Intelligent scheduling and distribution method for protein complex detection task in cloud platform

Abstract

The application discloses a kind of intelligent scheduling and distribution method of protein complex detection task in cloud platform, it is related to big data management technical field;By constructing task feature model;Real-time sequencing is carried out to task using dynamic priority evaluation function containing vulnerability index, and dynamic scheduling sequence is generated;Using improved space-time graph convolutional neural network to construct resource matching prediction model, and generating resource-task matching matrix;Using improved adaptive genetic algorithm, with minimizing completion time and maximizing equipment utilization as target, under the constraint of resource-task matching matrix, the optimal allocation strategy and execution instruction set are generated;Through abnormal hierarchical judgment and dynamic adjustment resource allocation, and generate exception traceability feedback report;Realize biological characteristic perception and physical resource state prediction deep coupling, improve the survival rate of easy-degradable sample and the utilization efficiency of experimental equipment, guarantee the intelligentization and high robustness of cloud platform detection whole process.

Description

Technical Field

[0001] This invention belongs to the field of big data management technology, and more specifically, it relates to an intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform. Background Technology

[0002] Protein complex detection is a core component of modern life science research and development, drug target discovery, and disease mechanism research. With the deep integration of biotechnology and cloud computing, the "cloud experiment" model—where testing requests are submitted through cloud platforms and executed by automated back-end laboratories—has become an industry trend. In this model, efficiently managing cross-regional, multi-category, and high-frequency testing tasks is crucial for improving research output efficiency.

[0003] Existing cloud platform scheduling methods largely borrow from general industrial manufacturing or IT resource scheduling logic, primarily employing First-Come, First-Served (FIFO) or simple priority algorithms based on deadlines. At the execution level, static scheduling typically relies on Laboratory Information Management Systems (LIMS). However, protein complex samples exhibit extremely high biological activity and are readily degradable. Their detection processes involve multiple coupled steps such as ultracentrifugation, cross-linking, enzyme digestion, and mass spectrometry, and place extremely high demands on equipment precision and real-time reagent availability. While existing models such as BiLSTM can process text, they suffer from gaps in closed-loop linkage with the real-time status of underlying hardware (such as IoT sensor data). In particular, existing resource matching is mostly based on static rules, failing to anticipate potential resource depletion or equipment fatigue failures during experiments. This results in a lack of foresight in scheduling schemes, making it highly susceptible to sample rejection due to resource bottlenecks during execution.

[0004] This reveals the following problems with existing technologies: 1. Lack of deep understanding of the biological characteristics of samples (such as half-life), leading to high-risk samples becoming invalid due to excessive queuing time; 2. Lack of dynamic adaptive capabilities in scheduling algorithms, making it impossible to achieve global optimization of equipment utilization and delivery cycle under multi-process constraints; 3. Lack of deep prediction of experimental resource status, resulting in a passive rescheduling mechanism that cannot provide early warning of potential risks. Summary of the Invention

[0005] (a) Technical problems to be solved To address the problems in related technologies, this invention provides an intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform, thereby overcoming the aforementioned technical problems existing in the prior art.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. Receive protein complex detection requests from cloud platform users, extract multi-dimensional core features including sample type, detection requirements, sample quantity, sample half-life, and delivery deadline; construct a task feature model based on the multi-dimensional core features; S2. Based on the task feature model, the dynamic priority evaluation function is used to score and prioritize all protein complex detection tasks in the cloud platform task pool in real time to obtain a dynamic task scheduling sequence. S3. Real-time acquisition platform background experimental detection resources' real-time status and processing capabilities to obtain real-time resource data; use historical resource data to train an improved spatiotemporal graph convolutional neural network to obtain a resource matching prediction model; Real-time resource data is input into the resource matching prediction model to obtain resource status prediction results; the resource status prediction results are correlated and mapped with the dynamic task scheduling sequence output by S2 and the matching degree is calculated to obtain the resource-task matching matrix; S4. Using an improved adaptive genetic algorithm, with the optimization objectives of minimizing the total completion time and maximizing equipment utilization, find the optimal detection task allocation strategy under the constraints of the resource-task matching matrix, and parse it into an execution instruction set; S5. Monitor the execution progress and resource occupancy rate of the execution instruction set on the experimental terminal in real time, and make anomaly classification judgment: if it is a minor anomaly, the execution local time window is extended and adjusted; if it is a severe anomaly, the affected task is frozen, the improved adaptive genetic algorithm in S4 is called to reconfigure the resources of the affected task, obtain the second task allocation strategy and update the execution instruction set, and generate an anomaly traceability feedback report. This invention achieves deep coupling between task requirements and physical resources by constructing a task feature-based model and a resource matching prediction model; it addresses the time-sensitive issue of easily degradable samples by utilizing a dynamic priority evaluation function, and seeks the globally optimal allocation scheme under multi-objective constraints by combining an improved adaptive genetic algorithm, while coping with the complex dynamic environment of the laboratory through closed-loop monitoring and rescheduling mechanisms; it improves the biological safety and resource utilization efficiency of protein complex detection; by quantifying sample vulnerability, it shortens the average waiting time for easily degradable samples, effectively ensuring the authenticity of experimental data; the application of the adaptive genetic algorithm improves the overall utilization rate of precision instruments; it introduces anomaly classification judgment and differentiated feedback routes to avoid invalid global rescheduling caused by minor fluctuations; it can quickly respond and reorganize resources in the face of sudden equipment failures, and provides anomaly tracing feedback, solving the problems of delayed manual intervention, high risk of sample failure, and unbalanced resource allocation in traditional scheduling modes.

[0007] Preferably, step S1 includes the following steps: S11. The raw data of the protein complex detection application received by the cloud platform, which includes unstructured text descriptions and structured forms; S12. Utilize a pre-trained biological domain language model to perform named entity recognition and semantic parsing on unstructured text descriptions, extracting sample type entities and detection requirement entities; extract sample quantity values, sample half-life, delivery deadline timestamps, and user level codes from structured forms. S13. Calculate the sample vulnerability index based on the sample half-life; S14. After normalizing the sample type, detection requirements, sample quantity, delivery deadline timestamp and user level code, the sample is encapsulated into a task feature model containing a unique task identifier. This invention introduces natural language processing technology and a rigorous exponential decay mathematical model to accurately quantify the user's vague description of easy degradation into a vulnerability feature value between 0 and 1. This design eliminates the subjectivity and lag of manual assessment, enabling the scheduling system to perceive the biological timeliness pressure of protein samples in a machine-readable way at the source of task input, providing data support for subsequent scheduling.

[0008] Preferably, step S13 includes the following steps: S131. Obtain the theoretical half-life of the sample and set the maximum tolerance period for the platform benchmark. S132. Based on the sample half-life, the theoretical half-life of the sample, and the maximum tolerable decay period of the platform benchmark, the sample vulnerability index is calculated using the exponential decay mapping formula. This invention achieves a digital mapping of the timeliness of biological samples through an exponential decay function; it transforms ambiguous descriptions of easy degradation into high-precision continuous values, providing quantitative weights for subsequent scheduling algorithms, eliminating the arbitrariness of manual assessment, and improving the survival rate of high-risk samples.

