A man-machine collaborative production efficiency optimization system based on big data analysis

Abstract

This invention discloses a human-machine collaborative production efficiency optimization system based on big data analysis, belonging to the field of industrial big data production optimization technology. It includes a data acquisition module that collects real-time data on human operation, equipment status, and material flow to form a raw production time-series big data stream; a preprocessing module that performs multimodal fusion cleaning and extracts three types of feature sequences, marking them with timestamps and process identifiers; a state modeling module that constructs a dynamic Bayesian network model of the human-machine collaborative production state; a bottleneck tracing module that traces the historical evolution path of bottlenecks and identifies key events through a state inversion algorithm; a causal analysis module that constructs a causal inference graph to quantify the contribution of events; and a strategy closed-loop module that generates optimization strategies and collects feedback data to form a closed-loop optimization. This invention achieves dynamic modeling of the human-machine collaborative production state and quantification of bottleneck reverse tracing, improving the accuracy of efficiency analysis and forming a closed-loop optimization mechanism.

Description

Technical Field

[0001] This invention belongs to the field of industrial big data production optimization technology, specifically a human-machine collaborative production efficiency optimization system based on big data analysis. Background Technology

[0002] Current performance analysis technologies in human-machine collaborative production scenarios mostly employ single-dimensional data collection methods, independently statistically analyzing data related to human operation, equipment operation, or material flow. They rely on static mathematical models or conventional machine learning models for production status assessment, primarily focusing on positive surface analysis for bottleneck identification. Anomalies in production processes are determined through simple data comparisons, without incorporating temporal features and process identifiers for deep multimodal data fusion processing. These technologies can only achieve superficial data cleaning and feature extraction, failing to construct state models that accurately reflect dynamic production changes, and lacking algorithms for state inversion and causal inference to support performance analysis.

[0003] Existing technologies cannot perform inference calculations on the characteristic sequences of human-machine interaction behavior marked with timestamps and process identifiers, equipment operation characteristic sequences, and material consumption characteristic sequences to construct dynamic Bayesian network models, making it difficult to adapt to the time-series dynamic evolution characteristics of human-machine collaborative production. Furthermore, existing solutions lack the technical path to input the production state model into the state inversion algorithm, making it impossible to trace the historical state evolution path of production efficiency bottlenecks, accurately identify key human-cause operations, equipment status, and material flow events, and construct causal inference graphs to quantify the contribution of various events to production efficiency indicators. Optimization strategy generation lacks precise targeted support, making it difficult to form a closed-loop efficiency optimization process. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a human-machine collaborative production efficiency optimization system based on big data analysis, comprising: The data acquisition module collects human operation data, equipment status data, and material flow data in real time during the human-machine collaborative production process, forming a raw production time-series big data stream; The preprocessing module performs multimodal data fusion and cleaning on the original production time-series big data stream, extracts human-machine interaction behavior feature sequences, equipment operation feature sequences and material consumption feature sequences, and simultaneously marks their timestamps and process identifiers; The state modeling module, based on the human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence, uses inference calculation to construct a dynamic Bayesian network model of the human-machine collaborative production state. The bottleneck tracing module inputs the dynamic Bayesian network model into the state inversion algorithm, and uses the state inversion algorithm to trace the historical state evolution path that led to the current production efficiency bottleneck, and identifies key human factor operation events, equipment status events and material flow events. The causal analysis module constructs a causal inference diagram based on the key human operation events, equipment status events, and material flow events, quantifying the contribution of each type of event to production efficiency indicators. The strategy closed-loop module generates a comprehensive optimization strategy based on the causal inference graph and the contribution degree, deploys the comprehensive optimization strategy to the production site, and collects feedback data after the strategy is executed in real time to form a new original production time-series big data stream, thus constituting a closed-loop optimization process.

[0005] Furthermore, the original production time-series big data stream is subjected to multimodal data fusion and cleaning to extract human-computer interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences, including: The video stream, sensor signals and operation logs in the human-computer interaction data are time-aligned, and standardized operation action units are segmented from them to form the human-computer interaction behavior feature sequence. The vibration, temperature, and current time-series signals in the equipment status data are subjected to frequency domain transformation and feature extraction to calculate the equipment health index and energy efficiency index, thereby forming the equipment operation feature sequence. Perform correlation analysis on the batch, quantity, and location change records in the material flow data, calculate the material completeness rate, turnover rate, and work-in-process inventory level, and form the material consumption characteristic sequence. The human-computer interaction behavior feature sequence, the equipment operation feature sequence, and the material consumption feature sequence are subjected to missing value imputation, outlier detection, and smoothing filtering to generate a cleaned standardized feature sequence. The cleaned standardized feature sequence is bound with a unified timestamp and process identifier and stored in a time-series feature database.

[0006] Furthermore, based on the aforementioned human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence, a dynamic Bayesian network model of the human-machine collaborative production state is constructed using inference computation, including: The nodes of the dynamic Bayesian network model represent production state variables of different granularities. Production efficiency is determined as the top-level target node, and the human-machine interaction behavior characteristics, the equipment operation characteristics, and the material consumption characteristics are used as the bottom-level observation nodes. Define intermediate layer status nodes, which include job cycle status, resource matching status, and collaboration smoothness status; Based on historical data learning or expert knowledge, initialize the conditional probability distribution between the top-level target node, the intermediate-level state node, and the bottom-level observation node; Using the temporal information of the human-computer interaction behavior feature sequence, the equipment operation feature sequence, and the material consumption feature sequence, the conditional probability distribution is dynamically updated through inference calculation, and the state transition probability of nodes between different time steps is learned; When the inference accuracy of the dynamic Bayesian network model reaches a preset threshold, the model construction is complete. The model responds to new feature sequence inputs and outputs the current production efficiency status and its probability distribution.

[0007] Furthermore, the dynamic Bayesian network model is input into the state inversion algorithm, which traces back the historical state evolution path leading to the current production efficiency bottleneck, including: In the dynamic Bayesian network model, the currently observed low productivity state is set as an evidence node, and its state is fixed. Run the state inversion algorithm, starting from the top-level target node, propagate the probability backward along the network structure, and calculate the posterior probability that each intermediate layer state node is in an abnormal state at each historical time step given evidence of low production efficiency. Based on the magnitude of the posterior probability, intermediate layer state nodes whose posterior probability exceeds the confidence threshold are selected and marked as abnormal historical states. By connecting the marked abnormal historical state nodes in chronological order, one or more state evolution paths from historical abnormal states to the current low productivity state are formed.

