ​​of the numerical differences between the first multidimensional transfer cost vector and the second multidimensional transfer cost vector in each dimension of processing time, energy consumption, and tool wear, and performing dimensionless processing; multiplying the absolute values ​​of each dimension after dimensionless processing by the weight values ​​of the corresponding dimensions in the weight vector to obtain the weighted absolute difference of each dimension; and extracting the maximum value among the weighted absolute differences of all dimensions as the weighted Chebyshev distance.

[0017] Preferably, the construction of the aggregated state model based on the state reduction set includes: mapping each state block in the state reduction set to a macro-state node in the aggregated state model; for any first macro-state node and second macro-state node, when at least one original state in the first macro-state node transitions to a certain original state in the second macro-state node after performing a specific operation, establishing a directed transition edge from the first macro-state node to the second macro-state node; and attaching the corresponding operation, the mean attribute of the multidimensional transition cost vector, and the transition probability distribution of the probabilistic multiple successor transitions to the directed transition edge to construct the aggregated state model.

[0018] By adopting the above technical solution, this invention maps state blocks to macro-state nodes and establishes weighted directed edges based on cross-domain transitions, achieving efficient dimensionality reduction from the underlying micro-state space to the macro-logical topology. Through arithmetic averaging and weighted summation, the underlying discrete data is aggregated into statistically equivalent macro-transition parameters, significantly compressing the search space without sacrificing core process features, thus laying a compact structural foundation for the efficient execution of subsequent heuristic searches.

[0019] Preferably, the improved A is used A heuristic search algorithm for generating production scheduling instructions includes: in the aggregated state model, setting the macro-state node where the current initial state is located as the search starting point, and setting the macro-state node where all production tasks are completed as the target node; constructing a heuristic evaluation function by minimizing the weighted comprehensive cost of processing time, energy consumption, and tool wear after dimensionless processing to reach the target node; and employing an improved A... The heuristic search algorithm starts from the search starting point, calculates the estimated cost of candidate paths according to the heuristic evaluation function, and prioritizes expanding the macro-state nodes with the smallest estimated cost until the target node is reached to obtain the main production scheduling path; based on the probabilistic multiple successor transitions on the main production scheduling path, including processing qualified, rework, and equipment shutdown branches, inserts conditional execution actions to generate a production scheduling instruction strategy containing conditional branches.

[0020] Secondly, the present invention provides an intelligent scheduling system for an automotive parts production workshop, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned intelligent scheduling method for an automotive parts production workshop is implemented.

[0021] By adopting the above technical solution, a computer program is generated from the intelligent scheduling method for an automotive parts production workshop and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of terminal equipment based on the memory and processor, making it convenient to use.

[0022] The technical solution of the present invention has the following beneficial technical effects:

[0023] This invention constructs a workshop state transition model that integrates equipment occupancy, work-in-process turnover, and a predictive resource availability matrix. It also simultaneously introduces a multi-dimensional transfer cost vector and probabilistic multi-successor transitions that include scenarios such as completed processing, rework, and equipment downtime. This modeling approach enables the production scheduling instruction generation process to deeply reflect the complex and ever-changing production environment of a real workshop, improving the accuracy and completeness of the model's mapping to the underlying physical entities.

[0024] A progressive composite criterion based on the difference in predictive resource availability matrix, the weighted Chebyshev distance between multidimensional transfer cost vectors, and the Bach distance between successor probability distribution vectors is established to perform iterative state reduction on the production workshop state transition model. This criterion effectively eliminates redundant isomorphic states and compresses the model size while improving the Aspect Ratio of subsequent improvements. The efficiency of heuristic search algorithms.

[0025] The system dynamically adjusts the weight distribution of the multi-dimensional transfer cost vector and related separation thresholds based on the changing trends of resource availability probabilities over multiple time periods. Under conditions of resource deterioration or supply shortages, by increasing the cost weights of related dimensions and simultaneously tightening the state separation scale, an optimal balance is achieved between refined division of production states and control of computational complexity.

[0026] An aggregated state model is constructed based on a state reduction set, and a production scheduling instruction strategy containing conditional execution logic is generated accordingly. This solution can effectively mitigate uncertainties such as sudden equipment failures or fluctuations in processing quality during the production process, and enhances the execution stability and on-site applicability of the production scheduling results. Attached Figure Description

[0027] Figure 1 This is a flowchart of an intelligent production scheduling method for an automotive parts manufacturing workshop according to the present invention; Figure 2 It is a graph showing how the expected probability of fault-free operation of different underlying devices changes over future time windows; Figure 3 This is a diagram illustrating the dynamic weighting adjustment of each cost dimension; Figure 4 This is a schematic diagram of the logic for calculating the weighted Chebyshev distance. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0029] This invention discloses an intelligent scheduling method for automotive parts manufacturing workshops, referring to... Figure 1 This includes steps S1-S3. S1: Construct a state transition model for the production workshop.

[0030] In an optional embodiment, a production workshop state transition model is constructed, wherein the state consists of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability in future multiple time periods. A multi-dimensional transition cost vector containing processing time, energy consumption, and tool wear is constructed for each state transition, and the deterministic transition is extended to represent probabilistic multiple successor transitions representing processing qualification, rework, and equipment downtime.

[0031] In practical implementation, the system employs a Markov decision process to construct a state transition model for the production workshop. The state in this model consists of equipment occupancy, work-in-process inventory flow, and a predictive resource availability matrix representing the probability of resource availability over multiple future time periods. Specifically, the system uses a Boolean array to record the occupancy and idle status of each piece of equipment in the workshop, and an integer queue to record the queue quantity and flow position of work-in-process at each process node, thereby mapping the underlying entity of the workshop into a computable state representation.

