Multi-task parallel processing method and system based on AI target identification

Abstract

The invention provides a multi-task parallel processing method and system based on AI target recognition, and the method comprises the steps: carrying out the target recognition analysis of input multi-mode data, obtaining an initial task list, and generating a parallel recognition task sequence through combining the context information of a system; performing adaptive feature extraction and optimization on the multi-modal data according to the parallel recognition task sequence to obtain a task specific feature set; based on the task specificity feature set and the system context information, optimizing a task execution sequence and resource allocation by using a deep reinforcement learning model to obtain a preliminary optimization scheme; performing task preheating and dynamic task merging on the preliminary optimization scheme to obtain and execute an optimization task execution scheme. According to the method, the processing strategy can be dynamically adjusted according to the multi-modal data features and the task requirements, the task execution sequence and resource allocation are effectively optimized, and the adaptability and efficiency of a target recognition system in a complex industrial environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of multi-task parallel technology, and in particular to a multi-task parallel method and system based on artificial intelligence target recognition. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, AI-based target recognition technology has been widely used in product quality inspection and production process monitoring. Traditional target recognition methods are usually optimized for a single task and have limitations when dealing with complex multi-task parallel scenarios. Especially when dealing with multimodal data, existing technologies often use fixed feature extraction and task processing processes, which are difficult to effectively adapt to the specific needs of different tasks and dynamically changing production environments.

[0003] This inflexible approach results in the system being unable to effectively allocate and utilize computing resources during multi-task parallel processing, resulting in resource waste or inefficient execution of certain tasks. At the same time, due to the lack of consideration of the mutual influence between tasks, interference between tasks may occur, affecting recognition accuracy and overall system performance. Summary of the invention

[0004] The main purpose of the present invention is to solve the technical problems of the lack of flexibility and low resource utilization efficiency of the existing artificial intelligence-based target recognition system when processing multi-task parallel scenarios; A first aspect of the present invention provides a multi-task parallel method based on artificial intelligence target recognition, and the multi-task parallel method based on artificial intelligence target recognition comprises: Perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; According to each task of the parallel recognition task sequence, adaptive feature extraction is performed on the multimodal data, and feature optimization processing is performed based on feature usage of the extracted adaptive features to obtain a task-specific feature set; Based on the task-specific feature set and system context information, the preset deep reinforcement learning model is used to optimize the task execution order and resource allocation to obtain a preliminary optimization plan. Task preheating and dynamic task merging are performed on the preliminary optimization plan to obtain and execute the optimized task execution plan.

[0005] Optionally, in a first implementation of the first aspect of the present invention, performing target recognition analysis on the input multimodal data, identifying product types and detection requirements, obtaining an initial task list, and generating a parallel recognition task sequence based on the initial task list and in combination with predefined system context information includes: Perform data preprocessing on the input multimodal data, and use a preset multimodal fusion network to perform feature fusion on the preprocessed multimodal data to obtain a fusion feature map; Use the preset target recognition model to perform product type recognition and defect detection based on the fused feature map to obtain an initial task list; For each task in the initial task list, a task priority score is calculated in combination with predefined system context information, wherein the system context information includes current production plan, equipment status, and historical quality data; The initial task list is sorted according to the task priority scores, and task dependency analysis is applied to generate a parallel identification task sequence.

[0006] Optionally, in a second implementation of the first aspect of the present invention, sorting the initial task list according to the task priority scores and applying task dependency analysis to generate a parallel recognition task sequence includes: Applying a quick sorting algorithm to the initial task list based on the task priority scores of each task in the initial task list to obtain a preliminary sorted task list; Topologically sort the tasks in the preliminary sorted task list to obtain a task sequence that takes into account dependencies; A parallelism analysis algorithm is applied to the task sequence considering dependencies to identify the set of tasks that can be executed in parallel and obtain the task parallel group. Then, a parallel identification task sequence is generated according to the task parallel group and the system resource constraints in the system context information.

[0007] Optionally, in a third implementation of the first aspect of the present invention, performing adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, performing feature optimization processing based on feature usage of the extracted adaptive features, and obtaining a task-specific feature set includes: According to the input multimodal data and predefined common features, the dynamic feature pool is updated, and from the updated dynamic feature pool, the attention mechanism is used to select and weight relevant features according to the task type of each task in the parallel recognition task sequence to obtain a task-related feature set; The adaptive feature selection algorithm determines the redundancy of the task-related feature set, selects the optimal feature subset based on the redundancy, and obtains the preferred feature set; Using transfer learning technology, the knowledge of the pre-trained model is transferred to the current task domain, and the model of the current task domain after migration is used for feature extraction and optimization to obtain a task-specific feature set.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the dynamic feature pool is updated according to the input multimodal data and the predefined common features, and from the updated dynamic feature pool, the attention mechanism is used to select and weight relevant features according to the task type of each task in the parallel recognition task sequence, so that the task-related feature set is obtained, including: Apply feature extraction algorithm to the input multimodal data to obtain the current data feature set, calculate the similarity between the current data feature set and the predefined common features, filter new features based on the similarity threshold, and obtain the candidate new feature set; Calculating feature importance scores for features in the dynamic feature pool and candidate new feature sets, and updating the dynamic feature pool based on the feature importance scores to obtain an updated dynamic feature pool; According to the task type of each task in the parallel recognition task sequence, relevant features are selected from the updated dynamic feature pool to construct a task feature matrix; A multi-head attention mechanism is applied to the task feature matrix, and the attention weights between features are calculated to obtain the weighted task-related feature set.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, based on the task-specific feature set and system context information, the task execution order and resource allocation are optimized using a preset deep reinforcement learning model to obtain a preliminary optimization plan, and task preheating and dynamic task merging are performed on the preliminary optimization plan, and obtaining and executing the optimized task execution plan includes: Perform feature fusion and dimension compression on the task-specific feature set and system context information to obtain a state vector, and construct a deep reinforcement learning model with a preset Markov decision process input based on the state vector; The deep reinforcement learning model is used to perform forward propagation calculations on the state vector to obtain the Q value matrix. The ε-greedy strategy is then applied to select the optimal action based on the Q value matrix to generate an action sequence that includes the task execution order and resource allocation ratio, and a preliminary optimization plan is obtained. Calculate the similarity of each task in the preliminary optimization plan, perform cluster analysis on similar tasks, obtain a set of tasks that can be merged, and perform parameter sharing and computing resource integration on the set of tasks that can be merged to obtain a dynamically merged task list; The model parameters and related data of the tasks in the dynamically merged task list are preloaded and cached in parallel to obtain the task preheating configuration, and based on the task preheating configuration and the dynamically merged task list, the preliminary optimization plan is adjusted and reordered to obtain the optimized task execution plan.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the deep reinforcement learning model is used to perform forward propagation calculation on the state vector to obtain a Q value matrix, and the ε-greedy strategy is applied according to the Q value matrix to select the optimal action, generate an action sequence including the task execution order and the resource allocation ratio, and obtain a preliminary optimization plan including: The input state vector is normalized to obtain a standardized state vector, and the standardized state vector is input into the multi-layer perceptron network of the deep reinforcement learning model. The original Q value output is obtained through continuous operation of the activation function and the weight matrix; The double Q-learning algorithm is applied to compare and fuse the original Q-value output with the output of the target Q network to obtain a revised Q-value matrix. The expected benefit of each possible action is calculated based on the revised Q-value matrix to obtain the action value evaluation result. Based on the action value evaluation results, combined with the exploration-exploitation balance mechanism of the ε-greedy strategy, the action selection probability distribution is generated, and the Monte Carlo sampling method is used to sample from the probability distribution to obtain a preliminary action sequence of task execution order and resource allocation ratio; A task dependency constraint check is applied to the preliminary action sequence to make the preliminary action sequence satisfy the logical dependency relationship between tasks, an optimized action sequence satisfying the constraint is obtained, and the optimized action sequence is decoded into a preliminary optimization solution.

