Production system holographic perception and task recommendation method, system, equipment and medium

Abstract

The invention discloses a production system holographic perception and task recommendation method, system and device and a medium, and the method comprises the steps: obtaining task attributes, resource states, personnel capabilities, task dependency relationships and historical execution statistical information, and generating a high-dimensional state vector; extracting low-order features of task-resource-personnel interaction by adopting a factorization machine, and extracting high-order nonlinear features in combination with a deep neural network to obtain fusion feature representation; inputting the fusion features into a deep reinforcement learning agent, performing strategy optimization by using a multi-target composite reward function, and outputting a plurality of candidate task scheduling schemes; and receiving feedback data generated by selection operation and adjustment operation of an operator, training the deep reinforcement learning agent based on the feedback data, and determining a final task scheduling scheme according to the selection operation and the adjustment operation of the operator. According to the method, the real-time performance, adaptability, resource balance and personnel satisfaction of scheduling are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power metering and verification technology, and in particular to a method, system, equipment and medium for holographic perception and task recommendation in production systems. Background Technology

[0002] Currently, the scheduling and management of power metering and verification equipment largely relies on manual experience for task allocation, a fixed set of rules for matching, or simply making decisions based on historical data. This traditional manual scheduling method often involves dispatchers assigning tasks based on their intuition and experience, with very limited integration of diverse information such as the current status of the equipment, the technical level of the personnel, and the urgency of the task. When faced with a large number of tasks, diverse types, and particularly complex on-site conditions, it becomes difficult to develop a scientific and efficient allocation plan. Manual scheduling is not only inefficient but also frequently problematic—for example, some tasks are left unattended, some resources are idle, and others are too busy to handle, leading to frequent task delays. Moreover, it cannot quickly respond to suddenly emerging urgent tasks or automatically adapt to changes in the actual situation on-site.

[0003] Later, as informatization and intelligentization levels gradually improved, some systems began to use auxiliary scheduling tools based on rule engines or simple optimization algorithms. However, these methods usually still rely on pre-written static rules or simply summarize past data roughly, making it difficult to truly understand the complex, multi-layered, and dynamically changing relationships between tasks, equipment, and personnel. Their ability to respond to emergencies is clearly inadequate; when task requirements are constantly changing, resources are scarce, or unexpected situations arise, it is difficult to guarantee that the resulting scheduling plan is both optimal and practical. Although some research has attempted to use intelligent optimization methods such as genetic algorithms and ant colony algorithms for scheduling modeling, most of these algorithms focus only on one objective, such as minimizing total working time, while neglecting the balance of multiple objectives such as efficiency, fairness, and personalization. They also rarely consider how humans and machines can collaborate, and they fail to establish a feedback loop for continuous optimization.

[0004] In recent years, machine learning technology has developed rapidly, and some people have begun to try applying methods such as collaborative filtering and matrix factorization to operational task recommendations. However, the problem is that these methods, originally designed for e-commerce recommendations and video interest-based scenarios, are somewhat unsuitable for industrial environments—they struggle to handle hard constraints such as resource limits, task execution in sequence, and the need for real-time feedback. Furthermore, many current scheduling systems operate as "black boxes," with little interaction with frontline staff, making it impossible to incorporate experienced workers' practical experience and new problems encountered on-site to update the model. As a result, the system's intelligence level remains low, its adaptive capabilities are weak, and it lacks room for continuous optimization. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for holographic perception and task recommendation in production systems to address the problems of existing power metering and verification task scheduling methods that rely excessively on manual experience or static rules, lack deep integration and dynamic modeling capabilities for multi-source heterogeneous information, and are difficult to achieve multi-objective collaboration, personalized recommendation, and human-machine closed-loop feedback, resulting in low scheduling efficiency, poor adaptability, and insufficient intelligence in complex and dynamic scenarios.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for holographic perception and task recommendation in a production system, comprising: Obtain task attributes, resource status, personnel capabilities, task dependencies, and historical execution statistics to generate a high-dimensional state vector representing the current production environment; Based on the state vector, a factorization machine is used to extract low-order features of task-resource-person interaction, and a deep neural network is combined to extract high-order nonlinear features to obtain a fused feature representation. The fused features are input into a deep reinforcement learning agent, and a multi-objective composite reward function is used to optimize the policy, outputting multiple candidate task scheduling schemes. The candidate task scheduling scheme is pushed to the operator, feedback data generated by the operator's selection and adjustment operations is received, the deep reinforcement learning agent is trained based on the feedback data, and the final task scheduling scheme is determined according to the operator's selection and adjustment operations.