[0032] For constructing the predictive resource availability matrix, the system establishes prediction branches according to resource type. For equipment resources, the system establishes a Weibull distribution reliability model based on the continuous normal operating time and historical failure time points in the system log to calculate the expected failure-free operation probability within multiple future time windows. The Weibull distribution reliability model is used to describe the probability change of equipment failure as operating time increases, and to predict the probability of equipment still operating normally within a certain period of time. For material resources, the system establishes a hidden Markov model based on historical material flow data, using on-time arrival, delayed arrival, and material shortage as the hidden state set, and the vector formed by discretizing the RFID arrival time difference and the supply buffer change as the observation sequence. The initial state distribution, state transition probability matrix, and observation probability matrix of this model are obtained through unsupervised training of historical delay records using the Baum-Welch algorithm, and then the on-time availability probability of materials within multiple future time windows is calculated using a forward algorithm. In the implementation where equipment resource availability is used as the basic resource row of the matrix, the system can also include material resources as an extended resource row in the predictive resource availability matrix. Based on this, the system arranges the probability of equipment operating without failure and the probability of materials being available on time according to a unified resource row and time window column to construct a predictive resource availability matrix. The rows of the matrix represent different types of equipment and material resources, and the columns of the matrix represent future time periods. For material resources, the expected probability of operating without failure is understood as the expected availability probability of materials being available on time in the future time period. For equipment resources, the expected probability of operating without failure is understood as the expected availability probability of equipment not failing and being able to participate in processing in the future time period.

[0033] Meanwhile, the system constructs a multi-dimensional transition cost vector for each state transition, including machining time, energy consumption, and tool wear. Specifically, the system calls the programmable logic controller data interface to obtain the standard machining time, calculates the energy consumption value using the energy integration algorithm, and calls a preset tool wear assessment model to calculate the tool wear amount. The system then stores these three dimensions of indicators into a floating-point array as the multi-dimensional transition cost vector. In this embodiment, the tool wear assessment model uses a random forest regression model trained based on spindle vibration characteristics, while the Taylor tool life formula can be used as a reference model for offline calibration or training sample generation, and is not executed concurrently with the random forest regression model in the same online calculation process.

[0034] Furthermore, the system extends deterministic transitions to represent probabilistic multi-successor transitions between processing completion, rework, and equipment downtime. This transformation is achieved using a Markov chain method, constructing a three-dimensional state transition probability tensor to store the probability values ​​of the current state transitioning to the processing completion, rework, and equipment downtime states respectively after executing scheduling actions. As a result, the system's production workshop state transition model better reflects the complex and ever-changing production environment of a real workshop, improving the accuracy and completeness of the model's description.

[0035] In an optional embodiment, the specific process of the above-mentioned state consisting of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability in multiple future time periods includes: acquiring the occupancy status data stream of the bottom-level equipment in the workshop to represent equipment occupancy; acquiring the turnover location status data of work-in-process to represent work-in-process turnover; statistically analyzing the continuous normal operation duration and historical failure time points of the bottom-level equipment, and using the Weibull distribution reliability model to calculate the expected fault-free operation probability of the bottom-level equipment in multiple future time periods; and combining the multiple future time periods with the corresponding expected fault-free operation probabilities to construct a predictive resource availability matrix.

[0036] More specifically, during the data acquisition phase for equipment occupancy and work-in-process (WIP) flow, the system deploys a periodic data acquisition gateway in the PLC control cabinet. This gateway reads the relay on / off signal status of the underlying equipment at a sampling frequency of once per second. This data, combined with the PLC's running / idle status bits, the current work order execution flag, or the fixture occupancy signal, is represented by 0 / 1 Boolean values ​​to generate a data stream indicating the equipment occupancy status. Simultaneously, the system deploys 900MHz UHF RFID read / write nodes at key workstations on the workbench. These nodes parse the EPC codes within the workpiece transfer tags in real time and generate workstation residency data with millisecond-level timestamps. This generates the location status data for WIP flow, ensuring that the underlying entity status is mapped to the state model.

[0037] During the reliability analysis phase, the system fits a Weibull distribution reliability model based on the historical failure time points of the equipment extracted from the system logs.

[0038] Specifically, the Weibull distribution reliability model satisfies the following relationship:

[0039] in, This indicates the normal operating time of the equipment. Reliability at that time; Indicates the equipment's uptime; The shape parameter is preferably in the range of 1.5 to 2.5 to indicate the wear-out failure period; The scale parameter is indicated, with a preferred range of 500h to 1000h.

[0040] Get the current continuous normal operating time of the device For example, 200h, and set multiple future time windows. Sets, for example Then, the expected probability of fault-free operation for each future time period is calculated.

[0041] Specifically, the probability of fault-free operation is expected to satisfy the following relationship:

[0042] in, This indicates the expected probability of trouble-free operation; This indicates the duration of continuous normal operation of the current underlying device; Indicates the set future time window; Indicates that the equipment is running to Reliability at that time; Indicates that the equipment is running to Reliability at that time.

[0043] exist , , Under the example parameters, The probability of fault-free operation for 2h, 4h, and 8h is approximately 0.9978, 0.9955, and 0.991, respectively. Therefore, the probability of equipment resource availability shows a slow downward trend within a short time window. To obtain a more pronounced downward curve, a smaller scale parameter and a larger scale parameter at the end of the equipment's lifespan should be used. Alternatively, degradation correction terms such as abnormal vibrations or overloads that have already been observed can be introduced.

[0044] During the matrix construction phase, the system establishes the Cartesian product of the set of equipment resources and the set of future time windows. For example, for three key processing equipment and the aforementioned three time windows, the calculated nine fault-free operation probability values, such as the probabilities of equipment A at 2h, 4h, and 8h being 0.998, 0.996, and 0.991 respectively, are arranged into a two-dimensional matrix according to a fixed row and column mapping rule, generating a matrix of size [missing information]. A multi-period predictive resource availability matrix. If material resources are also included, the on-time arrival probability of materials is treated as an independent resource row in the matrix. For example, the on-time availability probability of material M1 at 2h, 4h, and 8h is output by the Hidden Markov Model as 0.94, 0.9, and 0.86, respectively, and is included in the same matrix along with the equipment resource probability.

[0045] In an optional embodiment, the specific process for constructing a multidimensional transfer cost vector including processing time, energy consumption, and tool wear for state transition includes: calculating the processing time based on the start and end times of a single processing action; collecting the power consumption data of the processing equipment during processing and deducting the standby basic energy consumption to obtain the net energy consumption value as energy consumption; acquiring the characteristic value of the spindle vibration signal and calculating the tool wear amount as tool wear through a preset tool wear evaluation model; and combining the processing time, energy consumption, and tool wear to generate a multidimensional transfer cost vector.

