Autonomous monitoring method and system for unmanned laboratory sites

Abstract

The application discloses an unmanned laboratory field autonomous monitoring method and system, and relates to the technical field of unmanned laboratory monitoring.The method comprises the following steps: carrying out space-time alignment and preprocessing on multi-source heterogeneous sensing data, generating a state characteristic vector through dynamic physical deduction of a digital twin space, comparing the state characteristic vector with a reference vector, and identifying potential loss of control of a physicochemical reaction or equipment implicit recession; combining a quantitative confidence output; a reasoning engine deeply fuses real-time resource constraint conditions of a laboratory when generating an autonomous monitoring decision instruction; through multi-level strategies such as process path re-planning, standby instrument takeover and parameter self-adaptive compensation, the instruction is issued to drive the collaborative action of multiple actuators to complete physical disposal, and meanwhile, state feedback data in the execution process are fed back to the digital twin space in real time.The application significantly reduces the false alarm and missed alarm rates, and realizes self-sensing, self-decision, self-execution and closed-loop control of the field anomaly.

Description

Technical Field

[0001] This invention relates to the technical field of unmanned laboratories, and in particular to an autonomous monitoring method and system for unmanned laboratory sites. Background Technology

[0002] Existing automated monitoring methods for unmanned laboratories typically suffer from the following technical limitations in practical operation, especially when facing complex and ever-changing testing tasks: These methods usually rely on surface sensors deployed outside the equipment for data acquisition, resulting in a superficial perception dimension. Furthermore, the lack of a unified spatiotemporal alignment and preprocessing mechanism for multi-source heterogeneous sensing data collected during sample transfer, pretreatment, and analysis leads to spatiotemporal fragmentation of the data. Additionally, existing technologies often employ qualitative alarm mechanisms based on fixed thresholds, lacking in-depth state feature comparison and quantitative confidence analysis based on dynamic physical deduction. This makes it impossible to identify potential physicochemical reaction runaways or hidden equipment degradation in advance, resulting in severely delayed anomaly warnings. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides an autonomous monitoring method and system for unmanned laboratory sites.

[0004] This invention provides an autonomous monitoring method for unmanned laboratory sites, comprising:

[0005] In the unmanned laboratory, based on the sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste liquid disposal, multi-source heterogeneous sensing data is collected in real time, the multi-source heterogeneous sensing data is spatiotemporally aligned, and the corresponding key data combination is determined by combining data preprocessing.

[0006] Each key data combination is input into a preset digital twin space, and dynamic simulation is performed in the digital twin space to generate state feature vectors for each process node. By performing deviation processing with the state baseline vector, potential abnormal working condition types and their corresponding confidence levels are identified.

[0007] Each abnormal operating condition type and its confidence level are input into the inference engine of the unmanned laboratory, and autonomous monitoring decision instructions are dynamically generated in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategy.

[0008] The autonomous monitoring and decision-making instructions are sent to the execution control layer of each process node to drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling. At the same time, the status feedback data during the handling process is transmitted back to the digital twin space in real time until each process node returns to the steady-state operation range.

[0009] This invention provides an autonomous monitoring system for unmanned laboratory sites, which is applied to the aforementioned autonomous monitoring method for unmanned laboratory sites; the autonomous monitoring system for unmanned laboratory sites includes:

[0010] The data module is used in the unmanned laboratory to collect multi-source heterogeneous sensing data in real time based on sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste disposal. It performs spatiotemporal alignment of the multi-source heterogeneous sensing data and determines the corresponding key data combination in combination with data preprocessing.

[0011] The digital twin module is used to input various key data combinations into a preset digital twin space, and to perform dynamic simulation in the digital twin space to generate state feature vectors for each process node. By processing the deviation from the state baseline vector, potential abnormal working condition types and their corresponding confidence levels are identified.

[0012] The autonomous monitoring module is used to input various abnormal operating conditions and their confidence levels into the inference engine of the unmanned laboratory, and dynamically generate autonomous monitoring decision instructions in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies.

[0013] The anomaly handling module is used to issue autonomous monitoring and decision-making instructions to the execution control layer of each process node, drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling, and at the same time transmit the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range.