[0008] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the step of extracting low-order features of task-resource-person interaction using a factorization machine and combining it with a deep neural network to extract high-order nonlinear features includes: One-hot encoding is performed on the discrete variables of task attributes, resource status, and personnel capabilities; The vector after one-hot encoding is input into the factorization machine to calculate the second-order feature interaction term and obtain the low-order interaction feature. The one-hot encoded vector is embedded and mapped to obtain a dense feature vector, and the dense feature vector is input into a multi-layer deep neural network. High-order combined features are extracted layer by layer through a nonlinear activation function. The output of the factorization machine is added to the output of the deep neural network, and then mapped by the activation function to obtain the fused feature representation.

[0009] The beneficial effects of this preferred technical solution are that by combining factor decomposition machine and deep neural network to jointly model the multi-order interaction features of tasks, resources and personnel, it effectively takes into account both the explicit low-order combination of sparse discrete features and the implicit expression of complex nonlinear high-order relationships, and significantly improves the comprehensiveness of production scheduling status representation and the accuracy of model recommendation.

[0010] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the multi-objective composite reward function is composed of a weighted combination of priority reward items, resource efficiency reward items, dependency constraint reward items, and subjective feedback reward items. The value of the priority reward item is determined based on the priority category of the current task and the deviation between the actual completion time and the expected completion time; The value of the resource efficiency reward item is calculated based on the utilization rate of equipment resources and the load balance of personnel occupied by the scheduling scheme. The value of the dependency constraint reward item is assigned according to whether the task execution order violates the preset task dependency relationship. A negative reward is given when it violates the relationship, and a positive reward or zero reward is given when it conforms to the relationship. The value of the subjective feedback reward item is generated based on the scoring operation and click selection behavior performed by the operator on the pushed candidate task scheduling scheme.

[0011] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the deep reinforcement learning agent adopts a soft actor-critic architecture, including: Based on the current fusion feature representation, the policy network acts as the actor to generate the probability distribution of task scheduling actions, and samples the specific task scheduling actions from the probability distribution. The current fused feature representation is combined with the specific task scheduling action to form a state-action pair. Two Q-value networks are input as commentators to calculate the long-term reward estimate corresponding to the state-action pair. Based on the long-term reward estimate and the probability distribution of task scheduling actions, a policy optimization objective function containing an entropy regularization term is constructed, and the parameters of the policy network are updated through the gradient ascent method. During the process of updating the policy network and the Q-value network, a target network is introduced to make a stable estimate of the state-action value at the next time step, and the gradient of the network parameters is pruned. Periodically perform soft updates on the corresponding target network parameters based on the main network parameters of the policy network and the Q-value network.

[0012] The beneficial effects of this preferred technical solution are that by adopting a soft actor-critic architecture, and through maximum entropy policy optimization, double Q network estimation, target network soft update and gradient pruning mechanism, the exploration capability, training stability and long-term decision performance of the scheduling policy are effectively improved.

[0013] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the step of training the deep reinforcement learning agent based on feedback data includes: Store the quadruple consisting of state, action, reward and next state into the experience replay pool; A batch of experience samples is randomly sampled from the experience replay pool to form the training batch; The parameters of the policy network and the Q-value network are updated using the training batches; Periodically synchronize the main network parameters to the target network.

[0014] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the step of pushing candidate task scheduling schemes to operators includes: Display three or more candidate task scheduling schemes side by side on the same interface; Each solution is labeled with task priority, required resource types, estimated completion time, and status of dependent tasks; Operators are allowed to perform task replacement, sequence adjustment, and resource reallocation operations on any candidate solution; Record the scheduling plan finally confirmed by the operators as valid feedback data.