[0011] A second aspect of the present invention provides a multi-task parallel system based on artificial intelligence target recognition, the multi-task parallel system based on artificial intelligence target recognition comprising: The task generation module is used to perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; A feature extraction module, configured to perform adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, and perform feature optimization processing based on feature usage of the extracted adaptive features to obtain a task-specific feature set; The task optimization module is used to optimize the task execution sequence and resource allocation based on the task-specific feature set and system context information using a preset deep reinforcement learning model to obtain a preliminary optimization plan, and perform task preheating and dynamic task merging on the preliminary optimization plan to obtain and execute the optimized task execution plan.

[0012] The multi-task parallel method and system based on artificial intelligence target recognition obtains the initial task list by performing target recognition analysis on the input multimodal data, and generates a parallel recognition task sequence in combination with the system context information; performs adaptive feature extraction and optimization on the multimodal data according to the parallel recognition task sequence to obtain a task-specific feature set; based on the task-specific feature set and system context information, uses a deep reinforcement learning model to optimize the task execution sequence and resource allocation to obtain a preliminary optimization plan; performs task preheating and dynamic task merging on the preliminary optimization plan to obtain and execute the optimized task execution plan. This method can dynamically adjust the processing strategy according to the multimodal data characteristics and task requirements, effectively optimize the task execution sequence and resource allocation, and improve the adaptability and efficiency of the target recognition system in complex industrial environments.

[0013] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of a first embodiment of a multi-task parallel method based on artificial intelligence target recognition in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of a multi-task parallel system based on artificial intelligence target recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0018] To facilitate understanding of this embodiment, a multi-task parallel method based on artificial intelligence target recognition disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 As shown, the method comprises the following steps: 101. Perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; In one embodiment of the present invention, the target recognition analysis is performed on the input multimodal data, the product type and detection requirements are identified, an initial task list is obtained, and based on the initial task list, combined with predefined system context information, a parallel recognition task sequence is generated, including: data preprocessing is performed on the input multimodal data, and feature fusion is performed on the preprocessed multimodal data using a preset multimodal fusion network to obtain a fusion feature map; product type recognition and defect detection are performed based on the fusion feature map using a preset target recognition model to obtain an initial task list; for each task in the initial task list, a task priority score is calculated in combination with predefined system context information, wherein the system context information includes the current production plan, equipment status, and historical quality data; the initial task list is sorted according to the task priority score, and task dependency analysis is applied to generate a parallel recognition task sequence.

[0019] Specifically, when preprocessing the input multimodal data, it is necessary to first perform differentiated cleaning and normalization operations on images, texts, and other sensor information. The image part usually includes the process of noise filtering and brightness correction, while the text and sensor data require the separation of invalid characters and the unification of measurement units or timestamps. Then, all modal data are aligned in space or time, and missing values ​​are filled by interpolation technology to avoid size mismatch or information truncation in the subsequent fusion stage. The multimodal fusion network receives these aligned input tensors and uses multi-channel convolution or attention-based grouping to extract features of different modalities respectively. In order to quantify the degree of coupling between the modalities, the network constructs an interaction matrix in the middle layer to complete element-level comparison of image feature maps and text or environmental sensor data in the same vector space, and then integrates significant intersections with nonlinear activation functions to generate fusion feature maps.