[0014] Compared with the prior art, the beneficial effects of the present invention are:

[0015] (1) By performing spatiotemporal alignment and preprocessing on multi-source heterogeneous sensing data throughout the entire process of sample transfer, pretreatment, analysis and detection, and by leveraging the dynamic physical deduction of digital twin space, the limitation of traditional surface sensors being unable to detect the microscopic physical state inside the equipment is overcome. By generating deep state feature vectors and comparing them with benchmark vectors, potential physicochemical reaction runaway or equipment latent degradation can be identified in advance. Combined with quantitative confidence output, anomaly identification is upgraded from traditional qualitative alarm to quantitative diagnosis, significantly reducing false alarm and missed alarm rates.

[0016] (2) When generating disposal decisions, the inference engine deeply integrates the real-time resource constraints of the laboratory. Through multi-level strategies such as process path replanning, backup instrument takeover and parameter adaptive compensation, it achieves seamless process connection and flexible resource scheduling under abnormal working conditions while meeting the dual requirements of safety and efficiency. It drives multiple actuators to work together to complete physical disposal through instruction issuance, and transmits the status feedback data during the execution process back to the digital twin space in real time, guiding the current abnormality to continuously converge until it is restored to the steady-state operating range. It realizes the self-sensing, self-decision-making, self-execution and closed-loop control of on-site abnormalities. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the autonomous monitoring method for an unmanned laboratory site according to an embodiment of the present invention.

[0018] Figure 2 This is a simplified flowchart illustrating a specific application scenario of the autonomous monitoring method for unmanned laboratory sites in this invention embodiment;

[0019] Figure 3 This is a schematic diagram of the structural composition of the autonomous monitoring system for an unmanned laboratory site in an embodiment of the present invention.

[0020] Figure label:

[0021] 21. Data module; 22. Digital twin module; 23. Autonomous monitoring module; 24. Anomaly handling module. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] Please see Figures 1 to 2 An autonomous monitoring method for unmanned laboratory sites, applied to unmanned laboratory scenarios; the autonomous monitoring method for unmanned laboratory sites includes:

[0024] Step S11: In the unmanned laboratory, based on the sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste liquid disposal, multi-source heterogeneous sensing data is collected in real time, the multi-source heterogeneous sensing data is spatiotemporally aligned, and the corresponding key data combination is determined by combining data preprocessing.

[0025] Step S12: Input the combination of key data into the preset digital twin space, perform dynamic simulation in the digital twin space to generate the state feature vector of each process node, and identify the potential abnormal working condition type and its corresponding confidence level by performing deviation processing with the state reference vector.

[0026] Step S13: Input each abnormal working condition type and its confidence level into the inference engine of the unmanned laboratory, and dynamically generate autonomous monitoring decision instructions in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic and parameter adjustment strategy.

[0027] Step S14: Send autonomous monitoring decision instructions to the execution control layer of each process node to drive the corresponding execution mechanism to coordinate actions to complete the abnormality handling. At the same time, transmit the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range.

[0028] In step S11, the specific steps are as follows:

[0029] S111: In the unmanned laboratory site, mark each process node of sample transfer, pretreatment, analysis and detection and waste liquid disposal in the unmanned laboratory site, deploy corresponding sensing units for each process node, and collect corresponding multi-source heterogeneous sensing data in real time through the sensing unit. The multi-source heterogeneous sensing data covers image stream, weight signal, position coordinate, environmental parameters and instrument status code.

[0030] S112: A spatiotemporal registration mechanism based on timestamp indexing and spatial coordinate transformation is adopted to map multi-source heterogeneous sensing data to a unified global spatiotemporal reference, and an outlier removal mechanism is used to clean the data, thereby eliminating environmental noise interference and outputting the corresponding target data stream.

[0031] S113: Obtain a preset water quality analysis process knowledge graph, extract time-series feature vectors representing the process operation status from the target data stream, and construct corresponding key data combinations through feature fusion. These key data combinations include environmental background data, equipment operating parameters, and sample attribute labels.

[0032] In step S111, the specific steps are as follows:

[0033] A 3D model of the physical space of the unmanned laboratory site is created. Based on the business logic of sample transfer, pretreatment, analysis and testing and waste disposal, the physical boundaries and spatial topological relationships of each process node are clearly defined in the 3D spatial coordinate system, and a unique logical identifier is assigned to each node.

[0034] Specifically, the sample transfer area is marked as a logical set encompassing the AGV (Automated Guided Vehicle) operating path, robotic arm operating radius, and material storage location; the pretreatment area is marked as a set of physical operation units including centrifugation, oscillation, digestion, and filtration; the analysis and detection area is marked as a set of precision detection units integrating a spectrometer, chromatograph, and multi-parameter water quality analyzer; and the waste liquid disposal area is marked as a set of end-of-pipe treatment units encompassing classified collection bins, neutralization reaction tanks, and discharge control valves. Through these markings, a two-way mapping relationship between physical spatial coordinates and logical process nodes is established, providing a spatial reference for the precise deployment of subsequent sensing units.