[0008] Furthermore, based on the aforementioned key human-caused operational events, equipment status events, and material flow events, a causal inference diagram is constructed to quantify the contribution of each type of event to production efficiency indicators, including: From the state evolution path, extract the bottom observation node that is directly associated with the abnormal change of the intermediate layer state node. The bottom observation node corresponds to the specific human operation event, equipment status event or material flow event. Using production efficiency indicators as outcome variables and the extracted human operation events, equipment status events, and material flow events as causal variables, a causal inference diagram is constructed. Using counterfactual reasoning, the difference between the expected value and the current actual value of the production efficiency indicator is calculated when one of the causal variables is missing. The difference values ​​are normalized and used as the contribution of the causal variable to the production efficiency index. The greater the contribution, the higher the impact of the event corresponding to the causal variable on the current efficiency bottleneck.

[0009] Furthermore, based on the causal inference graph and the contribution degree, a comprehensive optimization strategy is generated, including: The comprehensive optimization strategy includes operational suggestions, equipment adjustment suggestions, and material scheduling suggestions; Events whose contribution exceeds a set threshold are categorized into sets for human factors operation improvement, equipment status adjustment, and material flow optimization. For each operation event in the human factors operation improvement set, standardized operation specifications are matched from a preset knowledge base, and a description of the differences compared to the current operation is generated to form specific operation suggestions; For each equipment event in the equipment status adjustment set, combined with the equipment maintenance manual and historical best operating parameters, a parameter adjustment range or preventive maintenance instruction is generated to form a specific equipment adjustment suggestion; For each material event in the material flow optimization set, an adjustment plan for material delivery time, quantity, or route is generated by combining production planning and inventory strategy, forming a specific material scheduling suggestion; The operation suggestions, equipment adjustment suggestions, and material scheduling suggestions are integrated and sorted according to the process sequence and priority to generate the comprehensive optimization strategy document.

[0010] Furthermore, the comprehensive optimization strategy is deployed to the production site, and feedback data after the strategy execution is collected in real time, including: The comprehensive optimization strategy document is converted into executable instructions and distributed to the corresponding workstations, equipment controllers, and warehouse management systems through the manufacturing execution system. During execution, new human factors operation data, equipment status data, and material flow data are continuously collected as feedback data for strategy execution.

[0011] Furthermore, it also includes a model update module for: In the closed-loop optimization process, new production data and corresponding actual production efficiency results are periodically used as training samples. The conditional probability distribution and state transition probability in the dynamic Bayesian network model are updated using the training samples through an incremental learning algorithm. When there are significant changes in the configuration, process, or product type of the production line, the model structure learning process is triggered to redetermine the network nodes and structure, and the dynamic Bayesian network model is reconstructed using data from the new scenario.

[0012] Furthermore, it also includes a simulation verification module for: Before implementing the comprehensive optimization strategy in the actual production site, the strategy is loaded into the corresponding production system digital twin model; Drive the digital twin model to run, simulate the running status of the entire production cycle after the strategy is executed, and predict changes in key performance indicators; The simulation prediction results are compared with the expected target of the strategy. If the deviation exceeds the allowable range, the parameters in the strategy are automatically adjusted and the simulation is repeated until the prediction results meet the requirements. The final strategy version, validated through simulation, will be deployed to the actual production site.

[0013] Furthermore, it also includes a cross-order coordination module for: Identify new constraints or changes in demand for downstream processes caused by the implementation of optimization strategies in upstream processes; Using the new constraints or changes in requirements as input, the inference calculation and state inversion algorithm are rerun to assess the impact on the production efficiency of downstream processes. Based on the evaluation results, compensatory or adaptive optimization strategies are generated for downstream processes; Coordinate the execution pace and resource allocation of optimization strategies between upstream and downstream processes to ensure overall optimal production efficiency, rather than local optimality of a single process.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Based on the human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence with timestamps and process identifiers, a dynamic Bayesian network model of human-machine collaborative production status is constructed through inference calculation. This model can fully depict the dynamic correlation between the multi-dimensional features of people, machines, and materials and the production status, conform to the continuous change characteristics of production time series data, restore the real-time evolution logic of human-machine collaborative production status, adapt to the dynamic change law of multi-factor interaction in the production process, get rid of the limitation of static models that cannot reflect the time series transmission relationship of production status, and achieve accurate dynamic representation of production status.

[0015] By inputting a dynamic Bayesian network model into a state inversion algorithm, the algorithm traces the historical evolution path of production efficiency bottlenecks in reverse, identifies key human-related operational events, equipment status events, and material flow events, constructs a causal inference graph based on these key events, and quantifies the contribution of each type of event to production efficiency indicators. It can deduce the transmission link of state evolution from the bottleneck result, locate the core event source that caused the efficiency bottleneck, clarify the impact weight of different types of events on production efficiency, and form a quantitative representation of the event impact. This avoids the problem that forward analysis only stays at the bottleneck surface, allowing optimization strategies to directly correspond to the core causal events. The feedback data after the strategy is executed can be reintegrated into the data stream, forming a continuously iterative closed-loop optimization logic. Attached Figure Description

[0016] Figure 1This is a timing diagram of a human-machine collaborative production efficiency optimization system based on big data analysis as described in this invention. Figure 2 A flowchart for state inversion and bottleneck path tracing; Figure 3 To generate a multi-dimensional probability change graph of abnormal states; Figure 4 To optimize the production efficiency improvement trend curve in a closed-loop manner; Figure 5 This is a performance evaluation chart of the core module of the human-machine collaborative system, showing both response time and prediction accuracy. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 This invention provides a human-machine collaborative production efficiency optimization system based on big data analysis, the system comprising: The data acquisition module operates continuously on the production floor. It collects human-caused operational data through visual sensors, wearable devices, and control panel log interfaces deployed at workstations; equipment status data through vibration, temperature, and current sensors installed on key equipment components and equipment controllers; and material flow data through RFID readers, warehouse management systems, and manufacturing execution systems. This data collectively constitutes the raw production time-series big data stream. The preprocessing module receives this big data stream and performs multimodal data fusion and cleaning operations, extracting human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences. Each feature set is then precisely timestamped and assigned a process identifier. The state modeling module uses inference computation to construct a dynamic Bayesian network model characterizing the human-machine collaborative production state based on the above three feature sequences. When a bottleneck in production efficiency is detected, the bottleneck tracing module inputs the current inefficient state into this dynamic Bayesian network model and drives the state inversion algorithm. The algorithm traces the historical evolution path leading to the current state, thereby identifying key human-caused operational events, equipment status events, and material flow events along the path. The causal analysis module uses these key events as input to construct a causal inference graph and quantifies the specific contribution of each type of event to production efficiency indicators. Based on the causal inference graph and the quantified contributions, the strategy closed-loop module generates a comprehensive optimization strategy including targeted measures. This strategy is deployed to the production site for execution, and the system immediately collects a new round of feedback data generated after the strategy's execution. This data enters the system as a new raw production time-series big data stream, thus initiating the next optimization cycle and forming a continuous closed-loop optimization process.

