Distributed industrial task collaborative scheduling method and system based on swarm intelligence

Abstract

The distributed industrial task collaborative scheduling method and system based on swarm intelligence provided by the embodiment of the present invention comprises: obtaining the target intelligent device cluster operation log for industrial production line collaborative scheduling analysis, the target intelligent device cluster operation log including the target distributed task status data; entering the target intelligent device cluster operation log into the distributed collaborative scheduling decision network that has completed debugging, and outputting the feasibility score of the target intelligent device cluster operation log matching the target scheduling strategy mode, the distributed collaborative scheduling decision network is optimized according to the linkage training cost algorithm. The method can adaptively match the target scheduling strategy mode and output the feasibility score, which enhances the flexibility and accuracy of the scheduling process. When the production line needs to be expanded or reduced, it is only necessary to adjust the decision network, showing good scalability.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a distributed industrial task collaborative scheduling method and system based on swarm intelligence. Background Art

[0002] As the complexity and scale of industrial production continue to increase, how to effectively schedule production lines to improve production efficiency, reduce costs and meet changing market demands has become an important challenge. Traditional scheduling methods often rely on manual experience and static planning strategies, lack flexibility and adaptability, and are difficult to cope with various emergencies in the production process.

[0003] In addition, with the in-depth promotion of the concept of Industry 4.0 and the implementation of new production models such as smart manufacturing and smart factories, the accuracy and real-time requirements for production line scheduling are becoming increasingly higher. Traditional scheduling methods often cannot meet these requirements, so new scheduling methods need to be developed. Summary of the invention

[0004] In order to improve the above problems, the present invention provides a distributed industrial task collaborative scheduling method and system based on swarm intelligence.

[0005] In a first aspect, a distributed industrial task collaborative scheduling method based on swarm intelligence is provided, which is applied to an AI collaborative scheduling system. The method comprises:

[0006] Obtaining an operation log of a target intelligent device cluster to be subjected to collaborative scheduling analysis of an industrial production line, wherein the operation log of the target intelligent device cluster includes target distributed task status data;

[0007] The target smart device cluster operation log is entered into a debugged distributed collaborative scheduling decision network, and a feasibility score of the target smart device cluster operation log matching the target scheduling strategy mode is output. The distributed collaborative scheduling decision network is optimized according to a linkage training cost algorithm.

[0008] In some possible embodiments, the step of optimizing the distributed collaborative scheduling decision network according to the linkage training cost algorithm includes:

[0009] Acquire a set of network optimization examples, each network optimization example comprising an intelligent device cluster operation log example, an initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example, and prior learning knowledge of the intelligent device cluster operation log example, the intelligent device cluster operation log example comprising a first intelligent device cluster operation log example and a second intelligent device cluster operation log example, the second intelligent device cluster operation log example matching a target scheduling strategy mode;

[0010] For each network optimization example in the network optimization example set, an intelligent device cluster operation log example in the network optimization example and an initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example are used as input information of a distributed collaborative scheduling decision network, and a feasibility score of the intelligent device cluster operation log example matching the target scheduling strategy mode is output; the distributed collaborative scheduling decision network is used to generate an operation log retelling text of the intelligent device cluster operation log example based on the intelligent device cluster operation log example and the initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example, and determine the feasibility score based on the intelligent device cluster operation log example and the operation log retelling text of the intelligent device cluster operation log example;

[0011] Generate a first neural network training cost algorithm based on the running state residual semantic vector of the smart device cluster running log example obtained in the retelling task and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example; generate a second neural network training cost algorithm based on the running log retelling text of the smart device cluster running log example and the smart device cluster running log example; the first neural network training cost algorithm is used to constrain the quantitative difference between the running state residual semantic vector of the first smart device cluster running log example in the network optimization example set and the running state residual semantic vector of the first smart device cluster running log example, and increase the quantitative difference between the running state residual semantic vector of the second smart device cluster running log example in the network optimization example set and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example; the second neural network training cost algorithm is used to constrain the text retelling training cost of the first smart device cluster running log example and increase the text retelling training cost of the second smart device cluster running log example;

[0012] Based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example, the network parameters of the distributed collaborative scheduling decision network are optimized until they meet the network optimization requirements, thereby obtaining a distributed collaborative scheduling decision network that has been debugged.

[0013] In some possible embodiments, the distributed collaborative scheduling decision network includes a first decision branch component and a second decision branch component, the first decision branch component includes a state feature mining core and a state feature processing core, the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example are used as input information of the distributed collaborative scheduling decision network, and the feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode is output, including:

[0014] Taking the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example as input information of the state feature mining core, outputting the operation state residual semantic vector of the smart device cluster operation log example and the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the smart device cluster operation log example;

[0015] Taking the running state residual semantic vector of the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example as the input information of the state feature processing core, outputting the u-th order running state residual semantic vector of the smart device cluster running log example and the u-th order running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and outputting the running log retelling text of the smart device cluster running log example, where u is an integer greater than or equal to 1 and less than or equal to X, and X is the total order of the pyramid reversible subnet of the state feature processing core;

[0016] Through the second decision branch component, based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example, a feasibility score of the smart device cluster operation log example matching the target scheduling strategy pattern is output.

[0017] In some possible embodiments, the first decision branch component further includes a semantic upgrade unit, and the method further includes:

[0018] Enter each smart device cluster operation log example in the network optimization example set into the semantic upgrading unit, and output the operation state residual enhanced semantic vector of the smart device cluster operation log example and the u-th order operation state residual enhanced semantic vector of the smart device cluster operation log example;

[0019] The operating state residual semantic vector of the second smart device cluster operation log example in the smart device cluster operation log example is updated to the operating state residual enhanced semantic vector of the second smart device cluster operation log example, and the u-th order operating state residual semantic vector of the second smart device cluster operation log example is updated to the u-th order operating state residual enhanced semantic vector of the second smart device cluster operation log example.

[0020] In some possible embodiments, the semantic upgrade unit is used to:

[0021] Extracting the target distributed task status highlight text of the smart device cluster operation log example, and adjusting the size of the target distributed task status highlight text of the smart device cluster operation log example according to the text box size of the highlight text of the corresponding order used to generate the first neural network training cost algorithm, to obtain the target highlight text;

[0022] Take any variable Q within the set operation value quadrant. If the Q is greater than the preset value, a first semantic upgrade rule is used to perform semantic upgrade based on the target highlighted text. If the Q is less than or equal to the preset value, a second semantic upgrade rule is used to perform semantic upgrade based on the target highlighted text.

[0023] In some possible embodiments, the second decision branch component includes a scheduling mode identification unit and a scheduling mode matching unit, and the second decision branch component outputs a feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example, including:

[0024] The smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example are entered into the scheduling mode determination unit, and the scheduling mode determination unit determines a first feasibility identification label based on the smart device cluster operation log example, the operation log retelling text of the smart device cluster operation log example, the pyramid state semantic relationship network extracted by the scheduling mode determination unit, the fully connected layer corresponding to the state feature mining core, and the fully connected layer corresponding to the pyramid reversible subnet of the state feature processing core;

[0025] Entering the running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid state semantic relationship network into the scheduling mode matching unit, the scheduling mode matching unit semantically downsamples the pyramid state semantic relationship network to obtain the pyramid downsampled semantic relationship network, and the scheduling mode matching unit determines a second feasibility identification label based on the running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid downsampled semantic relationship network;

[0026] The first feasibility identification tag and the second feasibility identification tag are processed based on weights to obtain a feasibility score for the smart device cluster operation log example matching the target scheduling strategy pattern and output it.