[0035] Based on the completion of process node marking, and according to the monitoring requirements of each node, multimodal sensing units with different physical principles are deployed to build a global sensing network; for sample transfer nodes, 2D / 3D vision industrial cameras and LiDAR are deployed to capture the shape image of the sample bottle and the real-time position of the transport carrier; RFID radio frequency identification readers or QR code scanning modules are deployed to read the identification tag information of the sample.

[0036] For the pretreatment nodes, high-precision weighing sensors are deployed at the reagent addition and sample weighing positions to collect the weight change signals of trace liquids; vibration sensors and temperature and humidity sensors are deployed in the outer shell or internal core area of ​​the centrifuge and digester to collect the physical state parameters of the equipment during operation.

[0037] For the analysis and testing nodes, smart meters and protocol conversion gateways are deployed to collect the instrument's operating status codes in real time via the instrument's digital communication interface, such as RS485, Modbus, or TCP / IP. At the same time, spectral probes or ion-selective electrodes are deployed to directly contact the liquid to be tested to obtain raw detection data. For the waste liquid disposal nodes, level gauges and pH sensors are deployed inside the waste liquid tank to collect the accumulated waste liquid volume and pH environmental parameters. All sensing units are connected to the edge computing gateway via industrial fieldbus or wireless network to achieve the aggregation of multi-source heterogeneous data.

[0038] Each sensing unit is activated to monitor and collect data on the operation status of the process nodes in real time according to the preset sampling frequency strategy, forming a multi-source heterogeneous sensing data stream. The specific acquisition process includes: the vision unit acquiring image stream data containing sample flow dynamics at a rate of 25 to 60 frames per second; the weighing sensor outputting analog or digital signals corresponding to the reagent weight at a sampling period of milliseconds; and the lidar or vision positioning system scanning and calculating the position coordinate data of the AGV or robotic arm end in the global coordinate system of the laboratory at a high frequency.

[0039] Environmental sensors monitor environmental parameters such as temperature, humidity, and light intensity in the laboratory in real time; the protocol conversion gateway continuously polls or listens to the analytical instruments, parses and outputs instrument status codes that indicate whether the equipment is idle, running, reporting an error, or shut down; during the acquisition process, the system ensures that each data source generates raw data packets with preliminary timestamps locally, and transmits them to the data preprocessing center in real time through a high-bandwidth network, thereby completing the acquisition of all elements of data, including image streams, weight signals, location coordinates, environmental parameters, and instrument status codes.

[0040] In step S112, the specific steps are as follows:

[0041] At this point, a globally unified clock synchronization system is constructed, using Precise Time Protocol (PTP) or Network Time Protocol (NTP) as the master clock source. Periodic time synchronization signals are sent to all sensing units, edge computing gateways, and data acquisition terminals in the unmanned laboratory to ensure that the local clock deviation of all devices in the system is controlled within the millisecond range. After receiving multi-source heterogeneous sensing data, the local timestamp carried in each piece of raw data is parsed and corrected according to the clock offset of each device relative to the master clock source, and converted into a globally unified standard timestamp.

[0042] For data sources with different sampling frequencies, a sliding time window mechanism based on a unified timestamp index is established. Taking the sampling time of high-frequency data as the reference point, the estimated value of low-frequency data at the reference time is calculated through linear interpolation or nearest neighbor interpolation algorithm. This ensures that the image stream, weight signal, position coordinates, environmental parameters and instrument status code are strictly aligned in the time dimension, eliminating the time phase difference caused by different sampling frequencies and forming a multidimensional data set under the same time slice.

[0043] A global three-dimensional spatial coordinate system for the unmanned laboratory is pre-established, and the extrinsic parameter matrices of each sensing unit, such as industrial cameras, LiDAR, and robotic arm base, in the global coordinate system are obtained through offline calibration, including rotation matrices and translation vectors. In the real-time data processing stage, local spatial data collected by the sensing units are extracted, such as two-dimensional pixel coordinates within the camera's field of view, joint coordinates at the end of the robotic arm, or local scanning point cloud data of the LiDAR. Rigid body transformation operations are performed on these data using the pre-calibrated extrinsic parameter matrices to convert the local coordinates into absolute position information in the global three-dimensional spatial coordinate system.