[0046] More specifically, the representation of processing time and energy consumption relies on edge metering. The system obtains the start and end millisecond-level system clock stamps of the process execution through the MES production execution system, and obtains the processing time of the first dimension, such as 125.5s, by subtracting them. By connecting to a smart meter with a resolution of 0.01kWh or an accuracy level that meets the preset metering requirements, the system obtains the total power consumption data within this time period, such as 0.45kWh. Based on the standby power specified on the equipment nameplate, such as 2.5kW, the theoretical standby base energy consumption within the same time period is calculated, which is 0.087kWh. The total power consumption is subtracted from this base energy consumption, and the result is taken as the maximum value between zero and the total power consumption, resulting in a net processing energy consumption value of 0.363kWh, which serves as the second dimension indicator.

[0047] In the tool wear dimension assessment, the system collects high-frequency vibration signals using a triaxial piezoelectric vibration sensor deployed on the spindle housing at a sampling frequency of 20kHz. The system uses Fast Fourier Transform (FFT) and filtering algorithms to extract the root mean square (RMS) value and peak factor at the spindle frequency as feature inputs, such as an RMS value of 0.5g. This feature matrix is ​​then input into a random forest regression model pre-trained based on historical data, outputting the corresponding tool cutting edge flank wear increment, for example, an estimated wear of 0.012mm, as the third dimension indicator. In this embodiment, the random forest regression model preferably contains 100 to 300 decision trees, using mean square error as the splitting quality assessment criterion, and setting a maximum tree depth constraint to prevent overfitting. The random forest regression model, as the preset tool wear assessment model in this embodiment, can have its training labels jointly calibrated by offline measured wear, microscopic measurement results, or Taylor's tool life formula conversion results. During the online execution phase, only vibration features are input and wear increments are output, avoiding inconsistencies in calculation methods caused by repeated online calculations with the Taylor formula. The system, following a fixed data structure framework of processing time, net energy consumption, and tool wear, concatenates the extracted index instances to generate a dataset similar to... A floating-point one-dimensional array is used to register the array as a multi-dimensional transition cost vector for the current state transition, in order to support multi-dimensional objective optimization in production scheduling assessment.

[0048] In an optional embodiment, the specific process of extending the deterministic transition to represent probabilistic multiple successor transitions of processing qualification, rework, and equipment downtime includes: calculating the equipment downtime branch probability based on the unit-time failure rate of the equipment during the duration of a single processing action; calculating the conditional qualification probability and conditional rework probability based on the number of qualified parts and the number of reworked parts in the same historical process, provided that no equipment downtime branch occurs; and calculating the processing qualification branch probability and rework branch probability based on the equipment downtime branch probability, conditional qualification probability, and conditional rework probability to generate probabilistic multiple successor transitions.

[0049] More specifically, the system first determines the equipment downtime probability based on the failure intensity of the equipment within the corresponding processing time.

[0050] Specifically, the probability of equipment downtime satisfies the following relationship:

[0051] in, Indicates the probability of equipment downtime; This indicates the equipment's failure rate per unit time. Indicates the duration of a single processing action, and when When measured in hours, It needs to be converted to hours before being substituted into the calculation.

[0052] Subsequently, assuming no equipment downtime, the system estimates the branch probability of the processing result based on the number of qualified parts and the number of rework parts in similar historical processes. Let the number of qualified parts in the completed samples of similar processes be... The number of returned workpieces is The probability of meeting the condition is then... The probability of rework under certain conditions is Based on this, the system calculates the probability of a successfully processed branch as follows: The probability of rework is The probability of equipment downtime remains at 0. .like If the value is 0, meaning there are no similar processing records in the historical samples, the system uses a preset prior probability or Laplace smoothing to determine the conditional pass probability and conditional rework probability to ensure the robustness of the probability estimation. In the specific software implementation, the system can call the Poisson distribution function, binomial distribution function, or multinomial distribution function to calculate the probability of random branches, and construct a three-dimensional state transition probability tensor to store the probability values ​​of the current state transitioning to the processing pass state, rework state, and equipment stop state after the execution of the scheduling action. In this way, the system extends the traditional deterministic state transition to a probabilistic multi-successor transition that includes multiple production outcome branches, enabling the scheduling model to characterize uncertain events such as processing pass, rework, and equipment stoppage in the real production process, and improving the model's ability to describe the actual working conditions of the workshop.

[0053] S2: Perform state reduction based on progressive composite criteria.

[0054] In an optional embodiment, the state transition model of the production workshop is reduced by iteratively subdividing the state blocks until convergence to obtain a reduced state set. The subdivision is based on a progressive composite criterion including criterion one, criterion two, and criterion three. Criterion one is to separate the two states if the difference between the predictive resource availability matrices of the two states is greater than a preset difference threshold. Criterion two is to calculate the weighted Chebyshev distance between the multidimensional transition cost vectors when the two states perform the same operation. If the distance is greater than the cost threshold, they are separated. The weight vector of the distance is generated by the predictive resource availability matrix. When the probability of a certain resource availability decreases in the future, the corresponding cost dimension weight is increased. Criterion three is to calculate the Bach distance between the probability distribution vectors of the two states transitioning to each successive state block. If the distance is greater than the probability threshold, they are separated. The cost threshold and the probability threshold are dynamically adjusted in reverse according to the generation result of the weight vector.

[0055] In the actual execution process, the system first initializes the state reduction set, dividing the state space into initial state blocks containing all states. Then, it uses a fixed-point iterative method to continuously subdivide the state blocks until the number of state blocks no longer increases or the state block partitioning result no longer changes, at which point convergence is determined. For any two states within the set, the system executes a progressive composite criterion for judgment.

[0056] When executing criterion one, the system calculates the difference in the predictive resource availability matrix between the two states using an element-by-element absolute deviation summation method. This involves summing the absolute values ​​of the differences between elements at the same row and column coordinate positions to obtain the total absolute deviation. If this total absolute deviation exceeds a preset difference threshold, the two states are assigned to different new state blocks. Before executing criterion two, the system compares the sets of executable operations for the two states. If the sets of executable operations differ, the two states are directly separated. If the sets of executable operations are the same, the weighted Chebyshev distance between the multidimensional transfer cost vectors is calculated for the same operation. When executing criterion two, the system calculates weight vectors based on the predictive resource availability matrix. When the average availability probability of a certain type of resource in the matrix over a future time period is less than a set value, the system calculates and increases the weight values ​​for the corresponding processing time, energy consumption, or tool wear dimensions using a linear incremental function.