[0020] Specifically, when using the preset target recognition model to perform product type recognition and defect detection based on the fused feature map, it is necessary to first separate the key points or regions that are highly relevant to target recognition from the fused feature map, and extract potential candidate blocks that meet the size and position requirements through the convolution layer or region proposal network, and then use the fully connected layer or multi-head attention mechanism to map these candidate blocks to the class probability distribution and detection box coordinates. Product type recognition usually calculates the difference between the predicted label and the true label based on the classification loss function, while the defect detection branch needs to predict the specific location and type of the defect at the same time, and measure the degree of deviation between the detection box and the true defect area through regression loss. After the recognition process is completed, the system will write the obtained product type and the label and coordinates of each defect into the data structure to form an initial task list, and indicate the corresponding image or sensor segment index, so that the specific input sample and inference result can be retrieved in the subsequent stage. The initial task list is a split display of the current recognition requirements, including the feature range and computational cost required for each detection target, so that the scheduling module can further implement priority measurement; Specifically, for each task in the initial task list, combined with the predefined system context information, the production plan, equipment status and historical quality data need to be mapped to the corresponding weight function to calculate the task priority score. The common way to record the production plan is to indicate the product quantity and delivery requirements in time periods and batches. If the total batch size is large or the deadline is tight, the corresponding inspection task will be magnified in weight. The equipment status is provided by the operation monitoring and self-checking module, including indicators such as processor load, storage capacity and failure probability. If the load is at peak time, the priority assessment of new tasks needs to be appropriately reduced to avoid resource blocking. The historical quality data stores the periodic distribution and severity of various defects. After this information is matched with the defect type in the initial task list through hash index, the priority score will be corrected to highlight the urgency of detecting fault-prone or serious defects. After the above interactive mapping, the system will give a list of all task priorities, and the corresponding correction origins will be attached in the data structure, which is convenient for the decision-making layer to transparently review and maintain the evaluation process later; Specifically, the process of sorting the initial task list according to the task priority score and applying task dependency analysis to generate a parallel recognition task sequence includes two steps: first, the priority scores of all tasks are arranged in descending order using a quick sort or heap sort algorithm, and items with the same or similar scores are temporarily stored in the same section for subsequent comparison; then the dependency analysis is performed, by scanning the mapping definition of task input and output, to identify that if the running result of task A is the input of task B, A must be executed before B is started to avoid data reference exceptions. After the dependency analysis is completed, the order requirements between tasks can be recorded in the form of an acyclic graph. The system will perform parallelism detection on independent branches in the graph structure, that is, unrelated nodes are grouped into the same group for parallel execution. Once the resource competition between groups is within an acceptable range, the scheduler will try to schedule these subtasks in parallel to improve the overall computing efficiency. The resulting parallel recognition task sequence will record the sorting position of each task and the way it cooperates with other tasks, and store it in the built-in scheduling system for subsequent execution. If real-time updates to dependency patterns or priorities are required, an incremental update strategy can be applied to this graph structure. Once significant changes in input data or context information are detected, the priorities will be recalculated and the dependency order and parallel grouping will be automatically adjusted.

[0021] Furthermore, the method of sorting the initial task list according to task priority scores and applying task dependency analysis to generate a parallel identification task sequence includes: applying a quick sorting algorithm to the initial task list according to the task priority scores of each task in the initial task list to obtain a preliminary sorted task list; topologically sorting the tasks in the preliminary sorted task list to obtain a task sequence that takes into account dependencies; applying a parallelism analysis algorithm to the task sequence that takes into account dependencies to identify a set of tasks that can be executed in parallel to obtain a task parallel group, and generating a parallel identification task sequence based on the task parallel group and system resource constraints in the system context information.

[0022] Specific, think, last a few seconds In the process of sorting the initial task list according to the task priority score, it is necessary to first comprehensively evaluate the priority score of each task, which is calculated based on multi-dimensional information such as production plan, equipment status, and historical quality data. The quick sort algorithm is suitable for processing the priority sorting of a large number of tasks because its average time complexity is O(nlogn) and it shows high efficiency in practical applications. In specific implementation, each task in the initial task list is first sorted according to its priority score as the sorting key, the task list is recursively divided into sublists, and high-priority tasks are moved to the front of the list and low-priority tasks are moved to the back by selecting pivot elements. During the sorting process, the quick sort algorithm ensures that the entire task list is arranged from high to low according to the priority score by constantly comparing and exchanging the positions of tasks. After the sorting is completed, the preliminary sorted task list obtained not only reflects the importance of each task, but also provides an orderly basis for the subsequent dependency analysis. This process ensures that high-priority tasks can be given priority in resource allocation and scheduling, thereby improving the response speed and processing efficiency of the overall system.

[0023] After the preliminary sorting of the task list is completed, it needs to be topologically sorted to ensure that the execution order of the tasks is consistent with their dependencies. Topological sorting relies on building a directed acyclic graph (DAG) between tasks, where nodes represent tasks and edges represent dependencies between tasks. The specific implementation steps include first scanning each task in the preliminary sorting list, identifying its dependent predecessor tasks, and establishing corresponding directed edges in the graph. Subsequently, using a topological sorting algorithm, such as the Kahn algorithm or the depth-first search (DFS) algorithm, nodes with zero in-degree are gradually removed and their execution order is recorded, while the in-degree information of other nodes in the graph is updated until all tasks are sorted. Through topological sorting, a linear sequence can be obtained to ensure that all tasks have completed their dependent predecessor tasks when they are executed, thereby avoiding data reference errors or execution conflicts caused by improper task order. Topological sorting not only ensures the logical consistency of tasks, but also provides a clear execution framework for subsequent parallelism analysis, so that the system can follow the established dependency order when processing multiple tasks, ensuring the stability and reliability of the overall process.

[0024] After obtaining the task sequence that takes dependencies into account, the parallelism analysis algorithm needs to be applied to identify the set of tasks that can be executed in parallel, so as to generate an efficient parallel identification task sequence. The core of parallelism analysis is to identify the independence between tasks, that is, to determine which tasks can be executed simultaneously without relying on each other. In the specific implementation process, first check the dependencies of each task one by one according to the topological sorting results. If all the predecessor tasks of some tasks have been completed and the current system resources allow, these tasks can be grouped into a parallel group. At the same time, the system needs to reasonably allocate resources to support parallel execution based on current resource constraints, such as CPU, memory, bandwidth, etc. In order to optimize resource utilization, optimization algorithms such as linear programming or integer programming can be used to ensure that resource allocation achieves maximum parallelism while meeting the requirements of each task. For example, set the resource constraint condition as , the task resource requirement is , then the objective function can be expressed as maximizing the sum of the number of parallel tasks: Subject to in, is a binary variable, indicating the task Whether it is selected for parallel execution. By solving the above optimization model. The system can efficiently identify the set of tasks that can be executed in parallel under the current resource constraints, and then form a task parallel group. Finally, the parallel identification task sequence is generated by combining the resource constraints in the task parallel group and the system context information. This process not only improves the efficiency of multi-tasking, but also ensures the rational use of system resources and the coordination of task execution, thereby optimizing and accelerating the overall production process.