[0044] For non-geometric attribute data, such as environmental parameters, the data is associated with the corresponding spatial node identifier in the global coordinate system based on the physical installation location of the sensor that collected the data. This achieves a unified mapping of multi-source heterogeneous data in the spatial dimension, ensuring that all data refer to the state at the same physical spatial location and eliminating spatial ambiguity caused by different perspectives or coordinate systems.

[0045] For the data stream after spatiotemporal registration, a multi-level data cleaning mechanism is deployed to eliminate environmental noise and abnormal interference; static threshold rules based on physical mechanisms are set to directly filter out "data points" that exceed the sensor's range or clearly violate physical laws, such as temperature sudden changes to several thousand degrees or negative weight signals.

[0046] A dynamic outlier detection algorithm based on statistical distribution is adopted to calculate the statistical characteristic value of the current data within the sliding time window. The 3σ principle or interquartile range (IQR) rule is used to determine whether the data point is an outlier. Furthermore, a filtering mechanism based on a time series prediction model is introduced. Kalman filtering or autoregressive moving average model (ARIMA) is used to predict the normal range of the data at the next time step. Residual analysis is performed between the actual collected value and the predicted value. When the residual exceeds a set threshold, the data is determined to be noise interference and is smoothed or removed. The cleaned, high-fidelity target data stream with spatiotemporal consistency is output, providing reliable data support for subsequent feature extraction and state inference.

[0047] In step S113, the specific steps are as follows:

[0048] At this point, a general knowledge graph of water quality analysis processes is retrieved from the preset knowledge base. This graph stores entities, attributes, and logical relationships between entities in the field of water quality testing in the form of a graph structure. Entities include pretreatment equipment, analytical instruments, chemical reagents, detection indicators, and environmental factors. Relationships include "equipment-indicator" dependencies, "reagent-reaction" associations, and "parameter-threshold" constraints. After obtaining the general graph, the system instantiates the graph according to the task configuration of the current autonomous monitoring scenario.

[0049] The system parses the input task instructions, identifies the type of water quality to be tested (e.g., type B water quality), and retrieves a specific sub-graph from the graph that matches this type of water quality. This sub-graph contains specific "detection process flow" and "parameter constraints" for type B water quality, such as specific digestion temperature curves, masking agent addition order, and interfering ion removal steps. These constraints include higher reaction time redundancy and specific sensor sensitivity requirements. Through the above instantiation process, a dedicated process knowledge graph is constructed, containing current task context information, such as the "type B water quality - high salt interference - UV digestion - total nitrogen detection" chain. This provides semantic-level prior knowledge and rule guidance for subsequent feature extraction.

[0050] Based on the instantiated process knowledge graph, the key feature dimensions that need to be monitored for each process node are determined, and the time-series feature vectors representing the process operation status are extracted from the target data stream output by S112. The specific extraction process includes: for continuously changing "analog signals, such as temperature, pressure, and speed", calculating the time-domain "statistical features, such as mean, variance, and peak value" and "frequency-domain features, such as main frequency energy and spectral entropy" within a sliding time window to form a numerical sequence that can reflect the dynamic evolution trend of the signal.

[0051] For unstructured image stream data, computer vision algorithms are used to extract the "geometric features" of the target region, such as liquid level height, bubble area ratio, and color histogram distribution, as well as "motion features, such as flow velocity and robotic arm trajectory smoothness," and transform them into structured vector forms. For discrete status code data, one-hot encoding or word embedding is performed to generate semantic vectors of equipment status. In this process, entity attributes in the knowledge graph are used to guide the focus of feature extraction. For example, if the graph defines a strong correlation between the "dissolution" process and "temperature," the system will assign higher weights to the derivative and volatility of the temperature signal when extracting features, thereby constructing a set of high-dimensional, multi-dimensional temporal feature vectors. These vectors accurately depict the operational details of the process nodes at the current moment.

[0052] The extracted time-series feature vectors are fused with environmental background data, equipment operating parameters, and sample attribute labels using multimodal feature fusion to construct the final key data combination. Specifically, current environmental background data, including laboratory atmospheric temperature, relative humidity, and cleanliness level, is acquired and appended to the head of the feature vector as global context information. Secondly, equipment operating parameters, including motor speed, pump and valve opening, and deviations between set and feedback parameter values, are acquired. These parameters reflect the equipment's execution capability and control precision and are mapped to the middle of the feature vector.