[0057] Specifically, the linear increment function satisfies the following relationship:

[0058] in, This represents the weight increment for the corresponding cost dimension. This indicates a decrease in the probability of obtaining the resource. This indicates the preset compensation coefficient; Subsequently, the system calculates the weighted Chebyshev distance between the multidimensional transition cost vectors of the two states. If the calculated result is greater than the cost threshold, the two states are separated and the isomorphism determination of the current two states ends; if the calculated result is not greater than the cost threshold, criterion three continues to be executed. In the specific software implementation, the system can call the chebyshev function of the spatial module and combine it with the weight vector to calculate the weighted Chebyshev distance.

[0059] When executing criterion three, the system constructs a probability distribution vector from the probability values ​​of the transitions between two states to the same successor state block. The Barthel coefficient is obtained by summing the square roots of the products of corresponding elements of the two probability distribution vectors. The Barthel coefficient is an indicator used to measure the similarity between two probability distributions; a larger value indicates a closer similarity. The negative logarithm is typically used to obtain the Barthel distance. When the Barthel coefficient is less than a preset small positive number... When, correct it to Then, the negative logarithm of the corrected Bartholomew's coefficient is taken to calculate the Bartholomew's distance; if the Bartholomew's distance is greater than the dynamic probability threshold, it is determined that the difference in the successor probability distribution of the two states exceeds the allowable range, and the two states are separated.

[0060] Both the cost threshold and the probability threshold are calculated using a reverse adjustment formula with a lower bound constraint. The system first calculates the weight increments of each cost dimension before normalization. Calculate the intensity of risk sensitivity Alternatively, the risk sensitivity intensity can be calculated based on the sum of the decreases in the expected availability probability of each resource compared to the baseline value. .

[0061] Specifically, the intensity of risk sensitivity satisfies the following relationship:

[0062] in, Indicates the intensity of risk sensitivity; This represents the weight increment of each cost dimension before normalization; This indicates that the increments of each dimension are summed.

[0063] Subsequently, based on the calculated risk sensitivity intensity, the system calculates the corresponding cost threshold and probability threshold.

[0064] Specifically, the cost threshold satisfies the following relationship:

[0065] in, Indicates the cost threshold; This represents the preset cost base threshold; This represents the preset positive lower limit of the cost threshold; The adjustment coefficient representing the cost threshold; Indicates the intensity of risk sensitivity; This indicates retrieving the maximum value among the elements in the set.

[0066] Specifically, the probability threshold satisfies the following relationship:

[0067] in, Indicates the probability threshold; This represents the preset probability threshold. This represents the preset positive lower limit of the probability threshold; The adjustment coefficient representing the probability threshold; Indicates the intensity of risk sensitivity; This indicates retrieving the maximum value among the elements in the set.

[0068] The above basic threshold , With adjustment coefficient , The lower limit of the positive limit is pre-calibrated by a person skilled in the art based on the statistical distribution of the reduction effect of historical production scheduling data in the workshop. , This is used to prevent legitimate state branches from being incorrectly discarded due to excessively tightening the threshold under extreme risk.

[0069] The adjusted weight vector, after normalization, is used for weighted Chebyshev distance calculation, while the total risk increment before normalization is used for threshold adjustment. Through this inverse adjustment mechanism, when the weight of any cost dimension increases relative to the initial weight, By increasing the threshold and decreasing it accordingly, the state separation scale is tightened when resources become scarce. At the same time, the threshold is prevented from becoming negative by using a positive lower limit, thus achieving a dynamic balance between fine state partitioning and computational scale control.

[0070] In an optional embodiment, the specific process of separating the two states when the difference between the predictive resource availability matrices of the two states is greater than a preset difference threshold includes: extracting the predictive resource availability matrices corresponding to the first state and the second state to be compared respectively; traversing the same row and column coordinate positions of the two predictive resource availability matrices and calculating the absolute value of the difference between the corresponding position element values ​​respectively; summing the absolute values ​​of all coordinate positions to obtain the sum of absolute deviations; when the sum of absolute deviations is greater than the difference threshold, determining that the first state and the second state do not meet the isomorphism condition, and separating the first state and the second state.

[0071] More specifically, the difference calculation is implemented based on the element-wise distance evaluation of the matrix elements. In each iteration of state reduction, the system retrieves the first state to be verified and the distances bound to the second state from memory. A predictive resource availability matrix, in which This refers to the total number of equipment or the total number of equipment and material resources. For the number of future time windows, for example, constructing A floating-point matrix of size. The system uses a nested loop structure to traverse all coordinate indices of the matrix. Extract two probability estimates for the same device in the same future time period. For example, in the first state, the probability of no failure for CNC device 1 in the next 2 hours is 0.9, while in the second state, the corresponding probability is 0.85. Subtract the values ​​and take the absolute value to calculate 0.05. Then, combine all values ​​in the entire matrix. The sum of the absolute differences at each location is used to calculate the total absolute deviation representing the difference in resource risk between the two.

[0072] Specifically, the sum of absolute deviations satisfies the following relationship:

[0073] in, This represents the sum of absolute deviations; The coordinate indices in the predictive resource availability matrix representing the first state are: The estimated availability at the location; Indices of the same coordinate in the predictive resource availability matrix representing the second state The estimated availability at the location; This indicates that for all of the entire matrix... The coordinates of each location are accumulated and summed.

[0074] In the separation decision mechanism, the preferred range for the unit element difference threshold is set between 0.01 and 0.06, and is determined based on the number of matrix elements. The final difference threshold is obtained through conversion. For example, when the matrix size is... Furthermore, when the unit element difference threshold is set to 0.033, the final difference threshold is approximately 0.3. The sum of the calculated absolute deviations... Assuming the calculated result is 0.35, a logical comparison is made with the final difference threshold of 0.3. Since 0.35 is greater than 0.3, it is determined that the two states differ in their resource degradation trends over future periods. Therefore, the merging relationship between the two states in the same state block is cancelled, and they are divided into different sub-state blocks, for example, from the parent block. Splits give rise to independent and This ensures that states retained within the same set are equivalent in predictive maintenance scheduling constraints.

