A sample driving scheduling method and system for a multi-pipeline cell preparation device

Abstract

This invention discloses a sample-driven scheduling method and system for multi-pipeline cell preparation equipment. The method includes the following steps: Step S1: Obtain historical cell preparation task samples and equipment operating parameters, and perform preprocessing; Step S2: Establish a cell preparation scheduling model; Step S3: Randomly divide the complete records of multiple historical scheduling cycles into two disjoint subsets; Step S4: Calculate the sample mean vector and sample covariance matrix; Step S5: Determine the ellipsoid size s; Step S6: Construct an ellipsoidal uncertainty set and solve for the first scheduling scheme; Step S7: Reconstruct the uncertainty set; Step S8: Solve the scheduling model again based on the reconstructed uncertainty set to obtain the second scheduling scheme; Step S9: Map the second scheduling scheme into a sequence of executable instructions for the equipment and send them to each cell machine controller for execution. This invention provides a robust scheduling scheme for multi-pipeline cell preparation under the condition of only having historical samples and no need for prior knowledge of distribution.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling and control technology for biomedical equipment, specifically to a sample-driven scheduling method and system for multi-pipeline cell preparation equipment. Background Technology

[0002] The cell and gene therapy (CGT) industry is at a critical juncture in its transformation from traditional manual preparation to Industry 4.0 intelligent manufacturing. Fully automated cell culture equipment (such as the Celles WK1 cell culture system) integrates automated operations throughout the entire process, including separation and sorting, activation and transduction, large-scale expansion, process analysis, and formulation dispensing, within a closed, clean chamber. This replaces the traditional manual preparation mode, elevating batch consistency, environmental control, and data traceability to GMP compliance levels.

[0003] In actual operation, a large amount of historical order sample data has been accumulated, but the uncertainty information contained in this data has not been effectively utilized by existing methods. The opportunity-constrained programming (CCP) method requires the probability distribution of the uncertain parameters to be known (such as assuming a normal distribution). However, in reality, the number of historical order samples is limited and often exhibits non-standard distributions such as multimodal and heavy-tailed distributions. The bias in the distribution assumption will cause the actual feasibility of the scheduling scheme to be far lower than the nominal guaranteed value. Distributed Optimization (DRO): Although it relaxes the assumption of the exact distribution, it still requires the assumption that the true distribution belongs to a certain ambiguous set (such as the Wasserstein sphere), and the construction of the ambiguous set still requires the assumption of the distribution shape (such as unimodality, sub-Gaussianity, etc.). At the same time, it is difficult to effectively handle joint opportunity constraints. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a sample-driven scheduling method and system for multi-pipeline cell preparation equipment. Under the condition of having only historical samples and no prior knowledge of distribution, it provides a robust scheduling scheme for multi-pipeline cell preparation that combines statistical feasibility guarantee and online solution efficiency.

[0005] The objective of this invention is achieved through the following technical solution: a sample-driven scheduling method for multi-pipeline cell preparation equipment, comprising the following steps:

[0006] Step S1: Obtain historical cell preparation task samples and equipment operating parameters, and perform preprocessing;

[0007] Step S2: Establish a cell preparation scheduling model with cross-contamination protection constraints;

[0008] Step S3: Randomly divide the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ;

[0009] Step S4: Based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ;

[0010] Step S5: Based on the calibration set Determining the ellipsoid size using preset statistical confidence conditions ;

[0011] Step S6: Construct the ellipsoidal uncertainty set The first scheduling scheme is obtained by converting it into a deterministic model through robust peer-to-peer transformation.

[0012] Step S7: Based on the solution results of step S6, reconstruct the uncertain set to obtain the reconstructed uncertain set. ;

[0013] Step S8: Solve the uncertainty set again based on the reconstructed set to obtain the re-solved scheduling model and the second scheduling scheme;

[0014] Step S9: Map the second scheduling scheme into a sequence of executable instructions for the device and send them to each cell controller for execution.

[0015] A sample-driven scheduling system for multi-pipeline cell preparation equipment includes:

[0016] The data acquisition and preprocessing module is used to acquire historical cell preparation task samples and equipment operating parameters, and to perform preprocessing.

[0017] The model building module establishes a cell preparation scheduling model with cross-contamination protection constraints.

[0018] The dataset partitioning module randomly divides the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ;

[0019] Uncertainty modeling module, based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ;

[0020] Size calibration module, based on calibration set The ellipsoid size s is determined using preset statistical confidence conditions;

[0021] The robust solution module constructs an ellipsoidal uncertainty set and solves it by robustly converting it into a deterministic model to obtain the first scheduling scheme.

[0022] The uncertain set reconstruction module reconstructs the uncertain set based on the solution results;

[0023] The reconstructed solution module solves the problem again based on the reconstructed uncertainty set, resulting in a resolvable scheduling model and a second scheduling scheme.

[0024] The instruction generation and distribution module maps the second scheduling scheme into a sequence of executable instructions for the device and distributes them to each cell controller for execution.