[0053] Sample attribute labels are obtained from the sample identification system and include sample number, origin, matrix type (e.g., Type B water quality - high salinity industrial wastewater), and expected detection indicators. These labels are embedded as semantic features at the end of the feature vector. Through the above splicing or cascading operations, discrete and heterogeneous feature information is fused into a unified high-dimensional feature vector, i.e., key data combination. This key data combination not only includes real-time monitoring data but also integrates environmental interference factors, equipment health status, and the inherent physicochemical properties of the sample, forming a holistic data structure with complete semantic information. This provides a standardized input interface for the subsequent dynamic extrapolation of the digital twin space.

[0054] For the key data set, it is composed of three sub-vectors concatenated in a fixed order: the first sub-vector is the environmental background data vector, with a fixed dimension of N1, which includes the current atmospheric temperature, relative humidity, and cleanliness level of the laboratory; the second sub-vector is the equipment operating parameter vector, with a fixed dimension of N2, which includes the motor speed feedback value, pump and valve opening value, deviation value between set parameters and feedback parameters, and semantic embedding value of equipment status code of each major actuator in the current process node; the third sub-vector is the sample attribute label vector, with a fixed dimension of N3, which includes the hash code of sample number, the primary thermal code of matrix type, the source code, and the identifier code of expected detection index.

[0055] The three sub-vectors mentioned above are merged by concatenating their beginning and end to form a single high-dimensional floating-point data vector with a total dimension of N = N1 + N2 + N3. This vector is the key data combination. In subsequent steps, this key data combination serves as standardized input interface data and is fed into the digital twin space for dynamic simulation. To eliminate the influence of different physical dimensions on subsequent calculations, before splicing, the system performs minimum-maximum normalization on the values ​​of each dimension, linearly mapping them to the interval [0, 1]. The historical extreme values ​​used for normalization are determined by offline statistical analysis of the data distribution of each process node during its historical normal operation cycle and are updated periodically.

[0056] In step S12, the specific steps are as follows:

[0057] S121: The key data combinations are mapped to the preset digital twin space in real time, and dynamic simulation is performed according to the physical engine and fluid dynamics rules. During the simulation, a state feature vector representing the physical state of the current process is generated. At the same time, the state reference vector is obtained, and the state feature vector is compared with the state reference vector. In the comparison process, multidimensional Euclidean distance is introduced to construct the corresponding deviation measurement matrix.

[0058] S122: If any dimension index in the deviation measurement matrix exceeds the preset threshold range, the abnormal pattern matching logic is triggered. The specific abnormal working condition type is identified by reverse tracing and deduction through Bayesian network combined with historical fault database, and the posterior probability of the abnormal working condition type is calculated as the confidence level.

[0059] In step S121, the specific steps are as follows:

[0060] At this point, a virtual mapping space corresponding to all elements of the physical site of the unmanned laboratory is constructed. This space is built based on a high-fidelity three-dimensional geometric model and a multiphysics solver, and can receive and parse external input data in real time. The key data combination generated in step S113 is injected into the digital twin space in real time through a high-speed data bus as boundary conditions and initial states to drive the operation of the virtual model. The specific mapping process includes mapping the sample attribute labels and position coordinates to the corresponding geometric entities in the virtual scene, driving the virtual sample model to move in three-dimensional space.

[0061] By mapping environmental background data and equipment operating parameters to the control interface of the virtual equipment model, the virtual sensor readings and the physical properties of the virtual actuators are updated. Temporal feature vectors are mapped to the field variables of the virtual environment, such as loading scalar field data such as temperature and pressure onto the fluid grid nodes of the virtual reactor. Through the above mapping, the virtual state in the digital twin space is synchronized with the real-time state of the physical site, forming a dynamic mirror that can reflect the overall picture of the current physical process.

[0062] Within the digital twin space, simulations are performed on the mapped real-time data based on a pre-set physics engine and fluid dynamics calculation rules to analyze deep-seated physical state changes. Specifically, for process nodes involving liquid mixing, chemical reactions, and heat transfer, the fluid dynamics solver is activated to calculate the velocity distribution, pressure gradient, and turbulence intensity within the flow field. For nodes involving heat transfer processes, the spatiotemporal evolution of the temperature field is calculated using the heat conduction equation and the convective heat transfer coefficient. For chemical reactions that may exist in type B water, the reaction kinetic equation is used to simulate the change of substance concentration over time.