[0025] 102. Perform adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, and perform feature optimization processing based on feature usage of the extracted adaptive features to obtain a task-specific feature set; In one embodiment of the present invention, adaptive feature extraction is performed on the multimodal data according to each task of the parallel recognition task sequence, and feature optimization processing is performed based on the feature usage of the extracted adaptive features to obtain a task-specific feature set, which includes: updating a dynamic feature pool according to the input multimodal data and predefined common features, and from the updated dynamic feature pool, using an attention mechanism to select and weight relevant features according to the task type of each task in the parallel recognition task sequence to obtain a task-related feature set; an adaptive feature selection algorithm determines the redundancy of the task-related feature set, and selects an optimal feature subset based on the redundancy to obtain a preferred feature set; using transfer learning technology to transfer the knowledge of the pre-trained model to the current task domain, and using the model of the current task domain after migration to perform feature extraction and optimization to obtain a task-specific feature set.

[0026] Specifically, in the process of performing adaptive feature extraction on multimodal data according to each task of the parallel recognition task sequence, it is first necessary to integrate the input multimodal data with the predefined common features to dynamically update the feature pool. Specifically, the system receives multimodal data from different sensors, such as images, texts, and environmental parameters, and matches these data with the predefined common feature library. The update of the dynamic feature pool involves fusing the newly extracted features with the existing features while eliminating outdated or irrelevant features. This process is achieved by calculating the timeliness and relevance scores of each feature, and features with high scores are retained in the pool, while features with low scores are eliminated. For example, a weighted moving average method can be used to dynamically adjust feature weights: ; in, represents the feature weight at the current moment, is the weight at the previous moment, is the smoothing coefficient, is the weight of the newly extracted feature. In this way, the feature pool can timely reflect the latest production environment and task requirements. Next, the attention mechanism is used to select and weight relevant features according to the task type of each task in the parallel recognition task sequence to generate a task-related feature set. The attention mechanism dynamically adjusts the importance of features by calculating the correlation weight between task type and feature. For example, if a task focuses on surface defect detection, the system will enhance the weight of features related to surface texture and reduce the influence of features not related to size measurement. In the specific implementation, the attention weight can be calculated by the following formula ; in, Is the query to It is the key to is a vector of values, is the dimension of the key. Through the multi-head attention mechanism, multiple attention distributions can be calculated in parallel, further improving the flexibility and accuracy of the selection of features. Finally, the task-related features are obtained through weighted summation. Each task corresponds to a specific feature subset, ensuring that the feature extraction process can accurately meet the needs of each task. After obtaining the task-related feature set, the adaptive feature selection method judges its redundancy and selects the optimal feature subset based on the redundancy to obtain the preferred feature set. The judgment of redundancy is mainly achieved by calculating the correlation coefficient matrix between features. Assume that the task-related feature set contains Features, correlation coefficient matrix Elements Indicator With features By setting a correlation threshold , the system can identify highly correlated feature pairs and select representative feature subsets through dimensionality reduction techniques such as principal component analysis (PCA) or independent component analysis (ICA). and Correlation coefficient , then the features with greater information are retained and redundant features are eliminated. In addition, the feature selection method based on information entropy is also applied to measure the information contribution of each feature and select features with higher information entropy as part of the preferred feature set. This process not only reduces feature redundancy and computational complexity, but also improves the generalization ability and recognition accuracy of the model.

[0027] Finally, the knowledge of the pre-trained model is transferred to the current task domain using transfer learning technology, and the transferred model is used for feature extraction and optimization to obtain a task-specific feature set. The implementation steps of transfer learning include selecting a pre-trained model that is highly relevant to the current task, such as a convolutional neural network (CNN) trained on a large-scale image dataset, and then fine-tuning the model to adapt it to a specific production environment and task requirements. In the specific operation, the first few layers of the pre-trained model are frozen, and only the last few layers are trained to retain the general feature extraction capabilities, while adjusting the high-level features to adapt to task-specific requirements. For example, for surface defect detection tasks, the system adjusts the last fully connected layer to enable it to capture tiny defect features more sensitively. In addition, regularization techniques such as L2 regularization or Dropout are introduced in the transfer learning process to prevent overfitting and improve the robustness of the model on new tasks. In this way, the pre-trained model not only speeds up the feature extraction process, but also improves the quality and adaptability of the task-specific feature set by transferring knowledge, thereby achieving more efficient and accurate multi-task parallel processing.

[0028] Furthermore, the method updates the dynamic feature pool according to the input multimodal data and predefined common features, and selects and weights relevant features from the updated dynamic feature pool according to the task type of each task in the parallel recognition task sequence using the attention mechanism to obtain the task-related feature set, including: applying a feature extraction algorithm to the input multimodal data to obtain the current data feature set, and calculating the similarity between the current data feature set and the predefined common features, screening new features based on the similarity threshold, and obtaining a candidate new feature set; calculating feature importance scores for the features in the dynamic feature pool and the candidate new feature set, and updating the dynamic feature pool based on the feature importance scores to obtain an updated dynamic feature pool; selecting relevant features from the updated dynamic feature pool according to the task type of each task in the parallel recognition task sequence, and constructing a task feature matrix; applying a multi-head attention mechanism to the task feature matrix, calculating the attention weights between features, and obtaining a weighted task-related feature set.

[0029] Specifically, in the process of updating the dynamic feature pool according to the input multimodal data and predefined common features, it is first necessary to apply a feature extraction algorithm to the multimodal data to obtain the feature set of the current data. Specifically, the system processes data of different modalities separately through deep learning models, such as using convolutional neural networks (CNN) to extract spatial features from image data, using transformer architecture to parse semantic information in text data, and extracting data features of other modalities such as environmental parameters through specially designed sensor data processing modules. Through these feature extraction algorithms, multimodal data is converted into high-dimensional feature vectors, which not only contain the basic information of each modality, but also integrate cross-modal correlation features, thereby enhancing the expressive power of the data. Next, the system calculates the similarity between the current data feature set and the predefined common features, usually using metrics such as cosine similarity or Euclidean distance to measure the similarity between the two. Let the current data feature set be F_current and the predefined common features be F_common. By calculating the similarity between the two, the system can identify those redundant features that are highly correlated with the common features. Based on the set similarity threshold θ, the system selects new features with similarities below the threshold to form a candidate new feature set F_candidate. The purpose of this screening process is to ensure that the features retained in the dynamic feature pool are unique and highly relevant, and to avoid incorporating redundant features that highly overlap with common features into the feature pool, thereby improving the diversity and effectiveness of the feature pool. In this way, the system can dynamically adjust and optimize the feature pool to ensure that it always contains the most valuable features for the current task, laying a solid foundation for subsequent feature selection and optimization.