[0025] The beneficial effects of this invention are as follows: This invention uses limited historical sample data as the sole source of information, and by introducing a statistical calibration mechanism (rather than relying on distribution assumptions) into a robust optimization framework, it constructs an uncertainty set whose size is precisely controlled by statistical confidence conditions, and adapts this uncertainty set to multi-dimensional constraints such as cross-contamination protection unique to the cell preparation scenario, and finally obtains an executable scheduling scheme that combines statistical feasibility assurance and practical economic efficiency. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0028] This invention uses limited historical sample data as the sole source of information. Uncertainty is no longer "errors that need to be eliminated," but rather "risks that need to be covered by statistical guarantees"—decision-making is not made on point estimates, but rather on optimization within a statistically calibrated feasible region. Specifically:

[0029] like Figure 1 As shown, a sample-driven scheduling method for multi-pipeline cell preparation equipment includes the following steps:

[0030] A sample-driven scheduling method for multi-pipeline cell preparation equipment includes the following steps:

[0031] Step S1: Obtain historical cell preparation task samples and equipment operating parameters, and perform preprocessing.

[0032] Step S1 includes:

[0033] S101. Obtain historical cell preparation task samples and equipment operating parameters.

[0034] The historical cell preparation task samples include: process information, including cell type; donor information, including desensitized donor ID; uncertainty parameters, including actual process duration and quality inspection time; constraint information, including delivery deadline; resource information, including allocated equipment number, workstation number, and pipeline number; and special markers, including whether it is an urgent task and whether it involves viral vectors.

[0035] The equipment operating parameters include: the total number of workstations B per cell culture machine and the set of cell types that each workstation can serve; the total number of pipelines P per cell culture machine and the process stages that each pipeline can handle; and the process compatibility matrix between cell types. The matrix is A matrix, where |Type| represents the number of cell types, and the element in the a-th row and b-th column of matrix C. Minimum cleaning isolation and pipeline switching time required to switch from type a to type b; Isolation matrix D between donors, which is a |Donor|×|Donor| matrix, where |Donor| represents the number of donors, and the element in the a-th row and b-th column of the matrix. Minimum washing time required to switch from donor a to donor b; duration of enhanced isolation window for high-risk cell types. Consumable inventory capacity and replenishment time; cleaning and decontamination procedures for each workstation / pipeline;

[0036] S102. Preprocess historical cell preparation task samples and equipment operating parameters, including:

[0037] Desensitizing donor and patient information: Using hash mapping or tokenization to replace real identifiers, preserving the distinguishing relationship between donors but preventing reverse tracing;

[0038] Marking and filling missing fields;

[0039] Detect and mark outliers;

[0040] Align timestamps to the scheduling time granularity;

[0041] By mapping the process steps of different cell types to a unified process code, the process stages are normalized.

[0042] Stratified by cell type, risk level, and equipment model.

[0043] S103. In multiple scheduling cycles, execute steps S101 to S102 respectively to obtain samples for each scheduling cycle and form a historical sample set.

[0044] Step S2: Establish a cell preparation scheduling model with cross-contamination protection constraints;

[0045] Step S2 includes:

[0046] S201. Suppose there are K cell preparation devices in total, index Each piece of equipment has B workstations and P independent pipelines. Workstation index Independent Piping Index ;

[0047] The scheduling time span is H discrete time periods, and the set of tasks to be scheduled is Each task j has the following attribute parameters:

[0048] , indicating cell type;

[0049] This indicates the donor identifier after desensitization;

[0050] , indicating the expected arrival time of the sample;

[0051] This indicates the actual processing time. ,in For reference duration, This is due to duration deviation;

[0052] Indicates the delivery deadline;

[0053] Indicates task priority: Normal = 0, Urgent = 1;

[0054] This indicates the risk level, whether it involves viral vectors, and whether it is high-risk.

[0055] S202. Define decision variables, including:

[0056] , indicating whether task j is assigned to device k;

[0057] , indicating whether task j is assigned to the b-th workstation of device k;

[0058] , indicating whether task j uses the p-th pipeline of device k;

[0059] , indicating the planned start time of task j;

[0060] , indicating the planned end time of task j;

[0061] This indicates whether workstation b of device k is occupied during time period t;

[0062] This indicates whether station b of equipment k is in a cleaning and isolation state during time period t;

[0063] This indicates whether task i precedes task j at the same workstation on the same equipment.

[0064] This indicates whether the pipe p of device k is occupied during time period t;

[0065] S203. Define the task uncertainty vector The random disturbances that affect the feasibility of equipment scheduling within the scheduling cycle are jointly characterized; the random disturbances include one or more of the following: sample arrival time deviation, process duration deviation, quality inspection release time deviation, cleaning and isolation time deviation, order quantity deviation, and emergency task arrival deviation; each sample in the historical sample set corresponds to an uncertainty vector within a scheduling cycle.

[0066] The definition methods include the following two:

[0067] (1) Construct at the granularity of all tasks within the current scheduling cycle:

[0068] ;

[0069] in, This represents the sample arrival time deviation from the first task to the Nth task; This represents the deviation in process time from the first task to the Nth task; This represents the time deviation between the check and release of the first and Nth tasks. This represents the time deviation for cleaning and isolation from task 1 to task N;

[0070] In the embodiments of this application, the order quantity deviation is not listed as a separate dimension in the above vector, but is reflected by the random fluctuation of the total number of tasks N within the scheduling period. That is, the natural variation of N in different historical samples characterizes the uncertainty of the order quantity.

[0071] (2) Construct a fixed-dimensional vector with cell type, process stage, or time window as the granularity:

[0072]

[0073] The subscripts MSC, NK, CIK, and CART represent different types of tasks, corresponding to the preparation of mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively. These represent the process time deviation vectors for mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively, and are formed by summing the process time deviations for the corresponding task types. These represent the quality inspection and release time deviation vectors for mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively, and are formed by summing the quality inspection and release time deviations for the corresponding task types. These represent the wash-isolation time deviation vectors for mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively, and are formed by summing the wash-isolation time deviations for the corresponding task types. These represent the sample arrival time deviation vectors for mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively, and are formed by summing the sample arrival time deviations for the corresponding task types. This indicates the emergency mission type deviation (i.e., the additional random perturbations in the number and time of emergency missions compared to regular missions).