[0027] In some possible embodiments, the first neural network training cost algorithm is generated based on the residual semantic vector of the running state of the smart device cluster running log example obtained in the retelling task and the residual semantic vector of the running state of the initial smart device cluster running log corresponding to the smart device cluster running log example, including:

[0028] Calculate a first semantic quantization difference based on the running state residual semantic vector of the second smart device cluster running log example in the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example;

[0029] Calculating a second semantic quantization difference based on the running state residual semantic vector of the first smart device cluster running log example in the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the first smart device cluster running log example, and the number of the first smart device cluster running log example in the smart device cluster running log example;

[0030] The ratio of the second semantic quantization difference to the first operation value is determined as the first neural network training cost algorithm, and the first operation value is the sum of the first semantic quantization difference and the set network parameter.

[0031] In some possible embodiments, the generating a second neural network training cost algorithm based on the operation log retelling text of the smart device cluster operation log example and the smart device cluster operation log example includes:

[0032] If the smart device cluster operation log example is the first smart device cluster operation log example, a cross entropy cost between the first smart device cluster operation log example and the operation log retelling text of the first smart device cluster operation log example is determined as a second neural network training cost algorithm;

[0033] If the smart device cluster operation log example is a second smart device cluster operation log example, a second neural network training cost algorithm is generated based on the operation state residual semantic vector of the second smart device cluster operation log example, the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the operation log retelling text of the second smart device cluster operation log example, the highlighted text of the state feature mining core, and the highlighted text of the pyramid reversible subnet of the state feature processing core.

[0034] In some possible embodiments, the optimizing of the network parameters of the distributed collaborative scheduling decision network based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example, and the prior learning knowledge of the smart device cluster operation log example includes:

[0035] Generate a third neural network training cost algorithm based on the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example;

[0036] Based on the global operation results of the first neural network training cost algorithm, the second neural network training cost algorithm and the third neural network training cost algorithm, a fourth neural network training cost algorithm is obtained;

[0037] Based on the fourth neural network training cost algorithm, network parameters of the distributed collaborative scheduling decision network are reverse optimized.

[0038] In a second aspect, an AI collaborative scheduling system is provided, comprising a processor and a memory communicating with each other, wherein the processor is used to call a computer program from the memory and implement the method described in the first aspect by running the computer program.

[0039] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the method described in the first aspect when running.

[0040] The distributed industrial task collaborative scheduling method and system based on swarm intelligence provided by the embodiment of the present invention can adaptively match the target scheduling strategy mode and output the feasibility score by entering the operation log into the debugged decision network, thereby enhancing the flexibility and accuracy of the scheduling process. The linkage training cost algorithm further optimizes the scheduling results, making the production line operation more efficient. When the production line needs to be expanded or reduced, it is only necessary to adjust the decision network, showing good scalability. It can be seen that the above technical solution improves the operating efficiency and flexibility of the industrial production line, which is of great significance to optimizing production management.

[0041] In addition, by optimizing the network parameters of the distributed collaborative scheduling decision network, the quantitative differences between the running state residual semantic vectors of all the first intelligent device cluster running log examples in the network optimization example set can be made as consistent as possible, while the quantitative differences between the running state residual semantic vectors of the second intelligent device cluster running log example and the running state residual semantic vectors of the initial intelligent device cluster running log corresponding to the second intelligent device cluster running log example can be distinguished as much as possible, and the text repetition difference between the running log retelling text of the second intelligent device cluster running log example and the second intelligent device cluster running log example can be expanded, so that the distributed collaborative scheduling decision network has a smaller text retelling difference for the first intelligent device cluster running log example and a larger text retelling difference for the second intelligent device cluster running log example, which can ensure the adaptability of the distributed collaborative scheduling decision network to new samples and ensure the matching analysis accuracy of the new target scheduling strategy pattern.

[0042] Furthermore, by adding a semantic upgrade unit, the generation of disturbance examples can be realized to help the distributed collaborative scheduling decision network to master the actual examples, thereby further expanding the quantitative difference between the operating state residual semantic vectors of the first / second smart device cluster operation log examples to ensure the rationality and explainability of distributed industrial task collaborative scheduling decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 A flow chart of a distributed industrial task collaborative scheduling method based on swarm intelligence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to better understand the above technical scheme, the technical scheme of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical scheme of the present invention, rather than limitations on the technical scheme of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0046] Figure 1 A distributed industrial task collaborative scheduling method based on swarm intelligence is shown, which is applied to an AI collaborative scheduling system. The method includes the following steps 110 to 120.

[0047] Step 110: Obtain an operation log of a target intelligent device cluster for collaborative scheduling analysis of an industrial production line, wherein the operation log of the target intelligent device cluster includes target distributed task status data.

[0048] Step 120, entering the target smart device cluster operation log into the debugged distributed collaborative scheduling decision network, outputting the feasibility score of the target smart device cluster operation log matching the target scheduling strategy model, and the distributed collaborative scheduling decision network is optimized according to the linkage training cost algorithm.

[0049] In the embodiment of the present invention, swarm intelligence is a new computational intelligence paradigm that simulates the collective behavior of animal groups such as insects, birds and fish in nature to solve complex problems. Distributed industrial task collaborative scheduling technology based on swarm intelligence uses this intelligent method to optimize the operation of decentralized industrial production lines.

[0050] In such systems, each industrial production unit can be regarded as an individual in a swarm intelligence system. They make task scheduling decisions independently and achieve the overall goal through mutual cooperation. Each unit has a certain degree of intelligence and can self-optimize strategies according to environmental changes and task requirements to achieve dynamic optimization.

[0051] First, collaborative scheduling based on swarm intelligence can improve timeliness. Swarm intelligence emphasizes parallelism and distributed processing, which enables the scheduling process to respond quickly, adapt to real-time changes in the production line, and reduce production delays.

[0052] Secondly, this technology increases the flexibility of production line scheduling. Since each individual in the swarm intelligence system can self-optimize and make decisions, the entire system can find the optimal solution under different conditions and schedule the production line efficiently.

[0053] In addition, the distributed industrial task collaborative scheduling technology based on swarm intelligence also has good scalability. Because each production unit is an independent individual, when the production line needs to be expanded, it only needs to add new production units and let them join the existing swarm intelligence system.

[0054] In general, distributed industrial task collaborative scheduling technology based on swarm intelligence achieves the optimal scheduling of decentralized industrial production lines by simulating the collective behavior of animals in nature, thereby improving the timeliness, flexibility and scalability of scheduling.

[0055] Based on the above content, step 110 and step 120 are exemplarily introduced below.

[0056] In step 110 and step 120, the target intelligent device cluster operation log is the operation information collected from the intelligent device cluster in the industrial production line, such as device status, fault records, workload, etc., which is used for real-time monitoring and subsequent analysis of the operation of the equipment. The target distributed task status data part records the status information of various tasks on the industrial production line, such as task completion status, resource occupancy status, etc., which can help the system understand the working status of the overall production line. The distributed collaborative scheduling decision network that has been debugged is a pre-trained and debugged neural network model that can receive the target intelligent device cluster operation log as input and then output a feasibility score that matches the target scheduling strategy mode. The target scheduling strategy mode is a predetermined industrial production line scheduling strategy, including but not limited to task priority setting, resource allocation strategy, equipment maintenance strategy, etc. The feasibility score reflects the degree of match between the current operation log and the target scheduling strategy mode. The higher the score, the greater the possibility of implementation of the strategy under the current circumstances.

