Machine learning based data driven perfusion process automatic regulation method

Abstract

The application discloses a data-driven perfusion process automatic adjustment method based on machine learning, comprising the following steps: collecting and preprocessing multi-source time series data to generate an aligned feature vector sequence; performing perfusion stage division and stage encoding vector generation based on the aligned feature vector sequence; constructing a three-layer liquid state machine model to determine the stage liquid pool activation sequence; performing stage gating encoding to drive the liquid pool to generate a dynamic state, complete cross-stage migration; integrating the dynamic state sequence to generate a joint state, input the joint state into a multi-task readout layer to output an adjustment parameter; performing distribution drift detection, topology update and parameter calibration to output a final perfusion adjustment parameter. Through the data-driven method based on stage topology modeling, liquid state machine dynamic evolution and multi-task readout mechanism, the application realizes accurate prediction, risk identification and adaptive adjustment of the perfusion process, and improves perfusion quality and long-term operation stability.

Description

Technical Field

[0001] This invention relates to the field of machine learning and data processing technology, and in particular to a data-driven automatic adjustment method for infusion processes based on machine learning. Background Technology

[0002] Injection processes are widely used in material mixing, resin encapsulation, molding, and chemical formulation. They typically require multi-source time-series data, including flow rate, pressure, volume, equipment operating status, and environmental parameters, to determine process stability. However, existing injection control methods often rely on fixed thresholds, empirical rules, or simple single-sequence prediction models, failing to capture the phased kinetic differences that occur over time. For example, the initial injection stage often exhibits strong pressure response sensitivity, the middle stage relies more on flow rate stability, and the final stage is significantly affected by factors such as equipment aging and changes in medium viscosity. Traditional control strategies generally treat the entire injection process as a single continuous process, ignoring phase characteristics, resulting in insufficient model expressive power.

[0003] While existing machine learning-based injection prediction or regulation methods can utilize historical data for trend forecasting, most employ fixed-structure neural networks, making it difficult to model complex, discontinuous behaviors across different stages. Machine learning models typically fit only a single stage or sequence, lacking the ability to characterize the switching logic or correlations between different stages, and thus failing to maintain stable performance under varying stage changes, operating condition fluctuations, or equipment operating conditions. Existing methods generally ignore the drift in data distribution caused by equipment wear, media aging, and changes in process conditions, resulting in a decline in model performance over time and a gradual deviation of regulation strategies from actual needs.

[0004] To address the complex, phased behavior of the perfusion process, some studies have attempted to improve it through rule-based phase division or cluster-based segmented prediction methods. However, these methods often rely on manual threshold settings and are difficult to adapt to dynamically changing operating environments. Existing technologies lack a unified framework that can form a closed loop between phase division, state modeling, and output regulation. They also fail to integrate distribution drift detection, adaptive topology updates, and incremental calibration of regulation parameters, resulting in a lack of self-correction capabilities in perfusion regulation methods during long-term operation.

[0005] Therefore, how to provide a data-driven, automated adjustment method for the perfusion process based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a data-driven automatic regulation method for the perfusion process based on machine learning. This invention comprehensively utilizes multi-source time-series data modeling, stage topology representation, liquid state machine reserve calculation, and a multi-task readout mechanism to construct a perfusion regulation framework capable of automatically identifying perfusion stages, characterizing cross-stage nonlinear dynamic behavior, and adaptively updating as data distribution changes. Throughout the perfusion process, this invention performs data alignment, stage division, dynamic state generation, joint state construction, and regulation parameter output. It also continuously corrects the regulation strategy through distribution drift detection and topology selection. This addresses the problems in existing technologies, such as difficulty in characterizing the dynamic differences between perfusion stages, difficulty in modeling discontinuous states across stages, and long-term performance degradation. It possesses advantages such as strong adaptability, high regulation accuracy, and good long-term stability.

[0007] The data-driven automatic adjustment method for perfusion process based on machine learning according to embodiments of the present invention includes:

[0008] Collect multi-source time-series data during the perfusion process, preprocess the multi-source time-series data, and generate an aligned feature vector sequence;

[0009] Injection phase partitioning is performed based on the aligned feature vector sequence, multiple phase nodes with phase features are identified, an injection phase topology graph is constructed, and a corresponding phase encoding vector is generated for each phase node.

[0010] For each stage node, a corresponding stage liquid pool is generated, a liquid state machine model is constructed, and the activation order of each stage liquid pool in the liquid state machine model is determined.

[0011] Stage-gated coding is performed based on the aligned feature vector sequence and the stage coding vector. The stage-gated code is then input into the activated stage liquid pool to perform cross-stage migration and obtain a continuous dynamic state sequence.

[0012] The continuous dynamic state sequence is integrated to generate a joint state representation. The joint state representation is then input to the multi-task readout layer, which outputs the future perfusion trend prediction results, perfusion deviation risk assessment results, and perfusion regulation parameters.

[0013] Based on the final deviation results of the infusion process, distribution drift detection is performed. When the input distribution drifts, the topology of the infusion stage is updated by topology selection, and incremental calibration is performed on the multi-task readout layer to generate calibrated infusion regulation parameters.

[0014] Optionally, the multi-source time-series data includes flow signals, pressure signals, cumulative injection volume signals, pump drive current signals, pump speed signals, equipment operating temperature signals, and ambient temperature and humidity signals that characterize the dynamic changes in the injection process.

[0015] Optionally, generating the aligned feature vector sequence refers to constructing a non-uniform time index sequence based on the gradient change rate and numerical fluctuation amplitude of multi-source time series data, using the non-uniform time index sequence to resample and synchronize multi-source time series data at a unified time step, and arranging the multi-dimensional features corresponding to each time step in chronological order to form a time series vector sequence composed of multiple time step feature vectors.

[0016] Optionally, the construction of the infusion stage topology graph, generating a corresponding stage encoding vector for each stage node, includes:

[0017] The alignment feature vectors corresponding to each time step in the alignment feature vector sequence are preprocessed to obtain the temporal feature input sequence according to the preset smoothing, denoising and scaling rules;

[0018] Based on the analysis of the change trend between adjacent time steps by inputting the time sequence with time series features, time intervals with similar change patterns and continuous occurrences are divided into the same candidate perfusion stage, and time intervals with different change patterns are divided into different candidate perfusion stages, resulting in multiple candidate stage segments arranged along the time axis.