[0074] In the embodiments of this application, it is assumed that the number of MSC, NK, CIK, and CART tasks are N1, N2, N3, and N4, and N1 + N2 + N3 + N4 = N; then The number of process time deviations included are N1, N2, N3, and N4, respectively. The number of deviations in quality inspection and release time included are N1, N2, N3, and N4, respectively. The number of deviations in cleaning and isolation time included are N1, N2, N3, and N4, respectively. The number of sample arrival time deviations included are N1, N2, N3, and N4, respectively.

[0075] S204. Define the objective function and deterministic constraints:

[0076] The objective is to minimize the total scheduling cost:

[0077] ;

[0078] in: The unit time period equipment occupancy cost for task j; The unit time delay penalty cost for task j; For delayed slack variables: for ordinary tasks, For urgent clinical missions, ; Cost of cleaning and isolation per unit time period; Cost of workstation vacancy per unit time period;

[0079] The deterministic constraints include:

[0080] (1) Task assignment constraint: Each task must be assigned to one and only one workstation of one device:

[0081] ;

[0082] (2) Workstation capacity constraint: Each workstation can be occupied by at most one task at any given time;

[0083] (3) Process non-preemption constraint: Once a task begins, it must be executed continuously until completion;

[0084] (4) Process sequence constraints:

[0085] ;

[0086] This constraint is a sequential constraint for different process steps within the same task, where the superscript sep represents the separation / sorting process, act represents the activation process, exp represents the amplification process, QA represents the quality inspection and release process, and form represents the formulation packaging process;

[0087] (5) Pipeline occupancy constraint: The same pipeline cannot be used simultaneously by different batches of incompatible tasks at the same time period;

[0088] (6) Donor isolation constraints:

[0089] For the same workstation of the same equipment If two tasks i and j are executed consecutively, and ,but:

[0090] ;

[0091] Where T1 is the minimum cleaning and isolation time between different donors;

[0092] (7) Cell type compatibility constraints;

[0093] For two tasks i and j executed consecutively on the same pipeline p of the same device k, if ≠ ,but:

[0094] ;

[0095] CompMatrix is ​​a process compatibility matrix, which includes the cleaning and pipeline switching time required to switch the pipeline from type a to type b.

[0096] (8) Strengthen isolation and constraints for high-risk tasks:

[0097] If task j involves viral vector operations. During the T3 time period before and after task j, other tasks are prohibited from occupying the same workstation and all shared pipelines.

[0098] (9) Consumable path conflict constraint: Tasks that share the same consumable path are subject to the exclusive constraint of that path.

[0099] (10) Uncertainty tolerance constraint;

[0100]

[0101] in, For the set of all constraints that need to be satisfied jointly, Indicates the first A constraint in uncertainty The following conditions are met: ε is the preset allowed violation probability, and 1-ε is the joint feasible probability of the objective.

[0102] Step S3: Randomly divide the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ;

[0103] Step S3 includes:

[0104] set up In a complete record of historical scheduling cycles, each sample corresponds to an uncertainty vector within a scheduling cycle. The subsets are randomly divided into two disjoint subsets:

[0105] Shape set : Contains 60% of the samples in the complete record, with a sample size of n1;

[0106] Calibration set : Contains 40% of the complete records, with a sample size of n²;

[0107] The randomness of the data partitioning ensures that the two sets are statistically identically distributed.

[0108] Step S4: Based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ;

[0109] Step S4 includes:

[0110] S401. Calculate the sample mean vector:

[0111] ;

[0112] S402. Calculate the sample covariance matrix:

[0113] ;

[0114] It is a d×d matrix, where d is the dimension of the uncertainty vector. Used to characterize the correlation structure and degree of dispersion between uncertainties of different dimensions;

[0115] S403. Covariance Matrix Regularization

[0116] When the sample covariance matrix is ​​not full rank, the condition number exceeds a preset threshold, or the number of samples is insufficient, regularization is performed to ensure numerical stability.

[0117] The regularization method is ;

[0118] Alternatively, one of the following methods can be used: Ledoit-Wolf shrinking of the covariance matrix, diagonal loading, or pseudo-inverse.

[0119] Step S5: Based on the calibration set Determining the ellipsoid size using preset statistical confidence conditions ;

[0120] Step S5 includes:

[0121] S501. For the calibration set Each sample in Calculate its Mahalanobis distance:

[0122] ;

[0123] S502. Since the calibration set contains n² samples, n² Mahalanobis distance values ​​are obtained. Arranged in ascending order, we get:

[0124] ;

[0125] S503. Based on the preset violation probability and confidence level The target index is determined by the binomial distribution sorting method. :

[0126] ;

[0127] This sorting index is used to ensure that the sorting index is not lower than... The confidence level of the sorted samples The determined size parameters cover at least 1-ε of the sample quality in the true distribution;

[0128] S504. Obtain dimensional parameters Construct the ellipsoidal uncertainty set:

[0129] ;

[0130] The ellipsoid Centered on the axis, each principal axis direction is from The eigenvectors are determined, and the length of each axis is determined by the corresponding eigenvalues ​​and the dimension parameter s.

[0131] The statistical feasibility theorem guarantees that: Then it shall be no less than confidence level Cover at least The samples are real and uncertain, ensuring that they do not depend on the distribution pattern.