[0057] For example, on an automobile production line, each smart device will generate an operation log, including its own status, completed tasks, and other information. These logs and task status data are entered into the collaborative scheduling decision network, which analyzes these data through a pre-trained algorithm and gives a feasibility score based on the preset scheduling strategy mode. If the score is high, it means that the current production status matches the preset strategy and can be scheduled according to this strategy.

[0058] In other exemplary application scenarios, in an automated factory, there are device clusters composed of various intelligent devices such as robot arms, conveyor belts, printers, etc. Each device will generate a target intelligent device cluster operation log during operation.

[0059] For example, a robot arm can record its speed, accuracy, power consumption, temperature and other information during the production process; a conveyor belt can record its operating speed, load, number of failures and other information; a printer can record its print quality, ink cartridge consumption, paper usage and other information. All of this data is collected and integrated into equipment operation logs for real-time monitoring and subsequent analysis.

[0060] At the same time, these devices will also generate target distributed task status data when completing their respective tasks. For example, after completing an assembly task, the robot arm will generate task status data including task start time, end time, whether it is successfully completed, problems encountered, etc.; when the conveyor belt is transporting products, it will also generate task status data including transmission start and end time, transmission quantity, whether the transmission is smooth, etc.

[0061] These target intelligent device cluster operation logs and target distributed task status data together provide input data for the collaborative scheduling decision network, helping it to schedule and optimize industrial production lines more effectively.

[0062] The above technical solution is based on swarm intelligence to optimize the coordinated scheduling of industrial production lines. Its beneficial effects are mainly reflected in the following aspects:

[0063] (1) Real-time: By directly obtaining the operation logs and distributed task status data of the intelligent device cluster, the solution can reflect the working status and task execution status of each device on the production line in real time, thereby making timely and effective scheduling decisions;

[0064] (2) Flexibility: Through the application of a distributed collaborative scheduling decision network, the solution can adaptively match the target scheduling strategy model and give a feasibility score, further increasing the flexibility of the scheduling process and being able to formulate the most appropriate scheduling strategy for different situations.

[0065] (3) Accuracy: With the help of the linkage training cost algorithm, the distributed collaborative scheduling decision network can optimize the scheduling results more accurately, so that the working efficiency of each device is maximized and the operation of the entire production line is more efficient;

[0066] (4) Scalability: When the production line needs to be expanded or reduced, only the distributed collaborative scheduling decision network needs to be adjusted without changing the entire system architecture, so it has good scalability;

[0067] (5) Improve production efficiency: Through real-time, flexible and accurate scheduling, the idle time and failure time of equipment can be minimized, thereby improving production efficiency.

[0068] In summary, the beneficial effect of the above technical solution is to provide real-time, flexible and accurate production line scheduling. By entering the operation log into the debugged decision network, the system can adaptively match the target scheduling strategy mode and output the feasibility score, which enhances the flexibility and accuracy of the scheduling process. The linkage training cost algorithm further optimizes the scheduling results, making the production line more efficient. When the production line needs to be expanded or reduced, only the decision network needs to be adjusted, showing good scalability. It can be seen that the above technical solution improves the operating efficiency and flexibility of industrial production lines, which is of great significance to optimizing production management.

[0069] In some possible embodiments, the step of optimizing the distributed collaborative scheduling decision network according to the linkage training cost algorithm includes steps 210 to 240.

[0070] Step 210: Obtain a set of network optimization examples, each network optimization example including an intelligent device cluster operation log example, an initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example, and prior learning knowledge of the intelligent device cluster operation log example, the intelligent device cluster operation log example including a first intelligent device cluster operation log example and a second intelligent device cluster operation log example, and the second intelligent device cluster operation log example matches a target scheduling strategy pattern.

[0071] In step 210, the network optimization example set is a set of data sets used to train and optimize the decision network. For example, it may include operation logs of multiple smart devices under different working conditions.

[0072] The smart device cluster operation log example refers to the operation log of smart devices collected during the actual production process. For example, the operation log of a robot arm can record its speed, accuracy, power consumption and other information.

[0073] The initial smart device cluster operation log refers to the smart device operation log when no scheduling strategy is applied. For example, before starting the production line, the recorded device operation status can be regarded as the initial operation log.

[0074] Prior knowledge refers to knowledge that has been learned from past experience or data, which can be used to predict or infer unknown situations. For example, by analyzing historical operation logs, it may be known that a certain equipment state may lead to a higher failure rate.

[0075] The first smart device cluster operation log example (positive sample) is an operation log example that conforms to the target scheduling strategy pattern and can be regarded as a positive training sample. For example, if the target scheduling strategy is to optimize energy consumption, then a device operation log with low energy consumption can be regarded as a positive sample.

[0076] Second smart device cluster operation log example (negative sample): On the contrary, this is an operation log example that does not conform to the target scheduling strategy pattern and can be regarded as a negative training sample. In the above example, a device operation log with high energy consumption can be regarded as a negative sample.

[0077] Step 220: for each network optimization example in the network optimization example set, use the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example as input information of the distributed collaborative scheduling decision network, output a feasibility score of the smart device cluster operation log example matching the target scheduling strategy pattern, and the distributed collaborative scheduling decision network is used to generate an operation log retelling text of the smart device cluster operation log example based on the smart device cluster operation log example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example, and determine the feasibility score based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example.

[0078] In step 220, the operation log retelling text is extracted and generated from the original operation log data to form a more readable and understandable descriptive text. This process usually involves semantic analysis and interpretation of the original log data.

[0079] For example, an original running log is as follows:

[0080] 2022-01-01, 12:00:00Device_1status: running; CPU load: 75%; Memory usage: 60%.

[0081] This record indicates that at 12:00 on January 1, 2022, the device named "Device_1" was running, with a CPU load of 75% and a memory usage of 60%.

[0082] The run log retelling text can convert this record into a more understandable form, for example: At 12:00 noon on January 1, 2022, device 1 is running normally. Its CPU load reached 75% and its memory usage was 60%.

[0083] Such retelling enables non-professionals to understand the operating status of the equipment, not just IT professionals or system administrators. In the management of large-scale intelligent device clusters, this operation log retelling text can help operators quickly understand the operating status of the equipment, thereby making effective decisions and scheduling.

[0084] Step 230: Generate a first neural network training cost algorithm based on the running state residual semantic vector of the smart device cluster running log example obtained in the retelling task and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example; generate a second neural network training cost algorithm based on the running log retelling text of the smart device cluster running log example and the smart device cluster running log example; the first neural network training cost algorithm is used to constrain the quantitative difference between the running state residual semantic vector of the first smart device cluster running log example in the network optimization example set and the running state residual semantic vector of the first smart device cluster running log example, and increase the quantitative difference between the running state residual semantic vector of the second smart device cluster running log example in the network optimization example set and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example; the second neural network training cost algorithm is used to constrain the text retelling training cost of the first smart device cluster running log example and increase the text retelling training cost of the second smart device cluster running log example.

[0085] In step 230, the operation state residual semantic vector is a semantic vector obtained by processing and analyzing the operation log example of the smart device cluster and its initial operation log. It can capture the changes in the operation state (i.e., "residuals") and express the semantic information of these changes in the form of vectors.

[0086] The first neural network training cost algorithm (first training loss function) is a function used to quantify the difference between the residual semantic vector of the running state predicted by the model and the residual semantic vector of the actual running state during the neural network training process. By minimizing this loss function, the residual semantic vector of the running state generated by the network can be made closer to the actual running state, thereby improving the prediction accuracy of the model.