[0030] Specifically, after obtaining the candidate new feature set, the system needs to calculate the feature importance scores of the existing features in the dynamic feature pool and the features in the candidate new feature set to determine which features should be retained and which features should be eliminated. This process is usually based on the contribution of the feature in the current task, and the value of each feature is quantified through gradient weighting, Shapley value or other feature importance evaluation methods. Specifically, the system evaluates the impact of each feature on the task loss function, calculates its corresponding gradient value, and thus obtains the feature importance score. Features with higher scores indicate that they contribute significantly to task performance and should be retained first; while features with lower scores or redundant features should be removed from the dynamic feature pool. This process can not only reduce redundant information in the feature pool and reduce computational complexity, but also improve the generalization ability and recognition accuracy of the model. Through the dynamic adjustment of feature importance scores, the system can adapt to changes in the production environment and task requirements in real time, ensuring that the feature pool always contains the most relevant and effective features. Finally, the updated dynamic feature pool F_updated will have higher relevance and information density, providing high-quality input for subsequent task feature selection and optimization, and further improving the overall performance and efficiency of the system in multi-task parallel processing.

[0031] Specifically, after the dynamic feature pool is updated, the system needs to select relevant features from the updated dynamic feature pool according to the task type of each task in the parallel recognition task sequence, and construct a task feature matrix. This process ensures that each task can obtain the most relevant features by mapping the correlation between the task type and the feature. For example, for the surface defect detection task, the system will give priority to features related to texture, color change, etc.; while for the size measurement task, it will select features related to structural contour and edge sharpness. The construction process of the task feature matrix T includes the following steps: first, according to the type of each task, select features with high correlation from the dynamic feature pool; second, weight these features through the weight allocation mechanism to reflect their importance in the specific task. The weight allocation can be based on factors such as historical data, task requirements, and feature importance scores to ensure that the feature set in each task feature matrix has both high relevance and can fully meet the specific needs of the task. In this way, the task feature matrix not only contains the key information required for the task, but also reflects the importance of each feature through the weighting mechanism, thereby improving the accuracy and pertinence of feature selection. This process ensures that each task can obtain the most suitable feature support in the feature extraction stage, improving the overall recognition capability and accuracy of the system in multi-task parallel processing.

[0032] Specifically, finally, the system applies a multi-head attention mechanism to the constructed task feature matrix, calculates the attention weights between features, and finally obtains a weighted task-related feature set. This step can capture the complex relationship and diverse dependency patterns between features through the multi-head attention mechanism, thereby further improving the expressiveness of the feature set and the recognition accuracy of the task. Specifically, the task feature matrix T is divided into multiple sub-matrices, each of which is processed by an independent attention head to calculate a different attention weight matrix A_h. These attention weights reflect the relative importance and relevance of different features in the current task. Through the parallel calculation of multiple attention heads, the system can simultaneously capture the relationship between features from multiple angles. The outputs of multiple attention heads are concatenated and linearly transformed to finally generate a weighted task-related feature set F_task. This multi-head attention mechanism not only enhances the diversity and expressiveness of the feature set, but also improves the overall quality of the feature set and the recognition ability of the task by introducing contextual information. The weighted task-related feature set provides high-quality input for subsequent feature optimization processing and the generation of task-specific feature sets, ensuring that each task can be efficiently executed with the best feature support, thereby significantly improving the recognition ability and accuracy of the entire system in multi-task parallel processing.

[0033] 103. Based on the task-specific feature set and system context information, the preset deep reinforcement learning model is used to optimize the task execution order and resource allocation to obtain a preliminary optimization plan, and the preliminary optimization plan is used to perform task preheating and dynamic task merging to obtain and execute the optimized task execution plan.

[0034] In one embodiment of the present invention, based on the task-specific feature set and system context information, the task execution order and resource allocation are optimized using a preset deep reinforcement learning model to obtain a preliminary optimization plan, and task preheating and dynamic task merging are performed on the preliminary optimization plan. The optimized task execution plan is obtained and executed, including: performing feature fusion and dimension compression processing on the task-specific feature set and system context information to obtain a state vector, and constructing a Markov decision process according to the state vector to input the preset deep reinforcement learning model; using the deep reinforcement learning model to perform forward propagation calculation on the state vector to obtain a Q value matrix, and applying the Q value matrix according to the Q value matrix The ε-greedy strategy selects the optimal action and generates an action sequence including the task execution order and resource allocation ratio to obtain a preliminary optimization plan; the similarity calculation is performed on each task in the preliminary optimization plan, and the similar tasks are clustered to obtain a set of tasks that can be merged, and parameter sharing and computing resource integration are performed on the set of tasks that can be merged to obtain a dynamically merged task list; the model parameters and related data of the tasks in the dynamically merged task list are preloaded and cached in parallel to obtain the task preheating configuration, and based on the task preheating configuration and the dynamically merged task list, the preliminary optimization plan is adjusted and reordered to obtain an optimized task execution plan.

[0035] Specifically, in the process of feature fusion and dimension compression of task-specific feature sets and system context information, it is first necessary to effectively integrate the task-specific feature sets with the system context information. The task-specific feature set contains high-dimensional feature vectors unique to each task, which reflect the specific information required by the task during execution. The system context information includes multi-dimensional data such as the current production plan, equipment status, and environmental parameters, which can dynamically reflect the real-time status of the production environment. In order to effectively fuse these two parts of information, the system adopts feature fusion technology to combine task-specific features with context information through splicing, weighted summation or other fusion methods to generate a comprehensive feature vector. This fusion process not only retains the uniqueness of task-specific features, but also introduces the dynamic changes of context information, so that the generated comprehensive feature vector can fully reflect the current production environment and task requirements.