[0019] The temporal features of each candidate stage segment are statistically analyzed, the comprehensive stage feature index of the candidate stage segment is calculated, and the candidate stage segments are merged or split to form multiple target stage segments with stable stage features, which serve as stage nodes in the perfusion stage topology graph.

[0020] The phase division is repeatedly performed on multiple infusion batches, and the temporal adjacency relationship of each target phase segment is recorded. Combined with the infusion deviation results of each infusion batch, the adjacency relationship between phase nodes is divided into three categories: normal phase transfer, corrected phase transfer, and abnormal phase transfer. Corresponding normal phase transfer edges, corrected phase transfer edges, and abnormal phase transfer edges are constructed respectively, forming an infusion phase topology graph containing a set of phase nodes and a set of phase transfer edges with category labels.

[0021] Based on the statistical results of the stage features of the target stage segment corresponding to each stage node and the connection relationship of the stage node in the infusion stage topology graph, a unique stage encoding vector is generated for each stage node.

[0022] Optionally, constructing the liquid state machine model includes:

[0023] For each stage node in the topology diagram of the infusion stage, the corresponding stage feature statistics and stage encoding vector are obtained. Each stage node is associated with a stage liquid pool to form a mapping relationship from stage node to stage liquid pool. The liquid state machine model is determined to consist of a stage input encoding layer, a stage reserve calculation layer and a global coordination and convergence layer.

[0024] In the stage input coding layer, based on the alignment feature vector sequence and the stage coding vector corresponding to each stage node, an input coding substructure is configured for each stage liquid pool. The combination method of each feature component in the alignment feature vector and each component in the stage coding vector at the input end of the stage liquid pool and the weight allocation rule are determined. The alignment feature vector and the stage coding vector are converted into input signals that adapt to the dynamic evolution inside the stage liquid pool.

[0025] In the stage reserve calculation layer, based on the stage characteristic statistical information of each stage node, the internal state dimension, state update time scale and internal connection sparsity are set for the corresponding stage liquid pool, and the internal structure of the stage liquid pool for reserve calculation of dynamic behavior of different injection stages is constructed. All stage liquid pools are combined to form the stage reserve calculation substructure of the liquid state machine model.

[0026] In the global coordination and aggregation layer, a global coordination unit is configured for all stage liquid pools. The dynamic states generated by each stage liquid pool during activation are collected and merged according to preset rules to generate a global coordination state. A fixed mapping relationship is established between the global coordination state and the multi-task readout layer.

[0027] Based on the adjacent relationships of stage nodes and the types of stage transition edges in the topology graph of the infusion stage, the activation order and cross-stage migration path of the liquid pool in each stage are determined, and the running order of the stage input coding layer, the stage reserve computing layer and the global coordination and convergence layer are fixed as the activation sequence.

[0028] Optionally, the step of performing cross-stage migration to obtain a continuous dynamic state sequence includes:

[0029] The aligned feature vector is combined with the stage coding vector to form a joint input for stage-gated coding;

[0030] Based on the joint input, the stage gating weights are calculated for the current time step in the stage input coding layer. The stage coding vector is used to adjust the weight distribution of each feature component in the aligned feature vector, resulting in a stage-gated coding vector containing the feature components after stage gating.

[0031] The stage gating encoding vector is input into the stage liquid pool in the active state of the liquid state machine model. The internal state is updated in the stage reserve calculation layer according to the dynamic parameters of the stage liquid pool to generate the stage liquid pool dynamic state, which is then sent to the global coordination and convergence layer to participate in the update of the global coordination state.

[0032] When it is detected that the stage node to which the current time step belongs is different from the stage node to which the previous time step belongs, the dynamic state of the liquid pool of the previous stage is called in the global coordination and convergence layer, and cross-stage migration is performed on the liquid pool of the next stage that is about to be activated. The dynamic state of the liquid pool of the previous stage is combined with the initial state of the liquid pool of the next stage to generate the transition dynamic state during stage switching.

[0033] The dynamic states of the liquid pool at each stage and the transitional dynamic states are connected in series and recorded in chronological order to form a continuous dynamic state sequence covering the entire injection process.

[0034] Optionally, the output of future perfusion trend prediction results, perfusion deviation risk assessment results, and perfusion adjustment parameters includes:

[0035] Obtain a continuous dynamic state sequence, organize the continuous dynamic state sequence in chronological order, and perform completion processing on abnormal interruptions and missing states to form a normalized continuous dynamic state sequence.

[0036] Based on the normalized continuous dynamic state sequence, the dynamic state is divided according to the preset perfusion stage boundary. The dynamic state is aggregated in each window or stage to extract multi-scale dynamic state features and associate the multi-scale dynamic state features with the corresponding time position.

[0037] We perform weighted integration on the multi-scale dynamic state features at different time scales, and fuse the multi-scale dynamic state features corresponding to the same time position into a set of joint state feature vectors. We then perform sequence-level aggregation on the joint state feature vectors in the time dimension to generate a joint state representation.

[0038] The joint state representation is input to the multi-task readout layer, which includes a shared feature transformation sublayer, a task-specific output sublayer, and a task cooperative constraint sublayer, wherein:

[0039] The shared feature transformation sublayer consists of feature channel clusters divided according to the category of infusion stage. Different channel clusters and hierarchical structures are selected according to the stage transition connection mode of the infusion stage topology graph, and grouped multi-layer feature transformation is performed on the joint state representation.

[0040] The task-specific output sublayer includes three task output paths corresponding to future perfusion trends, perfusion deviation risks, and perfusion adjustment parameters, respectively. Each path selects a subset of associated features from the feature channel cluster and forms the task output structure through independently connected feature mapping units and sparse lateral coupling units.

[0041] The task coordination constraint sublayer is set after the three task output paths. It consists of a group of coordination nodes connected in a preset topological order and cross-task connection edges. The coordination nodes receive the intermediate representations of each task output path in sequence and perform cross-task state updates within a limited number of iterations, providing the future perfusion trend prediction results, perfusion deviation risk assessment results, and perfusion adjustment parameters after coordination constraints.