[0087] The second neural network training cost algorithm (second training loss function) is a function used to quantify the difference between the run log retelling text generated by the model and the actual run log example during the neural network training process. By minimizing this loss function, the run log retelling text generated by the model can more accurately reflect the actual running situation, thereby improving the model's retelling quality.

[0088] Quantifying the difference is to quantify the difference between two objects (such as the residual semantic vectors of the running state) through mathematical calculation. For example, the difference between the two semantic vectors can be quantified by calculating the Euclidean distance or the cosine similarity.

[0089] The text retelling training cost refers to the cost that the neural network model needs to bear when performing text retelling tasks. Usually, this cost is related to the performance of the model (such as the accuracy and consistency of the retelling, etc.) and the training efficiency (such as convergence speed, computing resource consumption, etc.). The text retelling training cost can also be understood as a function used to quantify the difference between the run log retelling text generated by the model and the actual run log example during the neural network training process. By minimizing this loss function, the run log retelling text generated by the model can more accurately reflect the actual operation situation, thereby improving the model's retelling quality. This loss function may contain multiple parts, such as considering the accuracy of the retelling (whether the content of the model retelling is correct) and fluency (whether the text retelled by the model is easy to understand).

[0090] Step 240: Based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example, and the prior learning knowledge of the smart device cluster operation log example, the network parameters of the distributed collaborative scheduling decision network are optimized until the network optimization requirements are met, thereby obtaining a distributed collaborative scheduling decision network that has been debugged.

[0091] In step 240, network parameters are internal variables of the neural network model, usually including weights and biases. By learning and optimizing these parameters through training data, the performance of the model can be improved. The network optimization compliance requirement is a pre-set network optimization goal, which can be a specific loss function value, accuracy index or other performance index. Only when the network model meets these requirements, the network optimization task is considered completed. For example, it can be set that the training loss needs to be lower than a certain threshold, or the accuracy of the retelling text needs to be higher than a certain percentage, etc.

[0092] The following describes steps 210 to 240 through a complete example.

[0093] First, you need to collect a set of network optimization examples. For example, collect operation log data from a smart device cluster. Assume that there are two types of devices, so two types of operation logs are collected, log example 1 and log example 2. At the same time, the initial logs of these devices when they start running are also recorded.

[0094] Then, for each network optimization example in the example set, it is used as input data and input into the distributed collaborative scheduling decision network. For example, a deep neural network model can be used to take log example 1 and the corresponding initial log as input, and the output is the feasibility score of matching the target scheduling strategy pattern (such as using a softmax function to calculate the probability). At the same time, the network also generates a paraphrased run log text. Then, based on this paraphrased text and the original log example, the feasibility score is determined.

[0095] In this step, two training loss functions are generated. The first training loss function is based on the running state residual semantic vector. For example, the mean square error (MSE) can be used to measure the difference between the predicted residual semantic vector and the actual residual semantic vector. The formula is as follows:

[0096] L1=1 / N*Σ(y_true-y_pred)^2;

[0097] Where y_true is the actual residual semantic vector, y_pred is the predicted residual semantic vector, and N is the total number of samples.

[0098] The second training loss function is based on the running log retelling text. The cross-entropy loss (Cross-EntropyLoss) can be used to measure the difference between the retelling text and the actual text. The formula is as follows:

[0099] L2=-1 / N*Σ[y_true*log(y_pred)+(1-y_true)*log(1-y_pred)];

[0100] Where y_true is the actual text label (usually represented by 0 and 1), and y_pred is the predicted label of the paraphrase text.

[0101] Finally, the two training loss functions, feasibility scores, and prior learning knowledge are combined to optimize the parameters of the network. The optimization method can use gradient descent or its variants, such as stochastic gradient descent (SGD), Adam, etc. When the performance of the model reaches the set standard (such as the loss function value is lower than a certain threshold or the repetition quality is higher than a certain threshold), it can be considered that the distributed collaborative scheduling decision network has been debugged.

[0102] By applying the above-mentioned embodiment, the network parameters of the distributed collaborative scheduling decision network can be optimized so that the quantitative differences between the running state residual semantic vectors of all the first smart device cluster running log examples in the network optimization example set are as consistent as possible, and the quantitative differences between the running state residual semantic vectors of the second smart device cluster running log example and the running state residual semantic vectors of the initial smart device cluster running log corresponding to the second smart device cluster running log example are distinguished as much as possible, and the text repetition difference between the running log retelling text of the second smart device cluster running log example and the second smart device cluster running log example is expanded, so that the distributed collaborative scheduling decision network has a smaller text retelling difference for the first smart device cluster running log example and a larger text retelling difference for the second smart device cluster running log example, which can ensure the adaptability of the distributed collaborative scheduling decision network to new samples and ensure the matching analysis accuracy of the new target scheduling strategy pattern.

[0103] It can be seen that by constraining the quantitative difference between the residual semantic vectors of the running state of the first smart device cluster running log example, and increasing the quantitative difference between the residual semantic vectors of the running state of the second smart device cluster running log example and the residual semantic vectors of the running state of the initial smart device cluster running log, the resolution of the neural network model for the running state of different types of devices can be improved. This allows the network to better adapt to new samples, thereby improving its accuracy in matching new target scheduling policy patterns. By increasing the difference in textual repetition between the running log retelling text of the second smart device cluster running log example and the original log example, it can be ensured that the network generates more characteristic and easier to understand retelling texts for the second type of device (devices that need to match the target scheduling policy pattern). At the same time, for the first type of device, the retelling text generated by the network is less different, which can avoid introducing too many retelling differences to those devices that have not undergone major changes, thereby ensuring the quality of the retelling text. The generalization ability of the network is improved by optimizing the network parameters through the linkage training cost algorithm.

[0104] In some exemplary embodiments, the distributed collaborative scheduling decision network includes a first decision branch component and a second decision branch component, and the first decision branch component includes a state feature mining core and a state feature processing core. Based on this, step 220 uses the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example as input information of the distributed collaborative scheduling decision network, and outputs the feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode, including steps 221-223.

[0105] Step 221: Using the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example as input information of the state feature mining core, output the operation state residual semantic vector of the smart device cluster operation log example and the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the smart device cluster operation log example.

[0106] Step 222: using the running state residual semantic vector of the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example as input information of the state feature processing core, outputting the u-th order running state residual semantic vector of the smart device cluster running log example and the u-th order running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and outputting the running log retelling text of the smart device cluster running log example, where u is an integer greater than or equal to 1 and less than or equal to X, and X is the total order of the pyramid reversible subnet of the state feature processing core.

[0107] Step 223: Outputting a feasibility score of the smart device cluster operation log example matching the target scheduling strategy pattern based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example through the second decision branch component.

[0108] Among them, the first decision branch component and the second decision branch component are the two main parts in the distributed collaborative scheduling decision network. The first decision branch component is mainly responsible for processing state features, while the second decision branch component makes decisions based on the processed features. The state feature mining core and the state feature processing core are the two core parts in the first decision branch component, which are responsible for mining state features from input information and processing these features. The operating state residual semantic vector is a vector output by the state feature mining core, which is used to represent the change of the operating state of the equipment. The u-th order operating state residual semantic vector is a vector output by the state feature processing core, which represents the change of the operating state of the equipment at a certain level (order).

[0109] Next, let's take a specific example to illustrate this technical solution: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. These operation logs and the initial operation logs are fed into the state feature mining core, which is responsible for extracting features that characterize the changes in the operation state of the device from the logs and generating the operation state residual semantic vector.