[0019] In one embodiment of the present invention, the specific process of performing multimodal data fusion and cleaning on the original production time-series big data stream to extract human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences includes multiple collaborative steps. First, time alignment processing is performed on the video stream, sensor signals from wearable devices, and operation log records contained in the human operation data. Based on predefined action templates, standardized operation action units are segmented from the aligned data, and these action units are arranged in chronological order to form the human-machine interaction behavior feature sequence. Second, frequency domain transformation and feature extraction are performed on the time-series signals such as vibration, temperature, and current in the equipment status data to calculate indicators such as equipment health and energy efficiency, forming the equipment operation feature sequence. Third, correlation analysis is performed on the material batch, quantity, and location change records in the material flow data to calculate indicators such as material availability rate, turnover rate, and work-in-process inventory level, forming the material consumption feature sequence. Finally, data quality improvement processing is performed on the initially extracted human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence. This processing includes missing value imputation, outlier detection, and smoothing filtering to generate a cleaned standardized feature sequence. The cleaned standardized feature sequences are bound to a unified high-precision timestamp and the corresponding process identifier, and stored in a dedicated time-series feature database for subsequent module calls.

[0020] In practical implementation, a preprocessing module of a human-machine collaborative production efficiency optimization system based on big data analytics receives raw production time-series big data streams from the data acquisition module. These raw production time-series big data streams include human operation data, equipment status data, and material flow data. The preprocessing module performs multimodal data fusion and cleaning operations on the raw production time-series big data streams to extract human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences. In specific implementation, the processing of human operation data involves video streams, sensor signals from wearable devices, and operation panel logs. The preprocessing module performs time alignment on the video streams, sensor signals, and operation logs. Time alignment is achieved based on a unified network time protocol timestamp, ensuring that events from different data sources are synchronized at the millisecond level. In practice, standardized operational action units are segmented from time-aligned human-computer interaction data. The segmentation process relies on a predefined action template library, which defines the visual and signal characteristics of basic operational units such as "grabbing a component," "aligning a mounting hole," and "tightening a bolt." The preprocessing module uses a pattern matching algorithm to identify the start and end points of these units from a continuous signal stream and arranges them chronologically to form a sequence of human-computer interaction behavior features. In some embodiments, each unit in the sequence records the action type, duration, peak operational force, and motion trajectory smoothness index.

[0021] In practical implementation, the processing of equipment status data focuses on vibration, temperature, and current time-series signals. The preprocessing module performs frequency domain transformation on each type of time-series signal using the Fast Fourier Transform (FFT) algorithm. Features are extracted from the transformed spectrum, including the energy of the vibration signal within a specific frequency band, the rise slope of the temperature signal, and the effective value and harmonic distortion rate of the current signal. Based on the extracted features, the preprocessing module calculates equipment health indicators and energy efficiency indicators. The equipment health indicator is obtained by comparing the current feature value with the baseline feature value in the equipment's new state. The energy efficiency indicator is obtained by calculating the ratio of effective output power to input power. These indicators, arranged in chronological order, constitute the equipment operating feature sequence. In some embodiments, the equipment operating feature sequence also includes equipment operating status labels, such as "normal," "warning," and "fault," which are automatically labeled by a rule-based classifier based on feature values.

[0022] In practical implementation, the processing of material flow data involves batch, quantity, and location change records. The preprocessing module performs correlation analysis on the material flow data. This correlation analysis uses the unique material identifier to link records scattered across the warehouse management system and the manufacturing execution system, tracing the complete path of a single material batch from warehousing, outbound, online to completion. Based on the correlated data flow, the preprocessing module calculates the material availability rate, turnover rate, and work-in-process inventory level. The material availability rate is the proportion of all materials required for a certain process at the planned start time. The material turnover rate is the average number of times materials are consumed or removed per unit time. The work-in-process inventory level is the total number of all unfinished products on the production line. These calculated indicators constitute a material consumption characteristic sequence in chronological order.

[0023] In practical implementation, the initially extracted human-computer interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences undergo data quality improvement processing, including missing value imputation, outlier detection, and smoothing filtering. Missing value imputation employs linear interpolation or prediction imputation based on sequence patterns. Outlier detection uses statistical methods to identify and mark data points deviating from the normal range in the sequence. Smoothing filtering uses a moving average filter to eliminate random noise in the data. Optionally, outlier detection can employ a Mahalanobis distance-based method for a given set of data. The Mahalanobis distance of sample points with each feature The calculation formula is: in: Represents sample points Mahalanobis distance, It is 3D feature vectors It is the characteristic sequence calculated under normal operating conditions. dimensional mean vector, It is a characteristic sequence under normal operating conditions. The covariance matrix is ​​dimensional. Represents the transpose of a vector. This represents the inverse of a matrix. When If the value exceeds a preset threshold, the sample point is determined to be an outlier. The preprocessing process generates a cleaned, standardized feature sequence. This standardization process maps feature values ​​of different dimensions to a unified numerical range.

[0024] In practical implementation, the preprocessing module binds the cleaned, standardized feature sequences with unified timestamps and process identifiers. The unified timestamps are provided by high-precision clock timestamps recorded during data acquisition, and the process identifiers come from the process codes defined by the manufacturing execution system. The bound data is stored in a time-series feature database, which adopts a time-series database architecture and is indexed using timestamps and process identifiers as a joint primary key, supporting efficient querying of feature sequences for specific processes within a specific time window. In essence, the time-series feature database provides a structured, high-quality data source for the dynamic modeling of human-machine collaborative production states.

[0025] In one embodiment of the present invention, a dynamic Bayesian network model of human-machine collaborative production state is constructed using inference computation based on human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences. This dynamic Bayesian network model consists of nodes representing production state variables of different granularities, with directed edges connecting the nodes to represent dependencies. Production efficiency is determined as the top-level target node, and human-machine interaction behavior features, equipment operation features, and material consumption features are used as bottom-level observation nodes. Intermediate-layer state nodes are defined between the top-level target node and the bottom-level observation nodes, typically including work cycle state, resource matching state, and collaborative smoothness state. Based on historical data learning or importing domain expert knowledge, the conditional probability distributions between the top-level target node, intermediate-layer state nodes, and bottom-level observation nodes are initialized. Utilizing the temporal information carried by the continuously input human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences, these conditional probability distributions are dynamically updated through inference computation, and the state transition probabilities of nodes between different time steps are learned. When the inference accuracy of the dynamic Bayesian network model, that is, the prediction accuracy of the production efficiency status, reaches the preset threshold, the model is considered to be completed. At this time, the model can respond to new feature sequence inputs and output the production efficiency status and its probability distribution at the current moment.