[0132] Step S6: Construct the ellipsoidal uncertainty set The first scheduling scheme is obtained by converting it into a deterministic model through robust peer-to-peer transformation.

[0133] Step S6 includes:

[0134] Embedding the ellipsoidal uncertainty set into the scheduling model, for any linear uncertainty constraint Robust equivalence is:

[0135]

[0136] The model is either MIP or MISOCP, and the first scheduling scheme is obtained by solving the problem using a solver. .

[0137] Step S7: Based on the solution results of step S6, reconstruct the uncertain set to obtain the reconstructed uncertain set. ;

[0138] Step S7 includes:

[0139] S701. Based on the first scheduling scheme Define a uniform constraint violation margin function:

[0140] ;

[0141] in For all constraints that need to be satisfied jointly, This indicates that the constraint is satisfied. These are normalization parameters;

[0142] S702. To Calculate for each sample The calibration is performed in the same manner as in step S5. ;

[0143] S703. Reconstructing the Uncertain Set: ;

[0144] in For the physical boundary or the original ellipsoid, when Also obtained from sample calibration, and merged into a unified scoring function:

[0145]

[0146] Then, a unified sorting and calibration is performed; the unified scoring function unifies the two heterogeneous risk measures, the basic ellipsoid deviation and the constraint violation margin, into a single scalar index. By performing a sorting and binomial confidence calibration on this scalar, the statistical coverage of the joint opportunity constraints is naturally maintained, avoiding the Bonferroni conservative effect caused by calibrating each constraint dimension separately.

[0147] Step S8: Solve the uncertainty set again based on the reconstructed set to obtain the re-solved scheduling model and the second scheduling scheme;

[0148] Based on the set of reconstruction uncertainties obtained in step S7 The cell preparation scheduling model is solved again to obtain the second scheduling scheme. Because... The second scheduling scheme more accurately depicts the real joint constraint risk area and has a lower total scheduling cost than the first scheduling scheme.

[0149] In the embodiments of this application,

[0150] While steps S2-S8 provide a statistically feasible and non-conservative scheduling scheme, the complete MIP / MISOCP problem needs to be solved again each time a new order arrives, with online solution time ranging from several minutes to tens of minutes. To meet the timeliness requirements of online scheduling, a machine learning-based solution acceleration method was designed.

[0151] Sa: Historical Task Pattern Clustering

[0152] Collect historical solution instances (scheduling data already solved by RSRO), and extract the task set feature vector f for each instance, including:

[0153] The proportion of tasks for each cell type;

[0154] Percentage of urgent tasks;

[0155] Average tightness of the task time window = Average (process time / available time until delivery deadline);

[0156] Task arrival times are distributed across different time periods (morning / noon / evening shifts).

[0157] Donor diversity (the proportion of tasks performed by different donors).

[0158] Clustering methods (k-means, spectral clustering, hierarchical clustering, or Gaussian mixture model, etc.) are used to divide historical instances into multiple task pattern clusters.

[0159] Sb: Training the scheduling state classifier

[0160] For each task pattern cluster, extract the discrete scheduling state labels and input features of all historical instances in the cluster, and train the scheduling state prediction model.

[0161] Discrete scheduling status labels include (in units of workstation / pipeline):

[0162] Are each workstation occupied at each time period?

[0163] What type of task is being performed at each time period and workstation?

[0164] Are all workstations in a cleaning and isolation state at each time period?

[0165] The usage status of each pipeline at different times.

[0166] The classifier model can employ gradient boosting trees (LightGBM / XGBoost), random forests, support vector machines, neural networks, or temporal Transformer models. The classifier can be trained separately for each device, time window, and resource type, or a multi-output classification model can be used to predict multiple discrete states simultaneously.

[0167] Sc: Online Prediction

[0168] When a new batch of tasks arrives:

[0169] 1. Extract the feature vectors of the task set;

[0170] 2. Determine the task pattern cluster to which the feature vector belongs through a clustering model;

[0171] 3. Call the scheduling status classifier corresponding to the cluster to predict the occupancy status, cleaning status and pipeline occupancy status of each workstation in each time period.

[0172] Sd: Feasibility Check and Remediation

[0173] The predicted discrete states may violate certain deterministic constraints (such as exceeding the upper limit of workstation capacity or failing to meet donor isolation requirements). Therefore, a feasibility check and remediation should be performed on the predicted states.

[0174] Se: Dimensionality Reduction Solution

[0175] By fixing the discrete state variables, which have undergone feasibility repair, to constants, the original MIP / MISOCP scheduling model degenerates into linear programming (LP), second-order cone programming (SOCP), or a smaller-scale mixed-integer model containing only continuous variables. Since LP / SOCP can be solved efficiently in polynomial time, the overall online solution time is significantly reduced.

[0176] Sf: Safety rollback mechanism

[0177] If the predicted state still fails to meet cross-contamination prevention constraints or hard delivery deadline constraints (especially for urgent clinical tasks) after feasibility correction, the machine learning prediction results are abandoned, and the process reverts to solving the full robust optimization model in step S8. The learning-based acceleration module only provides candidate discrete states and does not change the feasibility criteria for scheduling constraints. The final scheduling scheme still needs to pass deterministic feasibility checks and robust feasibility checks.

[0178] Step S9: Map the second scheduling scheme into a sequence of executable instructions for the device and send them to each cell controller for execution.

[0179] In the embodiments of this application, the decision variables in the second scheduling scheme are converted into a sequence of control instructions executable by the device through instruction generation.