[0110] Then, these running state residual semantic vectors are sent to the state feature processing core. Assuming that the total order X of the pyramid reversible subnet is set to 3, the processing core will output three-order running state residual semantic vectors, namely the 1st order, 2nd order, and 3rd order running state residual semantic vectors. At the same time, the processing core will also generate a retelling text of the running log.

[0111] Finally, this information (original operation log, paraphrase text, and residual semantic vectors of operation status of each order) is fed into the second decision branch component. Based on this information, this component will output the feasibility score of each smart device cluster operation log example matching the target scheduling strategy pattern. This score can be calculated by the softmax function, as follows:

[0112] score=exp(scores) / sum(exp(scores));

[0113] Among them, scores is the original score output by the second decision branch component. In this way, the feasibility score of each device can be obtained for further decision-making or optimization operations.

[0114] In other examples, the pyramid reversible subnet can be understood as a deconvolution subnet. This is because the deconvolution operation is often used for upsampling, which increases the feature map in spatial dimensions (height and width). In the field of deep learning, deconvolution is often used in fully convolutional networks (FCN) to achieve tasks such as semantic segmentation.

[0115] Deconvolution, or transposed convolution, is not actually the inverse operation of convolution, but a variant of convolution. It first upsamples the input (usually inserting zeros) and then performs a normal convolution operation.

[0116] For example, there is a 2x2 input:

[0117] [1,2];

[0118] [3,4].

[0119] Before deconvolution, upsampling is performed, for example, by filling a circle of 0s around each element:

[0120] [0, 0, 0, 0];

[0121] [0, 1, 0, 2];

[0122] [0, 0, 0, 0];

[0123] [0, 3, 0, 4];

[0124] [0, 0, 0, 0].

[0125] Then use a 3x3 convolution kernel to perform the convolution operation to get a larger output.

[0126] In this way, deconvolution can map low-resolution feature maps to high-resolution space while maintaining spatial information, so that the network can have a better understanding of the spatial structure of the original input. This is very important in many tasks, such as object detection, semantic segmentation, etc. In the above embodiment, the state feature processing kernel uses deconvolution operations to generate running state residual semantic vectors of different orders. These vectors capture state change information at different scales, thereby providing a richer basis for decision-making.

[0127] By applying the above embodiment, by introducing the state feature mining core and the state feature processing core, the scheme can extract meaningful features from the device operation log and convert these features into operation state residual semantic vectors. Based on these vectors, the second decision branch component can generate a feasibility score that matches the target scheduling strategy mode, thereby assisting in making more accurate decisions. The state feature processing core uses a pyramid reversible subnet (deconvolution subnet) to generate operation state residual semantic vectors of different orders, which capture the change information of the device operation state at different scales. Therefore, even if the information at a certain scale is interfered by noise, the information at other scales can still provide a useful basis for decision-making, thereby enhancing the robustness of the model. In addition to generating the operation state residual semantic vector, the state feature processing core also generates a paraphrase text of the operation log. This paraphrase text can help better understand the operation state of the device, thereby improving the interpretability of the model. By learning to generate feasibility scores, the model can not only process the device operation logs seen in the training phase, but also make effective decisions on new device operation logs. This makes the model have good generalization ability and can cope with a variety of unknown device operation conditions. It can be seen that by comprehensively utilizing deep learning and convolutional neural network technology, the distributed collaborative scheduling problem is effectively solved, providing strong support for the management and optimization of smart device clusters.

[0128] In some examples, the first decision branch component further includes a semantic upgrade unit. Based on this, the method further includes steps 310 to 320.

[0129] Step 310: Enter each smart device cluster operation log example in the network optimization example set into the semantic upgrade unit, and output the operation state residual enhanced semantic vector of the smart device cluster operation log example and the u-th order operation state residual enhanced semantic vector of the smart device cluster operation log example.

[0130] Step 320: Update the running state residual semantic vector of the second smart device cluster running log example in the smart device cluster running log example into the running state residual enhanced semantic vector of the second smart device cluster running log example, and update the u-th order running state residual semantic vector of the second smart device cluster running log example into the u-th order running state residual enhanced semantic vector of the second smart device cluster running log example.

[0131] In the above example, the first decision branch component is a main part of the distributed collaborative scheduling decision network, which includes a state feature mining core, a state feature processing core and a newly introduced semantic upgrade unit. The semantic upgrade unit is a newly introduced module in the first decision branch component, and its function is to enhance the running state residual semantic vector and generate a running state residual enhanced semantic vector. The running state residual enhanced semantic vector is a vector output by the semantic upgrade unit, which enhances the original running state residual semantic vector to better represent the changes in the equipment running state. The u-th order running state residual enhanced semantic vector is a vector output by the semantic upgrade unit, which enhances the original u-th order running state residual semantic vector.

[0132] Next, we will use a specific example to illustrate this technical solution: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. These operation logs are sent to the state feature mining core and the state feature processing core. After processing, the operation state residual semantic vector and the operation state residual semantic vectors of different orders are obtained.

[0133] Then, these semantic vectors are fed into the semantic upgrading unit. Assume that a strengthening method is chosen, such as regularizing the original semantic vector (i.e. dividing the vector by its L2 norm). The formula can be expressed as: v_strong = v / ||v||;

[0134] Among them, v is the original running state residual semantic vector, v_strong is the enhanced running state residual enhanced semantic vector, and ||·|| represents the L2 norm.

[0135] Through such processing, the operation state residual enhanced semantic vector and the operation state residual enhanced semantic vectors of different orders are obtained. These vectors can better express the changes in the operation state of the equipment.

[0136] Finally, the running state residual semantic vector of the second smart device cluster running log example and the corresponding running state residual semantic vectors of each order are updated to the enhanced semantic vector.

[0137] With this design, the semantic upgrade unit strengthens the original residual semantic vector of the operating status so that the new vector can better express the changes in the operating status of the device, thereby improving the accuracy of the decision. The strengthened semantic vector regularizes the original information, which can reduce unnecessary noise interference and enable the model to maintain stable performance in various complex environments. By updating the residual semantic vector of the operating status of the second smart device cluster operation log example to a strengthened semantic vector, the model can better adapt to the new device operating conditions and further improve its generalization ability.

[0138] In some other possible embodiments, the semantic upgrade unit is used to: extract the target distributed task status highlighted text of the smart device cluster operation log example, adjust the size of the target distributed task status highlighted text of the smart device cluster operation log example according to the text box size of the highlighted text of the corresponding order used to generate the first neural network training cost algorithm, and obtain the target highlighted text; randomly select a variable Q within the set operation value quadrant, if the Q is greater than the preset value, then use the first semantic upgrade rule to perform semantic upgrade based on the target highlighted text, if the Q is less than or equal to the preset value, then use the second semantic upgrade rule to perform semantic upgrade based on the target highlighted text.

[0139] Among them, the target distributed task status highlight text is a part of the intelligent device cluster operation log example, which particularly highlights the status of the target distributed task. The first semantic upgrade rule and the second semantic upgrade rule are two different semantic upgrade methods in the semantic upgrade unit. Which one to use depends on the relationship between a variable Q taken at random in the set operation value quadrant and the pre-configured value.

[0140] Next, this technical solution will be illustrated with a specific example: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. In each operation log example, a highlighted text describing the state of the target distributed task is included. These highlighted texts are first sent to the semantic upgrade unit for resizing so that their size matches the text box size of the highlighted text of the corresponding order used by the first neural network training cost algorithm.