[0026] In practical implementation, a state modeling module of a human-machine collaborative production efficiency optimization system based on big data analysis constructs a dynamic Bayesian network model of the human-machine collaborative production state using inference computation based on the human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence output by the preprocessing module. The dynamic Bayesian network model consists of nodes representing production state variables of different granularities, with directed edges between nodes representing probabilistic dependencies between variables. In practical implementation, production efficiency is determined as the top-level target node, which is a comprehensive variable whose state can be quantified as discrete levels such as "high efficiency," "average," and "bottleneck," or as a continuous numerical range. In practical implementation, human-machine interaction behavior features, equipment operation features, and material consumption features are used as bottom-level observation nodes. These bottom-level observation nodes directly receive input from the standardized feature sequence from the preprocessing module, and the state of each bottom-level observation node corresponds to the observed value or quantized range of the feature sequence at a specific time point.

[0027] In practical implementation, intermediate-layer state nodes are defined, located between the top-level target node and the bottom-level observation nodes. These intermediate-layer state nodes include work cycle state, resource matching state, and collaborative smoothness state. Work cycle state represents the degree of matching between the actual time and standard time for a production unit to complete a work cycle. Resource matching state represents the degree of alignment between the configuration and demand of personnel, equipment, and materials between processes. Collaborative smoothness state represents the frequency of waiting, interruptions, and conflicts during human-equipment interaction. In some embodiments, intermediate-layer state nodes may further include finer-grained hidden state variables such as equipment health trends and personnel fatigue states. The structure of the dynamic Bayesian network model, i.e., the connections between nodes, is initialized based on prior knowledge of the production process. For example, the work cycle state node is simultaneously influenced by multiple bottom-level observation nodes, such as personnel operation characteristics, equipment operation characteristics, and material availability characteristics.

[0028] In practical implementation, based on historical data learning or expert knowledge, the conditional probability distribution among the top-level target node, intermediate-level state nodes, and bottom-level observation nodes is initialized. The conditional probability distribution is defined in the form of a conditional probability table or a parameterized function. For example, it defines the probability that a production efficiency node is in a "bottleneck" state given a work cycle state of "delay" and a resource matching state of "mismatch." Optionally, the initial conditional probability distribution can be set based on the experience and judgment of domain experts. In practical implementation, the conditional probability distribution is dynamically updated through inference calculation using the temporal information of continuously input human-computer interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences. The inference calculation employs algorithms such as expectation-maximization or variational inference to learn more accurate conditional probability parameters from continuous time-series data. In practical implementation, the system simultaneously learns the state transition probabilities of nodes between different time steps. The state transition probability describes the probability of transitioning from the "work cycle state" of the previous moment to the "work cycle state" of the current moment. The learning process relies on the analysis of historical state sequences. In some embodiments, the state transition probability can be estimated by calculating the co-occurrence frequency and transition frequency between states. The state modeling module iteratively performs inference calculations and parameter learning, enabling the dynamic Bayesian network model to continuously adapt to actual production data.

[0029] In practical implementation, the model is considered complete when the inference accuracy of the dynamic Bayesian network model reaches a preset threshold. Inference accuracy is evaluated using a reserved test dataset, calculating the model's prediction accuracy for production efficiency status. When the prediction accuracy is not lower than a preset threshold, such as 95%, the model is considered complete. It can be understood that a completed dynamic Bayesian network model can respond to new feature sequence inputs. After receiving new time-step human-computer interaction characteristics, equipment operation characteristics, and material consumption characteristics, the model performs forward inference calculations. This process involves calculating the posterior probability distribution of the top-level target node and each intermediate-level state node, given evidence from all bottom-level observation nodes. For a state node containing N possible states... At a given time All observed sequences Under the given conditions, the posterior probability distribution of its state It can be obtained through recursive updates, and its core update formula involves the likelihood of the belief state at the previous time step and the current observation: in: Indicates time State The belief, i.e., the posterior probability distribution , It is a normalization constant to ensure that the sum of the probabilities of all states is one. It is an observation model, representing the state The following observations The probability, It is a state transition model. It is time The state of belief. Summation symbol. This indicates the state node. Summing all possible states.

[0030] In one embodiment of the present invention, the implementation of state inversion and causal inference includes two progressive stages. In the first stage, the constructed dynamic Bayesian network model is input into the state inversion algorithm. The state inversion algorithm then traces back the historical state evolution path leading to the current production efficiency bottleneck. (See [reference]). Figure 2 In the dynamic Bayesian network model, the currently observed low productivity state is set as an evidence node and its state is fixed. The state inversion algorithm is run. The algorithm starts from the top target node and propagates the probability backward along the network structure. It calculates the posterior probability that each intermediate layer state node in each historical time step is in an abnormal state under the given evidence of low productivity. Based on the calculated posterior probability, intermediate layer state nodes with posterior probabilities exceeding a preset confidence threshold are selected and marked as abnormal historical states. The marked abnormal historical state nodes are connected in chronological order to form one or more state evolution paths from historical abnormal states to the current low productivity state. In the second stage, a causal inference graph is constructed based on key events and their contributions are quantified. From the identified state evolution paths, the underlying observation nodes directly related to the abnormal changes in intermediate state nodes are extracted. These underlying observation nodes correspond to specific human operation events, equipment status events, or material flow events. Using production efficiency indicators as outcome variables and the extracted events as causal variables, a causal inference graph is constructed. Using counterfactual reasoning, the difference between the expected value and the current actual value of the production efficiency indicator when a certain causal variable is missing is calculated. The obtained difference value is normalized, and the result is taken as the contribution of the causal variable to the production efficiency indicator. The larger the contribution value, the higher the degree of influence of the event corresponding to the causal variable on the current efficiency bottleneck.

[0031] In its implementation, a bottleneck tracing module and a causal analysis module of a human-machine collaborative production efficiency optimization system based on big data analysis work together. The implementation process includes two logical stages: state inversion and causal inference. In the first stage, the bottleneck tracing module inputs the dynamic Bayesian network model constructed by the state modeling module into the state inversion algorithm. The algorithm traces the historical state evolution path leading to the current production efficiency bottleneck. The dynamic Bayesian network model identifies a production efficiency node as being in a "low" or "bottleneck" state in the current observation. The bottleneck tracing module sets this low production efficiency state as an evidence node and fixes its state value as the observation result. The state inversion algorithm then runs, starting from the top-level target node (the production efficiency node) and propagating probabilities backward along the network structure of the dynamic Bayesian network model. In practice, the state inversion algorithm calculates the posterior probability that each intermediate-level state node in each historical time step is in an abnormal state, given evidence of low production efficiency. This calculation involves backward probability inference within the dynamic Bayesian network, considering the state transitions and observation sequences from historical moments to the current moment. In some embodiments, the state inversion algorithm can employ algorithms such as particle filtering smoothing or exact inference to calculate the posterior probability. Based on the calculated posterior probability, the bottleneck tracing module filters out intermediate-level state nodes whose posterior probabilities exceed a preset confidence threshold. The confidence threshold is a configurable parameter, for example, 0.7. The bottleneck tracing module marks intermediate-level state nodes with posterior probabilities exceeding the confidence threshold as abnormal historical states. It then connects these marked abnormal historical state nodes in chronological order to form one or more state evolution paths from historical abnormal states to the current low-productivity state. These state evolution paths visually demonstrate the key time points and stages of production state degradation.