[0015] As a preferred embodiment of the holographic perception and task recommendation method for the production system described in this invention, the high-dimensional state vector is composed of a basic label sub-vector, a task dependency sub-vector, a resource status sub-vector, a system load sub-vector, and a historical statistics sub-vector concatenated together. The basic tag subvector includes task priority, task type, and estimated completion time; The task-dependent subvectors include pre- and post-order constraints and parallel restrictions between tasks; The resource status subvector includes the availability of the verification device and the health status of the equipment. The system load subvector includes the current resource utilization rate and personnel allocation status; The historical statistical subvectors include the historical average time taken, resource consumption, and success rate of similar tasks.

[0016] Secondly, the present invention provides a holographic perception and task recommendation system for a production system, comprising: The holographic state perception module is used to acquire task attributes, resource status, personnel capabilities, task dependencies, and historical execution statistics to generate a high-dimensional state vector representing the current production environment. The multi-order feature fusion module is used to extract low-order features of task-resource-person interaction based on the state vector using a factorization machine, and to extract high-order nonlinear features by combining a deep neural network to obtain a fused feature representation. The enhanced decision recommendation module is used to input the fused features into the deep reinforcement learning agent, optimize the policy using a multi-objective composite reward function, and output multiple candidate task scheduling schemes. The human-machine collaborative feedback module is used to push the candidate task scheduling scheme to the operator, receive feedback data generated by the operator's selection and adjustment operations, train the deep reinforcement learning agent based on the feedback data, and determine the final task scheduling scheme according to the operator's selection and adjustment operations.

[0017] Thirdly, the present invention provides an electronic device, comprising: Memory, used to store programs; A processor is configured to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the holographic perception and task recommendation method for the production system.

[0018] Fourthly, the present invention provides a computer-readable storage medium, comprising: when the program is executed by a processor, the steps of implementing the production system holographic perception and task recommendation method.

[0019] The beneficial effects of this invention are as follows: This invention constructs a high-dimensional holographic state vector that integrates task attributes, resource status, personnel capabilities, task dependencies, and historical execution statistics. It employs a factorization machine and a deep neural network to jointly model low-order explicit interactions and high-order nonlinear features, achieving accurate perception of complex production environments and deep fusion of multi-dimensional features. By introducing a soft actor-critic deep reinforcement learning agent based on a multi-objective composite reward function, it integrates task priority, resource efficiency, dependency compliance, and operator subjective feedback into the strategy optimization process, achieving adaptive decision-making and multi-objective collaborative optimization of scheduling schemes under dynamic constraints. Furthermore, by designing a closed-loop mechanism for multi-candidate scheme recommendation, manual adjustment interaction, and feedback, it transforms operator selection and modification behaviors into effective training signals for continuous model iteration, realizing personalized recommendations and online evolution of system intelligence under human-machine collaboration. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a basic flowchart illustrating a holographic perception and task recommendation method for a production system, provided as an embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for holographic perception and task recommendation in a production system is provided, comprising: S100: Obtain task attributes, resource status, personnel capabilities, task dependencies, and historical execution statistics to generate a high-dimensional state vector representing the current production environment; S200: Based on the state vector, a factorization machine is used to extract low-order features of task-resource-person interaction, and a deep neural network is combined to extract high-order nonlinear features to obtain a fused feature representation; S300: Input the fused features into the deep reinforcement learning agent, optimize the policy using a multi-objective composite reward function, and output multiple candidate task scheduling schemes; S400: Push the candidate task scheduling scheme to the operator, receive feedback data generated by the operator's selection and adjustment operations, train the deep reinforcement learning agent based on the feedback data, and determine the final task scheduling scheme according to the operator's selection and adjustment operations.

[0023] It should be noted that existing power metering verification task scheduling methods face a series of challenges during operation, including: scheduling decisions relying heavily on human experience or static rules, making it difficult to dynamically respond to real-time changes such as sudden equipment status changes, temporary staff absences, or task priority adjustments, leading to scheduling scheme delays or even failures; a lack of modeling capabilities for the complex interactive relationships between tasks, resources, and personnel, failing to effectively capture implicit correlations under high-dimensional sparse features, resulting in uneven resource allocation or delays in critical tasks; existing systems generally lack human-machine collaboration mechanisms, making it difficult for operators to conveniently intervene or optimize recommendation results, and their professional judgments are difficult to feed back into the model to form a closed-loop learning, resulting in rigid recommendation strategies and poor adaptability; traditional methods usually only focus on a single optimization objective (such as the shortest construction period or the lowest cost), ignoring multi-dimensional business demands such as task dependency constraints, resource balance, and personnel satisfaction, making it difficult to meet the comprehensive scheduling needs of efficient, reliable, and flexible power metering verification scenarios.