[0075] In an optional embodiment, the specific process of generating the weight vector from the predictive resource availability matrix includes: obtaining the average value of the expected availability probability of each resource in the predictive resource availability matrix over multiple future time periods; when the average value is lower than a preset normal availability probability benchmark value, calculating the probability decrease deviation between the average value and the normal availability probability benchmark value; calculating the corresponding weight increment based on the probability decrease deviation and a preset compensation coefficient; and, according to a preset mapping relationship, superimposing the weight increment onto the corresponding cost dimension weight and performing normalization processing to generate the weight vector.

[0076] More specifically, the weight adjustment process is driven by the deterioration trend of resource health. The system calculates the mean along the row vectors of the predictive resource availability matrix, extracting the average expected availability probability of each independent resource within multiple set future time windows. Equipment resources correspond to the expected fault-free operation probability, and material resources correspond to the expected on-time availability probability. For example, the calculated average probability for the spindle motor is 0.78. This average probability is compared with a preset normal operation probability benchmark value in the workshop process specifications. The preferred range for the benchmark value is set between 0.85 and 0.95, with an example value of 0.9. When the average value of 0.78 is identified as lower than the benchmark value of 0.9, the difference between the two is calculated to obtain a positive probability decrease deviation. Input the preset compensation coefficient for adjusting sensitivity. The preferred value is between 1.2 and 2. For example, if the value is 1.5, it can be obtained through linear multiplication. The calculated weight increment is 0.18.

[0077] In the multi-dimensional cost weight allocation and reconstruction stage, the system reads a pre-built mapping table in the knowledge base to determine the transmission path of resource failure risk to production scheduling costs. The mapping table is generated from expert experience or historical failure correlation analysis, and internally records the directional transmission relationship from the underlying physical resource nodes to the multi-dimensional transfer cost vector (processing time, energy consumption, tool wear). For example, the degradation of the spindle motor will lead to a deterioration in the tool wear cost dimension; based on this mapping, the calculated increment of 0.18 is added to the initial weight of 0.3 for the third dimension, i.e., the tool wear dimension, updating the weight of this dimension to 0.48, while keeping the initial weights of the first dimension (processing time) at 0.4 and the second dimension (energy consumption) at 0.3 unchanged. The corrected weight array... Perform normalized division with a sum of L1 norms of 1.18, output the result, and send it down. As the adjusted weight vector, the cost weight of high-risk operations is increased. In this example, the unnormalized weight increment for the spindle motor corresponding to the tool wear dimension is 0.18, therefore the risk sensitivity intensity can be taken as... Or take the deviation based on the expected decrease in the probability of resource availability. Subsequently, both the cost threshold and probability threshold are reduced using a reverse adjustment formula with a lower limit, ensuring that the state separation criterion aligns with the direction of increasing resource risk. Through this dynamic weight adjustment mechanism, the system can automatically increase the weight of the cost dimension associated with a resource in the state separation determination when it predicts a decrease in the availability probability of that resource. This makes the reduction process more sensitive to resource deterioration trends and enables more refined state division when resources become scarce. Consequently, the resulting production scheduling strategy can better avoid resource bottlenecks.

[0078] In an optional embodiment, the specific process of calculating the weighted Chebyshev distance between the multidimensional transfer cost vectors when two states perform the same operation includes: obtaining the first multidimensional transfer cost vector and the second multidimensional transfer cost vector corresponding to the two states when performing the same operation; calculating the absolute values ​​of the numerical differences between the first multidimensional transfer cost vector and the second multidimensional transfer cost vector in each dimension of processing time, energy consumption, and tool wear, and performing dimensionless processing; multiplying the absolute values ​​of each dimension after dimensionless processing by the weight values ​​of the corresponding dimensions in the weight vector to obtain the weighted absolute difference of each dimension; and extracting the maximum value among the weighted absolute differences of all dimensions as the weighted Chebyshev distance.

[0079] More specifically, the deviation measure of multidimensional cost follows the Chebyshev distance paradigm combined with the range standardization method. When evaluating the second state separation criterion, the system extracts in parallel the cost vectors bound to state A and state B for performing specific processes such as milling; for example, the first vector is... The second vector is To reduce the scale interference of different physical dimensions on distance calculation, the system pre-calculates the range values ​​of each dimension in the historical database, such as the time range of 50s, the energy consumption range of 0.20kWh, and the wear range of 0.020mm. After taking the absolute value of the difference between the two vectors in the corresponding dimension, the system divides them by the range to perform dimensionless mapping. The normalized difference value is calculated to be 0.3 for the time dimension, 0.25 for the energy consumption dimension, and 0.25 for the tool wear dimension.

[0080] In the step of extracting the maximum cost dimension, the system compares the aforementioned dimensionless absolute difference with the dynamically generated weight vector, for example... Performing a dot product of each term yields a weighted absolute difference of 0.102 for the time dimension, 0.0625 for the energy dimension, and 0.1025 for the wear dimension. This is based on the Chebyshev distance's infinite norm. The principle of taking the supremum of the differences between the coordinate axes is established, for example, by calling the maximum value function. The value 0.1025 is extracted as the weighted Chebyshev distance between the two states. This maximum value indicates the degree of difference between the two states in the most unfavorable cost dimension, and becomes the threshold comparison index for triggering the state block stripping operation. The weighted Chebyshev distance calculation is illustrated in the figure below. Figure 4 As shown. By employing Chebyshev distance, which takes the maximum value of the weighted differences of each dimension, the system can use the most sensitive cost dimension as the criterion for state separation. This ensures that when a significant difference occurs in any key cost dimension, state subdivision is performed in a timely manner, avoiding the obscuring of substantial differences in key dimensions due to smaller differences in other dimensions. This improves the ability of the reduction results to distinguish multidimensional production cost differences.

[0081] S3: Construct an aggregated state model and generate a production scheduling instruction strategy.

[0082] In an optional embodiment, an aggregated state model is constructed based on the state reduction set, using an improved A... Heuristic search algorithms generate production scheduling strategies that include conditional branches.