[0009] Preferably, step S2 includes the following steps: S21. Input the detection requirements into the preset task standard working time mapping table, obtain the standard time consumption of each preprocessing and computer step required for the task, add up the standard time consumption of all steps and normalize it to obtain the task complexity. S22. Calculate the difference between the delivery deadline timestamp and the current system timestamp, and take its reciprocal to obtain the time urgency index; S23. Construct a dynamic priority evaluation function; the dynamic priority evaluation function includes a weight vector; the weight vector includes time weight, vulnerability weight, user level weight, and task complexity weight; S24. Adjust the weight vector according to the sample vulnerability index to obtain an adaptive weight vector; S25. The time urgency index, sample vulnerability index, user level and task complexity are weighted and calculated using the adaptive weight vector to obtain the dynamic priority score; all tasks in the task pool are sorted in descending order according to the dynamic priority score to obtain the dynamic task scheduling sequence. This invention designs a dynamic priority evaluation function that includes four dimensions: time, vulnerability, complexity, and user, and introduces a weight adaptive adjustment mechanism. When a sample is extremely vulnerable, the vulnerability weight is automatically amplified, making this feature dominant in the scoring. This non-linear dynamic scoring mechanism fully simulates the decision-making logic of a senior laboratory supervisor when facing sudden emergency samples, ensuring the survival rate of highly degradable samples and avoiding major scientific research losses caused by rigid rules.

[0010] Preferably, step S24 includes the following steps: S241. Define an initial weight vector, which includes an initial time weight, an initial vulnerability weight, an initial user level weight, and an initial task complexity weight. S242. Determine whether the sample vulnerability index exceeds the preset high-risk threshold. If it does, increase the proportion of vulnerability weight according to the exponential enhancement rule, and simultaneously reduce the proportion of other weights proportionally to obtain an adaptive weight vector. This invention achieves adaptive weight redistribution and establishes a non-linear priority response mechanism that automatically increases the weight when a sample faces the risk of degradation, ensuring that core biological constraints have the highest say in scheduling decisions and reducing the failure rate of experiments.

[0011] Preferably, step S3 includes the following steps: S31. The real-time resource data is obtained by collecting the operating status, estimated idle time and real-time reagent balance of the experimental testing equipment in real time through the MQTT protocol and API interface. S32. Based on the characteristics of laboratory resource data, the spatiotemporal graph convolutional neural network is improved to obtain an improved spatiotemporal graph convolutional neural network; the improvements include: S321. Improve the static spatial topology graph of the traditional spatiotemporal graph convolutional neural network into a dynamic spatial topology graph based on process coupling degree; S322. Improve the standard temporal convolutional layer of the traditional spatiotemporal graph convolutional neural network into a multi-scale dilated temporal convolutional layer; S323. Improve the traditional spatiotemporal graph convolutional neural network's single output layer into a multi-task joint output layer that includes idle time branches, reagent consumption branches, and fault warning branches; S33. Obtain historical laboratory operation log data to construct a training set, and use the training set to train the improved spatiotemporal graph convolutional neural network to obtain the resource matching prediction model; S34. Input the real-time resource data into the resource matching prediction model, and output the predicted idle time point, reagent depletion risk value and equipment reliability score of each device in the future execution cycle to obtain the resource status prediction result. S35. Perform a Cartesian product operation on the resource status prediction result and the dynamic task scheduling sequence output by S2, eliminate combinations that do not meet reagent constraints and whose reliability scores are lower than the safety threshold, calculate the matching degree score between each task and each available device, and obtain the resource-task matching matrix. This invention introduces an improved spatiotemporal graph convolutional neural network to construct a resource matching prediction model, achieving a leap from perceiving the current situation to predicting future moments. Compared with the traditional static resource matrix, this invention can predict equipment failure risks and reagent shortage trends in advance, and the generated resource-task matching matrix effectively filters high-risk resource allocation paths. This not only avoids sample scrapping due to equipment downtime or reagent depletion during the experiment, but also provides high-quality and safe search space constraints for subsequent genetic algorithms, improving the execution stability and security of the scheduling plan.

[0012] Preferably, step S4 includes the following steps: S41. Encode the scheduling scheme into a real number matrix chromosome, where rows of the matrix represent tasks, columns represent detection procedures, and matrix elements represent the allocated device numbers; based on the resource-task matching matrix, prioritize the allocation of devices with high matching scores to high-priority tasks in the dynamic task scheduling sequence, and generate an initial population containing multiple allocation schemes. S42. Define a bi-objective fitness function. The first objective is to minimize the completion time of all tasks on each device, and the second objective is to maximize the proportion of effective working time of all devices during task execution. S43. Perform adaptive genetic evolution operation, dynamically adjust the crossover probability and mutation probability according to the fitness variance of the current population, and retain the top-ranked individuals with the highest fitness scores in the next generation through a bi-objective fitness function. S44. Repeat S43 until the preset maximum number of iterations is reached. Output the Pareto optimal solution with the highest fitness as the detection task allocation strategy. Parse the strategy to generate machine-readable execution instructions, and send them to the control terminal of the corresponding device through the cloud platform. This invention transforms complex laboratory operations research problems into multi-objective optimization mathematical models. By introducing an adaptive genetic algorithm, the crossover and mutation probabilities are dynamically adjusted according to the population evolution state, effectively avoiding the algorithm from getting stuck in local optima (such as all tasks being crammed onto a slightly better mass spectrometer, causing the other one to be idle). The bi-objective fitness function finds a mathematical balance between "getting customers to receive reports as soon as possible" and "making expensive mass spectrometers run at full capacity".

[0013] Preferably, the dynamic adjustment of crossover probability and mutation probability in S43 includes the following steps: S431. Calculate the average fitness value and maximum fitness value of all individuals in the current population; S432. For crossover operations, when the fitness value of the two parent individuals participating in the crossover is higher than the average fitness of the population, the crossover probability is reduced by the crossover probability reduction formula; otherwise, the upper limit of the crossover probability is taken. For mutation operations, when the fitness of the individual to be mutated is higher than the average fitness of the population, the crossover probability is reduced by the mutation reduction formula; otherwise, the upper limit of the mutation probability is taken. Preferably, step S5 includes the following steps: S51. Real-time feedback of the completion signals of each process, compare with the expected end time in the execution instruction set, and calculate the execution time offset; S52. Extract the execution time offset and device alarm level, and perform anomaly classification judgment: if the execution time offset is less than or equal to the tolerance threshold and there is no device fault alarm signal, it is judged as a mild anomaly; if the execution time offset is greater than the preset tolerance threshold, or a device fault alarm signal is received, or an emergency request with a priority score higher than the median of the current execution sequence is received, it is judged as a severe anomaly. S53. For minor anomalies, perform local time window extension adjustment and update the expected end time in the execution instruction set; for severe anomalies, trigger rescheduling logic, lock the currently executing and uninterruptible task process, remove the resource time period occupied by it from the resource-task matching matrix, and obtain a snapshot of available resources. S54. Treat the frozen affected tasks as tasks to be assigned, call the improved adaptive genetic algorithm in S4 to search in the available resource snapshot, obtain the second task allocation strategy by minimizing the time schedule offset and completion time after rescheduling; and trace the source based on the trigger type of severe anomaly, generate an anomaly tracing feedback report containing anomaly nodes, processing time and resource loss assessment, and push it to the cloud platform management terminal. This invention constructs a self-healing closed loop for cloud platform scheduling. In real biological laboratories, equipment failures and reagent jams are common occurrences. Traditional static scheduling software often crashes when faced with such emergencies, requiring manual rescheduling for several hours. This invention, by constructing a local rescheduling algorithm and introducing an offset penalty term, can achieve disaster recovery and resource reorganization while limiting the scope of disturbances. It enhances the robustness of the cloud platform in the face of uncertainties in the physical world, ensuring the smooth, efficient, and safe operation of the entire protein complex detection process.