[0063] During the simulation, the system extracts numerical values ​​that characterize the system's microscopic physical state from the monitoring points of the virtual sensors in real time, including flow field shear force, local vorticity characteristics, reaction rate constant, and component concentration gradient. These discrete numerical values ​​are then serialized and vectorized according to preset dimensions to generate a state feature vector that characterizes the physical evolution mechanism within the current process. This state feature vector is not merely a simple reproduction of the input data, but a deep state description that reflects the essential operating characteristics of the system and is derived from the physical model.

[0064] Optionally, the state feature vector refers to the standardized numerical sequence extracted from the virtual model after dynamic simulation in the digital twin space, used to characterize the deep physical state of the current process node. Specifically, the state feature vector is a floating-point vector with a fixed dimension of M, and the values ​​of each dimension are derived from the monitoring points of virtual sensors inside the digital twin space. Different process nodes correspond to different feature dimensions.

[0065] Taking the digestion reaction node in the analysis and testing process as an example, the state feature vector sequentially includes the following M dimensions: the current temperature value in the reaction chamber, the first derivative of temperature with respect to time (i.e., the heating rate), the pressure value, the pressure fluctuation amplitude, the estimated value of the reactant concentration, the solution pH value, the stirring speed feedback value, and the dimensionless number of local turbulence intensity calculated by the fluid dynamics solver within the reaction system. All the above dimensions are generated by the digital twin space based on the physical model, rather than being directly read from sensors in the physical field, and their data refresh frequency is consistent with the simulation step size of the digital twin space.

[0066] The system retrieves a state baseline vector from a pre-defined standard process database that matches the current process node and the current task type, such as Type B water quality testing. This state baseline vector is generated based on historical best operating data or theoretical ideal models, representing the ideal physical state characteristics that the process node should have under standard operating conditions and without faults. After obtaining the current state feature vector and the state baseline vector, a deviation measurement model in a multi-dimensional feature space is constructed. Specifically, the two vectors are placed in the same multi-dimensional Euclidean space, the difference between them in each feature dimension is calculated, and the spatial distance between the vectors is calculated based on the multi-dimensional Euclidean distance formula. This distance quantifies the degree to which the current actual simulated state deviates from the ideal standard state.

[0067] Optionally, the state baseline vector refers to a sequence of reference values ​​that has the same dimensional definition and dimensional order as the aforementioned state feature vector. This baseline vector is pre-stored in the system database and used as a comparison benchmark to determine whether the current process is operating normally. The state baseline vector is obtained as follows: For each type of process node and its corresponding detection task type, the system selects no less than 30 continuously stable process batches with qualified final detection results from the equipment acceptance and commissioning phase or historical normal batches. It extracts state feature vector samples of the same process node in the same process stage from these batches, calculates the arithmetic mean of each dimension, and arranges the average values ​​of each dimension in their original order to form the standard state baseline vector under this condition. At the same time, the system synchronously records the standard deviation of each dimension of this vector for subsequent threshold setting. For the Type B water quality detection task listed in the current specification, the state baseline vector of its digestion reaction node is generated statistically from the normal digestion operation data of at least 30 batches of Type B standard samples in the above manner, and is periodically recalibrated as the process is optimized or the equipment is replaced.

[0068] Considering that different feature dimensions have different weights in influencing the overall state of the process, the system introduces a weighting coefficient when calculating the Euclidean distance, assigning higher weights to key "sensitive features, such as reaction temperature and core pressure". The calculated differences of each dimension and the weighted distance values ​​are used to construct a deviation measurement matrix. The element values ​​of this deviation measurement matrix intuitively reflect the degree of operational deviation of the current process at different physical levels, providing a quantitative basis for subsequent anomaly identification.

[0069] Optionally, the deviation metric matrix refers to a two-dimensional data table used to quantify the degree of deviation of the current process's actual operating state from the standard baseline state in various physical dimensions. The row and column dimensions of this matrix are defined as follows: each row of the matrix corresponds to an independent physical feature dimension, such as temperature deviation, heating rate deviation, pressure deviation, pH value deviation, etc.; the matrix contains two columns, the first column is the absolute deviation value of this dimension, and the second column is the relative deviation rate of this dimension.