[0042] Optionally, generating the calibrated perfusion regulation parameters includes:

[0043] Based on the final deviation results of the infusion process, the alignment feature vector sequence of the corresponding infusion batch is divided into key detection interval and benchmark reference interval, and the distribution characteristics of the alignment feature vector in the two types of intervals are extracted respectively.

[0044] By comparing the distribution characteristics of the key detection interval with the benchmark reference interval, it is determined whether the input distribution has drifted. When it is determined that a drift has occurred, the set of key stage nodes that caused the drift is determined based on the attribution relationship of the alignment feature vectors at each stage node.

[0045] For the set of key stage nodes, based on the connection method of each stage node in the injection stage topology graph and the category of the corresponding stage transfer edge, multiple topology perturbation candidate structures are constructed by adjusting the type, number and order of the stage transfer edges related to the key stage nodes. The topology perturbation candidate structures form several candidate injection stage topology graphs by adjusting the type, number and order of the stage transfer edges related to the key stage nodes.

[0046] The aligned feature vector sequence containing the key detection interval and the benchmark reference interval is replayed and calculated using the topology map of each candidate perfusion stage. Based on the output results of the multi-task readout layer under different candidate topologies, the topology with better performance is selected as the updated perfusion stage topology map. Incremental calibration is performed on the parameters of the multi-task readout layer to generate calibrated perfusion adjustment parameters.

[0047] The beneficial effects of this invention are:

[0048] This invention introduces a stage division mechanism and stage topology representation oriented towards the characteristics of the perfusion process, enabling the system to automatically identify different stages of the perfusion process without relying on manual rules, thereby depicting the widespread stage-specific dynamic differences in the perfusion process. By combining aligned feature vector sequences with stage encoding vectors, the system can accurately capture the changing trends and key features of each stage, making the division of perfusion stages more robust and the stage transition logic clearer, thus effectively improving the accuracy and stability of perfusion process modeling.

[0049] This invention further constructs a liquid state machine model consisting of a staged liquid pool, a staged input encoding layer, a staged reserve computing layer, and a global coordination and convergence layer. This enables the model to achieve continuous and smooth dynamic state transitions between different stages, effectively characterizing the complex nonlinear behavior and cross-stage state discontinuities during the perfusion process. Through a unified feature transformation structure and inter-task collaborative constraint mechanism in the multi-task readout layer, this invention can simultaneously output future trend predictions, deviation risk assessments, and adjustment parameters, achieving comprehensive judgment and precise adjustment of the perfusion process. This allows the adjustment strategy to respond to changes in the perfusion state in real time, improving adjustment efficiency and accuracy.

[0050] This invention introduces a distribution drift detection, topology perturbation candidate structure construction, and topology selection update mechanism. This enables the system to automatically identify key stage nodes and adaptively update the topology map of the injection stage when data distribution drifts due to equipment aging, changes in media properties, or changes in environmental conditions. At the same time, it performs incremental calibration on the parameters of the multi-task readout layer, enabling the model to maintain stable performance over a long period of time. This avoids the problem of traditional fixed models gradually degrading over time, thus enabling the injection process to maintain high precision adjustment capability and reliability during long-term operation, achieving the beneficial effects of long-term self-evolution and self-optimization. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of the data-driven automatic adjustment method for perfusion process based on machine learning proposed in this invention;

[0053] Figure 2 This is a schematic diagram of the liquid state machine model structure of the data-driven automatic adjustment method for perfusion process based on machine learning proposed in this invention.

[0054] Figure 3 This is a schematic diagram illustrating the operation of the stage-gated encoding input to the stage liquid pool and the generation of dynamic states in the data-driven automatic adjustment method for the perfusion process based on machine learning proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 , Figure 2 and Figure 3 A data-driven automatic regulation method for the perfusion process based on machine learning includes:

[0057] Collect multi-source time-series data during the perfusion process, preprocess the multi-source time-series data, and generate an aligned feature vector sequence;

[0058] Injection phase partitioning is performed based on the aligned feature vector sequence, multiple phase nodes with phase features are identified, an injection phase topology graph is constructed, and a corresponding phase encoding vector is generated for each phase node.

[0059] For each stage node, a corresponding stage liquid pool is generated, a liquid state machine model is constructed, and the activation order of each stage liquid pool in the liquid state machine model is determined.

[0060] Stage-gated coding is performed based on the aligned feature vector sequence and the stage coding vector. The stage-gated code is then input into the activated stage liquid pool to perform cross-stage migration and obtain a continuous dynamic state sequence.

[0061] The continuous dynamic state sequence is integrated to generate a joint state representation. The joint state representation is then input to the multi-task readout layer, which outputs the future perfusion trend prediction results, perfusion deviation risk assessment results, and perfusion regulation parameters.

[0062] Based on the final deviation results of the infusion process, distribution drift detection is performed. When the input distribution drifts, the topology of the infusion stage is updated by topology selection, and incremental calibration is performed on the multi-task readout layer to generate calibrated infusion regulation parameters.

[0063] In this embodiment, the multi-source time-series data includes flow rate signal, pressure signal, cumulative injection volume signal, pump drive current signal, pump speed signal, equipment operating temperature signal, and ambient temperature and humidity signal, which characterize the dynamic changes of the injection process.

[0064] In this embodiment, generating the aligned feature vector sequence refers to constructing a non-uniform time index sequence based on the gradient change rate and numerical fluctuation amplitude of multi-source time series data, using the non-uniform time index sequence to resample and synchronize multi-source time series data at a unified time step, and arranging the multi-dimensional features corresponding to each time step in chronological order to form a time series vector sequence composed of multiple time step feature vectors.

[0065] In this embodiment, the step of constructing the infusion stage topology graph and generating a corresponding stage encoding vector for each stage node includes:

[0066] The aligned feature vectors at each time step in the aligned feature vector sequence are preprocessed to obtain the temporal feature input sequence according to preset smoothing, denoising, and scaling rules. The preset smoothing, denoising, and scaling rules are as follows:

[0067] Smoothing is performed using a sliding window of size 5, noise suppression is performed using a bandpass noise reduction rule with a cutoff frequency of 0.35, and scale normalization is performed using a linear scale unification rule based on the range of minimum and maximum feature dimensions in the [0, 1] interval.