[0024] Therefore, addressing the aforementioned issues of existing power metering verification task scheduling methods over-reliance on manual experience or static rules, lack of deep fusion and dynamic modeling capabilities for multi-source heterogeneous information, difficulty in achieving multi-objective collaboration, personalized recommendation, and human-machine closed-loop feedback, resulting in low scheduling efficiency, poor adaptability, and insufficient intelligence in complex and dynamic scenarios, this paper addresses these problems through steps S100-S400. By constructing a holographic perception state representation, a deep recommendation model integrating low-order and high-order features, a soft actor-critic reinforcement learning mechanism based on multi-objective rewards, and a human-machine collaborative feedback closed loop, this paper achieves accurate perception, intelligent decision-making, multi-objective optimization, and continuous self-evolution of power metering verification task scheduling in dynamic environments.

[0025] Example 2, this is an embodiment of the present invention, which provides a holographic perception and task recommendation method for a production system based on the previous embodiment, including: In this embodiment of the application, the high-dimensional state vector in step S100 is composed of a concatenation of a basic label subvector, a task dependency subvector, a resource status subvector, a system load subvector, and a historical statistics subvector; mathematically represented as: in, Based on the sub-vector of the label, For task-dependent sub-vectors, For resource status subvectors, For the system load subvector, This is a historical statistics subvector. The basic label subvector includes task priority, task type, and estimated completion time; the task dependency subvector includes pre- and post-task constraints and parallel limitations; the resource status subvector includes the availability of the calibration device and the health status of the equipment; the system load subvector includes the current resource utilization rate and personnel allocation status; and the historical statistics subvector includes the historical average time, resource consumption, and success rate of similar tasks.

[0026] It should be noted that, given that one-hot encoding is prone to feature dimension expansion and sparsity problems, which in turn lead to the curse of dimensionality in neural networks, this invention further introduces an embedding layer to perform low-dimensional dense mapping on discrete features, so as to compress the feature space and retain semantic information.

[0027] In this embodiment of the application, the construction method of the high-dimensional state vector in step S100 includes obtaining task attributes, resource status, personnel capabilities, task dependencies and historical execution statistics, which are respectively encoded into five types of sub-vectors: basic labels, task dependencies, resource status, system load and historical statistics. The discrete features are mapped into dense vectors through an embedding layer after one-hot encoding. Finally, the sub-vectors are concatenated to form a high-dimensional state vector representing the current production environment.

[0028] In an optional implementation, the high-dimensional state vector in step S100 can also be constructed by using tasks, resources and personnel as nodes, task dependencies, resource allocation and historical interaction relationships as edges, constructing a heterogeneous scheduling graph, a graph neural network (such as GAT or GCN) to aggregate and update node features, and a global readout function to generate a graph-level embedding as the current state representation.

[0029] In an optional implementation, the construction method of the high-dimensional state vector in step S100 can also be to serialize the historical interaction records between the operator and the system (including tasks, resources, timestamps, etc.) into an input sequence with time location encoding, extract context-aware temporal features through the Transformer encoder, and use the output of the final hidden state or [CLS] tag as the current high-dimensional state vector.

[0030] In this embodiment of the application, step S200 uses a factorization machine to extract low-order features of task-resource-person interaction and combines it with a deep neural network to extract high-order nonlinear features, including: One-hot encoding is performed on the discrete variables of task attributes, resource status, and personnel capabilities; The vector after one-hot encoding is input into the factorization machine to calculate the second-order feature interaction term and obtain the low-order interaction feature. The vector after one-hot encoding is embedded and mapped to obtain a dense feature vector, and the dense feature vector is input into a multi-layer deep neural network. High-order combined features are extracted layer by layer through a nonlinear activation function. The output of the factorization machine is added to the output of the deep neural network, and then mapped by the activation function to obtain the fused feature representation.

[0031] In this embodiment of the application, the low-order and high-order feature fusion method in step S200 includes feeding the one-hot encoded input features into a factorization machine (FM) to extract second-order interactive features and a deep neural network (DNN) to extract high-order nonlinear features, and then adding the outputs of the two and mapping them to the final fused feature representation through the sigmoid function.