[0180] Task start / end time → Device controller scheduling queue;

[0181] Equipment allocation and workstation allocation → MES work instructions;

[0182] Piping selection and switching time → Equipment piping control commands;

[0183] Cleaning and isolation period → Automatic cleaning program trigger time.

[0184] (2) Command issuance: Commands are issued to each cell machine for execution through the MES / SCADA system or by communicating directly with the device controller.

[0185] In the embodiments of this application, monitoring and dynamic rescheduling can also be performed: the actual execution status of each cell machine is continuously monitored. Rescheduling is triggered when the following events are detected:

[0186] (a) Emergency insertion of single dispatch

[0187] When receiving an urgent clinical preparation task:

[0188] Freeze tasks that have already begun execution and tasks that are in an uninterrupted process phase (such as virus transduction);

[0189] Release unstarted and adjustable tasks;

[0190] Update the task uncertainty vector (incorporating the uncertainty of urgent tasks);

[0191] Recalibrate the uncertainty set based on the updated data (historical calibration results can be reused to speed up the process).

[0192] Reschedule the released task set;

[0193] Assess the impact on delivery deadlines for affected tasks and notify operators if necessary.

[0194] (b) Equipment failure / process deviation rescheduling

[0195] When a cell culture machine malfunctions or the actual processing time of a batch of cells deviates significantly from the reference value:

[0196] The freeze has begun / completed.

[0197] Remove the device from the available resource pool (failure) or update its available time window (deviation);

[0198] Reassign affected tasks to other available devices;

[0199] Update the cleaning and isolation arrangements.

[0200] A sample-driven scheduling system for multi-pipeline cell preparation equipment includes:

[0201] The data acquisition and preprocessing module is used to acquire historical cell preparation task samples and equipment operating parameters, and to perform preprocessing.

[0202] The model building module establishes a cell preparation scheduling model with cross-contamination protection constraints.

[0203] The dataset partitioning module randomly divides the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ;

[0204] Uncertainty modeling module, based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ;

[0205] Size calibration module, based on calibration set The ellipsoid size s is determined using preset statistical confidence conditions;

[0206] The robust solution module constructs an ellipsoidal uncertainty set and solves it by robustly converting it into a deterministic model to obtain the first scheduling scheme.

[0207] The uncertain set reconstruction module reconstructs the uncertain set based on the solution results;

[0208] The reconstructed solution module solves the problem again based on the reconstructed uncertainty set, resulting in a resolvable scheduling model and a second scheduling scheme.

[0209] The instruction generation and distribution module maps the second scheduling scheme into a sequence of executable instructions for the device and distributes them to each cell controller for execution.

[0210] The solution of this application will be described in detail below with reference to specific embodiments:

[0211] Example 1: Routine Daily Cell Preparation Scheduling

[0212] This embodiment focuses on key operational tasks within a scheduling day, including sample reception, separation, activation, fluid exchange, transduction, sampling, and pre-quality control processing. Complete cell preparation batches can span multiple scheduling cycles; for cross-cycle tasks, the occupied workstations and pipeline status are input into the scheduling model as the initial frozen state.

[0213] Scene settings

[0214]

[0215] Detailed steps

[0216] (1) Data preparation and preprocessing

[0217] Complete records for 500 historical scheduling days were extracted from the historical database. Each historical scheduling day includes data such as process time deviation, quality inspection time deviation, and sample arrival deviation for all tasks on that day, forming a complete uncertainty vector. .

[0218] Preprocessing is performed on the data for each historical scheduling day: donor information is desensitized, missing values ​​are filled with the median of similar tasks, outliers (exceeding the 99th percentile) are marked, timestamps are aligned to 15-minute time intervals, and process stages are normalized.

[0219] Verification: n2=200 ≥ log(0.05) / log(0.95) ≈ 58.4, which satisfies the premise of statistical feasibility.

[0220] (2) Data partitioning

[0221] Randomly sample 300 historical scheduling days as shape sets The remaining 200 are used as a calibration set. .

[0222] (3) Shape estimation

[0223] based on Calculate the mean vector Covariance Matrix (Including regularization). This embodiment uses a fixed-dimensional uncertainty vector aggregated by cell type × time period (early / mid / evening shift) to ensure that the sample dimension is consistent across different historical scheduling days (the number of tasks may vary). The off-diagonal blocks revealed the correlation of uncertainty between different cell types and different time periods.

[0224] (4) Dimensional calibration

[0225] right Mahalanobis distances were calculated for each of the 200 samples, and the samples were sorted in ascending order. When n²=200, ε=0.05, and δ=0.05, the target index was calculated according to the binomial sorting rule: X ~ Binomial(200, 0.95), with a mean of 190, a standard deviation of approximately 3.08, and the 95th percentile corresponding to... Approximately 196 (the exact value is based on the cumulative probability calculation of the precise binomial distribution). Take... .

[0226] (5) Robust Equivalence Model Solving (SRO)

[0227] Construct an ellipsoidal uncertainty set, embed each constraint in a general robust equivalence form, and solve the MISOCP / MIP model with a solver to obtain the first scheduling scheme.

[0228] (6) Reconstruction of Uncertain Sets (RSRO)

[0229] Based on the first scheduling scheme, calculate the uniform constraint violation margin function for 200 samples in D2. The normalization coefficients for each constraint ensure comparability across different dimensions (hours, number of time periods, number of times occupied).