[0036] Specifically, after completing feature fusion, in order to reduce the feature dimension and extract more representative features, the system further applies dimensionality compression techniques, such as principal component analysis (PCA) or autoencoder, to compress the high-dimensional comprehensive feature vector into a low-dimensional state vector. The purpose of this step is to reduce the computational complexity while retaining the main information of the features, thereby improving the processing efficiency and effect of subsequent models. The compressed state vector not only contains the key information in the fused features, but also removes redundancy and noise through the dimensionality reduction process to ensure the compactness and effectiveness of the state vector. Finally, the system uses these state vectors as the input of the Markov Decision Process (MDP) to construct an environmental description of the deep reinforcement learning model. In this way, the state vector can accurately reflect the overall state of the current system, provide a solid foundation for the decision-making of the deep reinforcement learning model, and enable it to perform effective task scheduling and resource allocation in a complex production environment.

[0037] Specifically, in the process of forward propagation calculation of the state vector using the deep reinforcement learning model, the system first inputs the state vector into the preset deep reinforcement learning network. The network is usually composed of a multi-layer neural network that can process high-dimensional inputs and extract deep feature representations. Through forward propagation, the model calculates the Q value corresponding to each possible action, that is, the expected benefit of each action in the current state. These Q values ​​form a Q value matrix, which reflects the pros and cons of different actions in the current state. Next, the system applies the ε-greedy strategy to select the optimal action sequence from the Q value matrix. The ε-greedy strategy is a method of balancing exploration and utilization, in which the action with the highest Q value is selected most of the time, while a small part of the time, the action is randomly selected to ensure that the model can explore more possibilities. After generating an action sequence containing the order of task execution and the ratio of resource allocation, the system obtains a preliminary optimization plan. The plan specifies in detail the execution order of each task and the ratio of resources allocated to each task, aiming to maximize the efficiency and output of the overall system.

[0038] Specifically, after obtaining the preliminary optimization plan, the system calculates the similarity of each task and performs cluster analysis based on the similarity to identify a set of tasks that can be merged. Similarity calculation is usually based on the feature vectors in the task-specific feature set, and quantifies the similarity between tasks by calculating the distance or similarity index between tasks, such as cosine similarity, Euclidean distance, etc. By setting a similarity threshold, the system can effectively classify tasks with similarity higher than the threshold into the same category to form a set of tasks that can be merged. Subsequently, the system performs parameter sharing and computing resource integration on these mergeable task sets, that is, multiple similar tasks share the same set of model parameters and computing resources to reduce computing redundancy and resource waste. This process not only improves the computing efficiency of the system, but also enhances the generalization ability of the model through parameter sharing. Finally, the system generates a dynamically merged task list, which contains the task set after cluster analysis and resource integration, providing a basis for subsequent task warm-up and optimization of task execution plans.

[0039] Specifically, in the dynamically merged task list, the system performs parallel preloading and cache optimization on the model parameters and related data of each task to achieve task preheating configuration. The purpose of task preheating configuration is to load the required model parameters and data into the cache in advance before the task is officially executed, reduce the delay when switching tasks, and improve execution efficiency. In the specific implementation process, the system dynamically loads the relevant models and data into efficient storage media, such as GPU cache or memory, according to the task type and resource requirements in the task list. By optimizing the cache strategy, it ensures that the tasks to be executed can quickly access the required resources and avoid execution delays caused by long resource loading time. In addition, the system will adjust and reorder the tasks according to the priority and resource allocation ratio of the tasks to ensure that high-priority tasks can obtain resources and execution opportunities first. Finally, based on the task preheating configuration and the dynamically merged task list, the system refines and optimizes the preliminary optimization plan to generate the final optimized task execution plan.

[0040] Furthermore, the method of using the deep reinforcement learning model to perform forward propagation calculation on the state vector to obtain a Q value matrix, and applying the ε-greedy strategy to select the optimal action according to the Q value matrix to generate an action sequence including the task execution order and the resource allocation ratio, and obtaining a preliminary optimization scheme includes: normalizing the input state vector to obtain a standardized state vector, and inputting the standardized state vector into the multi-layer perceptron network of the deep reinforcement learning model, and obtaining the original Q value output through continuous operation of the activation function and the weight matrix; applying the double Q learning algorithm to compare and fuse the original Q value output with the output of the target Q network to obtain a revised Q value matrix, and calculating the expected benefit of each possible action according to the revised Q value matrix to obtain an action value evaluation result; based on the action value evaluation result, combined with the exploration-utilization balance mechanism of the ε-greedy strategy, generating an action selection probability distribution, and using the Monte Carlo sampling method to sample from the probability distribution to obtain a preliminary action sequence of the task execution order and the resource allocation ratio; applying task dependency constraint check to the preliminary action sequence so that the preliminary action sequence satisfies the logical dependency relationship between tasks, obtaining an optimized action sequence that satisfies the constraints, and decoding the optimized action sequence into a preliminary optimization scheme.

[0041] Specifically, in the process of using the deep reinforcement learning model to perform forward propagation calculations on the state vector and generate a preliminary optimization plan, the input state vector must first be normalized to ensure that each feature is within the same scale range, so as to avoid some features having an adverse effect on model training due to excessive or small values. Normalization usually uses a standardization method to subtract the mean of each feature value and divide it by the standard deviation to obtain a standardized state vector. This process not only improves the training stability of the model, but also accelerates the convergence speed. The standardized state vector is input into the multi-layer perceptron (MLP) network of the deep reinforcement learning model, and feature information is extracted and combined layer by layer through multi-layer linear transformations and nonlinear activation functions such as ReLU or Sigmoid. The neurons in each layer are linearly combined with the output of the previous layer through the weight matrix, and then nonlinearity is introduced through the activation function, so that the network can learn complex feature representations. After being processed by the multi-layer perceptron, the model generates raw Q value outputs, which represent the expected benefits of each possible action in the current state. The key to this stage lies in the depth and width design of the multi-layer perceptron network, which determines the model's ability to express the state vector and its level of complexity processing.