[0083] In the specific execution process, the system maps each state block in the converged state reduction set to a macro-state node in the aggregated state model. It then traverses the state transition probability tensor and multi-dimensional transition cost vector in the original state transition model, using a weighted average algorithm to calculate the transition probabilities and expected costs between state blocks, thereby generating the macro-state node transition model. The aggregated state model transforms the state reduction set into a macro-state transition model, where each state reduction set corresponds to a macro-state node, and the directed transition edges between macro-state nodes correspond to executable production operations. These directed transition edges carry the aggregated multi-dimensional transition costs and successor probability distributions. The system adopts a Markov decision process with action constraints as the aggregated state model architecture, and an improved A* algorithm is applied to this aggregated state model. A heuristic search algorithm is used to generate production scheduling instructions. Specifically, the system sets A... The algorithm's cost function is the sum of the accumulated expected transfer cost and the estimated heuristic cost from the current macrostate node to the target macrostate node. Both the accumulated expected transfer cost and the estimated heuristic cost are calculated using dynamically generated cost dimension weights in S2, instead of a fixed weight allocation. The estimated heuristic cost is estimated using the minimum dynamically weighted sum of the remaining incomplete processes under ideal fault-free conditions, and this estimate is required to be no higher than the actual reachable cost from the current macrostate to the target macrostate, in order to maintain A... The system assesses the acceptability of the search results. During the expansion of the search tree, when multiple probabilistic successor transitions (processing qualified, rework, and equipment downtime) are detected, the system generates a scheduling tree structure with multi-directional conditional branches: for the processing qualified branch, it continues to execute subsequent processing instructions; for the rework branch, it inserts rework instructions; and for the equipment downtime branch, it inserts instructions to wait or reallocate equipment. When the probability of an abnormal successor branch is higher than a preset abnormal probability threshold, or the probability of successful processing is lower than a preset success rate threshold, the system re-calls A for the corresponding rework branch or equipment downtime branch. The search algorithm finds local correction paths and concatenates the main path with the local correction paths to form a tree-shaped scheduling strategy with lower overall cost under the current evaluation function. Based on the reduced aggregate state model, a scheduling strategy with conditional branches is generated, which can better cope with unexpected events and uncertain operating conditions in the production process, enhancing the stability and on-site executability of the scheduling plan.

[0084] In an optional embodiment, the specific process of constructing the aggregated state model based on the state reduction set includes: mapping each state block in the state reduction set to a macro-state node in the aggregated state model; for any first macro-state node and second macro-state node, when at least one original state in the first macro-state node transitions to a certain original state in the second macro-state node after performing a specific operation, establishing a directed transition edge from the first macro-state node to the second macro-state node; and attaching the corresponding operation, the mean attribute of the multidimensional transition cost vector, and the transition probability distribution of the probabilistic multiple successor transitions to the directed transition edge to construct the aggregated state model.

[0085] More specifically, the dimensionality reduction and reconstruction process of the aggregated state model essentially involves aggregating the equivalence classes of the underlying microstates and establishing topological connectivity. After the reduction algorithm terminates its iterations (i.e., the set no longer splits), the system will generate each independent set of states, for example, containing the underlying states. , , The collection is encapsulated and instantiated as a macro-state node with no internal structure, denoted as . This node carries the core technological features shared by all its original states. Subsequently, the system traverses all node combinations to perform a reachability scan; assuming a reachability scan is detected... The underlying state When a specific operation is triggered, such as a workpiece clamping operation. After that, the underlying state transitions to the successor state. It happens to belong to another macro state node Then add a path to the directed graph data structure. Starting point A directed transition edge with the endpoint as the destination.

[0086] In the parameter aggregation and assignment phase of directed edges, the system backtracks all factors that contribute to the aggregation and assignment of directed edges. Domain crossing to The underlying transition events of the domain. By statistically accumulating and calculating the arithmetic mean, a comprehensive transition attribute is calculated: for example, averaging the time taken for multiple clamping operations (10s, 12s) results in 11s; averaging the energy consumption (0.05kWh, 0.07kWh) results in 0.06kWh. The transition probability distribution is then weighted according to state weights; for example, the weighted averages of the processing qualification probability, rework probability, and equipment downtime probability in the underlying transition events are calculated separately, forming an aggregated probability distribution vector. When the branches of "processing qualified," "rework," and "equipment shutdown" correspond to different subsequent macrostate nodes, the system aggregates the probabilities of these branches according to their corresponding subsequent macrostate nodes, forming a probability distribution vector for transitioning to each subsequent state block. Furthermore, for the transition probabilities of the same macrostate node pointing to each subsequent macrostate node under the same operation, the system performs unified normalization processing, ensuring that the sum of the corresponding successor probabilities is 1. Directed edges corresponding to different operations are stored separately, without mixing the costs and probabilities of different operations to calculate the average. The system encapsulates the operation label, aggregated cost vector, and aggregated probability distribution vector together as a data attribute dictionary for this directed transition edge. By performing the aforementioned dimensionality reduction topological connection and mean parameter binding on all graph nodes, a workshop aggregated state model is constructed. Through this aggregation process, the system significantly reduces the size of the state space while retaining key differences in the underlying states, laying the foundation for the efficient execution of subsequent heuristic search.

[0087] In an optional embodiment, the above employs an improved A The specific process of generating production scheduling strategies using a heuristic search algorithm includes: in the aggregated state model, setting the macro-state node where the current initial state is located as the search starting point, and setting the macro-state node where all production tasks are completed as the target node; constructing a heuristic evaluation function by minimizing the weighted comprehensive cost of processing time, energy consumption, and tool wear after dimensionless processing to reach the target node; and using an improved A... The heuristic search algorithm starts from the search starting point, calculates the estimated cost of candidate paths based on the heuristic evaluation function, and prioritizes expanding the macro-state nodes with the smallest estimated cost until the target node is reached, thus obtaining the main production scheduling path. Based on the probabilistic multiple successor transitions on the main production scheduling path, including processing qualified, rework, and equipment shutdown branches, conditional execution actions are inserted to generate a production scheduling instruction strategy containing conditional branches.