[0014] An intelligent scheduling and allocation system for protein complex detection tasks in a cloud platform is provided to implement the aforementioned intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform. The system includes: a task feature extraction module, a real-time resource perception module, a dynamic priority evaluation module, an intelligent scheduling and allocation module, and a dynamic monitoring and adjustment module. Task feature extraction module: used to receive multi-source data from detection applications, extract core features, and quantify and construct a task feature model; Real-time resource perception module: used to collect equipment status and reagent inventory, and to build a resource matching prediction model and generate a resource-task matching matrix using an improved spatiotemporal graph convolutional neural network; Dynamic priority evaluation module: used to calculate sample vulnerability and time urgency, and obtain dynamic task scheduling sequence through adaptive weight adjustment; Intelligent scheduling and allocation module: used to optimize the candidate resource subset using an adaptive genetic algorithm and generate an execution instruction set; Dynamic monitoring and adjustment module: Used to monitor execution progress, classify and judge anomalies, extend the execution time window for minor anomalies, trigger rescheduling logic for severe anomalies and update the task allocation strategy using a local rescheduling algorithm, and generate an anomaly tracing feedback report.

[0015] (III) Beneficial Effects The present invention has the following beneficial effects: This invention achieves precise quantification of the activity state of protein samples by constructing a task feature model that includes sample half-life and vulnerability indicators. With the help of a dynamic weight adaptive adjustment mechanism, the system can automatically identify and prioritize the scheduling of high-risk samples, improving the experimental success rate of easily degradable samples and solving the problem of precious biological samples becoming invalid due to scheduling delays in the cloud platform environment. This design avoids the one-size-fits-all approach in traditional scheduling and provides a refined and differentiated guarantee channel for scientific research samples with different timeliness requirements, thereby improving the detection success rate and the transformation rate of results for high-value scientific research projects.

[0016] This invention utilizes an improved adaptive genetic algorithm to perform multi-objective optimization under the constraints of a resource-task matching matrix, effectively balancing task completion time and equipment utilization. Through real-time data feedback from the Internet of Things and LIMS systems, combined with an improved spatiotemporal graph convolutional neural network for forward-looking prediction of resource status, it can accurately avoid equipment that may experience reagent shortages or be under maintenance in the future, thereby increasing the average daily throughput of core laboratory equipment (such as high-resolution mass spectrometers) and significantly reducing the unit detection cost of the cloud platform. It effectively reduces equipment idling and energy waste caused by temporary material shortages, achieving lean flow of laboratory reagents and consumables and globally optimal configuration of underlying hardware resources.

[0017] This invention establishes a closed-loop monitoring and rescheduling system for the entire process. By introducing a local rescheduling algorithm with an offset penalty term, the system possesses extremely strong anti-interference capabilities. Combined with an anomaly classification and judgment mechanism, the system can accurately distinguish between minor delays and serious faults, avoiding the ineffective global rescheduling computational consumption caused by minor fluctuations. In the event of equipment failures or emergency order insertions, the system can achieve rapid automatic response and scheme reorganization, reducing the need for manual intervention, and simultaneously automatically generating anomaly traceability feedback reports. This provides a reliable data foundation for the laboratory's long-term equipment operation and maintenance planning and process optimization, ensuring the continuity and stability of cloud platform testing tasks in complex physical environments.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform according to the present invention. Figure 2 This is a flowchart illustrating the process of dynamically obtaining adaptive weight vectors in the intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform according to the present invention. Figure 3 This is a flowchart illustrating the dynamic adjustment of crossover and mutation probabilities in an intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform according to the present invention. Figure 4 This is a flowchart illustrating the process of obtaining the detection task allocation strategy in the intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform according to the present invention. Figure 5 This is a schematic diagram of a module of an intelligent scheduling and allocation system for protein complex detection tasks in a cloud platform according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0022] Please see Figure 1 This invention discloses an intelligent scheduling and allocation method for protein complex detection tasks in a cloud platform, comprising the following steps: S1. Receive protein complex detection applications from cloud platform users, and extract multi-dimensional core features including sample type, detection requirements, sample quantity, sample half-life, and delivery deadline using natural language processing technology; construct a task feature model based on the multi-dimensional core features, including task attribute features, timeliness features, and priority features. S1 includes the following steps: S11. The raw data of the protein complex detection application received by the cloud platform, which includes unstructured text descriptions and structured forms; In specific implementation, S11 is as follows: the user submits an application through the cloud platform; the fields filled in the structured form include: sample quantity (e.g., 100 tubes), delivery deadline (e.g., before 17:00 on January 31, 2026), and user ID (associated with their VIP level of Level 3 according to the backend database); the user fills in the unstructured text as follows: "The sample submitted for testing is a crude cell lysate extracted from mouse liver tissue, which contains a large amount of protease and is very easy to degrade. Please be sure to perform cross-linked mass spectrometry (CXMS) detection as soon as possible to find potential protein-protein interaction (PPI) networks." S12. Utilize a pre-trained biological domain language model to perform named entity recognition and semantic parsing on unstructured text descriptions, extracting sample type entities and detection requirement entities; extract sample quantity values, sample half-life, delivery deadline timestamps, and user level codes from structured forms. In specific implementation, S12 involves: calling a BiLSTM-CRF model fine-tuned based on medical and biological corpus to parse the above text; extracting the "sample type" entity as "cell lysate"; extracting the "detection requirement" entity as "crosslinked mass spectrometry (CXMS)"; extracting potential risk features such as "easily degradable" and "contains protease"; and combining the knowledge graph, the system automatically infers the effective half-life of this type of unpurified cell lysate under conventional 4℃ storage conditions. T half Approximately 4 hours; Furthermore, in this embodiment, the BiLSTM-CRF model structure includes: an input layer, a bidirectional long short-term memory network layer, and a conditional random field layer; wherein, the input layer uses pre-trained biomedical-specific word vectors, with a dimension set to 300; the bidirectional long short-term memory network layer contains two hidden layers, each with 256 hidden units, and the dropout rate is set to 0.5 to prevent overfitting; the CRF layer is used to output the globally optimal label sequence; before use, a biomedical entity annotation corpus is constructed, covering core terms such as protein type, detection method, and reagent name; PubMed and Uni... Incremental pre-training was performed on approximately 100,000 text data entries from the UniProt database to enable the model to recognize biological contextual semantics. Supervised fine-tuning was conducted using 5,000 historical application texts from the cloud platform, verified by experts, to ensure that the model's recognition accuracy for detection requirements reached over 95%. PubMed, a biomedical literature database, contains a vast amount of academic papers and abstracts in the life sciences field, used to train the model to understand biological semantics. The UniProt database, a protein sequence and function knowledge base, provides high-confidence protein annotation information to enhance the model's ability to recognize protein-related entities. S13. Calculate the sample vulnerability index based on the sample half-life; S13 includes the following steps: S131. Obtain the theoretical half-life of the sample. T half Set the platform's maximum tolerance decay period. T max ; S132. Based on the sample half-life, the theoretical half-life of the sample, and the platform's maximum tolerable decay period, the sample vulnerability index is calculated using the exponential decay mapping formula. V fra ; In specific implementation, S13 specifically involves: converting the extracted business information into a numerical vector that the scheduling algorithm can calculate; and setting the platform's baseline maximum tolerance decay period. T maxThe threshold is 72 hours (meaning samples that do not degrade after 72 hours are considered extremely stable, with fragility approaching 0). Based on the gradient discrimination requirements of extremely sensitive laboratory samples (such as those with a 4-hour half-life in this example), and by fitting multiple sets of historical degradation data, a decay control coefficient λ=5.0 was determined to ensure that even when samples are highly susceptible to degradation, the degradation control coefficient remains stable. V fra It can quickly converge to above 0.99, therefore the decay control coefficient λ is set to 5.0; Calculated using the exponential decay mapping formula: Substitute data V fra =1− e −5.0*(72 / 4) =1− e −90 ;because e −90 Extremely low, with a value of approximately 0.9999; this is quantified as extremely high vulnerability (out of 1.0) in the system; in contrast, if the sample submitted is purified lyophilized protein powder... T half =720 hours), then V fra =1− e −5.0*(72 / 720) =1− e −0.5 ≈0.393; This invention achieves a digital mapping of the timeliness of biological samples through an exponential decay function; it transforms ambiguous descriptions of easy degradation into high-precision continuous values, providing quantitative weights for subsequent scheduling algorithms, eliminating the arbitrariness of manual assessment, and improving the survival rate of high-risk samples.