[0042] Specifically, after obtaining the original Q-value output, the system applies the dual Q-learning algorithm to correct the Q-value matrix to reduce the deviation and over-estimation problems in the Q-value estimation. Dual Q-learning introduces the target Q-network to compare and fuse the output of the original Q-network with the output of the target network to obtain a more accurate Q-value matrix. Specifically, the system first uses the original Q-network to forward propagate the state vector to obtain the Q-value estimate under the current strategy, and at the same time, the target Q-network independently forward propagates the same state vector to obtain the Q-value estimate under the target strategy. Subsequently, the system weightedly fuses the outputs of the two to generate a corrected Q-value matrix. This process effectively separates the process of action selection and action evaluation, and reduces the Q-value deviation caused by action selection. In addition, the system also calculates the expected benefit of each possible action based on the corrected Q-value matrix to form an action value evaluation result. This evaluation result not only reflects the potential benefit of the action in the current state, but also provides a scientific basis for subsequent action selection, ensuring that the generated action sequence is optimal and efficient in terms of resource allocation and task execution order.

[0043] Specifically, based on the action value evaluation results, the system combines the exploration and utilization balance mechanism of the ε-greedy strategy to generate an action selection probability distribution, and samples from it through the Monte Carlo sampling method to obtain a preliminary action sequence including the task execution order and resource allocation ratio. The ε-greedy strategy plays a key role in this process. By setting an exploration rate ε, the system selects the action with the highest current value (utilization) most of the time, and randomly selects other actions (exploration) in a small part of the time to prevent falling into the local optimal solution. Specifically, the system first constructs a probability distribution based on the action value evaluation results, in which high-value actions have a higher selection probability, while low-value actions have a lower probability. Then, the system uses the Monte Carlo sampling method to randomly extract actions from the probability distribution to generate a diverse and exploratory action sequence. This action sequence specifies in detail the execution order of each task and the resource ratio allocated to each task, forming a preliminary optimization plan. In this way, the system can not only make full use of the high-yield characteristics of the current optimal action, but also discover potential better strategies through random exploration to improve the overall optimization effect.

[0044] Specifically, after generating the preliminary optimization plan, the system calculates the similarity of each task in it, and performs cluster analysis based on the similarity to identify the task sets that can be merged. This process quantifies the similarity between tasks by evaluating the feature vectors in the task-specific feature set, and then classifies highly similar tasks into the same category. For example, the system identifies tasks that are similar in resource requirements, execution steps, or target results by calculating the cosine similarity or other similarity indicators between tasks. These similar task sets are then clustered to form mergeable task groups. The system performs parameter sharing and computing resource integration on these mergeable task sets, reduces computing redundancy and resource waste by sharing model parameters and uniformly allocating computing resources, and improves the overall system operation efficiency. The dynamically merged task list not only optimizes resource utilization, but also improves the generalization ability of the model and the consistency of task execution through parameter sharing. Finally, the system generates an optimized task execution plan, which not only includes the task sequence after cluster analysis and resource integration, but also further improves the response speed and stability of task execution through parallel preloading and cache optimization. By refining and adjusting the preliminary optimization plan, the system can ensure that the generated optimized task execution plan satisfies the logical dependencies between tasks while achieving efficient and coordinated multi-task parallel processing, significantly improving the performance and production efficiency of the overall system.

[0045] In this embodiment, the input multimodal data is subjected to target recognition analysis to obtain an initial task list, and a parallel recognition task sequence is generated in combination with system context information; adaptive feature extraction and optimization are performed on the multimodal data according to the parallel recognition task sequence to obtain a task-specific feature set; based on the task-specific feature set and system context information, a deep reinforcement learning model is used to optimize the task execution order and resource allocation to obtain a preliminary optimization plan; task preheating and dynamic task merging are performed on the preliminary optimization plan to obtain and execute the optimized task execution plan. This method can dynamically adjust the processing strategy according to the multimodal data characteristics and task requirements, effectively optimize the task execution order and resource allocation, and improve the adaptability and efficiency of the target recognition system in a complex industrial environment.

[0046] The above describes the multi-task parallel method based on artificial intelligence target recognition in the embodiment of the present invention. The following describes the multi-task parallel system based on artificial intelligence target recognition in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a multi-task parallel system based on artificial intelligence target recognition includes: The task generation module 201 is used to perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; A feature extraction module 202 is used to perform adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, and perform feature optimization processing based on feature usage of the extracted adaptive features to obtain a task-specific feature set; The task optimization module 203 is used to optimize the task execution sequence and resource allocation based on the task-specific feature set and system context information using a preset deep reinforcement learning model to obtain a preliminary optimization plan, and perform task preheating and dynamic task merging on the preliminary optimization plan to obtain and execute the optimized task execution plan.

[0047] In an embodiment of the present invention, the multi-task parallel system based on artificial intelligence target recognition runs the multi-task parallel method based on artificial intelligence target recognition, and the multi-task parallel system based on artificial intelligence target recognition obtains an initial task list by performing target recognition analysis on the input multimodal data, and generates a parallel recognition task sequence in combination with system context information; performs adaptive feature extraction and optimization on the multimodal data according to the parallel recognition task sequence to obtain a task-specific feature set; based on the task-specific feature set and system context information, uses a deep reinforcement learning model to optimize the task execution sequence and resource allocation to obtain a preliminary optimization plan; performs task preheating and dynamic task merging on the preliminary optimization plan to obtain and execute the optimized task execution plan. This method can dynamically adjust the processing strategy according to the multimodal data characteristics and task requirements, effectively optimize the task execution sequence and resource allocation, and improve the adaptability and efficiency of the target recognition system in a complex industrial environment.

[0048] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0049] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0050] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-task parallel processing method based on AI target recognition, characterized in that: The multi-task parallel processing method based on AI target recognition includes: Perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; According to each task of the parallel recognition task sequence, adaptive feature extraction is performed on the multimodal data, and feature optimization processing is performed based on feature usage of the extracted adaptive features to obtain a task-specific feature set; Based on the task-specific feature set and system context information, the preset deep reinforcement learning model is used to optimize the task execution order and resource allocation to obtain a preliminary optimization plan. Task preheating and dynamic task merging are performed on the preliminary optimization plan to obtain and execute the optimized task execution plan.

2. The multi-task parallel processing method based on AI target recognition according to claim 1 is characterized in that: The target recognition analysis is performed on the input multimodal data to identify the product type and the detection requirements, and an initial task list is obtained. Based on the initial task list and in combination with predefined system context information, a parallel recognition task sequence is generated, including: Perform data preprocessing on the input multimodal data, and use a preset multimodal fusion network to perform feature fusion on the preprocessed multimodal data to obtain a fusion feature map; Use the preset target recognition model to perform product type recognition and defect detection based on the fused feature map to obtain an initial task list; For each task in the initial task list, a task priority score is calculated in combination with predefined system context information, wherein the system context information includes current production plan, equipment status, and historical quality data; The initial task list is sorted according to the task priority scores, and task dependency analysis is applied to generate a parallel identification task sequence.