[0032] In an optional implementation, the low-order and high-order feature fusion method in step S200 can also generate dense vectors for discrete features through a shared embedding layer. The embedding is simultaneously input into the factorization machine (FM) branch to calculate second-order cross features and the deep neural network (DNN) branch to extract high-order nonlinear features. The two outputs are then concatenated and fused through a fully connected layer to generate the final scheduling scheme score.

[0033] In an optional implementation, the low-order and high-order feature fusion method in step S200 can also input the embedded feature vector sequence into a multi-head self-attention layer, generate a context-aware cross-feature representation by calculating the dynamic interaction weights between features, and then add or concatenate it with the high-order features extracted by the deep neural network to form a fused feature for scheduling score prediction.

[0034] In this embodiment, the forward propagation process of the deep neural network is as follows: , in, For the first Embedding vectors of each feature For the first Hidden unit output of layer, and These are the weight matrix and the bias term, respectively. It is a non-linear activation function; the final DNN output is a scalar. .

[0035] In this embodiment of the application, the predicted output p of the fused feature representation is given by the following formula: in, This represents the predicted comprehensive score of the scheduling scheme.

[0036] In this embodiment of the application, the output of the factorization machine Calculated by the following formula: in, The input feature vector after one-hot encoding. For first-order feature weights, For the first The latent vector of each feature For the latent vector dimension, This indicates the inner product operation.

[0037] By algebraically reconstructing the second-order cross terms, the computational complexity can be reduced from... Reduce to Specifically: This equation enables the factorization machine to efficiently compute feature interactions in high-dimensional sparse scenarios, significantly improving the speed of model training and inference, and meeting the real-time requirements of power metering verification task scheduling.

[0038] In this embodiment, the task scheduling recommendation process is modeled as a Markov decision process, where the state is the current fused feature representation, the action is the set of candidate scheduling schemes, the transition probability is driven by operator feedback, and the reward function consists of multi-objective composite rewards.

[0039] In this embodiment, the multi-objective composite reward function in step S300 is composed of a weighted combination of priority reward items, resource efficiency reward items, dependency constraint reward items, and subjective feedback reward items. The value of the priority reward item is determined based on the priority category of the current task and the deviation between the actual completion time and the expected completion time. The value of the resource efficiency reward item is calculated based on the utilization rate of equipment resources and the load balancing of personnel in the scheduling scheme. The value of the dependency constraint reward item is assigned based on whether the task execution order violates the preset task dependency relationship; a negative reward is given if it violates the relationship, and a positive reward or zero reward is given if it conforms. The value of the subjective feedback reward item is generated based on the operator's rating operation and click selection behavior on the pushed candidate task scheduling scheme.

[0040] In this embodiment of the application, the reinforcement learning policy optimization objective in step S300 includes using the soft actor-critic (SAC) algorithm to construct an optimization objective function that includes a multi-objective composite reward and a policy entropy regularization term, and co-training the policy network and the double Q network to maximize the weighted sum of expected cumulative reward and exploration diversity.

[0041] In an optional implementation, the reinforcement learning policy optimization objective in step S300 can also be based on calculating the advantage estimate using a multi-objective composite reward function, generating the probability distribution of scheduling actions using a policy network, and updating the policy using a PPO objective function with a pruning mechanism, so as to optimize task scheduling performance while ensuring training stability.

[0042] In an optional implementation, the reinforcement learning policy optimization objective in step S300 can also utilize historical scheduling decision data of operators to train a behavior clone model as an initial policy, and then input its output action and state into a deep Q-network (DQN), combined with multi-objective composite rewards for online fine-tuning, and gradually optimize the scheduling policy through an ε-greedy exploration mechanism.

[0043] In the embodiments of this application, a multi-objective composite reward function Defined as: in, As a priority reward item, For resource efficiency rewards, For dependency-constrained reward items, For subjective feedback reward items, , , , These are the corresponding weighting coefficients.