[0070] The specific construction process is as follows: Let the state feature vector corresponding to the current process node be V_actual, with dimension M, and its dimensions arranged in a fixed order; let the state baseline vector obtained from the database be V_base, also with dimension M, and the order of its dimensions strictly consistent with V_actual. The system first calculates the difference vector V_diff = V_actual - V_base to obtain the absolute difference of each dimension. Based on the physical dimensions and numerical range of each dimension, the system performs dimensionless processing to form the entries for each row of the deviation metric matrix.

[0071] For temperature-related features, the absolute deviation is the difference between the actual temperature and the reference temperature, in degrees Celsius. The relative deviation rate is the absolute deviation divided by the sum of the reference temperature and 273.15, then multiplied by 100%. For pressure-related features, the absolute deviation is the difference between the actual pressure and the reference pressure, in megapascals (MPA). The relative deviation rate is the absolute deviation divided by the reference pressure, then multiplied by 100%. For rate-related features (such as heating rate), the absolute deviation is the difference between the actual rate and the reference rate, in degrees Celsius per minute. The relative deviation rate is the absolute deviation divided by the reference rate, then multiplied by 100%. After dimensionless processing of all feature dimensions, the absolute deviation values ​​and relative deviation rates of each dimension are arranged row by row according to the original arrangement of the feature dimensions, forming an M-row, two-column two-dimensional matrix, which is the deviation measurement matrix.

[0072] Based on this, in order to facilitate comprehensive judgment in the subsequent anomaly identification process, the system also presets a weight coefficient for each dimension according to the degree of influence of each feature dimension on the safety of the process and the accuracy of the detection results. The value of the weight coefficient ranges from 0 to 1, and the sum of the weights of all dimensions is equal to 1.

[0073] Taking the digestion reaction node as an example, the weighting coefficient for the temperature dimension is set to 0.25, the weighting coefficient for the pressure dimension is set to 0.20, the weighting coefficient for the heating rate dimension is set to 0.15, the weighting coefficient for the pH value dimension is set to 0.15, the weighting coefficient for the concentration estimate is set to 0.15, and the weighting coefficient for the turbulence intensity dimension is set to 0.10.

[0074] The system multiplies each absolute deviation value in the first column of the deviation metric matrix by its corresponding weight coefficient, calculates the weighted sum of squares, and takes the square root of this sum to obtain a comprehensive scalar value, called the comprehensive deviation distance. This comprehensive deviation distance, along with the deviation metric matrix, is output and used in subsequent steps to compare with a preset threshold range to trigger anomaly pattern matching logic. The comprehensive deviation distance is used for rapid determination of the overall degree of deviation, while the complete deviation metric matrix is ​​used for fine-grained reverse tracing in the subsequent Bayesian network. The two complement each other, jointly constituting the quantitative input basis for anomaly identification.

[0075] In step S122, the specific steps are as follows:

[0076] At this point, a multi-dimensional threshold monitoring model for the deviation measurement matrix is ​​constructed. This model stores the safety fluctuation range of each dimension of the state feature vector of each process node under different operating modes. The system receives the deviation measurement matrix generated in step S121 in real time and extracts the deviation values ​​in the matrix dimension by dimension, and compares them with the preset threshold range. The preset threshold range is determined based on the "statistical distribution, such as the 3σ principle" or the safety limit of the physical mechanism of historical normal operation data.

[0077] If the values ​​of all dimensions in the deviation measurement matrix are within the corresponding preset threshold range, the current process is determined to be in a steady state, and the system maintains the normal monitoring mode. If the value of any dimension index in the deviation measurement matrix exceeds its corresponding preset threshold range, it indicates that the current physical state has deviated significantly from the ideal state. At this time, the system immediately triggers the abnormal pattern matching logic, activates the fault diagnosis module, and enters the abnormal condition identification and tracing process to ensure an immediate response to potential risks.

[0078] A Bayesian network inference engine based on a probabilistic graphical model was introduced. This Bayesian network inference engine constructs its topology based on prior knowledge in a historical fault database. The root node of the Bayesian network represents various potential abnormal operating conditions, such as pipeline blockage, sensor drift, reagent deterioration, and temperature control failure. The leaf nodes represent specific feature dimensions in the deviation metric matrix generated by digital twin inference, such as pressure deviation, temperature rise rate deviation, and flow rate deviation. The directed edge nodes in the network store the conditional probability parameters of the root node (cause) leading to the occurrence of the leaf node (effect).