[0023] S14. After normalizing the sample type, detection requirements, sample quantity, delivery deadline timestamp and user level code, the sample is encapsulated into a task feature model containing a unique task identifier. In specific implementation, S14 is as follows: The task feature model finally constructed in this embodiment is: {Sample type code: 02 (cell lysate), detection requirement code: 105 (CXMS), quantity: 100, vulnerability: 0.9999, delivery deadline timestamp: 1764579600, user priority: 3}; S2. Based on the task feature model, a dynamic priority evaluation function with adaptive weights is used to score and prioritize all protein complex detection tasks in the cloud platform task pool in real time to obtain a dynamic task scheduling sequence. S2 includes the following steps: S21, Testing requirements D reqInput a preset task standard time mapping table to obtain the standard time consumption of each preprocessing and computer operation step required for the task. Add up the standard time consumption of all steps and normalize it to obtain the task complexity. C com ; S22. Calculate the difference between the delivery deadline timestamp and the current system timestamp, and take its reciprocal to obtain the time urgency index; S23. Construct a dynamic priority evaluation function; the dynamic priority evaluation function includes a weight vector; the weight vector has time weights. α Vulnerability weights β User level weight c and task complexity weights d ; S24. Adjust the weight vector according to the sample vulnerability index to obtain an adaptive weight vector; S24 includes the following steps: S241. Define an initial weight vector, which includes an initial time weight, an initial vulnerability weight, an initial user level weight, and an initial task complexity weight. S242. Determine whether the sample vulnerability index exceeds the preset high-risk threshold. If it does, increase the proportion of vulnerability weight according to the exponential enhancement rule, and simultaneously reduce the proportion of other weights proportionally to obtain an adaptive weight vector. For specific implementation details, please refer to [link / reference]. Figure 2 Specifically, S24 refers to the weighting coefficient in S23. α , β , c , d The adaptive adjustment rules are specifically as follows: [Limitation] α , β , c , d The sum is 1; by constructing a judgment matrix, five senior laboratory directors were invited to compare the importance of the four dimensions pairwise. Since sample activity (vulnerability) is the key to the success or failure of the experiment in protein complex detection, its relative importance is the highest; followed by delivery deadline (timeliness); task complexity determines the resource consumption time, ranking third; user level, as a business auxiliary indicator, ranks last; therefore, the initial basic weight values ​​are set. α 0 = 0.3 β 0 = 0.4 c 0 = 0.1 d 0 = 0.2; Vulnerability indicators for this task were detected. V fra =0.9999 is greater than the set high-risk threshold. V high=0.8; β Values ​​are updated dynamically β 0 * (1+ k 2* V fra );in, k 2 represents the vulnerability sensitivity coefficient; in this embodiment, a threshold is set when a sample enters an "extremely high risk" state ( V fra When the value is greater than 0.8, the system must ensure its weight. β It can generate an increase of over 150%, achieving absolute dominance over routine tasks; when k When 2=1.5, the vulnerability score increases from 0.4 to 0.99, and its weighted contribution increases from 0.16 to 0.625, which is sufficient to make the task jump from the bottom to the top in the ranking, which is consistent with the business logic of emergency avoidance in the laboratory; therefore, the following setting is made: k 2 = 1.5; then the new β =0.4*(1+1.5*0.9999)≈0.4*2.5=1.0; At this time α 0 = 0.3 β 0 = 1.0 c 0 = 0.1 d 0 = 0.2; the sum is 1.6; normalization is then performed: α =0.3 / 1.6=0.1875; β =1.0 / 1.6=0.625; c =0.1 / 1.6=0.0625; d =0.2 / 1.6=0.125; Meanwhile, the user's VIP level is 3, with a maximum score of 5, which is normalized to 0.6; S25. The time urgency index, sample vulnerability index, user level and task complexity are weighted and calculated using the adaptive weight vector to obtain the dynamic priority score; all tasks in the task pool are sorted in descending order according to the dynamic priority score to obtain the dynamic task scheduling sequence. In specific implementation, S25 specifically involves: substituting the values ​​into the dynamic priority evaluation function to calculate, thus obtaining... P score =0.1875*(1 / 300)+0.625*0.9999+0.0625*(3 / 5)+0.125*0.432=0.000625+0.6249+0.0375+0.054=0.717; where, P score This indicates a dynamic priority score; simultaneously, there is also a routine molecular weight determination task for purified protein in the task pool, which has low fragility and short processing time, and the calculated score is... P score=0.152; Therefore, although the CXMS task was submitted late in this embodiment, it was assigned to the head of the scheduling queue with a dynamic score as high as 0.717. S3. Real-time acquisition platform background experimental detection resources' real-time status and processing capabilities to obtain real-time resource data; use historical resource data to train an improved spatiotemporal graph convolutional neural network to obtain a resource matching prediction model; Real-time resource data is input into the resource matching prediction model to obtain resource status prediction results; the resource status prediction results are correlated and mapped with the dynamic task scheduling sequence output by S2 and the matching degree is calculated to obtain the resource-task matching matrix; S3 includes the following steps: S31. The real-time resource data is obtained by collecting the operating status, estimated idle time and real-time reagent balance of the experimental testing equipment in real time through the MQTT protocol and API interface. In specific implementation, S31 is as follows: The laboratory backend has 3 automated pipetting workstations (numbered R1, R2, R3) and 2 high-resolution mass spectrometers (numbered M1, M2); each device's motherboard is connected to an IoT edge computing gateway; the gateway publishes a JSON-formatted status message to the cloud every 5 seconds via the MQTT protocol; the message content includes: device ID, current operating current, robotic arm coordinates, currently executing Task ID, and estimated remaining completion time; for example, the current message of mass spectrometer M1 shows that it is executing Task 089, and according to the mass spectrometry acquisition progress, the estimated remaining time is 120 minutes; the system calls the LIMS database through API; for CXMS detection, the key reagents DSSO (crosslinking agent) and Trypsin are required; LIMS returns the following data: the current stock of DSSO is 500 reaction portions, and the remaining Trypsin is 50 reaction portions; S32. Based on the characteristics of laboratory resource data, the spatiotemporal graph convolutional neural network is improved to obtain an improved spatiotemporal graph convolutional neural network; the improvements include: S321. Improve the static spatial topology graph of the traditional spatiotemporal graph convolutional neural network into a dynamic spatial topology graph based on process coupling degree; S322. Improve the standard temporal convolutional layer of the traditional spatiotemporal graph convolutional neural network into a multi-scale dilated temporal convolutional layer; S323. Improve the traditional spatiotemporal graph convolutional neural network's single output layer into a multi-task joint output layer that includes idle time branches, reagent consumption branches, and fault warning branches; In specific implementation, S32 is as follows: based on the strong and weak coupling characteristics of laboratory equipment in the process flow, laboratory equipment is used as graph nodes, and the edge weights are dynamically calculated according to the flow frequency of historical tasks between equipment to construct a dynamic spatial topology graph, so that equipment with tight process coupling has a greater amount of information interaction. Based on the multi-scale time-dependent characteristics of laboratory equipment sensor data, the expansion rate is introduced. d =1, 2, 4 1D dilated convolutions were used to capture short-term instantaneous mechanical action features and long-term reagent consumption trend features, respectively. Based on the multidimensional concurrency characteristics of abnormal laboratory conditions, the network endpoint was improved to a multi-task structure containing three independent fully connected layer branches; The improved spatiotemporal graph convolutional neural network architecture and parameters are as follows: The network consists of an input layer, two dynamic spatiotemporal graph convolutional modules, and a multi-task output layer; the input feature dimension is a concatenated vector of device sensor and inventory features (64-dimensional); in the dynamic spatiotemporal graph convolutional module, the number of hidden units in the graph convolutional layer is set to 128, and the temporal convolutional layer uses a 1D dilated convolution with a kernel size of 3; the multi-task output layer contains three independent branches, which are mapped to continuous idle time prediction values ​​(using the MSE loss function), reagent