[0044] In this embodiment, the task scheduling recommendation process is modeled as a Markov decision process, consisting of a quintuple. express: state space Current state in Defined as the operator's most recent The vector formed by the historical records of interactions with the system, i.e. Each of them Characteristic representation of a scheduling scheme; Action space Given a set of all feasible candidate task scheduling schemes, the agent at time... Output action This is a recommendation list containing multiple candidate solutions; State transition probability Driven by operator feedback on the recommended list, specifically, the system selects the option based on the operator's final confirmation or preference. Update the state and get the next state. ; reward function The calculation is based on a multi-objective composite reward function, reflecting the overall performance of the scheduling scheme in terms of priority achievement, resource efficiency, dependency compliance, and user satisfaction. Discount factor [0,1] is used to weigh the short-term versus long-term scheduling benefits.

[0045] The goal of a deep reinforcement learning agent is to learn an optimal policy. This maximizes the expected cumulative reward obtained from task scheduling under this strategy, i.e.: in, It is a multi-objective composite reward function that comprehensively reflects the business performance and user satisfaction of the scheduling scheme.

[0046] To further enhance the policy's exploration capability and training stability, this invention employs a soft actor-commentator (SAC) architecture, introducing a policy entropy regularization term into the aforementioned objective. Therefore, the policy optimization objective function is expressed as: in, For policy entropy, This is a temperature coefficient used to balance cumulative rewards and exploration diversity.

[0047] In this embodiment of the application, the deep reinforcement learning agent in step S300 adopts a soft actor-critic architecture, including: Based on the current fusion feature representation, the policy network acts as the actor to generate the probability distribution of task scheduling actions, and samples the specific task scheduling actions from the probability distribution. The current fused feature representation is combined with the specific task scheduling action to form a state-action pair. Two Q-value networks are input as commentators to calculate the long-term reward estimate corresponding to the state-action pair. Based on the long-term reward estimate and the probability distribution of task scheduling actions, a policy optimization objective function containing an entropy regularization term is constructed, and the parameters of the policy network are updated through the gradient ascent method. During the process of updating the policy network and the Q-value network, a target network is introduced to make a stable estimate of the state-action value at the next time step, and the gradient of the network parameters is pruned. Periodically perform soft updates on the corresponding target network parameters based on the main network parameters of the policy network and the Q-value network.

[0048] In this embodiment, the gradient update direction of the policy network is determined by the following formula: in, The temporal difference error or advantage estimate based on the output of the Critic network is used to measure the action. In relation to the merits and demerits of the current strategy, For policy network parameters; in the soft actor-commentator architecture, Further integrate the policy entropy term to encourage exploration and improve training stability.

[0049] In this embodiment of the application, policy entropy Defined as the expected value of the negative log probability under the policy distribution, i.e.: In this embodiment of the application, step S400, which trains the deep reinforcement learning agent based on feedback data, includes: Store the quadruple consisting of state, action, reward and next state into the experience replay pool; A batch of experience samples is randomly sampled from the experience replay pool to form the training batch; The parameters of the policy network and the Q-value network are updated using the training batches; Periodically synchronize the main network parameters to the target network.

[0050] In this embodiment of the application, the utilization of human-machine feedback data in step S400 includes taking the operator's selection or adjustment result of the candidate scheduling scheme as the actual execution action, combining the current state, multi-objective composite reward and the next state to form an experience quadruple and storing it in the replay pool, and using it to train the soft actor-critic (SAC) reinforcement learning agent online to achieve continuous optimization of the strategy.

[0051] In an optional implementation, the human-machine feedback data in step S400 can also be used to record the operator's relative preferences (such as ranking or rating) for multiple candidate scheduling schemes, infer the implicit reward function through a preference learning model (such as Bradley-Terry or RankNet), and then use this reward signal to train the policy network.

[0052] In an optional implementation, the human-machine feedback data utilization method in step S400 can also be to systematically evaluate the uncertainty of the current scheduling state (such as policy entropy or Q-value variance) in real time, and actively request operator intervention only when there is high uncertainty, and prioritize the use of the obtained high-quality feedback for local updates or key training of the policy network.

[0053] In this embodiment, the loss function of the Q-value network is expressed in temporal difference form as follows: In this embodiment, the parameter update of the Q-value network is based on the gradient of the temporal difference loss function, the gradient expression of which is: in, Main Q network, For the target Q network, Its parameters, For instant rewards, This is the discount factor; in actual updates, this gradient is pruned to constrain the magnitude of parameter changes, and the network parameters are iteratively optimized using gradient descent.