[0032] In the second stage, the causal analysis module constructs a causal inference graph and quantifies the contribution of key events identified in the state evolution path. The module extracts the underlying observation nodes directly associated with the abnormal changes in intermediate state nodes from the identified state evolution paths. In specific implementations, the extraction process uses the directed edges connecting intermediate state nodes and underlying observation nodes in the dynamic Bayesian network model and their conditional probability distributions to identify which probability changes best explain the abnormal intermediate state at the underlying observation nodes. These underlying observation nodes correspond to specific human-related operational events, equipment status events, or material flow events. The causal analysis module uses production efficiency indicators as the outcome variable and the extracted human-related operational events, equipment status events, and material flow events as causal variables to construct a causal inference graph. This causal inference graph is a directed acyclic graph, where nodes represent variables and directed edges represent potential causal relationships. In some embodiments, the structure of the causal inference graph can partially inherit the connection relationships of related nodes in the dynamic Bayesian network model and be simplified to focus on key causal chains. The causal analysis module uses counterfactual reasoning to quantify contribution. Counterfactual reasoning calculates the difference between the expected value and the current actual value of the production efficiency indicator when a certain causal variable is missing.

[0033] In practical implementation, for a certain causal variable Its impact on production efficiency indicators Contribution This can be quantified by calculating the counterfactual expectation difference, expressed by the formula: in: Describing the cause variable For outcome variables Contribution Represents the expectation operator. It is a production efficiency indicator. Indicates setting the cause variable For its actual observed value Intervention operations, Indicates setting the cause variable Its counterfactual value Intervention operations, counterfactual values The baseline value of the causal variable is usually taken when the event is normal or when no event occurs. That is, the expected production efficiency as currently observed. It is a hypothetical cause variable. Expected production efficiency under normal circumstances (based on baseline values). Expected value. The calculation is based on causal inference graphs and observational data, and is estimated by adjusting the confusion factor.

[0034] See Figure 3In the analysis of the probability changes of multi-dimensional abnormal states in production, the dynamic Bayesian network state inversion algorithm was used to calculate the posterior probabilities of abnormal states for three types of underlying observed variables—human factors operation, equipment status, and material flow—at each process time step, and compared them with a preset anomaly confidence threshold of 0.7. From the temporal evolution trend, the probabilities of all three types of anomalies peaked in the interval between process 5 and process 7. Specifically, the probability of human factors operation anomalies reached approximately 0.75 in process 6, the probability of equipment status anomalies reached approximately 0.70 in process 7, and the probability of material flow anomalies reached approximately 0.65 in process 7. The posterior probabilities of human factors operation and equipment status anomalies both exceeded the confidence threshold and were marked as key historical anomaly states. The curves intuitively present the temporal evolution path of the production efficiency bottleneck: from process 1 to process 4, the probabilities of the three types of anomalies gradually increased, forming a core anomaly outbreak window in process 5-7. The state degradation in this stage directly drove the subsequent low production efficiency; from process 8 onwards, the anomaly probabilities continuously declined, gradually approaching the baseline level by process 12. By cross-analyzing the peak intervals of the three types of anomalies with confidence thresholds, process 6-7 can be identified as the critical time window for bottleneck tracing. Among them, abnormal human operation and abnormal equipment status are the core contributing factors to the decline in efficiency, providing a quantitative basis for subsequent causal inference and optimization strategy generation.

[0035] In one embodiment of the present invention, a comprehensive optimization strategy is generated based on a causal inference graph and contribution level, and deployed to the production site. The comprehensive optimization strategy includes operational suggestions, equipment adjustment suggestions, and material scheduling suggestions. First, events with contribution levels exceeding a set threshold are categorized into a human factors operation improvement set, an equipment status adjustment set, and a material flow optimization set. For each operational event in the human factors operation improvement set, standardized operating procedures are matched from a preset knowledge base, and a description of the differences compared to the current operation is generated, forming a specific operational suggestion. For each equipment event in the equipment status adjustment set, parameter adjustment ranges or preventative maintenance instructions are generated based on the equipment maintenance manual and historical best operating parameters, forming a specific equipment adjustment suggestion. For each material event in the material flow optimization set, adjustment schemes for material delivery time, quantity, or path are generated based on the production plan and inventory strategy, forming a specific material scheduling suggestion. The generated operational suggestions, equipment adjustment suggestions, and material scheduling suggestions are integrated and sorted according to the process sequence and preset priority, ultimately generating a comprehensive optimization strategy document. Deploy comprehensive optimization strategies to the production site and collect feedback data. Transform the comprehensive optimization strategy document into executable instructions that can be recognized by the Manufacturing Execution System (MES), Equipment Controllers (ECS), and Warehouse Management System (WMS). Issue these instructions to the corresponding workstations, ECS, and WMS through the MES. During strategy execution, the data acquisition module continuously collects new human factor operation data, equipment status data, and material flow data. These data are then fed into the system as feedback data for strategy execution.

[0036] In practical implementation, a strategy closed-loop module of a human-machine collaborative production efficiency optimization system based on big data analysis generates a comprehensive optimization strategy and deploys it to the production site based on the causal inference graph and contribution degree output by the causal analysis module. The comprehensive optimization strategy includes operation suggestions, equipment adjustment suggestions, and material scheduling suggestions. In practical implementation, the strategy closed-loop module receives a quantitative contribution degree list output by the causal analysis module and categorizes events with contribution degrees exceeding a set threshold. The set threshold is a configurable numerical parameter; for example, events with a contribution degree greater than 10% are considered key impact events. The categorization operation forms a human factor operation improvement set, an equipment status adjustment set, and a material flow optimization set. Each set contains detailed information and contribution degrees of similar events. In some embodiments, referring to Table 1, event categorization can be represented logically as follows.