[0068] Based on the analysis of the time-series input sequence, the changing trends between adjacent time steps are analyzed. Time intervals with similar changing patterns and continuous occurrences are divided into the same candidate injection stage, and time intervals without similar changing patterns are divided into different candidate injection stages, resulting in multiple candidate stage segments arranged along the time axis. The similar changing pattern refers to the time series behavior in which adjacent time steps have the same direction of change, similar magnitude of change, and continuous distribution of change rate in terms of flow rate change, pressure response, medium flow stability and equipment load level.

[0069] The temporal features input within each candidate stage segment are statistically analyzed, and a comprehensive stage feature index within the candidate stage segment is calculated. The candidate stage segments are then merged or split to form multiple target stage segments with stable stage features, which serve as stage nodes in the perfusion stage topology graph. Specifically, the calculation of the comprehensive stage feature index within the candidate stage segment involves:

[0070] The flow time series within the candidate stage segment is subtracted step by step to obtain the flow change value of adjacent time steps. All change values ​​are averaged to obtain the average change amplitude. The consistency ratio of the change value signs is statistically analyzed to determine the continuity of the change direction. At the same time, the variance of the change value is calculated to characterize the change stability, forming the flow change pattern within the segment.

[0071] The maximum and minimum values ​​of the pressure time series within the candidate stage segment are calculated, and the difference between the two is used as the pressure fluctuation amplitude. At the same time, the length of the continuous time period in which the pressure change rate is lower than the preset threshold is counted on the time axis as the length of the pressure stability interval, and the main pressure change trend is counted based on the monotonicity of pressure over time.

[0072] The perfusion volume data within the candidate stage segment is subtracted step by step, and the average value of multiple time differences is calculated as the average growth rate of perfusion volume. The standard deviation of the equipment temperature or load and other operating status data within the segment is calculated, and the inverse index is taken as the equipment stability performance. The growth rate and stability performance are incorporated into the comprehensive index.

[0073] The phase division is repeatedly performed on multiple infusion batches, and the temporal adjacency relationship of each target phase segment is recorded. Based on the infusion deviation results of each infusion batch, the adjacency relationship between phase nodes is classified into three categories: normal phase transition, corrected phase transition, and abnormal phase transition. Corresponding normal phase transition edges, corrected phase transition edges, and abnormal phase transition edges are constructed respectively, forming an infusion phase topology graph containing a set of phase nodes and a set of phase transition edges with category labels. Specifically, the construction of the corresponding normal phase transition edges, corrected phase transition edges, and abnormal phase transition edges involves:

[0074] Match the stage nodes that appear sequentially in adjacent batches and whose irrigation deviation results are within the normal range, mark the relationship between the stage nodes as normal stage transitions, and establish normal stage transition edges between the stage nodes.

[0075] For stage nodes in adjacent batches that maintain the same order but have slightly higher than normal injection deviations, the relationship between the stage nodes is re-marked as a corrected stage transition, and corrected stage transition edges are established between the stage nodes.

[0076] For stage nodes in adjacent batches that jump, repeat, or occur simultaneously with a significant increase in deviation in their sequential order, mark the sequential relationship of the stage nodes as an abnormal stage transition, and establish an abnormal stage transition edge between the stage nodes.

[0077] Based on the statistical results of the stage features of the target stage segment corresponding to each stage node and the connection relationship of the stage node in the infusion stage topology graph, a unique stage encoding vector is generated for each stage node.

[0078] In this embodiment, constructing the liquid state machine model includes:

[0079] For each stage node in the topology diagram of the infusion stage, the corresponding stage feature statistics and stage encoding vector are obtained. Each stage node is associated with a stage liquid pool to form a mapping relationship from stage node to stage liquid pool. The liquid state machine model is determined to consist of a stage input encoding layer, a stage reserve calculation layer and a global coordination and convergence layer.

[0080] In the stage input coding layer, based on the alignment feature vector sequence and the stage coding vector corresponding to each stage node, an input coding substructure is configured for each stage liquid pool. The combination method of each feature component in the alignment feature vector and each component in the stage coding vector at the input end of the stage liquid pool and the weight allocation rule are determined. The alignment feature vector and the stage coding vector are converted into input signals that adapt to the dynamic evolution inside the stage liquid pool.

[0081] In the stage reserve calculation layer, based on the stage characteristic statistical information of each stage node, the internal state dimension, state update time scale, and internal connection sparsity are set for the corresponding stage liquid pool. This constructs the internal structure of the stage liquid pool for reserve calculation of the dynamic behavior of different infusion stages. All stage liquid pools are combined to form the stage reserve calculation substructure of the liquid state machine model, where:

[0082] The internal state dimension refers to the capacity of dynamic features that the liquid pool can express during the processing stage, so that different stages can be matched with different dynamic complexities.

[0083] The state update timescale refers to the rate at which the internal state of the liquid pool in the control stage is updated over time.

[0084] Internal connectivity sparsity refers to the density of connections between neurons within the control phase liquid pool.

[0085] In the global coordination and aggregation layer, a global coordination unit is configured for all stage liquid pools. This unit collects and merges the dynamic states generated by each stage liquid pool during activation according to preset rules, generating a global coordination state. A fixed mapping relationship is established between the global coordination state and the multi-task readout layer, where the preset rules are:

[0086] Priority rules based on stage transition edge categories:

[0087] During the normal phase, the transition edges should be read first and the continuous structure should be maintained.

[0088] The transition edges in the correction phase correspond to fusion after the normal sequence;

[0089] The transition edges in the abnormal phase correspond to those that are only added when the state changes abruptly.

[0090] Integration rules based on the activation order of the phased liquid pool:

[0091] Dynamic state merging is performed according to the path order in the topology graph to ensure the structural consistency of states across stages;

[0092] Fusion rules based on time continuity:

[0093] Align the dynamic state sequences across stages in chronological order to ensure that there are no jumps or omissions in the time dimension;

[0094] Based on the adjacent relationships of stage nodes and the types of stage transition edges in the topology graph of the infusion stage, the activation order and cross-stage migration path of the liquid pool in each stage are determined, and the running order of the stage input coding layer, the stage reserve computing layer and the global coordination and convergence layer are fixed as the activation sequence.