consumption rate (using the MSE loss function), and discrete fault probability classification values ​​(using the cross-entropy loss function), respectively. To prevent overfitting, the dropout rate in the network is set to 0.3; S33. Obtain historical laboratory operation log data to construct a training set, and use the training set to train the improved spatiotemporal graph convolutional neural network to obtain the resource matching prediction model; In specific implementation, S33 specifically involves: extracting the laboratory operation logs from the cloud platform over the past year as a dataset, including historical status sequences of equipment sensors (such as operating current and robotic arm position), historical LIMS inventory consumption sequences, historical task process flow records, and historical equipment fault tags. During training, a multi-task hybrid loss function (weighted sum of MSE and cross-entropy) is used, and the Adam optimizer is used for backpropagation to update network parameters. The initial learning rate is set to 0.001, and the batch size is set to 64. When the overall prediction accuracy of the model on the validation set stabilizes above 95%, and the loss on the validation set no longer decreases for 10 consecutive epochs, the early stopping mechanism is triggered, the model training is determined to be complete, and the resource matching prediction model is obtained. S34. Input the real-time resource data into the resource matching prediction model, and output the predicted idle time point, reagent depletion risk value and equipment reliability score of each device in the future execution cycle to obtain the resource status prediction result. In specific implementation, S34 is as follows: The system uses the data collected at the current moment on the M1 mass spectrometer, centrifuge C2, and LIMS reagent inventory as input node features and inputs them into the trained resource matching prediction model; the model outputs the prediction results: the M1 mass spectrometer will be idle in 1.5 hours, and its failure probability in the next 24 hours is only 0.05 (within a safe range); however, the model predicts that the Trypsin enzyme digestion reagent will be exhausted in 2 hours based on the current consumption trend; at the same time, it predicts that the vibration characteristics of centrifuge C2 have abnormal accumulation, and the failure probability is as high as 0.75; S35. Perform a Cartesian product operation on the resource status prediction result and the dynamic task scheduling sequence output by S2, eliminate combinations that do not meet reagent constraints and whose reliability scores are lower than the safety threshold, calculate the matching degree score between each task and each available device, and obtain the resource-task matching matrix. In specific implementation, S35 specifically involves: associating and matching the prediction results with the dynamic task scheduling sequence; such as Task A 100 doses of Trypsin reagent are needed. Because the model predicts the reagents are about to run out, and the predicted failure rate of centrifuge C2 (0.75) is higher than the preset safety threshold (0.50), the system will assign the Task to the appropriate match during the association process. A With C2, Task A The matching score with the current reagent cabinet is forcibly set to 0 (i.e., the high-risk combination is removed), and a reagent replenishment warning is triggered; for available equipment combinations that meet the predicted safety constraints, the system calculates the matching score by weighting the reciprocal of the predicted idle time and the reliability score, and generates a resource-task matching matrix that records the feasibility and recommendation scores of all task-equipment matching. S4. Using an improved adaptive genetic algorithm, with the optimization objectives of minimizing the total completion time and maximizing equipment utilization, under the constraints of the resource-task matching matrix, a heuristic search and iterative optimization are performed to obtain the Pareto optimal detection task allocation strategy, and the strategy is parsed into an execution instruction set containing equipment number, process start time, expected end time and required reagent list. Please see Figure 4 S4 includes the following steps: S41. Encode the scheduling scheme into a real number matrix chromosome, where rows of the matrix represent tasks, columns represent detection procedures, and matrix elements represent the allocated device numbers; based on the resource-task matching matrix, prioritize the allocation of devices with high matching scores to high-priority tasks in the dynamic task scheduling sequence, and generate an initial population containing multiple allocation schemes. In specific implementation, S41 is as follows: the chromosome length is equal to the total number of processes in all tasks to be scheduled; the value of the gene bit represents the assigned device ID; for example, the gene sequence [R1, M1, R2, M2] represents that the preprocessing of task 1 is assigned to R1, and the on-machine processing is assigned to M1; the system generates 100 such chromosomes to form the initial population; the system reads the resource-task matching matrix and finds the Task A The highest matching score was found with mass spectrometer M2, while the lower matching score was found with M1 (due to a later expected idle time). Therefore, when performing gene encoding, the task was prioritized with an 80% probability. A The machine operation is assigned to M2, and then randomly assigned to other available devices with a 20% probability. This heuristic guidance based on the matching matrix to generate the initial population effectively improves the quality of the initial population and speeds up the convergence of the algorithm. S42. Define a bi-objective fitness function, where the first objective is the completion time of all tasks on each device. M ak Minimize, the second objective is the percentage of effective working time for all devices during task execution. U ti maximize; In specific implementation, S42 specifically refers to: the bi-objective fitness function. F = w 1*1 / ( M ak )+ w 2* U ti ;in, w 1 and w 2 represents the weighting of completion time and the percentage weighting of working time, respectively. In cloud platform operation, the depreciation and operating costs of high-precision equipment such as mass spectrometers are extremely high (accounting for approximately 70% of the total cost), while time cost (customer satisfaction and sample failure risk) is the core KPI; through backtesting of operational data from 1000 historical tasks, when w When 1 is set to 0.7, the system can shorten the average delivery cycle by 25% while ensuring that the equipment utilization rate is not less than 80%, achieving the optimal balance between operational efficiency and user experience. Therefore, this embodiment sets 1 to 0.7. w 1 = 0.7 and w 2 = 0.3; When calculating fitness, a resource-task matching matrix constraint check mechanism is introduced. If an individual's chromosome contains a gene pair with a score of 0 in the matching matrix (i.e., violating the safety constraint of future equipment failure or reagent depletion), an infinitely large penalty value is applied to the individual's fitness, causing its fitness to approach 0, thus allowing it to be naturally eliminated during the evolutionary process. For the individual chromosome Chro1, the latest completion time of all its tasks was calculated using Gantt chart simulation. M ak =48 hours; during this period, the average utilization rate is obtained by dividing the total operating time of the M1 and M2 mass spectrometers by the total available time. U ti =0.85; then the fitness of this individual is FChro1=0.7*(1 / 48)+0.3*0.85=0.0145+0.255=0.2695; Furthermore, the Gantt chart simulation process is as follows: The device IDs in the chromosome encoding are mapped to the task steps, and arranged on the timeline according to the logical order of the steps (e.g., centrifugation, enzyme digestion, mass spectrometry); the estimated idle time of each device in the resource-task matching matrix is ​​read as the earliest possible start time for each step of the task; standard working hours are accumulated, and a preset sample transport time is added between adjacent steps; conflicts are checked for the same device in the same time period, and if a conflict occurs, the lower priority task is shifted backward; finally, the end time of the last node in all step sequences is extracted as the end time of the task. M ak value; S43. Perform adaptive genetic evolution operation, dynamically adjust the crossover probability and mutation probability according to the fitness variance of the current population, and retain the top-ranked individuals with the highest fitness scores in the next generation through a bi-objective fitness function. The dynamic adjustment of crossover and mutation probabilities in S43 includes the following steps: S431. Calculate the average fitness value and maximum fitness value of all individuals in the current population; S432. For crossover operations, when the fitness value of the two parent individuals participating in the crossover is higher than the average fitness of the population, the crossover probability is reduced by the crossover probability reduction formula. The higher the fitness, the lower the crossover probability; otherwise, the upper limit of the crossover probability