[0088] More specifically, production scheduling optimization based on the aggregation state model is mainly guided by the evaluation function A. Tree search is used to achieve this. The system locates the corresponding starting macrostate node in the graph by matching the initial state of the equipment with the workpiece blank information. And set the macro status of marking all order lists as completed as the target node. The system constructs a heuristic evaluation function. ,in This represents the cumulative expected comprehensive cost of the candidate path under the current probabilistic multiple successor transitions. For a single transition edge, the system first linearly weights the dimensionless mean values ​​of time, energy consumption, and wear using a dynamically generated weight vector based on S2 to obtain the comprehensive cost of that edge under the corresponding successor branch. Then, it performs expected weighting according to the branch probabilities of processing qualification, rework, and equipment downtime, and accumulates this value along the candidate path to obtain the expected comprehensive cost. For example, the dynamic weights generated under the current resource availability matrix are: Calculations yielded Heuristic functions A fast lower bound estimate is performed using the minimum weighted sum of costs required for the remaining unfinished processes under ideal, fault-free conditions. Specifically, the system ignores resource conflicts and queuing times, independently assigning each remaining process to the ideal equipment with the shortest theoretical processing time for cumulative calculation. For example, the calculated... To ensure If the admissibility is no higher than the actual cost, then the estimated cost of the current node is calculated as follows: If it cannot be proven that a certain heuristic estimate is not higher than the actual cost, it is used as a non-strictly admissible engineering heuristic, and the output is expressed as a better scheduling instruction strategy, rather than as the globally optimal strategy.

[0089] During the iterative expansion phase of the algorithm search, the system maintains a list based on... The priority queue OpenList is sorted in ascending order. Each time, the macro-state node with the lowest estimated cost value at the top is popped to explore adjacent edges, and the derived legal child nodes are re-enqueued after evaluation, until the target node is reached. .

[0090] In the production scheduling instruction tree generation stage, the system backtracks A in reverse. The system extracts the main operation sequence from the directed edges attached to the optimal search path obtained from the search, such as CNC1 performing milling. For the previously generated rework branches or equipment stoppage branches, the system assembles the corresponding local correction paths into the main operation sequence and outputs a tree-structured scheduling instruction strategy containing If-Then conditional branch control logic. For example, if CNC1 milling is successful, the AGV is switched to transport; if an exception occurs, a pre-assigned work order is called and reassigned to CNC2. This tree-structured strategy is a low-cost executable strategy obtained under the constraints of the aggregated state model and heuristic evaluation function, and is not limited to an unproven globally optimal strategy.

[0091] The test set consisted of 30 consecutive days of historical operating data from a heavy machinery processing workshop, involving 15 key CNC machining machines and 500 manufacturing work orders containing multiple processes. All equipment was initially calibrated as normal, and the test server was configured with a 64-core processor and 256GB of memory. The control group used a traditional static, fixed-weight state transition model and conventional distance evaluation criteria for production scheduling. The experimental group used the predictive resource availability matrix generation weight vector mechanism of this invention, combined with weighted Chebyshev distance to construct an aggregated state model for conditional branch production scheduling.

[0092] In processing 500 production tasks, the scheduling strategy generated by the control group resulted in a total processing time of 4200 hours, a total net energy consumption of 15000 kWh, and 35 instances of sudden tool failures, with an average time of 45 seconds per scheduling solution. The scheduling strategy generated by the experimental group reduced the total processing time to 3850 hours, the total net energy consumption to 13200 kWh, reduced the number of sudden tool failures to 8, and shortened the average time of a single scheduling algorithm solution to 8.5 seconds.

[0093] The number of sudden tool failures decreased by 77.1%, and the average time per algorithm solution decreased by 81.1%. This indicates that adjusting multidimensional cost weights based on the predictive availability matrix can identify and avoid high-risk equipment degradation states, reducing the failure rate. The dimensionality reduction and reconstruction technique combining weighted Chebyshev distance and aggregated state models removes a large number of redundant isomorphic states, reducing the search space complexity of the scheduling algorithm and enabling the production workshop to achieve a better overall balance between processing efficiency and energy consumption management.

[0094] Figure 2 This graph shows the expected fault-free operation probability of different underlying devices as a function of future time windows. In the graph, the solid line with dots represents device A; the dashed line with square dots represents device B; and the dotted line with triangles represents device C. The graph shows that all curves decrease as the time window lengthens, proving that the Weibull distribution reliability model successfully characterizes the wear and tear failure pattern of devices over time. Observing the dotted line corresponding to device C, we can see that its downward slope is significantly higher than that of device A. This corresponds to the greater failure risk characteristic of devices at the end of their lifespan or experiencing abnormal disturbances, as described in the specific implementation.

[0095] Figure 3 This diagram illustrates the dynamic weight adjustment for each cost dimension. Each group in the diagram contains three bars: the left bar represents the initial weight; the middle bar represents the weight after adding the weight increment; and the right bar represents the final generated weight vector. The diagram shows that the middle and right bars for the tool wear dimension have increased weights relative to the initial weights on the left side of the group. This demonstrates that the weight generation mechanism successfully captures the risk shift caused by the decrease in resource availability probability. Observing the bar distribution for processing time and energy consumption dimensions reveals that the height of the right bar decreases after normalization. This corresponds to the adjustment logic described in the specific implementation where the system automatically focuses the search on the high-risk resource association dimension.

[0096] Figure 4This is a schematic diagram illustrating the logic of weighted Chebyshev distance calculation. Each group in the diagram contains two bars: the left bar represents the normalized difference after dimensionless processing; the right bar represents the weighted absolute difference; and the horizontal dashed line running through the diagram represents the maximum weighted absolute difference among all extracted dimensions. The image shows that the right bar for the wear difference dimension is significantly taller than the right bars for the other two evaluation dimensions, and the height of the horizontal dashed line aligns with the top of the right bar in the wear difference group. This demonstrates that the algorithm successfully identified the most significant deviation in the cost vector's weakest dimension and used the difference in this dimension as the criterion for state separation. This corresponds to the algorithmic feature described in the specific implementation that utilizes the infinite norm to extract the difference in the most unfavorable cost dimension to trigger state subdivision.

[0097] This invention also discloses an intelligent scheduling system for an automotive parts production workshop, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent scheduling method for an automotive parts production workshop according to this invention is implemented.