[0054] In this embodiment of the application, step S400, which involves pushing candidate task scheduling schemes to the operator, includes: Display three or more candidate task scheduling schemes side by side on the same interface; Each solution is labeled with task priority, required resource types, estimated completion time, and status of dependent tasks; Operators are allowed to perform task replacement, sequence adjustment, and resource reallocation operations on any candidate solution; Record the scheduling plan finally confirmed by the operators as valid feedback data.

[0055] In this embodiment, the reward in the quadruple includes a weighted combination of task priority achievement, resource utilization efficiency, dependency constraint compliance, and operator subjective evaluation. Both the current state and the next state are represented by the aforementioned high-dimensional state vector.

[0056] In this embodiment, the feedback data is not only used for online policy updates, but also for building offline training datasets to support continuous learning of the model without environmental interaction. At the same time, through the policy transfer mechanism, the scheduling knowledge learned in historical scenarios can be transferred to the new task environment to improve the recommendation performance in the cold start phase.

[0057] Example 3 is an embodiment of the present invention. This embodiment differs from the first embodiment in that it provides a holographic perception and task recommendation system for a production system.

[0058] It should be noted that the technical solution of the production system holographic perception and task recommendation system is based on the same concept as the technical solution of the production system holographic perception and task recommendation method described above. For details not described in detail in the technical solution of the production system holographic perception and task recommendation system in this embodiment, please refer to the description of the technical solution of the production system holographic perception and task recommendation method described above.

[0059] This embodiment describes a holographic perception and task recommendation system for a production system, comprising: The holographic state perception module is used to acquire task attributes, resource status, personnel capabilities, task dependencies, and historical execution statistics to generate a high-dimensional state vector representing the current production environment. The multi-order feature fusion module is used to extract low-order features of task-resource-person interaction based on the state vector using a factorization machine, and to extract high-order nonlinear features by combining a deep neural network to obtain a fused feature representation. The enhanced decision recommendation module is used to input the fused features into the deep reinforcement learning agent, optimize the policy using a multi-objective composite reward function, and output multiple candidate task scheduling schemes. The human-machine collaborative feedback module is used to push the candidate task scheduling scheme to the operator, receive feedback data generated by the operator's selection and adjustment operations, train the deep reinforcement learning agent based on the feedback data, and determine the final task scheduling scheme according to the operator's selection and adjustment operations.

[0060] This embodiment also provides an electronic device applicable to a holographic perception and task recommendation method for a production system, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a holographic perception and task recommendation method for a production system, as proposed in the above embodiments.

[0061] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a method for holographic perception and task recommendation of a production system as proposed in the above embodiments.

[0062] The storage medium proposed in this embodiment and the method for implementing holographic perception and task recommendation in a production system proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0063] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0064] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A production system holographic awareness and task recommendation method, characterized by, The method comprises the following steps: acquiring task attributes, resource states, personnel capabilities, task dependency relationships and historical execution statistical information to generate a high-dimensional state vector representing the current production environment; based on the state vector, using a factorization machine to extract low-order features of task-resource-personnel interaction, and combining a deep neural network to extract high-order nonlinear features to obtain a fusion feature representation; inputting the fusion feature into a deep reinforcement learning agent, using a multi-objective compound reward function to optimize the strategy, and outputting multiple candidate task scheduling schemes; pushing the candidate task scheduling schemes to the operator, receiving feedback data generated by the selection operation and adjustment operation of the operator, training the deep reinforcement learning agent based on the feedback data, and determining the final task scheduling scheme according to the selection operation and adjustment operation of the operator.

2. The production system holographic awareness and task recommendation method of claim 1, wherein: The method comprises the following steps: one-hot encoding discrete variables of task attributes, resource states and personnel capabilities; inputting the one-hot encoded vector into the factorization machine to calculate the second-order feature cross terms and obtain the low-order interaction features; embedding and mapping the one-hot encoded vector to obtain a dense feature vector, and inputting the dense feature vector into a multi-layer deep neural network to extract high-order combined features through a nonlinear activation function layer by layer; adding the factorization machine output and the deep neural network output, and mapping the result through an activation function to obtain the fusion feature representation.