[0095] In this embodiment, the step of performing cross-stage migration to obtain a continuous dynamic state sequence includes:

[0096] The aligned feature vector and the stage coding vector are combined to form a joint input for stage-gated coding. Specifically, the combination of the aligned feature vector and the stage coding vector involves:

[0097] Based on the correspondence between the two in the time dimension, the feature components of the aligned feature vector at the current time step and the stage identifier components of the stage coding vector are concatenated in the feature dimension, and the concatenated feature sequence is normalized in a preset order.

[0098] Based on the joint input, stage gating weights are calculated for the current time step in the stage input coding layer. The stage coding vector is used to adjust the weight distribution of each feature component in the aligned feature vector, resulting in a stage-gated coding vector containing the stage-gated feature components. The calculation of the stage gating weights specifically involves:

[0099] Based on the stage identifier component of the stage coding vector in the joint input, determine the target stage to which the current time step belongs, and extract the stage feature statistics corresponding to the target stage;

[0100] Based on the weight indications of features including flow rate, pressure, volume and equipment stability in the stage feature statistics, the basic weights of each feature component of the aligned feature vector are paired with the feature indication values ​​of that stage to form the initial adjustment relationship of the feature weights.

[0101] According to the initial adjustment relationship, the basic weights of each feature component are proportionally adjusted so that the weights of key features in the target stage are increased and the weights of non-key features are suppressed, and finally the stage gating weight set used for the current time step is obtained.

[0102] The stage gating encoding vector is input into the stage liquid pool in the active state of the liquid state machine model. The internal state is updated in the stage reserve calculation layer according to the dynamic parameters of the stage liquid pool to generate the stage liquid pool dynamic state, which is then sent to the global coordination and convergence layer to participate in the update of the global coordination state.

[0103] When it is detected that the stage node to which the current time step belongs is different from the stage node to which the previous time step belongs, the dynamic state of the liquid pool of the previous stage is called in the global coordination and convergence layer, and cross-stage migration is performed on the liquid pool of the next stage that is about to be activated. The dynamic state of the liquid pool of the previous stage is combined with the initial state of the liquid pool of the next stage to generate the transition dynamic state during stage switching.

[0104] The dynamic states of the liquid pool at each stage and the transitional dynamic states are connected in series and recorded in chronological order to form a continuous dynamic state sequence covering the entire injection process.

[0105] In this embodiment, the output of future perfusion trend prediction results, perfusion deviation risk assessment results, and perfusion adjustment parameters includes:

[0106] Obtain a continuous dynamic state sequence, organize the continuous dynamic state sequence in chronological order, and perform completion processing on abnormal interruptions and missing states to form a normalized continuous dynamic state sequence.

[0107] Based on the normalized continuous dynamic state sequence, the infusion stage is divided according to the preset infusion stage boundary. The dynamic state is aggregated in each window or stage to extract multi-scale dynamic state features. The multi-scale dynamic state features are associated with the corresponding time position. The preset infusion stage boundary is the stage division boundary determined according to the time start position, time end position of the stage node in the infusion stage topology graph and the connection relationship between the stage node and the stage transition edge set.

[0108] Weighted integration is performed on multi-scale dynamic state features at different time scales, fusing multi-scale dynamic state features corresponding to the same time position into a set of joint state feature vectors. Sequence-level aggregation of the joint state feature vectors is then performed in the time dimension to generate a joint state representation. Specifically, the sequence-level aggregation of the joint state feature vectors in the time dimension involves:

[0109] According to the natural order of the joint state feature vectors on the time axis, the joint state feature vectors of all time steps are arranged in sequence to form a time series structure, maintaining a one-to-one correspondence between time position and feature content;

[0110] While maintaining the temporal order, the joint state feature vectors of adjacent time steps are checked for temporal continuity to eliminate time abrupt segments caused by stage switching, so that the sequence remains continuous in the time dimension.

[0111] The corrected joint state feature vectors are accumulated and integrated sequentially in the time dimension to form a joint state representation that fully covers the entire perfusion process;

[0112] The joint state representation is input to the multi-task readout layer, which includes a shared feature transformation sublayer, a task-specific output sublayer, and a task cooperative constraint sublayer, wherein:

[0113] The shared feature transformation sublayer consists of feature channel clusters categorized according to the infusion stage. Different channel clusters and hierarchical structures are selected based on the stage transition connection mode of the infusion stage topology. Grouped multi-layer feature transformation is performed on the joint state representation. Specifically, the grouped multi-layer feature transformation on the joint state representation is as follows:

[0114] Based on the perfusion stage category to which each feature component in the joint state representation belongs, the joint state representation is divided into several feature subgroups along the feature dimension, and each feature subgroup is matched with a feature channel cluster corresponding to its stage category.

[0115] According to the stage transition connection pattern between adjacent stages in the topology diagram of the infusion stage, the feature transformation path in the channel cluster that matches the current feature subgroup is selected, and a hierarchical multi-layer feature transformation is performed on the feature subgroup to generate the intermediate feature expression corresponding to the stage.

[0116] The intermediate feature expressions obtained by each feature subgroup through multi-layer feature transformation are grouped and merged according to the arrangement order in the topology map of the infusion stage, forming a shared feature transformation result with consistent structure, stage correspondence and cross-stage correlation.

[0117] The task-specific output sublayer includes three task output paths corresponding to future perfusion trends, perfusion deviation risks, and perfusion adjustment parameters, respectively. Each path selects a subset of associated features from the feature channel cluster and forms the task output structure through independently connected feature mapping units and sparse lateral coupling units.

[0118] The task coordination constraint sublayer is set after the three task output paths. It consists of a set of coordination nodes connected in a preset topological order and cross-task connection edges. The coordination nodes sequentially receive the intermediate representations of each task output path and perform cross-task state updates within a limited number of iterations. It provides the future perfusion trend prediction results, perfusion bias risk assessment results, and perfusion adjustment parameters after coordination constraints. The preset topological order is as follows:

[0119] Based on the sequential relationship of stage nodes on the three types of stage transfer paths formed by the normal stage transfer edge, the correction stage transfer edge, and the abnormal stage transfer edge in the topology diagram of the injection stage, and arranged according to the task priority of the injection deviation risk assessment results, the deviation risk path is placed at the beginning of the collaborative node link, the future injection trend prediction path is placed in the middle position, and the injection adjustment parameter path is placed at the end position, so as to ensure that the collaborative nodes execute the collaborative constraints in the order of risk priority, trend following, and adjustment convergence when updating the status across tasks.