[0141] Next, a variable Q is randomly selected within the set operation value quadrant. Assume that the preset value is set to 0.5. If Q is greater than 0.5, the first semantic upgrade rule is used for semantic upgrade; if Q is less than or equal to 0.5, the second semantic upgrade rule is used for semantic upgrade. The specific upgrade rule may involve some transformation of the original text. For example, the first semantic upgrade rule may include adding specific words or phrases, while the second semantic upgrade rule may include replacing certain words or phrases.

[0142] In this way, by processing highlighted text, the semantic upgrade unit can present key information more intuitively, thereby helping to better understand the operating status and decision-making process of the device. Two different semantic upgrade rules are introduced, allowing the model to flexibly adjust its processing methods according to specific circumstances, so that it can still maintain stable performance in the face of various complex environments. The semantic upgrade unit can extract more useful information by processing and strengthening highlighted text, thereby helping the model make more accurate decisions.

[0143] In some other examples, the second decision branch component includes a scheduling mode discrimination unit and a scheduling mode matching unit. Based on this, step 223 outputs a feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode through the second decision branch component based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example, including steps 2231-2233.

[0144] Step 2231: enter the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example into the scheduling mode determination unit, and the scheduling mode determination unit determines a first feasibility identification label based on the smart device cluster operation log example, the operation log retelling text of the smart device cluster operation log example, the pyramid state semantic relationship network extracted by the scheduling mode determination unit, the fully connected layer corresponding to the state feature mining core, and the fully connected layer corresponding to the pyramid reversible subnet of the state feature processing core.

[0145] Step 2232: Enter the running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid state semantic relationship network into the scheduling pattern matching unit; the scheduling pattern matching unit semantically downsamples the pyramid state semantic relationship network to obtain the pyramid downsampled semantic relationship network; the scheduling pattern matching unit determines a second feasibility identification label based on the running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid downsampled semantic relationship network.

[0146] Step 2233: Perform a weight-based operation on the first feasibility identification tag and the second feasibility identification tag to obtain a feasibility score for the smart device cluster operation log example matching the target scheduling strategy pattern and output it.

[0147] In the above embodiment, the scheduling mode discrimination unit is a module in the second decision branch component, and its function is to determine the first feasibility identification label based on the input information. The scheduling mode matching unit is also a module in the second decision branch component, and its function is to determine the second feasibility identification label based on the input information. The pyramid state semantic relationship network is a structure extracted by the scheduling mode discrimination unit, which is used to express the semantic relationship between the operating states of the equipment. The pyramid downsampling semantic relationship network is a structure obtained by the scheduling mode matching unit by semantically downsampling the pyramid state semantic relationship network.

[0148] Next, a specific example is used to illustrate this technical solution: there is a cluster of intelligent devices whose operation logs are recorded and used as network optimization examples. These operation logs and their retelling texts are first sent to the scheduling mode discrimination unit. In this unit, based on the input information and the fully connected layer obtained from the state feature mining core and the state feature processing core, the pyramid state semantic relationship network is extracted and the first feasible identification label is determined.

[0149] Next, the running state residual semantic vector, the running state residual semantic vector of the initial running log, and the pyramid state semantic relationship network are sent to the scheduling pattern matching unit. In this unit, the pyramid state semantic relationship network is first semantically downsampled to obtain the pyramid downsampled semantic relationship network. Then, based on the input information and the pyramid downsampled semantic relationship network, the second feasibility identification label is determined.

[0150] Finally, the first feasibility identification tag and the second feasibility identification tag are processed based on the weight to obtain the feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode. Assuming the weight is w, the calculation formula can be expressed as:

[0151] score=w*label_1+(1-w)*label_2;

[0152] Among them, label_1 is the first feasibility identification label, label_2 is the second feasibility identification label, and score is the final feasibility score.

[0153] In this way, by introducing the scheduling mode discrimination unit and the scheduling mode matching unit, the scheme can evaluate the matching degree between the equipment operation log examples and the target scheduling strategy mode from different angles, thereby improving the accuracy of decision-making. The scheduling mode matching unit downsamples the pyramid state semantic relationship network, so that the model can still maintain stable performance when facing complex semantic relationships. The weight-based operation processing makes the decision process clearer and improves the interpretability of the model. The scheme processes the two steps of discrimination and matching separately, so that the model has better generalization ability and can cope with a variety of unknown equipment operation conditions.

[0154] Furthermore, the pyramid down-sampling semantic relationship network can also be understood as a pyramid down-sampling semantic feature map. This is a common concept in the field of computer vision and deep learning, which is usually used to extract and retain important information at different scales.

[0155] The word pyramid comes from its structure: at the bottom of the network, there is the original input, which contains the most detailed and sophisticated information; then as the network layers increase, the size of the feature map will gradually decrease, and at the same time each pixel will contain more and more surrounding environment information, which forms a pyramid-shaped structure.

[0156] In this context, downsampling is to reduce the resolution of the feature map in some way (such as max pooling or convolution step size greater than 1), thereby reducing the amount of computation and model complexity. Although this process will lose some detailed information, if designed properly, it can allow the model to learn to retain the most important semantic information. Therefore, the pyramid downsampled semantic relationship network or pyramid downsampled semantic feature map is a structure that can capture and retain key semantic information at different scales. For example, in the above example, the scheduling pattern matching unit obtains the pyramid downsampled semantic relationship network by downsampling the pyramid state semantic relationship network. This network retains the key information of the device operation log examples at different scales, thereby helping the model make better decisions.

[0157] In some alternative embodiments, step 230 generates a first neural network training cost algorithm based on the running state residual semantic vector of the smart device cluster running log example obtained in the retelling task and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, including steps 231-233.

[0158] Step 231: Calculate a first semantic quantization difference based on the running state residual semantic vector of the second smart device cluster running log example in the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example.

[0159] Step 232: Calculate a second semantic quantization difference based on the running state residual semantic vector of the first smart device cluster running log example in the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the first smart device cluster running log example, and the number of the first smart device cluster running log example in the smart device cluster running log example.

[0160] Step 233: Determine the ratio of the second semantic quantization difference to the first operation value as the first neural network training cost algorithm, and the first operation value is the sum of the first semantic quantization difference and the set network parameter.

[0161] In the above example, the semantic quantitative difference is a measure of the difference between two semantic vectors. It can be calculated based on any suitable distance or similarity measure, such as Euclidean distance, cosine similarity, etc.

[0162] Next, let's take a specific example to illustrate this technical solution: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. These operation logs and their corresponding initial operation logs all contain the operation state residual semantic vector.

[0163] In step 231, a first semantic quantitative difference is calculated based on the residual semantic vector of the running state of the second smart device cluster running log example and the residual semantic vector of the running state of the corresponding initial running log. Assuming that the Euclidean distance is selected as the metric, the calculation formula can be expressed as: d1 = ||v2-v2_initial||;

[0164] Wherein, v2 is the running state residual semantic vector of the second smart device cluster running log example, v2_initial is the running state residual semantic vector of the corresponding initial running log, and ||·|| represents the L2 norm.

[0165] In step 232, a second semantic quantitative difference is calculated based on the residual semantic vector of the running state of the first smart device cluster running log example, the residual semantic vector of the running state of the corresponding initial running log, and the number of the first smart device cluster running log examples. The specific calculation method can be selected according to actual needs. For example, if you want to consider the influence of the number of samples, you can use the average Euclidean distance as a metric.