is taken to preserve superior genes. For mutation operations, when the fitness of the individual to be mutated is higher than the average fitness of the population, the crossover probability is reduced by the mutation reduction formula. The higher the fitness, the lower the mutation probability; otherwise, the upper limit of the mutation probability is taken to promote the generation of new schemes, thereby obtaining dynamically adjusted evolutionary parameters. For specific implementation details, please refer to [link / reference]. Figure 3 S432 specifically refers to: when F 1 ≥ F avg hour, ;when F 1 < F avg hour, P c =P cmax ;when F ≥ F avg hour, ; F < F avg hour, P m = P mmax ;in, P c , P m These represent the crossover probability and the mutation probability, respectively. F max For the maximum fitness of the population, F avg For average fitness, F 1 The maximum fitness of the two parent individuals in the crossover operation. F For the fitness of the variant individual; P cmax , P cmin These are the upper and lower limits of the set crossover probability, respectively. P mmax , P mmin These are the upper and lower limits of the mutation probability, respectively; The system calculates the maximum fitness of the current population. F max =0.31, average fitness F avg =0.22; For the two selected parent individuals, their greater fitness is... F 1 =0.28; because F 1 ≥ F avg The dynamic crossover probability formula is used: [Setting...] P cmax =0.9, P cmin =0.6; P c =0.9-[(0.9-0.6)*(0.28-0.22) / (0.31-0.22)]=0.9−0.2=0.7; The system performs a multi-point crossover operation on these two individuals with a probability of 0.7, exchanging the equipment allocation genes for some tasks; Similarly, the mutation probability is calculated using the formula. P m And perform single-point mutation (e.g., randomly change the mass spectrometer for a task from M1 to M2). S44. Repeat S43 until the preset maximum number of iterations is reached. Output the Pareto optimal solution with the highest fitness as the detection task allocation strategy. Parse the strategy to generate machine-readable execution instructions, and send them to the control terminal of the corresponding device through the cloud platform. In specific implementation, S44 is as follows: After 200 generations of evolution, the algorithm converges and outputs the optimal chromosome; the chromosome is analyzed: the batch A preprocessing task is assigned to the currently idle pipetting workstation R1, and the right to use the M1 mass spectrometer is locked for 15 hours; the system converts this strategy into JSON control instructions and sends them to the control gateway of R1, and the R1 robotic arm is then started to process the liver cell lysate sample. S5. Monitor the execution progress and resource occupancy rate of the execution instruction set on the experimental terminal in real time, and make anomaly classification judgment: if it is a minor anomaly, the execution local time window is extended and adjusted; if it is a severe anomaly, the affected task is frozen, the improved adaptive genetic algorithm in S4 is called to reconfigure the resources of the affected task, obtain the second task allocation strategy and update the execution instruction set, and generate an anomaly traceability feedback report. S5 includes the following steps: S51. The completion signals of each process are fed back in real time through the Internet of Things terminal, and the execution time offset is calculated by comparing them with the expected end time in the execution instruction set. In specific implementation, S51 is as follows: When batch A sample is subjected to enzymatic digestion on the R1 pipetting workstation, the cloud platform suddenly receives an MQTT alarm message from the R1 gateway: "Error Code 404: Z-axis stepper motor overload shutdown"; at this time point T tri The time is 14:35:00; at the same time, another centrifuge C1 sends out a task completion signal, and compared with the expected end time, the execution time offset is calculated to be a delay of 3 minutes; S52. Extract the execution time offset and device alarm level, and perform anomaly classification judgment: if the execution time offset is less than or equal to the tolerance threshold and there is no device fault alarm signal, it is judged as a mild anomaly; if the execution time offset is greater than the preset tolerance threshold, or a device fault alarm signal is received, or an emergency request with a priority score higher than the median of the current execution sequence is received, it is judged as a severe anomaly. In specific implementation, S52 is as follows: the system sets the tolerance threshold to 10 minutes; for centrifuge C1, if the delay is 3 minutes, which is less than the tolerance threshold and there is no fault alarm, the system determines it as "mild abnormality"; for R1 pipetting workstation, if a clear fault alarm signal of "Z-axis stepper motor overload shutdown" is received, the system directly determines it as "severe abnormality". S53. For minor anomalies, perform local time window extension adjustment and update the expected end time in the execution instruction set; for severe anomalies, trigger rescheduling logic, lock the currently executing and uninterruptible task process, remove the resource time period occupied by it from the resource-task matching matrix, and obtain a snapshot of available resources. In specific implementation, S53 is as follows: For a minor anomaly in C1, the system does not trigger complex rescheduling calculations, but directly postpones the estimated start and end times of subsequent processes in C1 by 3 minutes and updates the instruction set; for a severe anomaly in R1, the system quickly scans the global Gantt chart, extracts the remaining processes (desalination) of batch A originally in R1, as well as the other 3 regular tasks behind R1, and forms a set of affected tasks. Task im The system extracts available resources and finds that R2 is executing an uninterrupted task, while R3 is idle. At the same time, the system re-inputs the real-time status of the currently extracted available resources into the resource matching prediction model in S3 to re-predict the failure probability and idle time of the remaining resources and update the resource-task matching matrix to ensure that the resources selected for rescheduling have no potential failure risk in the future cycle. Therefore, the available resource subset Res only includes R3 and subsequent mass spectrometer resources. S54. Treat the frozen affected tasks as tasks to be assigned, call the improved adaptive genetic algorithm in S4 to search in the available resource snapshot, obtain the second task allocation strategy by minimizing the time schedule offset and completion time after rescheduling; and trace the source based on the trigger type of severe anomaly, generate an anomaly tracing feedback report containing anomaly nodes, processing time and resource loss assessment, and push it to the cloud platform management terminal. In specific implementation, S54 is as follows: the system initiates local rescheduling; to avoid the impact of rescheduling on the progress of tasks in the laboratory, the system introduces an offset penalty term in the fitness function. To minimize the disruption to the original schedule caused by rescheduling, where m The penalty coefficient is... T ns and T os Let be the start times of the old and new plans for the task, respectively; then the fitness function for local rescheduling is modified as follows: F lo = w 1*1 / ( M ak )+ w 2* U ti - Pen ; Furthermore, m The calculation formula is μ=( Csw+Cre ) / CDU ;in, Csw The risks and costs associated with transferring samples across devices, such as the risk of temperature fluctuations; Cre The time cost of re-washing and recalibrating the equipment; CDU The unit delay cost is calculated; based on calculations, the comprehensive risk cost of a single unplanned equipment switch in the protein detection process is approximately five times the unit delay cost, therefore, a threshold is set. m =5.0 The guided algorithm prioritizes fine-tuning the original path and only performs cross-device switching when the original path is no longer viable; The algorithm rearranges on R3 Task im The task in question; due to the extremely high vulnerability of batch A ( V fra =0.9999), its priority in local scheduling is still the highest; the algorithm instantly decides to move the desalination process of batch A to workstation R3, the new plan starts at 14:40, the old plan starts at 14:35, the offset is minimal and the penalty is low; while the other 3 regular tasks are postponed. The system automatically issued instructions, and the AGV in the laboratory received the instructions to transfer the microplate of batch A from the paralyzed R1 to R3 to continue the operation; throughout the entire fault response process, the exposure time of the sample at room temperature did not exceed 5 minutes, thus preserving the easily degradable protein complex. Meanwhile, the system automatically generates a fault rescheduling feedback report for the R1 workstation, which includes: the fault node (R1 Z-axis motor), the number of affected tasks (4), the rescheduling calculation time (1.2 seconds), the sample transfer risk assessment (low risk), and the recommended maintenance time. The report is then pushed to the laboratory supervisor's mobile device via the cloud platform message interface, completing the closed loop of anomaly handling.