[0120] In this embodiment, generating calibrated perfusion regulation parameters includes:

[0121] Based on the final deviation results of the infusion process, the alignment feature vector sequence of the corresponding infusion batch is divided into a key detection interval and a benchmark reference interval. The distribution characteristics of the alignment feature vectors in the two intervals are extracted respectively. Specifically, the division of the alignment feature vector sequence of the corresponding infusion batch into the key detection interval and the benchmark reference interval is as follows:

[0122] Based on the final deviation results of the injection process, the time period with the largest deviation in this batch is identified, and the start and end positions of the time period on the time axis are determined as the core interval of the key detection interval.

[0123] Centered on the core interval, a preset number of time steps are extended forward and backward along the time axis to cover the transitional behavior before and after the deviation occurs, forming a complete key detection interval;

[0124] All remaining time steps that do not belong to the key detection interval are included in the benchmark reference interval, so that the benchmark reference interval can cover the normal behavior segment of the perfusion process and provide stable reference distribution characteristics.

[0125] By comparing the distribution characteristics of the key detection interval with the benchmark reference interval, it is determined whether the input distribution has drifted. When drift is determined, the set of key stage nodes causing the drift is determined based on the attribution relationship of the alignment feature vectors at each stage node. Specifically, the determination of whether the input distribution has drifted is as follows:

[0126] The statistical distributions of the alignment feature vectors in the key detection interval and the benchmark reference interval in terms of flow rate change pattern, pressure response level, injection volume growth rate and equipment operating status are compared dimension by dimension.

[0127] During the comparison process, when the feature distribution of the key detection interval does not overlap with the statistical range of the corresponding dimension of the benchmark reference interval in one or more dimensions, or when there is a significant deviation in the statistical mean, magnitude of change, and trend direction, the deviation is regarded as a difference in the input distribution.

[0128] When the degree of difference in any feature dimension reaches the preset distribution deviation threshold, the alignment feature vector sequence of this batch is determined to have experienced distribution drift, and the stage node corresponding to the relevant time step is marked as a drift-related node.

[0129] For the set of key stage nodes, based on the connection method of each stage node in the injection stage topology graph and the category of the corresponding stage transfer edge, multiple topology perturbation candidate structures are constructed by adjusting the type, number and order of the stage transfer edges related to the key stage nodes. The topology perturbation candidate structures form several candidate injection stage topology graphs by adjusting the type, number and order of the stage transfer edges related to the key stage nodes.

[0130] The aligned feature vector sequence containing the key detection interval and the benchmark reference interval is replayed and calculated using the topology map of each candidate perfusion stage. Based on the output results of the multi-task readout layer under different candidate topologies, the topology with better performance is selected as the updated perfusion stage topology map. Incremental calibration is performed on the parameters of the multi-task readout layer to generate calibrated perfusion adjustment parameters.

[0131] Example 1:

[0132] To verify the feasibility of this invention in practice, it was applied to a resin composite material injection production line. The company had long used an injection strategy based on fixed thresholds for production control. The viscosity of the injection medium used in the production process fluctuated with changes in temperature and humidity. The equipment pumps had been in use for over four years, and slight wear caused a slower pressure response. Mechanical seal decay caused periodic fluctuations in flow rate. From March to June 2025, data from 42 injection batches were collected, including multi-source time-series characteristics such as flow rate, pressure, cumulative injection volume, pump speed, motor temperature, and ambient humidity, with a sampling frequency of 1 Hz. The traditional injection strategy resulted in 11.8% of deviation alarms, with a significant increase in deviations, particularly in the later stages of injection, leading to multiple quality problems such as artificially high pressure, insufficient replenishment, and incomplete injection.

[0133] After deploying the method of this invention in the production line, the multi-source time-series data is first aligned, unifying the timestamps of different sensors and filling in a small number of missing values ​​to form an aligned feature vector sequence. Subsequently, this invention automatically performs stage division based on this sequence, without requiring manual threshold setting. Through automatic analysis of the time series change trends, it identifies three stages: the initial stage of perfusion, steady-state perfusion, and the final stage of fluid replenishment. For some batches with significant temperature variations, the pre-steady-state stage is also automatically identified. The perfusion stage topology constructed by this invention not only records the temporal order between stages but also labels three types of edges: normal stage transitions, corrected stage transitions, and abnormal transitions, thus forming a topological structure that can express complex stage behaviors. Compared to traditional methods that can only fix two segments, the stage division model of this invention significantly improves the accuracy of stage identification and can more realistically reflect the stage differences caused by changes in operating conditions.

[0134] During the infusion process, the liquid state machine model of this invention achieves cross-stage dynamic behavior modeling based on the activation of the stage liquid pool. Taking batch number 20250603-14 as an example, in the steady-state infusion stage, the system detects slight fluctuations in flow rate between 6.4 and 6.9 L / min and a rise in motor temperature from 39°C to 43°C. The dynamic state within the liquid pool remains stable but shows a slow upward trend. When entering the final replenishment stage, the pressure fluctuation increases. The stage gating code of this invention immediately activates the corresponding stage liquid pool, and the internal state shows a significant transition, thus avoiding the stage abrupt change misjudgment problem common in traditional single-model methods. Based on the joint state representation generated by the multi-task readout layer, the mean square error of this invention in predicting the infusion trend in the steady-state stage is approximately 0.13, while that of the traditional LSTM is 0.22, indicating that it has a stronger ability to capture dynamic behavior within stages.

[0135] It is worth noting that this invention can handle the data distribution drift problem caused by long-term operation. Taking batch 20250528-09 as an example, when the ambient humidity rises sharply from 54% to 68%, the increased viscosity of the injection medium leads to a higher pressure response than historical batches. The distribution drift detection module of this invention accurately identifies the input distribution shift in the steady-state stage and automatically generates four sets of topology perturbation candidate structures for playback evaluation. The finally selected topology structure can better reflect the stage transition logic under the condition of increased humidity. Its multi-task readout layer prediction error is reduced from 0.19 under the original topology to 0.11. After this topology update and parameter incremental calibration, the injection deviation of this batch is successfully controlled within 3%, while the deviation rate of this batch is expected to exceed 9% under the traditional strategy.