3. The production system holographic awareness and task recommendation method of claim 1 or 2, wherein: The multi-objective compound reward function is composed of a priority reward item, a resource efficiency reward item, a dependency constraint reward item and a subjective feedback reward item; The value of the priority reward item is determined according to the priority category of the current task and the deviation between the actual completion time and the expected completion time; The value of the resource efficiency reward item is calculated according to the device resource utilization rate and the personnel load balancing degree occupied by the scheduling scheme; The value of the dependency constraint reward item is assigned according to whether the task execution order violates the preset task dependency relationship, and negative reward is given when the task execution order violates the preset task dependency relationship, positive reward and zero reward are given when the task execution order meets the preset task dependency relationship; The value of the subjective feedback reward item is generated according to the scoring operation and click selection behavior of the operator on the pushed candidate task scheduling scheme.

4. The production system holographic awareness and task recommendation method of claim 3, wherein: The deep reinforcement learning agent adopts a soft actor-critic architecture, which comprises the following steps: based on the current fusion feature representation, a policy network is used as an actor to generate a probability distribution of task scheduling actions, and a specific task scheduling action is sampled from the probability distribution; combining the current fusion feature representation and the specific task scheduling action into a state-action pair, inputting the state-action pair into two Q-value networks as critics to calculate the long-term return estimate value corresponding to the state-action pair; based on the long-term return estimate value and the probability distribution of the task scheduling action, a policy optimization objective function containing an entropy regularization term is constructed, and the parameters of the policy network are updated through a gradient ascent method; In the process of updating the policy network and the Q-value network, a target network is introduced to stably estimate the state-action value at the next moment, and the gradient of the network parameters is clipped. Periodically, the main network parameters of the policy network and the Q-value network are updated to the target network parameters.

5. The production system holographic awareness and task recommendation method of claim 4, wherein: The training of the deep reinforcement learning agent based on the feedback data comprises: storing a four-tuple of state, action, reward and next state into an experience replay pool; randomly sampling a batch of experience samples from the experience replay pool to form a training batch; updating the parameters of the policy network and the Q-value network using the training batch; periodically synchronizing the main network parameters to the target network.

6. The production system holographic awareness and task recommendation method of claim 5, wherein: The pushing of the candidate task scheduling schemes to the operator comprises: displaying more than three candidate task scheduling schemes in parallel on the same interface; annotating each scheme with task priority, required resource type, estimated completion time and dependent task status; allowing the operator to perform task replacement, sequence adjustment and resource reallocation operations on any candidate scheme; recording the scheduling scheme finally confirmed by the operator as valid feedback data.

7. The production system holographic awareness and task recommendation method of claim 6, wherein: The high-dimensional state vector is composed of a basic label sub-vector, a task dependency sub-vector, a resource status sub-vector, a system load sub-vector and a historical statistics sub-vector; The basic label sub-vector includes task priority, task type and estimated completion time; The task dependency sub-vector includes precedence constraints and parallel restrictions between tasks; The resource status sub-vector includes availability of detection devices and health status of equipment; The system load sub-vector includes current resource utilization and personnel allocation status; The historical statistics sub-vector includes historical average time consumption, resource consumption and completion success rate of similar tasks.

8. A production system holographic awareness and task recommendation system applying the method of any of claims 1-7, characterized in that, Comprise: a holographic state perception module for acquiring task attributes, resource status, personnel capabilities, task dependency relationships and historical execution statistics information, and generating a high-dimensional state vector representing the current production environment; a multi-order feature fusion module for extracting low-order features of task-resource-personnel interaction based on the state vector using a factorization machine, and combining deep neural networks to extract high-order nonlinear features to obtain a fusion feature representation; a reinforcement decision recommendation module for inputting the fusion features into a deep reinforcement learning agent, using a multi-objective compound reward function to optimize the policy, and outputting multiple candidate task scheduling schemes; a man-machine collaborative feedback module for pushing the candidate task scheduling schemes to the operator, receiving feedback data generated by the selection and adjustment operations of the operator, training the deep reinforcement learning agent based on the feedback data, and determining the final task scheduling scheme according to the selection and adjustment operations of the operator.

9. An electronic device, comprising: Comprise: a memory for storing a program; a processor for loading the program to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a program, characterized in that, The program is executed by the processor to implement the steps of the method according to any one of claims 1-7. The program is executed by the processor to implement the steps of the method according to any one of claims 1-7.

Images