[0024] For further details, please refer to Figure 5 A smart scheduling and allocation system for protein complex detection tasks in a cloud platform is provided to implement the aforementioned smart scheduling and allocation method for protein complex detection tasks in a cloud platform. The system includes: a task feature extraction module, a real-time resource perception module, a dynamic priority evaluation module, a smart scheduling and allocation module, and a dynamic monitoring and adjustment module. Task feature extraction module: used to receive multi-source data from detection applications, extract core features, and quantify and construct a task feature model; Real-time resource perception module: used to collect equipment status and reagent inventory, and to build a resource matching prediction model and generate a resource-task matching matrix using an improved spatiotemporal graph convolutional neural network; Dynamic priority evaluation module: used to calculate sample vulnerability and time urgency, and obtain dynamic task scheduling sequence through adaptive weight adjustment; Intelligent scheduling and allocation module: used to optimize the candidate resource subset using an adaptive genetic algorithm and generate an execution instruction set; Dynamic monitoring and adjustment module: Used to monitor execution progress, classify and judge anomalies, extend the execution time window for minor anomalies, trigger rescheduling logic for severe anomalies and update the task allocation strategy using a local rescheduling algorithm, and generate an anomaly tracing feedback report.

[0025] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0026] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for intelligent scheduling and allocation of protein complex detection tasks in a cloud platform, characterized in that, Includes the following steps: S1. Receive protein complex detection applications from cloud platform users, extract information including sample type, detection requirements, sample quantity, sample half-life, delivery deadline, and sample vulnerability indicators, and construct a task feature model. S2. Based on the task feature model, the dynamic priority evaluation function adjusted by the sample vulnerability index is used to score and prioritize all protein complex detection tasks in the cloud platform task pool in real time to obtain a dynamic task scheduling sequence. S3. Real-time acquisition platform background experimental detection resources' real-time status and processing capabilities to obtain real-time resource data; use historical resource data to train an improved spatiotemporal graph convolutional neural network to obtain a resource matching prediction model; Real-time resource data is input into the resource matching prediction model to obtain resource status prediction results; the resource status prediction results are correlated and mapped with the dynamic task scheduling sequence output by S2 and the matching degree is calculated to obtain the resource-task matching matrix; S4. Using an improved adaptive genetic algorithm, with the optimization objectives of minimizing the total completion time and maximizing equipment utilization, find the optimal detection task allocation strategy under the constraints of the resource-task matching matrix, and parse it into an execution instruction set; S5. Monitor the execution progress and resource occupancy rate of the execution instruction set on the experimental terminal in real time, and make anomaly classification judgment: if it is a minor anomaly, the execution local time window is extended and adjusted; if it is a severe anomaly, the affected task is frozen, the improved adaptive genetic algorithm in S4 is called to reconfigure the resources of the affected task, obtain the second task allocation strategy and update the execution instruction set, and generate an anomaly traceability feedback report. 2.The intelligent scheduling and distribution method of protein complex detection tasks in a cloud platform according to claim 1, characterized in that, S1 includes the following steps: S11. The raw data of the protein complex detection application received by the cloud platform, which includes unstructured text descriptions and structured forms; S12. Utilize a pre-trained biological domain language model to perform named entity recognition and semantic parsing on unstructured text descriptions, extracting sample type entities and detection requirement entities; extract sample quantity values, sample half-life, delivery deadline timestamps, and user level codes from structured forms. S13. Calculate the sample vulnerability index based on the sample half-life; S14. After normalizing the sample type, detection requirements, sample quantity, delivery deadline timestamp, and user level code, the data is encapsulated into a task feature model containing a unique task identifier. 3.The method of claim 2, wherein, S13 includes the following steps: S131. Obtain the theoretical half-life of the sample and set the maximum tolerance period for the platform benchmark. S132. Based on the sample half-life, the theoretical half-life of the sample, and the maximum tolerable decay period of the platform benchmark, the sample vulnerability index is calculated using the exponential decay mapping formula. 4.The method of claim 2, wherein, S2 includes the following steps: S21. Input the detection requirements into the preset task standard working time mapping table, obtain the standard time consumption of each preprocessing and computer step required for the task, add up the standard time consumption of all steps and normalize it to obtain the task complexity. S22. Calculate the difference between the delivery deadline timestamp and the current system timestamp, and take its reciprocal to obtain the time urgency index; S23. Construct a dynamic priority evaluation function; the dynamic priority evaluation function includes a weight vector; the weight vector includes time weight, vulnerability weight, user level weight, and task complexity weight; S24. Adjust the weight vector according to the sample vulnerability index to obtain an adaptive weight vector; S25. Using the adaptive weight vector, the time urgency index, sample vulnerability index, user level, and task complexity are weighted and calculated using a dynamic priority evaluation function to obtain the dynamic priority score; based on the dynamic priority score, all tasks in the task pool are sorted in descending order to obtain a dynamic task scheduling sequence. 