[0098] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0099] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An intelligent scheduling method for an automotive parts manufacturing workshop, characterized in that, include: S1: Construct a state transition model for the production workshop. Each state consists of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability over multiple future time periods. A multi-dimensional transition cost vector is constructed for state transitions, incorporating processing time, energy consumption, and tool wear. Deterministic transitions are extended to represent probabilistic multi-successor transitions representing processing completion, rework, and equipment downtime. S2: Perform state reduction on the production workshop state transition model. Iteratively subdivide the state blocks until convergence to obtain a reduced state set. Subdivision is based on a progressive composite criterion, which includes: when the difference between the predictive resource availability matrices of two states exceeds a preset difference threshold... The process involves separating two states; calculating the weighted Chebyshev distance between the multidimensional transition cost vectors when the two states perform the same operation; separating the two states when the weighted Chebyshev distance is greater than the cost threshold; calculating the Barthel distance between the probability distribution vectors of the two states transitioning to each subsequent state block; separating the two states when the Barthel distance is greater than the probability threshold; where the weight vector corresponding to the weighted Chebyshev distance is generated by the predictive resource availability matrix, and when the predicted future availability probability of a certain resource decreases, the corresponding cost dimension weight is increased, and the cost threshold and probability threshold are dynamically adjusted inversely according to the weight vector; S3: Constructing an aggregated state model based on the state reduction set, using an improved A Heuristic search algorithms are used to generate production scheduling instruction strategies.

2. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The status is composed of equipment occupancy, work-in-process turnover, and a predictive resource availability matrix representing the probability of resource availability in multiple future time periods. This includes: acquiring occupancy status data streams of equipment at the bottom of the workshop to represent equipment occupancy; acquiring turnover location status data of work-in-process to represent work-in-process turnover; statistically analyzing the continuous normal operating time and historical failure time points of the bottom-level equipment, and using a Weibull distribution reliability model to calculate the expected fault-free operating probability of the bottom-level equipment in multiple future time periods; and combining the multiple future time periods with the corresponding expected fault-free operating probabilities to construct the predictive resource availability matrix.

3. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The process of constructing a multidimensional transfer cost vector for state transition, which includes processing time, energy consumption, and tool wear, includes: calculating the processing time based on the start and end times of a single processing action; collecting the power consumption data of the processing equipment during processing and deducting the standby basic energy consumption to obtain the net energy consumption value as the energy consumption; acquiring the characteristic value of the spindle vibration signal and calculating the tool wear amount through a preset tool wear evaluation model as the tool wear; and combining the processing time, the energy consumption, and the tool wear to generate the multidimensional transfer cost vector.

4. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The extension of deterministic transitions to probabilistic multiple successor transitions representing processing qualification, rework, and equipment downtime includes: calculating the equipment downtime branch probability based on the unit-time failure rate of the equipment during the duration of a single processing action; calculating the conditional qualification probability and conditional rework probability based on the number of qualified parts and rework parts of similar historical processes, provided that no equipment downtime branch occurs; and calculating the processing qualification branch probability and rework branch probability based on the equipment downtime branch probability, the conditional qualification probability, and the conditional rework probability to generate the probabilistic multiple successor transitions.

5. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, When the difference between the predictive resource availability matrices of the two states is greater than a preset difference threshold, the two states are separated, which includes: extracting the predictive resource availability matrices corresponding to the first state and the second state to be compared; traversing the same row and column coordinate positions of the two predictive resource availability matrices and calculating the absolute value of the difference between the values ​​of the corresponding position elements; summing the absolute values ​​of all coordinate positions to obtain the sum of absolute deviations; when the sum of absolute deviations is greater than the difference threshold, determining that the first state and the second state do not meet the isomorphism condition, and separating the first state and the second state.

6. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The weight vector is generated from the predictive resource availability matrix, including: obtaining the average value of the expected availability probability of each resource in the predictive resource availability matrix over multiple future time periods; when the average value is lower than a preset normal availability probability benchmark value, calculating the probability decrease deviation between the average value and the normal availability probability benchmark value; calculating the corresponding weight increment based on the probability decrease deviation and a preset compensation coefficient; and superimposing the weight increment onto the corresponding cost dimension weight according to a preset mapping relationship and performing normalization processing to generate the weight vector.

7. The intelligent scheduling method for an automotive parts production workshop according to claim 1 or 6, characterized in that, The calculation of the weighted Chebyshev distance between the multidimensional transfer cost vectors when two states perform the same operation includes: obtaining the first multidimensional transfer cost vector and the second multidimensional transfer cost vector corresponding to the two states when performing the same operation; calculating the absolute values ​​of the numerical differences between the first multidimensional transfer cost vector and the second multidimensional transfer cost vector in each dimension of processing time, energy consumption, and tool wear, and performing dimensionless processing; multiplying the absolute values ​​of each dimension after dimensionless processing by the weight values ​​of the corresponding dimensions in the weight vector to obtain the weighted absolute difference of each dimension; and extracting the maximum value among the weighted absolute differences of all dimensions as the weighted Chebyshev distance.

8. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The method of constructing an aggregated state model based on a state reduction set includes: mapping each state block in the state reduction set to a macro-state node in the aggregated state model; for any first macro-state node and second macro-state node, when at least one original state in the first macro-state node transitions to a certain original state in the second macro-state node after performing a specific operation, establishing a directed transition edge from the first macro-state node to the second macro-state node; and attaching the corresponding operation, the mean attribute of the multidimensional transition cost vector, and the transition probability distribution of the probabilistic multiple successor transitions to the directed transition edge to construct the aggregated state model.

9. The intelligent scheduling method for an automotive parts production workshop according to claim 1, characterized in that, The improved A is adopted A heuristic search algorithm for generating production scheduling instructions includes: in the aggregated state model, setting the macro-state node where the current initial state is located as the search starting point, and setting the macro-state node where all production tasks are completed as the target node; constructing a heuristic evaluation function by minimizing the weighted comprehensive cost of processing time, energy consumption, and tool wear after dimensionless processing to reach the target node; and employing an improved A... The heuristic search algorithm starts from the search starting point, calculates the estimated cost of candidate paths according to the heuristic evaluation function, and prioritizes expanding the macro-state nodes with the smallest estimated cost until the target node is reached to obtain the main production scheduling path; based on the probabilistic multiple successor transitions on the main production scheduling path, including processing qualified, rework, and equipment shutdown branches, inserts conditional execution actions to generate a production scheduling instruction strategy containing conditional branches.

10. An intelligent production scheduling system for an automotive parts manufacturing workshop, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement an intelligent scheduling method for an automotive parts production workshop according to any one of claims 1-9.

Images