[0037] Table 1: Contribution and Classification of Key Events For each operational event in the human factors operation improvement set, the strategy closed-loop module matches standardized operating procedures from a pre-set knowledge base. This knowledge base stores verified standard operating procedures, best practice videos, and illustrated guides. The strategy closed-loop module generates a description of the differences compared to the current operation. This description is derived by comparing the actual collected operational feature sequence with the feature templates of the standardized operation in the knowledge base. For example, it might be described as "The current average tightening time is 3.5 seconds, which is lower than the standard requirement of 4.0 ± 0.2 seconds." Based on the matched standard and the description of the differences, specific operational suggestions are generated, such as "Adjust the electric screwdriver torque setting to the standard value of 8 N·m and ensure a continuous tightening time of 4.0 seconds."

[0038] For each equipment event in the equipment status adjustment set, the strategy closed-loop module generates suggestions by combining the equipment maintenance manual and historical best operating parameters. The equipment maintenance manual provides the standard range and maintenance cycle of equipment parameters, while historical best operating parameters are statistical values ​​of the equipment's operating parameters during high-efficiency periods, retrieved from the time-series characteristic database. The strategy closed-loop module generates parameter adjustment ranges or preventive maintenance instructions, forming specific equipment adjustment suggestions. The equipment adjustment suggestions may include, for example, "Adjust the servo motor gain parameter Kp from the current value of 120 to the optimal range of 135-145," or "It is recommended to preventively replace the shaft arm bearing during the next planned shutdown."

[0039] For each material event in the material flow optimization center, the strategy closed-loop module combines production planning and inventory strategy to generate adjustment plans. Production planning provides the material demand time and quantity for future processes, while inventory strategy defines safety stock levels and replenishment points. The strategy closed-loop module generates adjustment plans for material delivery time, quantity, or route, forming specific material scheduling suggestions. Material scheduling suggestions include, for example, "advance the delivery time of material B to assembly station 3 by 20 minutes from the planned start time," or "increase the safety stock of material C in the line-side warehouse from 5 units to 8 units."

[0040] In practical implementation, the strategy closed-loop module integrates and sorts the generated operation suggestions, equipment adjustment suggestions, and material scheduling suggestions according to the process sequence and priority. The process sequence is based on the process flow definition in the manufacturing execution system, and the priority ranking is based on the contribution of events and the urgency of event handling, using a priority score. The calculation method is as follows: in: The priority score represents the event. This represents the normalized value of the contribution of the event. This represents an urgency coefficient determined by the event type and the current stage of production. and These are the weighting coefficients for contribution and urgency, respectively. The strategy closed-loop module is based on priority scores. All suggestions are sorted in descending order, and then integrated and ranked to generate a structured, comprehensive optimization strategy document. This strategy document clearly lists the actions to be performed, their execution locations, standard parameters, and execution order.

[0041] The strategy closed-loop module deploys comprehensive optimization strategies to the production floor, collects feedback data in real time after strategy execution, and transforms the comprehensive optimization strategy document into executable instructions. This transformation process follows the application programming interface (API) specifications of the Manufacturing Execution System (MES), Equipment Controllers (ECS), and Warehouse Management System (WMS), mapping natural language descriptions of suggestions into system-recognizable instruction codes or parameter files. In practice, operational suggestions are converted into standardized work instruction update instructions that pop up on the MES workstation terminals; equipment adjustment suggestions are converted into parameter modification instructions sent to the equipment controllers; and material scheduling suggestions are converted into delivery task update instructions sent to the WMS. The strategy closed-loop module then distributes these executable instructions to the corresponding workstations, ECS, and WMS through the MES. During execution, the data acquisition module continuously collects new human factor operation data, equipment status data, and material flow data. This newly collected data serves as feedback data for strategy execution. The feedback data is input into the system preprocessing module, forming a new raw production time-series big data stream, thus initiating the next closed-loop optimization cycle.

[0042] See Figure 4 In the closed-loop optimization process of the human-machine collaborative production efficiency optimization system based on big data analysis, production efficiency and the number of bottlenecks show a significant inverse synergistic evolution trend. Throughout the entire cycle from before optimization to the final state: Production efficiency (dot curve, left axis): continuously climbed from an initial 72% to a final 96%, a cumulative increase of approximately 33.3%. Each round of optimization achieved a stable gain, demonstrating the continuous pulling effect of the closed-loop optimization strategy on production efficiency. The number of production bottlenecks (box curve, right axis): continuously decreased from an initial 4 to 0. After three rounds of optimization, the bottlenecks were completely eliminated, intuitively reflecting the accurate identification and resolution capabilities of the bottleneck tracing and causal analysis module for production constraints. The trend curve fully confirms the effectiveness of the system's closed-loop optimization logic: through dynamic Bayesian network modeling, state inversion tracing, and causal inference analysis, key bottleneck events are accurately located and eliminated, thereby driving continuous improvement in production efficiency and ultimately achieving a bottleneck-free, high-efficiency production state.

[0043] In one embodiment of the present invention, the model update module operates within a closed-loop optimization process. It periodically uses new production data and corresponding actual production efficiency results as training samples, employing an incremental learning algorithm to update the conditional probability distribution and state transition probabilities in the dynamic Bayesian network model using these training samples. When significant changes occur in the production line configuration, process, or product type, a model structure learning process is triggered to redetermine network nodes and connection structures, and the dynamic Bayesian network model is reconstructed using data accumulated under the new scenario. The simulation verification module is initiated before implementing the comprehensive optimization strategy on the actual production site. It loads the optimization strategy to be implemented into the corresponding production system digital twin model, drives the digital twin model to run, simulates the operating status of one or more complete production cycles after strategy execution, predicts the changing trends of key performance indicators, compares the simulation prediction results with the expected goals of the strategy, and if the deviation exceeds the allowable range, automatically adjusts the relevant parameters in the strategy and re-runs the simulation until the prediction results meet the requirements. The final strategy version that has passed simulation verification is then deployed to the actual production site. The cross-order collaboration module operates during multi-process collaborative optimization in the system. It identifies new constraints or demand changes in downstream processes caused by the execution of optimization strategies in upstream processes. Using these new constraints or demand changes as input, it re-runs the inference computation and state inversion algorithm of the dynamic Bayesian network to evaluate their impact on the production efficiency of downstream processes. Based on the evaluation results, it generates compensatory or adaptive optimization strategies for downstream processes. This module coordinates the execution rhythm and resource allocation of optimization strategies between upstream and downstream processes to ensure overall optimal production efficiency, rather than local optima for a single process.