[0230] Sort Values, selected according to the same confidence conditions. ≈196, calculate The calibration is performed in the same manner as in step S5. Construct a reconstructed set. The constraint violation margin threshold of the reconstructed set is usually near zero or slightly less than zero (indicating that most historical samples did not actually touch the joint constraint red line under the first scheduling scheme), so that the uncertainty set shrinks in the non-risk direction.

[0231] (7) Reconstruction and Solving

[0232] The second scheduling scheme is obtained by solving the set of uncertainties through reconstruction.

[0233] (8) Learning Acceleration (Optional)

[0234] Clustering the feature vectors of 500 historical scheduling instances yields 5 task pattern clusters. A scheduling state classifier is trained for each cluster (classifiers are trained separately for device, time window, and resource type; multi-output models can also be used). When a new task arrives: extract features → determine cluster affiliation → call the corresponding model for prediction → feasibility repair → dimensionality reduction solution. If repair fails, it reverts to the RSRO full solution.

[0235] (9) Command generation and issuance

[0236] Map the second scheduling scheme to a sequence of executable instructions. Example:

[0237] Partial Instruction Sequence for Equipment WK1-A: 06:00 - 06:30 Power-on Self-test; 06:30 - 07:00 Prepare MSC Operating Consumables; 07:00 - 07:15 Receive Sample MSC-J001 (Donor D001, Station 1); 07:15 - 07:30 Add MSC-J001 Culture Medium to Tube A; 07:30 - 09:30 MSC-J001 Isolation Operation (Station 1, Tube A Release); 07:45 - 08:00 Receive Sample MSC-J003 (Donor D002, Station 2); 08:00 - 08:15 Add MSC-J003 Culture Medium to Tube B; 08:15 - 08:30 [Station 2 Cleaning and Isolation: Donor Switching D001→D002 Not Applicable (Different Stations)] 08:15 - 10:15 MSC-J003 Separation Operation (Station 2, Pipeline B Release)...

[0238] 10:30 - 10:45 Pipeline A cleaning switch: MSC → NK 10:45 - 12:45 NK-J007 activation operation (Station 1, Pipeline A release)

[0239] The instruction sequence is sent to the industrial control computer of each WK1 device via TCP / IP or OPC-UA protocol, and the device controller parses and executes it one by one.

[0240] Example 2: Emergency Single-Symbol Dispatch

[0241] Scenario: When the scheduling scheme of Example 1 is executed for 12 hours (i.e., 18:00), an urgent clinical CAR-T preparation task is received (donor D099, reference process time of 16 hours, delivery deadline of 36 hours - hard deadline, slack variable). =0).

[0242] Processing flow:

[0243] 1. Freeze: Identify tasks that have started at the current moment (14 in total) and mark them as unmodifiable; tasks in the virus transduction stage (2 in total) that have not yet been completed but cannot be interrupted are also frozen.

[0244] 2. Release: The remaining 32 unstarted and adjustable tasks are released.

[0245] 3. Update the uncertainty vector: Add the duration deviation information of emergency mission D099. .

[0246] 4. Recalibrate: Reuse historical calibration results (time < 1 second).

[0247] 5. Rescheduling: Remove the workstations and pipelines locked by the frozen tasks from the available resource pool, and perform RSRO solution on 33 tasks including release tasks and emergency tasks (including learning acceleration, total time is about 45 seconds).

[0248] 6. Assessment: Two MSC tasks in the original mission were delayed by 2.5 hours (soft deadline, within the tolerance range); the emergency CAR-T task was scheduled for WK1-C and is expected to be completed in 31 hours (meeting the 36-hour hard deadline).

[0249] 7. Command Update: Generate incremental update commands and send them to the controllers of affected devices.

[0250] In summary, this invention uses a binomial distribution ranking method for size calibration, providing a theoretical guarantee of statistical feasibility under finite sample conditions: with a confidence level of not less than 1-δ, the resulting scheduling scheme satisfies all constraints with a joint probability of not less than 1-ε. This guarantee does not depend on the true distribution of uncertainty parameters. Furthermore, by reconstructing the uncertainty set based on a unified constraint margin function (RSRO) and intersecting it with the basic uncertainty set, redundant coverage in non-risk directions is removed while maintaining the statistical guarantee of joint opportunity constraints, thereby reducing scheduling conservatism and improving equipment utilization and economy. The scheduling model explicitly models a hierarchical cross-contamination protection constraint system for equipment-workstation-pipeline-consumable paths—including workstation-level donor isolation, pipeline-level process compatibility matrix, and cross-resource-level high-risk task enhanced isolation—ensuring that the scheduling scheme meets GMP compliance requirements and reduces the risk of isolation window conflicts in multi-pipeline mixed-run scenarios.

[0251] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A sample-driven scheduling method for multi-pipeline cell preparation equipment, characterized in that: Includes the following steps: Step S1: Obtain historical cell preparation task samples and equipment operating parameters, and perform preprocessing; Step S2: Establish a cell preparation scheduling model with cross-contamination protection constraints; Step S3: Randomly divide the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ; Step S4: Based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ; Step S5: Based on the calibration set Determining the ellipsoid size using preset statistical confidence conditions ; Step S6: Construct the ellipsoidal uncertainty set The first scheduling scheme is obtained by converting it into a deterministic model through robust peer-to-peer transformation. Step S7: Based on the solution results of step S6, reconstruct the uncertain set to obtain the reconstructed uncertain set. ; Step S8: Solve the uncertainty set again based on the reconstructed set to obtain the re-solved scheduling model and the second scheduling scheme; Step S9: Map the second scheduling scheme into a sequence of executable instructions for the device and send them to each cell controller for execution.

2. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S1 includes: S101. Obtain historical cell preparation task samples and equipment operating parameters. The historical cell preparation task samples include: process information, including cell type; donor information, including desensitized donor ID; uncertainty parameters, including actual process duration and quality inspection time; constraint information, including delivery deadline; resource information, including allocated equipment number, workstation number, and pipeline number; and special markers, including whether it is an urgent task and whether it involves viral vectors. The equipment operating parameters include: the total number of workstations B per cell culture machine and the set of cell types that each workstation can serve; the total number of pipelines P per cell culture machine and the process stages that each pipeline can handle; and the process compatibility matrix between cell types. The matrix is A matrix, where |Type| represents the number of cell types, and the element in the a-th row and b-th column of matrix C. Minimum cleaning isolation and pipeline switching time required to switch from type a to type b; Isolation matrix D between donors, which is a |Donor|×|Donor| matrix, where |Donor| represents the number of donors, and the element in the a-th row and b-th column of the matrix. Minimum washing time required to switch from donor a to donor b; duration of enhanced isolation window for high-risk cell types. Consumable inventory capacity and replenishment time; cleaning and decontamination procedures for each workstation / pipeline; S102. Preprocess historical cell preparation task samples and equipment operating parameters, including: Desensitizing donor and patient information: Using hash mapping or tokenization to replace real identifiers, preserving the distinguishing relationship between donors but preventing reverse tracing; Marking and filling missing fields; Detect and mark outliers; Align timestamps to the scheduling time granularity; By mapping the process steps of different cell types to a unified process code, the process stages are normalized. Stratified by cell type, risk level, and equipment model; S103. In multiple scheduling cycles, execute steps S101 to S102 respectively to obtain samples for each scheduling cycle and form a historical sample set.

3. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S2 includes: S201. Suppose there are K cell preparation devices in total, index Each piece of equipment has B workstations and P independent pipelines. Workstation index Independent Piping Index ; The scheduling time span is H discrete time periods, and the set of tasks to be scheduled is Each task j has the following attribute parameters: , indicating cell type; This indicates the donor identifier after desensitization; , indicating the expected arrival time of the sample; This indicates the actual processing time. ,in For reference duration, This is due to duration deviation; Indicates the delivery deadline; Indicates task priority: Normal = 0, Urgent = 1; This indicates the risk level, whether it involves viral vectors, and whether it is high-risk. S202. Define decision variables, including: , indicating whether task j is assigned to device k; , indicating whether task j is assigned to the b-th workstation of device k; , indicating whether task j uses the p-th pipeline of device k; , indicating the planned start time of task j; , indicating the planned end time of task j; This indicates whether workstation b of device k is occupied during time period t; This indicates whether station b of equipment k is in a cleaning and isolation state during time period t; This indicates whether task i precedes task j at the same workstation on the same equipment. This indicates whether the pipe p of device k is occupied during time period t; S203. Define the task uncertainty vector The random disturbances that affect the feasibility of equipment scheduling within the scheduling cycle are jointly characterized; the random disturbances include one or more of the following: sample arrival time deviation, process duration deviation, quality inspection release time deviation, cleaning and isolation time deviation, order quantity deviation, and emergency task arrival deviation; each sample in the historical sample set corresponds to an uncertainty vector within a scheduling cycle. The definition methods include the following two: (1) Construct at the granularity of all tasks within the current scheduling cycle: ； in, This represents the sample arrival time deviation from the first task to the Nth task; This represents the deviation in process time from the first task to the Nth task; This represents the time deviation between the check and release of the first and Nth tasks. This represents the time deviation for cleaning and isolation from task 1 to task N; (2) Construct a fixed-dimensional vector with cell type, process stage, or time window as the granularity: ； The subscripts MSC, NK, CIK, and CART represent different types of tasks, corresponding to the preparation of mesenchymal stem cells, natural killer cells, cytokine-induced killer cells, and chimeric antigen receptor T cells, respectively. These represent the process time deviation vectors for each type of task, which are formed by summing the process time deviations for the corresponding task types. These represent the quality inspection and release time deviation vectors for each type of task, which are formed by summing the quality inspection and release time deviations for the corresponding task types. These represent the cleaning and isolation time deviation vectors for each type of task, which are formed by summing the cleaning and isolation time deviations for the corresponding task types. These represent the sample arrival time deviation vectors for each type of task, which are formed by summing the sample arrival time deviations for the corresponding task types. This indicates a deviation in the type of emergency mission; S204. Define the objective function and deterministic constraints: The objective is to minimize the total scheduling cost: ； in: The unit time period equipment occupancy cost for task j; The unit time delay penalty cost for task j; For delayed slack variables: for ordinary tasks, For urgent clinical missions, ; Cost of cleaning and isolating per unit time period; Cost of vacant workstations per unit time period; The deterministic constraints include: (1) Task assignment constraint: Each task must be assigned to one and only one workstation of one device: ； (2) Workstation capacity constraint: Each workstation can be occupied by at most one task at any given time; (3) Process non-preemption constraint: Once a task begins, it must be executed continuously until completion; (4) Process sequence constraints: ； This constraint is a sequential constraint for different process steps within the same task, where the superscript sep represents the separation / sorting process, act represents the activation process, exp represents the amplification process, QA represents the quality inspection and release process, and form represents the formulation packaging process; (5) Pipeline occupancy constraints: The same pipeline cannot be used simultaneously by different batches of incompatible tasks at the same time period; (6) Donor isolation constraints: For the same workstation of the same equipment If two tasks i and j are executed consecutively, and ,but: ； Where T1 is the minimum cleaning and isolation time between different donors; (7) Cell type compatibility constraints; For two tasks i and j executed consecutively on the same pipeline p of the same device k, if ≠ ,but: ； CompMatrix is ​​a process compatibility matrix, which includes the cleaning and pipeline switching time required to switch the pipeline from type a to type b. (8) Strengthen isolation and constraints for high-risk tasks: If task j involves viral vector operations. During the T3 time period before and after task j, other tasks are prohibited from occupying the same workstation and all shared pipelines. (9) Consumable path conflict constraint: Tasks that share the same consumable path are subject to the exclusive constraint of that path. (10) Uncertainty tolerance constraint; ； in, For the set of all constraints that need to be satisfied jointly, This indicates that the ℓth constraint is under uncertainty. The following conditions are met: ε is the preset allowed violation probability, and 1-ε is the joint feasible probability of the objective.

4. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S3 includes: set up In a complete record of historical scheduling cycles, each sample corresponds to an uncertainty vector within a scheduling cycle. The subsets are randomly divided into two disjoint subsets: Shape set : Contains 60% of the samples in the complete record, with a sample size of n1; Calibration set : Contains 40% of the complete records, with a sample size of n²; The randomness of the data partitioning ensures that the two sets are statistically identically distributed.

5. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 4, characterized in that: Step S4 includes: S401. Calculate the sample mean vector: ； S402. Calculate the sample covariance matrix: ； It is a d×d matrix, where d is the dimension of the uncertainty vector. Used to characterize the correlation structure and degree of dispersion between uncertainties of different dimensions; S403. Covariance Matrix Regularization When the sample covariance matrix is ​​not full rank, the condition number exceeds a preset threshold, or the number of samples is insufficient, regularization is performed to ensure numerical stability. The regularization method is ; Alternatively, one of the following methods can be used: Ledoit-Wolf shrinking of the covariance matrix, diagonal loading, or pseudo-inverse.

6. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S5 includes: S501. For the calibration set Each sample in Calculate its Mahalanobis distance: ； S502. Since the calibration set contains n² samples, n² Mahalanobis distance values ​​are obtained. Arranged in ascending order, we get: ； S503. Based on the preset violation probability and confidence level The target index is determined by the binomial distribution sorting method. : ； This sorting index is used to ensure that the sorting index is not lower than... The confidence level of the sorted samples The determined size parameters cover at least 1-ε of the sample quality in the true distribution; S504. Obtain dimensional parameters Construct the ellipsoidal uncertainty set: ； The ellipsoid Centered on the axis, each principal axis direction is from The eigenvectors are determined, and the length of each axis is determined by the corresponding eigenvalues ​​and the dimension parameter s. The statistical feasibility theorem guarantees that: Then it shall be no less than Confidence level, Cover at least The samples are real and uncertain, ensuring that they do not depend on the distribution pattern.

7. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S6 includes: Embedding the ellipsoidal uncertainty set into the scheduling model, for any linear uncertainty constraint Robust equivalence is: ； The model is either MIP or MISOCP, and the first scheduling scheme is obtained by solving the problem using a solver. ; For the deterministic part that is only related to the decision variables, The coefficient vector is multiplied by the uncertainty vector and used to map the uncertainty disturbance to the constraint space.

8. The sample-driven scheduling method for multi-pipeline cell preparation equipment according to claim 1, characterized in that: Step S7 includes: S701. Based on the first scheduling scheme Define a uniform constraint violation margin function: ； in For all constraints that need to be satisfied jointly, This indicates that the constraint is satisfied. These are normalization parameters; S702. To Calculate for each sample The calibration is performed in the same manner as in step S5. ; S703. Reconstructing the Uncertain Set: ; in For the physical boundary or the original ellipsoid, when Also obtained from sample calibration, and merged into a unified scoring function for unified ranking calibration. The unified scoring function is: ； The unified scoring function unifies the two heterogeneous risk measures, basic ellipsoid deviation and constraint violation margin, into a single scalar index. By performing a sorting and binomial confidence calibration on this scalar, the statistical coverage of the joint chance constraints is naturally maintained, avoiding the Bonferroni conservatism effect caused by calibrating each constraint dimension separately.

9. A sample-driven scheduling system for multi-pipeline cell preparation equipment, based on the method described in any one of claims 1 to 8, characterized in that: include: The data acquisition and preprocessing module is used to acquire historical cell preparation task samples and equipment operating parameters, and to perform preprocessing. The model building module establishes a cell preparation scheduling model with cross-contamination protection constraints. The dataset partitioning module randomly divides the complete records of multiple historical scheduling cycles into two disjoint subsets, denoted as shape sets respectively. and calibration set ; Uncertainty modeling module, based on shape set Calculate the uncertainty vector of the task sample mean vector and sample covariance matrix ; Size calibration module, based on calibration set The ellipsoid size s is determined using preset statistical confidence conditions; The robust solution module constructs an ellipsoidal uncertainty set and solves it by robustly converting it into a deterministic model to obtain the first scheduling scheme. The uncertain set reconstruction module reconstructs the uncertain set based on the solution results; The reconstructed solution module is used to solve the problem again based on the reconstructed uncertainty set, resulting in a resolvable scheduling model and a second scheduling scheme. The instruction generation and distribution module maps the second scheduling scheme into a sequence of executable instructions for the device and distributes them to each cell controller for execution.

Images