5.The method of claim 4, wherein, S24 includes the following steps: S241. Define an initial weight vector, which includes an initial time weight, an initial vulnerability weight, an initial user level weight, and an initial task complexity weight. S242. Determine whether the sample vulnerability index exceeds the preset high-risk threshold. If it does, increase the proportion of vulnerability weight according to the exponential enhancement rule, and simultaneously reduce the proportion of other weights proportionally to obtain an adaptive weight vector. 6.The method of claim 4, wherein, S3 includes the following steps: S31. The real-time resource data is obtained by collecting the operating status, estimated idle time and real-time reagent balance of the experimental testing equipment in real time through the MQTT protocol and API interface. S32. Based on the characteristics of laboratory resource data, the spatiotemporal graph convolutional neural network is improved to obtain an improved spatiotemporal graph convolutional neural network. The improvement includes: changing the static spatial topology graph of the traditional spatiotemporal graph convolutional neural network into a dynamic spatial topology graph based on the process coupling degree. The standard temporal convolutional layer of the traditional spatiotemporal graph convolutional neural network is improved into a multi-scale dilated temporal convolutional layer; The traditional spatiotemporal graph convolutional neural network's single output layer is improved into a multi-task joint output layer that includes idle time branches, reagent consumption branches, and fault warning branches; S33. Obtain historical laboratory operation log data to construct a training set, and use the training set to train the improved spatiotemporal graph convolutional neural network to obtain the resource matching prediction model; S34. Input the real-time resource data into the resource matching prediction model, and output the predicted idle time point, reagent depletion risk value and equipment reliability score of each device in the future execution cycle to obtain the resource status prediction result. S35. Perform a Cartesian product operation on the resource status prediction result and the dynamic task scheduling sequence output by S2, eliminate combinations that do not meet the reagent constraints and whose reliability scores are lower than the safety threshold, calculate the matching degree score between each task and each available device, and obtain the resource-task matching matrix. 7.The method of claim 6, wherein, S4 includes the following steps: S41. Encode the scheduling scheme into a real number matrix chromosome, where rows of the matrix represent tasks, columns represent detection procedures, and matrix elements represent the allocated device numbers; based on the resource-task matching matrix, prioritize the allocation of devices with high matching scores to high-priority tasks in the dynamic task scheduling sequence, and generate an initial population containing multiple allocation schemes. S42. Define a bi-objective fitness function. The first objective is to minimize the completion time of all tasks on each device, and the second objective is to maximize the proportion of effective working time of all devices during task execution. S43. Perform adaptive genetic evolution operation, dynamically adjust the crossover probability and mutation probability according to the fitness variance of the current population, and retain the top-ranked individuals with the highest fitness scores in the next generation through a bi-objective fitness function. S44. Repeat S43 until the preset maximum number of iterations is reached, output the Pareto optimal solution with the highest fitness as the detection task allocation strategy, and parse and generate machine-readable execution instructions, which are then sent to the control terminal of the corresponding device through the cloud platform. 8.The method of claim 7, wherein, The dynamic adjustment of crossover and mutation probabilities in S43 includes the following steps: S431. Calculate the average fitness value and maximum fitness value of all individuals in the current population; S432. For crossover operations, when the fitness value of the two parent individuals participating in the crossover is higher than the average fitness of the population, the crossover probability is reduced by the crossover probability reduction formula; otherwise, the upper limit of the crossover probability is taken. For mutation operations, when the fitness of the individual to be mutated is higher than the average fitness of the population, the crossover probability is reduced by the mutation reduction formula; otherwise, the upper limit of the mutation probability is taken. 9.The method of claim 7, wherein, S5 includes the following steps: S51. Real-time feedback of the completion signals of each process, compare with the expected end time in the execution instruction set, and calculate the execution time offset; S52. Extract the execution time offset and device alarm level, and perform anomaly classification judgment: if the execution time offset is less than or equal to the tolerance threshold and there is no device fault alarm signal, it is judged as a mild anomaly; if the execution time offset is greater than the preset tolerance threshold, or a device fault alarm signal is received, or an emergency request with a priority score higher than the median of the current execution sequence is received, it is judged as a severe anomaly. S53. For minor anomalies, perform local time window extension adjustment and update the expected end time in the execution instruction set; for severe anomalies, trigger rescheduling logic, lock the currently executing and uninterruptible task process, remove the resource time period occupied by it from the resource-task matching matrix, and obtain a snapshot of available resources. S54. Treat the frozen affected tasks as tasks to be assigned, call the improved adaptive genetic algorithm in S4 to search in the available resource snapshot, obtain the second task allocation strategy by minimizing the schedule offset and completion time after rescheduling; and trace the source based on the trigger type of severe anomaly, generate an anomaly tracing feedback report containing the abnormal node, processing time and resource loss assessment, and push it to the cloud platform management terminal. 10.A system for intelligent scheduling and allocation of protein complex detection tasks in a cloud platform, characterized in that, A method for intelligent scheduling and allocation of protein complex detection tasks in a cloud platform as described in any one of claims 1-9 is provided, wherein the system includes a task feature extraction module, a real-time resource perception module, a dynamic priority evaluation module, an intelligent scheduling and allocation module, and a dynamic monitoring and adjustment module.

Images