[0044] In practical implementation, a human-machine collaborative production efficiency optimization system based on big data analysis employs a model update module, a simulation verification module, and a cross-sequence collaboration module. The model update module operates within a closed-loop optimization process, periodically using new production data and corresponding actual production efficiency results as training samples. These training samples include standardized feature sequences obtained from the preprocessing module and verified actual production efficiency index values ​​acquired from the manufacturing execution system. The model update module uses an incremental learning algorithm to update the conditional probability distribution and state transition probabilities in the dynamic Bayesian network model using the training samples. The incremental learning algorithm progressively adjusts the model parameters based on new samples while retaining existing knowledge. In some embodiments, incremental learning can employ a variant of the online expectation-maximization algorithm, where the parameter vector of the dynamic Bayesian network model... The updates follow these rules: in: This represents the updated model parameter vector. This represents the model parameter vector before the update. It is the learning rate, used to control the step size for each update. It is a loss function Regarding parameters In the latest batch of training data gradient on, It is the regularization coefficient. This is a reference parameter vector used to prevent catastrophic forgetting, typically a moving average of historical parameters. The model update module performs incremental learning when triggering conditions are met. These triggering conditions can be a fixed time period or the accumulation of a certain number of new samples. When there are significant changes in the production line configuration, process, or product type, the model update module triggers the model structure learning process. This process redefines the network nodes and structure of the dynamic Bayesian network model and reconstructs it using data collected in the new scenario. Structure learning can employ score-based or constraint-based methods to find the network structure that best matches the data.

[0045] In practical implementation, the simulation verification module is activated before the comprehensive optimization strategy implemented by the strategy closed-loop module is applied to the actual production site. The simulation verification module loads the optimization strategy to be implemented into the corresponding digital twin model of the production system. This digital twin model is a high-fidelity virtual mapping of the physical production line and its operational logic. The simulation verification module drives the digital twin model to run, simulating the operational status of the complete production cycle after the strategy execution. The duration of the complete production cycle can be configured according to the scope of the strategy's impact. During the simulation process, the behavioral logic of virtual personnel, the control parameters of virtual equipment, or the scheduling rules of virtual materials in the digital twin model are adjusted according to the strategy content, and changes in key performance indicators are predicted. Key performance indicators include, but are not limited to, output rate, overall equipment efficiency, and order delivery cycle. The simulation verification module compares the simulation prediction results with the expected goals of the strategy. If the deviation exceeds the allowable range, the simulation verification module automatically adjusts the parameters in the strategy and re-simulates until the prediction results meet the requirements.

[0046] The cross-sequence collaboration module operates during multi-process collaborative optimization in the system. It identifies new constraints or demand changes in downstream processes caused by the execution of optimization strategies in upstream processes. This identification process is achieved by analyzing the simulated output of the digital twin model or the instructions of deployed strategies. For example, if an upstream process increases cycle time by adjusting equipment parameters, this may lead to new requirements on the material supply rate of downstream processes. The cross-sequence collaboration module takes these new constraints or demand changes as input and re-runs the inference computation and state inversion algorithm of the dynamic Bayesian network to assess the impact on the production efficiency of downstream processes. In practice, the assessment process transforms the new constraints into evidence settings for the corresponding observation nodes or state nodes in the dynamic Bayesian network model of the downstream process, and then runs inference computation to predict changes in the production efficiency of the downstream process. Based on the assessment results, the cross-sequence collaboration module generates compensatory or adaptive optimization strategies for downstream processes. Compensatory strategies aim to offset the negative impact of upstream changes, while adaptive strategies aim to help downstream processes utilize new opportunities brought about by upstream changes. It is understandable that the cross-order collaboration module coordinates the execution rhythm and resource allocation of optimization strategies between upstream and downstream processes. The coordination mechanism may involve setting timestamps or dependencies for strategy instructions of different processes to ensure the overall optimization of global production efficiency, rather than the local optimization of a single process.

[0047] See Figure 5 In the performance evaluation system of the core modules of the human-machine collaborative system, this chart quantifies the dual-indicator performance of five core modules—model updating, simulation verification, cross-order collaboration, strategy closure, and state modeling—in terms of response time and prediction accuracy. Specifically, a bar chart represents the response time (ms) of each module, while a line chart corresponds to the prediction accuracy (%), constructing a two-dimensional performance mapping relationship of "time-accuracy." In terms of module performance characterization, the simulation verification module exhibits a dual optimal state of peak response time (approximately 180ms) and peak prediction accuracy (96%). The cross-order collaboration module, however, shows a performance inflection point, with its response time dropping to around 150ms, while the prediction accuracy falls to 92%, representing the lowest accuracy point in the entire system. The state modeling module reaches its maximum response time (200ms) across the entire process, with the prediction accuracy simultaneously dropping to 90%, forming a dual high point of time consumption and accuracy loss. In the performance correlation analysis, each module exhibits a characteristic pattern of "inverse fluctuation between response time and prediction accuracy": when the response time is low (such as cross-order collaboration and strategy closure), the prediction accuracy is relatively improved; however, a significant improvement in response time (such as simulation verification and state modeling) does not bring a synchronous gain in accuracy, but instead causes a stepwise decline in accuracy. This chart clearly reveals the trade-off between computational efficiency and prediction performance of different modules in the human-machine collaborative system, providing quantitative data support for subsequent module performance optimization and resource scheduling strategy formulation.

[0048] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A man-machine collaborative production efficiency optimization system based on big data analysis, characterized in that, include: The data acquisition module collects human operation data, equipment status data, and material flow data in real time during the human-machine collaborative production process, forming a raw production time-series big data stream; The preprocessing module performs multimodal data fusion and cleaning on the original production time-series big data stream, extracts human-machine interaction behavior feature sequences, equipment operation feature sequences and material consumption feature sequences, and simultaneously marks their timestamps and process identifiers; The state modeling module, based on the human-machine interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence, uses inference calculation to construct a dynamic Bayesian network model of the human-machine collaborative production state. The bottleneck tracing module inputs the dynamic Bayesian network model into the state inversion algorithm, and uses the state inversion algorithm to trace the historical state evolution path that led to the current production efficiency bottleneck, and identifies key human factor operation events, equipment status events and material flow events. The causal analysis module constructs a causal inference diagram based on the key human operation events, equipment status events, and material flow events, quantifying the contribution of each type of event to production efficiency indicators. The strategy closed-loop module generates a comprehensive optimization strategy based on the causal inference graph and the contribution degree, deploys the comprehensive optimization strategy to the production site, and collects feedback data after the strategy is executed in real time to form a new original production time-series big data stream, thus constituting a closed-loop optimization process.

2. The human-machine collaborative production efficiency optimization system based on big data analysis according to claim 1, wherein, The original production time-series big data stream is subjected to multimodal data fusion and cleaning to extract human-machine interaction behavior feature sequences, equipment operation feature sequences, and material consumption feature sequences, including: The video stream, sensor signals and operation logs in the human-computer interaction data are time-aligned, and standardized operation action units are segmented from them to form the human-computer interaction behavior feature sequence. The vibration, temperature, and current time-series signals in the equipment status data are subjected to frequency domain transformation and feature extraction to calculate the equipment health index and energy efficiency index, thereby forming the equipment operation feature sequence. Perform correlation analysis on the batch, quantity, and location change records in the material flow data, calculate the material completeness rate, turnover rate, and work-in-process inventory level, and form the material consumption characteristic sequence. The human-computer interaction behavior feature sequence, the equipment operation feature sequence, and the material consumption feature sequence are subjected to missing value imputation, outlier detection, and smoothing filtering to generate a cleaned standardized feature sequence. The cleaned standardized feature sequence is bound with a unified timestamp and process identifier and stored in a time-series feature database.