3. The multi-task parallel processing method based on AI target recognition according to claim 2 is characterized in that: The step of sorting the initial task list according to the task priority scores and applying task dependency analysis to generate a parallel recognition task sequence includes: Applying a quick sorting algorithm to the initial task list based on the task priority scores of each task in the initial task list to obtain a preliminary sorted task list; Topologically sort the tasks in the preliminary sorted task list to obtain a task sequence that takes into account dependencies; A parallelism analysis algorithm is applied to the task sequence considering dependencies to identify the set of tasks that can be executed in parallel and obtain the task parallel group. Then, a parallel identification task sequence is generated according to the task parallel group and the system resource constraints in the system context information.

4. The multi-task parallel processing method based on AI target recognition according to claim 2 is characterized in that: The step of performing adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, and performing feature optimization processing based on feature usage of the extracted adaptive features to obtain a task-specific feature set includes: According to the input multimodal data and predefined common features, the dynamic feature pool is updated, and from the updated dynamic feature pool, the attention mechanism is used to select and weight relevant features according to the task type of each task in the parallel recognition task sequence to obtain a task-related feature set; The adaptive feature selection algorithm determines the redundancy of the task-related feature set, selects the optimal feature subset based on the redundancy, and obtains the preferred feature set; Using transfer learning technology, the knowledge of the pre-trained model is transferred to the current task domain, and the model of the current task domain after migration is used for feature extraction and optimization to obtain a task-specific feature set.

5. The multi-task parallel processing method based on AI target recognition according to claim 4 is characterized in that: The method updates the dynamic feature pool according to the input multimodal data and the predefined common features, and selects and weights relevant features from the updated dynamic feature pool according to the task type of each task in the parallel recognition task sequence using the attention mechanism, so as to obtain a task-related feature set including: Apply feature extraction algorithm to the input multimodal data to obtain the current data feature set, calculate the similarity between the current data feature set and the predefined common features, filter new features based on the similarity threshold, and obtain the candidate new feature set; Calculating feature importance scores for features in the dynamic feature pool and candidate new feature sets, and updating the dynamic feature pool based on the feature importance scores to obtain an updated dynamic feature pool; According to the task type of each task in the parallel recognition task sequence, relevant features are selected from the updated dynamic feature pool to construct a task feature matrix; A multi-head attention mechanism is applied to the task feature matrix, and the attention weights between features are calculated to obtain the weighted task-related feature set.

6. The multi-task parallel processing method based on AI target recognition according to claim 1 is characterized in that: Based on the task-specific feature set and system context information, the preset deep reinforcement learning model is used to optimize the task execution order and resource allocation to obtain a preliminary optimization plan, and task preheating and dynamic task merging are performed on the preliminary optimization plan to obtain and execute the optimized task execution plan, including: Perform feature fusion and dimension compression on the task-specific feature set and system context information to obtain a state vector, and construct a deep reinforcement learning model with a preset Markov decision process input based on the state vector; The deep reinforcement learning model is used to perform forward propagation calculations on the state vector to obtain the Q value matrix. The ε-greedy strategy is then applied to select the optimal action based on the Q value matrix to generate an action sequence that includes the task execution order and resource allocation ratio, and a preliminary optimization plan is obtained. Calculate the similarity of each task in the preliminary optimization plan, perform cluster analysis on similar tasks, obtain a set of tasks that can be merged, and perform parameter sharing and computing resource integration on the set of tasks that can be merged to obtain a dynamically merged task list; The model parameters and related data of the tasks in the dynamically merged task list are preloaded and cached in parallel to obtain the task preheating configuration, and based on the task preheating configuration and the dynamically merged task list, the preliminary optimization plan is adjusted and reordered to obtain the optimized task execution plan.

7. The multi-task parallel processing method based on AI target recognition according to claim 6 is characterized in that: The deep reinforcement learning model is used to perform forward propagation calculation on the state vector to obtain the Q value matrix, and the ε-greedy strategy is applied according to the Q value matrix to select the optimal action, generate an action sequence including the task execution order and resource allocation ratio, and obtain the preliminary optimization plan including: The input state vector is normalized to obtain a standardized state vector, and the standardized state vector is input into the multi-layer perceptron network of the deep reinforcement learning model. The original Q value output is obtained through continuous operation of the activation function and the weight matrix; The double Q-learning algorithm is applied to compare and fuse the original Q-value output with the output of the target Q network to obtain a revised Q-value matrix. The expected benefit of each possible action is calculated based on the revised Q-value matrix to obtain the action value evaluation result. Based on the action value evaluation results, combined with the exploration-exploitation balance mechanism of the ε-greedy strategy, the action selection probability distribution is generated, and the Monte Carlo sampling method is used to sample from the probability distribution to obtain a preliminary action sequence of task execution order and resource allocation ratio; A task dependency constraint check is applied to the preliminary action sequence to make the preliminary action sequence satisfy the logical dependency relationship between tasks, an optimized action sequence satisfying the constraint is obtained, and the optimized action sequence is decoded into a preliminary optimization solution.

8. A multi-task parallel processing system based on AI target recognition, characterized in that: The multi-task parallel processing system based on AI target recognition includes: The task generation module is used to perform target recognition analysis on the input multimodal data, identify product types and detection requirements, obtain an initial task list, and generate a parallel recognition task sequence based on the initial task list and predefined system context information; A feature extraction module, configured to perform adaptive feature extraction on the multimodal data according to each task of the parallel recognition task sequence, and perform feature optimization processing based on feature usage of the extracted adaptive features to obtain a task-specific feature set; The task optimization module is used to optimize the task execution sequence and resource allocation based on the task-specific feature set and system context information using a preset deep reinforcement learning model to obtain a preliminary optimization plan, and perform task preheating and dynamic task merging on the preliminary optimization plan to obtain and execute the optimized task execution plan.

Images