[0166] Finally, in step 233, the ratio of the second semantic quantization difference to the first operation value is determined as the first neural network training cost algorithm. The first operation value is the sum of the first semantic quantization difference and the set network parameter. The formula can be expressed as:

[0167] cost = d2 / (d1+p);

[0168] Among them, d2 is the second semantic quantization difference, d1 is the first semantic quantization difference, p is the set network parameter, and cost is the final cost value.

[0169] In this way, by calculating the semantic quantization difference and using it in the cost function, the performance of the model can be measured more directly, thereby improving the training efficiency. This solution takes into account the influence of multiple factors (such as the number of device operation log examples), so that the model can still maintain stable performance in various complex environments. This solution introduces semantic quantization differences, which enables the model to better capture and learn changes in the device operation status, thereby improving the accuracy of the model.

[0170] In some other possible embodiments, the step 230 described in which a second neural network training cost algorithm is generated based on the operation log retelling text of the smart device cluster operation log example and the smart device cluster operation log example includes step 230a or step 230b.

[0171] Step 230a: If the smart device cluster operation log example is the first smart device cluster operation log example, a cross entropy cost between the first smart device cluster operation log example and the operation log retelling text of the first smart device cluster operation log example is determined as a second neural network training cost algorithm.

[0172] Step 230b: If the smart device cluster operation log example is a second smart device cluster operation log example, a second neural network training cost algorithm is generated based on the operation state residual semantic vector of the second smart device cluster operation log example, the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the operation log retelling text of the second smart device cluster operation log example, the highlighted text of the state feature mining core, and the highlighted text of the pyramid reversible subnet of the state feature processing core.

[0173] In the above embodiment, the cross entropy cost is a loss function applied to classification problems. It can measure the difference between the probability distribution predicted by the model and the actual distribution. The operating state residual semantic vector is a vector representation that describes the operating state of the device. The residual refers to the difference between the predicted value and the actual value, while the semantics emphasizes that this vector captures some important features of the operating state of the device.

[0174] Next, we will use a specific example to illustrate this technical solution: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. These operation logs and their retelling texts contain the residual semantic vector of the operation status.

[0175] In step 230a, if the first smart device cluster operation log example is being processed, the cross entropy cost between the log example and the corresponding operation log retelling text is determined as the second neural network training cost algorithm. The cross entropy can be calculated by the following formula: H(p, q) = -Σp(x)log(q(x));

[0176] Here, p is the actual distribution and q is the predicted distribution.

[0177] In step 230b, if the second smart device cluster operation log example is being processed, a second neural network training cost algorithm is generated based on various information of the log example (including the operation state residual semantic vector, the initial operation log, the retelling text, and the highlighted text of the state feature mining core and the state feature processing core, etc.). The specific generation method may involve comprehensive consideration of multiple factors, for example, the loss of each part may be calculated separately first, and then combined according to a certain weight.

[0178] In this way, the solution uses different cost functions according to different types of log examples, which enables the model to evaluate its performance more accurately, thereby improving training efficiency. The solution can automatically adjust the cost function to adapt it to different types of log examples, which enhances the adaptability of the model. The solution introduces the running state residual semantic vector, which enables the model to better capture and learn the changes in the device running state, thereby improving the accuracy of the model.

[0179] Under some possible design ideas, step 240 optimizes the network parameters of the distributed collaborative scheduling decision network based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example, including steps 241-243.

[0180] Step 241: Generate a third neural network training cost algorithm based on the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example.

[0181] Step 242: Obtain a fourth neural network training cost algorithm based on the global operation results of the first neural network training cost algorithm, the second neural network training cost algorithm, and the third neural network training cost algorithm.

[0182] Step 243: Based on the fourth neural network training cost algorithm, reversely optimize the network parameters of the distributed collaborative scheduling decision network.

[0183] In the above embodiment, the third neural network training cost algorithm is a method for calculating the loss function or cost function of the model during the training process. It evaluates the performance of the model by comparing the difference between the predicted value and the actual value. The fourth neural network training cost algorithm is a new cost algorithm obtained based on the global operation results of the previous cost algorithms.

[0184] Next, we will use a specific example to illustrate this technical solution: there is a cluster of smart devices whose operation logs are recorded and used as network optimization examples. These operation logs and their retelling texts contain the residual semantic vector of the operation status.

[0185] In step 241, a third neural network training cost algorithm is generated based on the feasibility score of the smart device cluster running log example and the prior learning knowledge. This may involve a process of defining a new cost function using the prior knowledge and the feasibility score.

[0186] In step 242, a fourth neural network training cost algorithm is obtained based on the global operation results of the first, second and third neural network training cost algorithms. This may involve a process of combining or optimizing the first three cost functions to obtain a new, more effective cost function.

[0187] Finally, in step 243, based on the fourth neural network training cost algorithm, the network parameters of the distributed collaborative scheduling decision network are reversely optimized. This means updating the network parameters through the back propagation algorithm to minimize the cost value obtained by the fourth neural network training cost algorithm. The specific update formula can be expressed as:

[0188] Among them, θ is the network parameter, η is the learning rate, is the gradient of the cost function J with respect to θ.

[0189] This design, by generating and using multiple cost algorithms, can measure the performance of the model from different perspectives, thereby improving training efficiency. This solution uses the feasibility scores and prior learning knowledge of the smart device cluster operation log examples, so that the model can still maintain stable performance in complex environments. This solution introduces the operation state residual semantic vector, which enables the model to better capture and learn the changes in the device operation state, thereby improving the accuracy of the model.

[0190] On the basis of the above, an AI collaborative scheduling system is provided, comprising a processor and a memory that communicate with each other, wherein the processor is used to retrieve a computer program from the memory and implement the above method by running the computer program.

[0191] Based on the above, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the above method when running.

[0192] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A distributed industrial task collaborative scheduling method based on swarm intelligence, characterized in that: Applied to an AI collaborative scheduling system, the method includes: Obtaining an operation log of a target intelligent device cluster to be subjected to collaborative scheduling analysis of an industrial production line, wherein the operation log of the target intelligent device cluster includes target distributed task status data; Entering the target smart device cluster operation log into the debugged distributed collaborative scheduling decision network, outputting the feasibility score of the target smart device cluster operation log matching the target scheduling strategy mode, the distributed collaborative scheduling decision network is optimized according to the linkage training cost algorithm; The step of optimizing the distributed collaborative scheduling decision network according to the linkage training cost algorithm includes: Acquire a set of network optimization examples, each network optimization example comprising an intelligent device cluster operation log example, an initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example, and prior learning knowledge of the intelligent device cluster operation log example, the intelligent device cluster operation log example comprising a first intelligent device cluster operation log example and a second intelligent device cluster operation log example, the second intelligent device cluster operation log example matching a target scheduling strategy mode; For each network optimization example in the network optimization example set, an intelligent device cluster operation log example in the network optimization example and an initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example are used as input information of a distributed collaborative scheduling decision network, and a feasibility score of the intelligent device cluster operation log example matching the target scheduling strategy mode is output; the distributed collaborative scheduling decision network is used to generate an operation log retelling text of the intelligent device cluster operation log example based on the intelligent device cluster operation log example and the initial intelligent device cluster operation log corresponding to the intelligent device cluster operation log example, and determine the feasibility score based on the intelligent device cluster operation log example and the operation log retelling text of the intelligent device cluster operation log example; A first neural network training cost algorithm is generated according to the running state residual semantic vector of the smart device cluster running log example obtained in the retelling task and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example. A second neural network training cost algorithm is generated based on the running log retelling text of the smart device cluster running log example and the smart device cluster running log example. The first neural network training cost algorithm is used to constrain the quantitative difference between the running state residual semantic vector of the first smart device cluster running log example in the network optimization example set and the running state residual semantic vector of the initial smart device cluster running log corresponding to the first smart device cluster running log example, and increase the running state residual semantic vector of the first smart device cluster running log example in the network optimization example set. The quantified difference between the running state residual semantic vector of the second smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example, the second neural network training cost algorithm is used to constrain the text repetition training cost of the first smart device cluster running log example and increase the text repetition training cost of the second smart device cluster running log example; wherein the first neural network training cost algorithm is a first training loss function, which is used to quantify the difference between the running state residual semantic vector predicted by the model and the actual running state residual semantic vector; the second neural network training cost algorithm is a second training loss function, which is a function of the difference between the running log repetition text generated by the quantified model and the actual running log example; Based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example, the network parameters of the distributed collaborative scheduling decision network are optimized until they meet the network optimization requirements, thereby obtaining a distributed collaborative scheduling decision network that has been debugged.