3. The human-machine collaborative production efficiency optimization system based on big data analysis as described in claim 2, characterized in that, Based on the aforementioned human-computer interaction behavior feature sequence, equipment operation feature sequence, and material consumption feature sequence, a dynamic Bayesian network model of the human-machine collaborative production state is constructed using inference computation, including: The nodes of the dynamic Bayesian network model represent production state variables of different granularities. Production efficiency is determined as the top-level target node, and the human-machine interaction behavior characteristics, the equipment operation characteristics, and the material consumption characteristics are used as the bottom-level observation nodes. Define intermediate layer status nodes, which include job cycle status, resource matching status, and collaboration smoothness status; Based on historical data learning or expert knowledge, initialize the conditional probability distribution between the top-level target node, the intermediate-level state node, and the bottom-level observation node; Using the temporal information of the human-computer interaction behavior feature sequence, the equipment operation feature sequence, and the material consumption feature sequence, the conditional probability distribution is dynamically updated through inference calculation, and the state transition probability of nodes between different time steps is learned; When the inference accuracy of the dynamic Bayesian network model reaches a preset threshold, the model construction is complete. The model responds to new feature sequence inputs and outputs the current production efficiency status and its probability distribution.

4. The human-machine collaborative production performance optimization system based on big data analysis according to claim 3, characterized in that, The dynamic Bayesian network model is input into the state inversion algorithm, which traces back the historical state evolution path leading to the current production efficiency bottleneck, including: In the dynamic Bayesian network model, the currently observed low productivity state is set as an evidence node, and its state is fixed. Run the state inversion algorithm, starting from the top-level target node, propagate the probability backward along the network structure, and calculate the posterior probability that each intermediate layer state node is in an abnormal state at each historical time step given evidence of low production efficiency. Based on the magnitude of the posterior probability, intermediate layer state nodes whose posterior probability exceeds the confidence threshold are selected and marked as abnormal historical states. By connecting the marked abnormal historical state nodes in chronological order, one or more state evolution paths from historical abnormal states to the current low productivity state are formed.

5. The human-machine collaborative production efficiency optimization system based on big data analysis as described in claim 4, characterized in that, Based on the aforementioned key human-caused operational events, equipment status events, and material flow events, a causal inference diagram is constructed to quantify the contribution of each type of event to production efficiency indicators, including: From the state evolution path, extract the bottom observation node that is directly associated with the abnormal change of the intermediate layer state node. The bottom observation node corresponds to the specific human operation event, equipment status event or material flow event. Using production efficiency indicators as outcome variables and the extracted human operation events, equipment status events, and material flow events as causal variables, a causal inference diagram is constructed. Using counterfactual reasoning, the difference between the expected value and the current actual value of the production efficiency indicator is calculated when one of the causal variables is missing. The difference values ​​are normalized and used as the contribution of the causal variable to the production efficiency index. The greater the contribution, the higher the impact of the event corresponding to the causal variable on the current efficiency bottleneck.

6. The human-machine collaborative production performance optimization system based on big data analysis according to claim 5, wherein, Based on the causal inference graph and the contribution level, a comprehensive optimization strategy is generated, including: The comprehensive optimization strategy includes operational suggestions, equipment adjustment suggestions, and material scheduling suggestions; Events whose contribution exceeds a set threshold are categorized into sets for human factors operation improvement, equipment status adjustment, and material flow optimization. For each operation event in the human factors operation improvement set, standardized operation specifications are matched from a preset knowledge base, and a description of the differences compared to the current operation is generated to form specific operation suggestions; For each equipment event in the equipment status adjustment set, combined with the equipment maintenance manual and historical best operating parameters, a parameter adjustment range or preventive maintenance instruction is generated to form a specific equipment adjustment suggestion; For each material event in the material flow optimization set, an adjustment plan for material delivery time, quantity, or route is generated by combining production planning and inventory strategy, forming a specific material scheduling suggestion; The operation suggestions, equipment adjustment suggestions, and material scheduling suggestions are integrated and sorted according to the process sequence and priority to generate the comprehensive optimization strategy document.

7. The human-machine collaborative production efficiency optimization system based on big data analysis as described in claim 6, characterized in that, Deploy the comprehensive optimization strategy to the production site and collect feedback data in real time after the strategy is executed, including: The comprehensive optimization strategy document is converted into executable instructions and distributed to the corresponding workstations, equipment controllers, and warehouse management systems through the manufacturing execution system. During execution, new human factors operation data, equipment status data, and material flow data are continuously collected as feedback data for strategy execution.

8. The human-machine collaborative production efficiency optimization system based on big data analysis as described in claim 7, characterized in that, It also includes a model update module, used for: In the closed-loop optimization process, new production data and corresponding actual production efficiency results are periodically used as training samples. The conditional probability distribution and state transition probability in the dynamic Bayesian network model are updated using the training samples through an incremental learning algorithm. When there are significant changes in the configuration, process, or product type of the production line, the model structure learning process is triggered to redetermine the network nodes and structure, and the dynamic Bayesian network model is reconstructed using data from the new scenario.

9. The human-machine collaborative production efficiency optimization system based on big data analysis as described in claim 8, characterized in that, It also includes a simulation verification module for: Before implementing the comprehensive optimization strategy in the actual production site, the strategy is loaded into the corresponding production system digital twin model; Drive the digital twin model to run, simulate the running status of the entire production cycle after the strategy is executed, and predict changes in key performance indicators; The simulation prediction results are compared with the expected target of the strategy. If the deviation exceeds the allowable range, the parameters in the strategy are automatically adjusted and the simulation is repeated until the prediction results meet the requirements. The final strategy version, validated through simulation, will be deployed to the actual production site. 10.The big data analysis based man-machine collaborative production efficiency optimization system according to claim 9, wherein, It also includes a cross-order coordination module, used for: Identify new constraints or changes in demand for downstream processes caused by the implementation of optimization strategies in upstream processes; Using the new constraints or changes in requirements as input, the inference calculation and state inversion algorithm are rerun to assess the impact on the production efficiency of downstream processes. Based on the evaluation results, compensatory or adaptive optimization strategies are generated for downstream processes; Coordinate the execution pace and resource allocation of optimization strategies between upstream and downstream processes to ensure overall optimal production efficiency, rather than local optimality of a single process.

Images