[0136] During the subsequent month of continuous operation, the invention was applied to 18 injection batches with automatic parameter adjustment. The injection deviation alarm rate dropped from 11.8% before implementation to 4.3%, the deviation rate in the final liquid replenishment stage decreased by about 51%, and the molding defect rate decreased by 18.7%. This demonstrates that the invention is effective in solving the problem of discontinuous state across stages, adapting to changes in data distribution, and improving adjustment accuracy.

[0137] Table 1. Comparison of Multi-Source Time-Series Statistical and Predictive Performance of Typical Injection Batches

[0138] Batch number Infusion duration (min) Average flow rate (L / min) Mean pressure (MPa) Ambient humidity (%) Motor temperature (°C) Traditional prediction error Prediction error of this invention Improvement in deviation rate 20250503-07 56 6.8 4.2 55 41 0.24 0.15 37.5% 20250515-27 61 6.7 4.9 56 51 0.28 0.17 39.3% 20250528-09 63 6.5 5.1 68 52 0.32 0.18 43.8% 20250601-11 58 7.0 4.5 53 44 0.22 0.14 36.4% 20250603-14 66 6.9 4.3 57 46 0.20 0.13 35.0%

[0139] As shown in Table 1, the present invention significantly reduces the prediction error of the injection trend under different operating conditions. The prediction error of traditional methods is generally between 0.20 and 0.32, especially in batches with higher pressure or humidity. For example, the prediction error of batch 20250528-09 increased to 0.32 when the humidity rose to 68% and the motor temperature reached 52℃. In contrast, the present invention reduced the error to 0.18 under the same conditions, an improvement of 43.8%. This indicates that the present invention, through staged topology modeling and staged liquid pool activation mechanisms, can more accurately reflect the dynamic coupling relationship between pressure, flow rate, and temperature during the injection process.

[0140] In batches 20250515-27 and 20250503-07, where equipment temperature rise was significant, the prediction errors of the traditional method were 0.28 and 0.24, respectively, while the present invention reduced the errors to 0.17 and 0.15. This demonstrates that under operating conditions caused by response delays and flow fluctuations due to temperature changes, the present invention utilizes joint state representation to more effectively capture stage characteristics, maintaining prediction stability and robustness.

[0141] Even in batches 20250601-11 and 20250603-14, where operating conditions were relatively stable, this invention still reduced the prediction error from 0.22 and 0.20 for traditional methods to 0.14 and 0.13, respectively, maintaining an improvement of approximately 35%. This indicates that this invention not only performs exceptionally well under abnormal operating conditions but also maintains a consistently stable performance advantage under normal conditions. Overall, the results of each batch verify that this invention can effectively address the problem of insufficient modeling of phased dynamic differences in existing technologies, while still exhibiting good prediction accuracy and adaptive adjustment capabilities when facing equipment aging and changes in data distribution.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic adjustment of a data-driven perfusion process based on machine learning, characterized in that, The method comprises the following steps: Collecting multi-source time series data in the perfusion process, preprocessing the multi-source time series data, and generating an aligned feature vector sequence; Based on the aligned feature vector sequence, performing perfusion stage division, determining a plurality of stage nodes with stage characteristics, constructing a perfusion stage topology graph, and generating a corresponding stage encoding vector for each stage node; Generating a corresponding stage liquid pool for each stage node, constructing a liquid state machine model, and determining the activation sequence of each stage liquid pool in the liquid state machine model; Based on the aligned feature vector sequence and the stage encoding vector, performing stage gating coding, inputting the stage gating coding into the activated stage liquid pool, performing cross-stage migration, and obtaining a continuous dynamic state sequence; Integrating the continuous dynamic state sequence to generate a joint state representation, inputting the joint state representation into a multi-task readout layer, and outputting future perfusion trend prediction results, perfusion bias risk assessment results, and perfusion adjustment parameters; According to the final deviation result of the perfusion process, performing distribution drift detection, when the input distribution drifts, updating the topology selection of the perfusion stage topology graph, and performing incremental calibration on the multi-task readout layer to generate calibrated perfusion adjustment parameters; The construction of the liquid state machine model comprises: For each stage node in the perfusion stage topology graph, obtain the corresponding stage feature statistical information and stage encoding vector, associate each stage node with a stage liquid pool in a one-to-one manner to form a mapping relationship from stage node to stage liquid pool, and determine that the liquid state machine model is composed of a stage input encoding layer, a stage reserve calculation layer, and a global coordination convergence layer; In the stage input encoding layer, based on the aligned feature vector sequence and the stage encoding vector corresponding to each stage node, configure an input encoding substructure for each stage liquid pool, determine the combination mode and weight distribution rule of each feature component in the aligned feature vector and each component in the stage encoding vector at the input end of the stage liquid pool, and convert the aligned feature vector and the stage encoding vector into input signals suitable for the internal dynamic evolution of the stage liquid pool; In the stage reserve calculation layer, according to the stage feature statistical information of each stage node, set the internal state dimension, state update time scale and internal connection sparsity of the corresponding stage liquid pool respectively, construct the internal structure of the stage liquid pool for reserve calculation of the dynamic behavior of different perfusion stages, and combine all stage liquid pools to form the stage reserve calculation substructure of the liquid state machine model; In the global coordination convergence layer, configure a global coordination unit for all stage liquid pools, collect and fuse the dynamic states generated by each stage liquid pool during activation according to a preset rule to generate a global coordination state, and establish a fixed mapping relationship between the global coordination state and the multi-task readout layer; According to the front and rear adjacent relationship of the stage nodes in the perfusion stage topology graph and the category of the stage transition edge, the activation sequence and the cross-stage migration path of each stage liquid pool are determined, and the running order of the stage input encoding layer, the stage reserve calculation layer and the global coordination convergence layer is fixed as the activation sequence.

2. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The multi-source time series data includes flow signals, pressure signals, cumulative perfusion volume signals, pump drive current signals, pump rotation speed signals, device operating temperature signals, and ambient temperature and humidity signals, which represent dynamic changes in the perfusion process.

3. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The generating of the aligned feature vector sequence refers to constructing a non-uniform time index sequence according to the gradient change rate and the numerical fluctuation amplitude of the multi-source time series data, resampling and synchronizing the multi-source time series data at a uniform time step using the non-uniform time index sequence, arranging the multi-dimensional features corresponding to each time step in chronological order to form a time series vector sequence composed of multiple time step feature vectors.

4. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The constructing of the perfusion phase topology includes generating a corresponding phase encoding vector for each phase node, including: Preprocessing the aligned feature vector corresponding to each time step in the aligned feature vector sequence to obtain a time series feature input sequence according to preset smoothing, denoising, and scale unification rules; Based on the time series feature input sequence, the change trend between adjacent time steps is analyzed, and time intervals with similar change patterns and continuous occurrence are divided into the same candidate perfusion phase, and time intervals without similar change patterns are divided into different candidate perfusion phases, to obtain multiple candidate phase segments arranged along the time axis; Statistical analysis of the time series features within each candidate phase segment is performed to calculate the comprehensive phase feature indicators within the candidate phase segment, and the candidate phase segments are merged or split to form multiple target phase segments with stable phase features as phase nodes in the perfusion phase topology; The phase division is repeatedly performed on multiple perfusion batches, the adjacent relationship of each target phase segment in time is recorded, and the adjacent relationship between the phase nodes is divided into three categories: normal phase transition, corrected phase transition, and abnormal phase transition, to construct corresponding normal phase transition edges, corrected phase transition edges, and abnormal phase transition edges, respectively, to form a perfusion phase topology containing a phase node set and a phase transition edge set with category labels; According to the phase feature statistical results of the target phase segment corresponding to each phase node and the connection relationship of the phase nodes in the perfusion phase topology, a unique phase encoding vector is generated for each phase node.

5. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The performing of the cross-phase migration to obtain a continuous dynamic state sequence includes: Combining the aligned feature vector and the phase encoding vector to form a joint input for phase gating encoding; According to the joint input, the phase gating weight for the current time step is calculated in the phase input encoding layer, and the phase encoding vector is used to adjust the weight distribution of each feature component in the aligned feature vector to obtain a phase gating encoding vector containing the feature components after phase gating; The phase gating encoding vector is input into the phase fluid pool in the active state in the liquid state machine model, and the internal state is updated according to the dynamic parameters of the phase fluid pool in the phase reserve calculation layer to generate a phase fluid pool dynamic state, which is sent to the global coordination convergence layer to participate in the update of the global coordination state. When it is detected that the stage node to which the current time step belongs is different from the stage node to which the last time step belongs, the dynamic state of the last stage liquid pool is called in the global coordination convergence layer, a cross-stage migration is performed on the next stage liquid pool to be activated, the dynamic state of the last stage liquid pool is combined with the initial state of the next stage liquid pool, and a transition dynamic state at the stage switching time is generated; The stage liquid pool dynamic states and the transition dynamic states are sequentially concatenated and recorded in time sequence to form a continuous dynamic state sequence covering the entire perfusion process.

6. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The output future perfusion trend prediction result, the perfusion deviation risk assessment result and the perfusion adjustment parameter include: The continuous dynamic state sequence is obtained, the continuous dynamic state sequence is arranged in time sequence, and the abnormal interruption and missing state in the continuous dynamic state sequence are completed to form a normalized continuous dynamic state sequence; Based on the normalized continuous dynamic state sequence, the dynamic state is aggregated in each window or each stage according to the preset perfusion stage boundary, and multi-scale dynamic state features are extracted; the multi-scale dynamic state features are associated with the corresponding time positions; The multi-scale dynamic state features at different scales are integrated by weighting, the multi-scale dynamic state features corresponding to the same time position are fused into a set of joint state feature vectors, and the joint state feature vectors are aggregated at the sequence level in the time dimension to generate joint state representations; The joint state representations are input into a multi-task reading layer, the multi-task reading layer includes a shared feature transformation sub-layer, a task-specific output sub-layer and a task coordination constraint sub-layer, wherein: The shared feature transformation sub-layer is composed of feature channel clusters divided according to perfusion stage categories, different channel clusters and hierarchical structures are selected according to the stage transition connection mode of the perfusion stage topology graph, and grouped multi-layer feature transformation is performed on the joint state representations; The task-specific output sub-layer includes three task output paths corresponding to the future perfusion trend, the perfusion deviation risk and the perfusion adjustment parameter respectively, each path selects an associated feature subset from the feature channel clusters and forms a task output structure through independently concatenated feature mapping units and sparse lateral coupling units; The task coordination constraint sub-layer is arranged after the three task output paths and is composed of a group of coordination nodes connected in a preset topological order and cross-task connection edges, the coordination nodes receive intermediate representations of each task output path in turn and perform cross-task state updating within a limited iteration round to give the future perfusion trend prediction result, the perfusion deviation risk assessment result and the perfusion adjustment parameter subjected to coordination constraints.

7. The machine learning based data driven perfusion process auto-tuning method of claim 1, wherein, The generation of the calibrated perfusion adjustment parameter includes: Based on the final deviation result of the perfusion process, the alignment feature vector sequence corresponding to the perfusion batch is divided into a key detection interval and a reference interval, and the distribution features of the alignment feature vectors in the two intervals are extracted respectively; The distribution features of the key detection interval and the reference interval are compared to determine whether the input distribution has drifted, and when it is determined that the distribution has drifted, the key stage node set causing the drift is determined according to the attribution relationship of the alignment feature vectors at the stage nodes; For the key stage node set, according to the connection mode of each stage node in the perfusion stage topology graph and the category of the corresponding stage transition edge, a plurality of topology disturbance candidate structures are constructed by adjusting the type, number and order of the key stage node related stage transition edge, and the topology disturbance candidate structure forms a plurality of candidate perfusion stage topology graphs by adjusting the type, number and order of the key stage node related stage transition edge; The alignment feature vector sequence containing the key detection interval and the reference interval is played back and calculated by using each candidate perfusion stage topology graph, and the topology structure with better performance is selected as the updated perfusion stage topology graph according to the output results of the multi-task readout layer under different candidate topology structures, and the incremental calibration of the multi-task readout layer parameters is performed to generate the calibrated perfusion adjustment parameters.

Images