2. The method according to claim 1, characterized in that The distributed collaborative scheduling decision network includes a first decision branch component and a second decision branch component, the first decision branch component includes a state feature mining core and a state feature processing core, the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example are used as input information of the distributed collaborative scheduling decision network, and the feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode is output, including: Taking the smart device cluster operation log example in the network optimization example and the initial smart device cluster operation log corresponding to the smart device cluster operation log example as input information of the state feature mining core, outputting the operation state residual semantic vector of the smart device cluster operation log example and the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the smart device cluster operation log example; Taking the running state residual semantic vector of the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example as the input information of the state feature processing core, outputting the u-th order running state residual semantic vector of the smart device cluster running log example and the u-th order running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and outputting the running log retelling text of the smart device cluster running log example, where u is an integer greater than or equal to 1 and less than or equal to X, and X is the total order of the pyramid reversible subnet of the state feature processing core; wherein the pyramid reversible subnet is a deconvolution subnet; Through the second decision branch component, based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example, a feasibility score of the smart device cluster operation log example matching the target scheduling strategy pattern is output.

3. The method according to claim 2, characterized in that The first decision branch component further includes a semantic upgrade unit, and the method further includes: Enter each smart device cluster operation log example in the network optimization example set into the semantic upgrading unit, and output the operation state residual enhanced semantic vector of the smart device cluster operation log example and the u-th order operation state residual enhanced semantic vector of the smart device cluster operation log example; The operating state residual semantic vector of the second smart device cluster operation log example in the smart device cluster operation log example is updated to the operating state residual enhanced semantic vector of the second smart device cluster operation log example, and the u-th order operating state residual semantic vector of the second smart device cluster operation log example is updated to the u-th order operating state residual enhanced semantic vector of the second smart device cluster operation log example.

4. The method according to claim 3, characterized in that The semantic upgrade unit is used for: Extracting the target distributed task status highlight text of the smart device cluster operation log example, and adjusting the size of the target distributed task status highlight text of the smart device cluster operation log example according to the text box size of the highlight text of the corresponding order used to generate the first neural network training cost algorithm, to obtain the target highlight text; Take any variable Q within the set operation value quadrant. If the Q is greater than the preset value, a first semantic upgrade rule is used to perform semantic upgrade based on the target highlighted text. If the Q is less than or equal to the preset value, a second semantic upgrade rule is used to perform semantic upgrade based on the target highlighted text.

5. The method according to claim 2, characterized in that: The second decision branch component includes a scheduling mode identification unit and a scheduling mode matching unit. The second decision branch component outputs a feasibility score of the smart device cluster operation log example matching the target scheduling strategy mode based on the smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example, including: The smart device cluster operation log example and the operation log retelling text of the smart device cluster operation log example are entered into the scheduling mode determination unit, and the scheduling mode determination unit determines a first feasibility identification label based on the smart device cluster operation log example, the operation log retelling text of the smart device cluster operation log example, the pyramid state semantic relationship network extracted by the scheduling mode determination unit, the fully connected layer corresponding to the state feature mining core, and the fully connected layer corresponding to the pyramid reversible subnet of the state feature processing core; The running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid state semantic relationship network are entered into the scheduling mode matching unit, the scheduling mode matching unit semantically downsamples the pyramid state semantic relationship network to obtain the pyramid downsampled semantic relationship network, and the scheduling mode matching unit determines a second feasibility identification label based on the running state residual semantic vector of the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example, and the pyramid downsampled semantic relationship network; wherein the pyramid state semantic relationship network is a structure extracted by the scheduling mode discrimination unit, and is used to express the semantic relationship between device running states; The first feasibility identification tag and the second feasibility identification tag are processed based on weights to obtain a feasibility score for the smart device cluster operation log example matching the target scheduling strategy pattern and output it.

6. The method according to claim 1, characterized in that The step of generating a first neural network training cost algorithm based on the running state residual semantic vector of the smart device cluster running log example obtained in the retelling task and the running state residual semantic vector of the initial smart device cluster running log corresponding to the smart device cluster running log example includes: Calculate a first semantic quantization difference based on the running state residual semantic vector of the second smart device cluster running log example in the smart device cluster running log example and the running state residual semantic vector of the initial smart device cluster running log corresponding to the second smart device cluster running log example; Calculating a second semantic quantization difference based on the running state residual semantic vector of the first smart device cluster running log example in the smart device cluster running log example, the running state residual semantic vector of the initial smart device cluster running log corresponding to the first smart device cluster running log example, and the number of the first smart device cluster running log example in the smart device cluster running log example; The ratio of the second semantic quantization difference to the first operation value is determined as the first neural network training cost algorithm, and the first operation value is the sum of the first semantic quantization difference and the set network parameter.

7. The method according to claim 2, characterized in that The generating a second neural network training cost algorithm based on the operation log retelling text of the smart device cluster operation log example and the smart device cluster operation log example includes: If the smart device cluster operation log example is the first smart device cluster operation log example, a cross entropy cost between the first smart device cluster operation log example and the operation log retelling text of the first smart device cluster operation log example is determined as a second neural network training cost algorithm; If the smart device cluster operation log example is a second smart device cluster operation log example, a second neural network training cost algorithm is generated based on the operation state residual semantic vector of the second smart device cluster operation log example, the operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the second smart device cluster operation log example, the u-th order operation state residual semantic vector of the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the initial smart device cluster operation log corresponding to the second smart device cluster operation log example, the operation log retelling text of the second smart device cluster operation log example, the highlighted text of the state feature mining core, and the highlighted text of the pyramid reversible subnet of the state feature processing core.

8. The method according to claim 1, characterized in that The optimizing of the network parameters of the distributed collaborative scheduling decision network based on the first neural network training cost algorithm, the second neural network training cost algorithm, the feasibility score of the smart device cluster operation log example, and the prior learning knowledge of the smart device cluster operation log example includes: Generate a third neural network training cost algorithm based on the feasibility score of the smart device cluster operation log example and the prior learning knowledge of the smart device cluster operation log example; Based on the global operation results of the first neural network training cost algorithm, the second neural network training cost algorithm and the third neural network training cost algorithm, a fourth neural network training cost algorithm is obtained; Based on the fourth neural network training cost algorithm, network parameters of the distributed collaborative scheduling decision network are reverse optimized.

9. An AI collaborative scheduling system, characterized in that: The AI ​​collaborative scheduling system includes a processor and a memory that communicate with each other, and the processor is used to call a computer program from the memory and implement the method described in any one of claims 1 to 8 by running the computer program.

Images