[0079] During the reverse tracing process, the system uses the deviation metric matrix calculated in step S121 that exceeds the threshold as the observation evidence of the leaf nodes of the Bayesian network to input into the inference engine. Using Bayesian inference algorithms, such as variable elimination or belief propagation algorithms, the network combines the prior probability distribution obtained from the historical fault database to calculate the conditional probability of "each root node: abnormal operating condition type" under the current observed deviation evidence. This process can not only identify the abnormal type most likely to cause the current deviation, but also eliminate false alarms caused by environmental noise or random interference based on the feature combination of deviation patterns, thereby achieving accurate reverse tracing from "phenomenon: deviation" to "essence: cause of failure".

[0080] After completing the backward inference of the Bayesian network, the system sorts and filters the conditional probabilities of each abnormal working condition type output by the network. The specific identification logic is as follows: select the type with the highest posterior probability among all candidate abnormal working condition types as the current identification result. This result represents the most likely cause of failure under known observation evidence. At the same time, the maximum posterior probability is directly defined as the confidence level of the abnormal working condition identification result.

[0081] The confidence level directly reflects the system's certainty in identifying the anomaly. If the confidence level is lower than the preset threshold, such as below 85%, it indicates insufficient or ambiguous data evidence, and the system will mark it as "suspected anomaly" and request auxiliary information. If the confidence level is higher than the preset threshold, the system will officially lock the abnormal condition type and package and output the type and its corresponding confidence level to form a structured anomaly diagnosis report, providing solid fault attribute basis for subsequent decision generation.

[0082] In step S13, the specific steps are as follows:

[0083] S131: Input each abnormal operating condition type and its confidence level into the inference engine of the unmanned laboratory, and simultaneously obtain the resource constraints of the unmanned laboratory. These resource constraints include the current instrument load rate, reagent balance, and sample queue length. Based on a multi-objective optimization strategy, the inference engine searches for the optimal response strategy in the digital twin space, thereby dynamically generating autonomous monitoring decision instructions. These autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies.

[0084] S132: Process path replanning is used to generate alternative flow paths when the current path is blocked; backup instrument switching logic is used to start the redundant equipment takeover task when the confidence level is higher than the first threshold; parameter adjustment strategy is used to compensate for the temperature control time, sample volume or digestion power when the confidence level is lower than the second threshold.

[0085] In step S131, the specific steps are as follows:

[0086] At this point, the abnormal operating condition type and its corresponding confidence level output in step S122 are used as the core trigger signal and input to the central inference engine of the unmanned laboratory in real time. At the same time, the inference engine synchronously queries and obtains the current real-time resource constraints of the unmanned laboratory through the middleware interface to construct the boundary environment for decision-making. The specific acquisition process includes: accessing the equipment management system, collecting the instrument load rate of the analysis instruments at each process node, and calculating the proportion of time that each device is in a busy or idle state.

[0087] Access the reagent storage management system to read the current inventory and consumption rate of available chemical reagents, paying particular attention to the remaining amount of special masking agents or strong acids required for type B water quality; access the task scheduling system to obtain the queue length of samples to be tested, analyze the current backlog of samples and their urgency. Through the above operations, the inference engine constructs a comprehensive state context that includes "fault severity - equipment availability - material sufficiency - task urgency", providing realistic constraint boundaries for subsequent strategy search.

[0088] In the digital twin space, the inference engine initiates a strategy search process based on a multi-objective optimization algorithm, aiming to find the optimal response strategy that can eliminate anomalies and maximize the overall system benefit under the current resource constraints; a weight function for the optimization objective is set, which takes into account multiple dimensions of indicators, including the success rate of anomaly handling, the overall efficiency of the detection task, the consumption cost of reagents and energy, and the safety of the equipment.

[0089] The inference engine constructs a virtual simulation sandbox in the digital twin space. For the identified abnormal operating conditions, it enumerates multiple feasible alternative response strategies. It uses a multi-objective optimization algorithm, such as NSGA-II or strength Pareto evolution algorithm, to iteratively evaluate and screen the alternative strategies. During the evaluation process, the algorithm inputs each strategy into the digital twin for pre-simulation, calculates the weighted contribution of the strategy to the above optimization objective function under the premise of satisfying resource constraints, and locks in an optimal response strategy from the solution space by comparing the Pareto front distribution of different strategies. This strategy can eliminate the current anomaly with the minimum resource cost and the shortest time delay within the confidence range, ensuring the maximization of laboratory operating efficiency.

[0090] Based on the optimal response strategy obtained from the search, the inference engine parses and transforms it into structured autonomous monitoring and decision-making instructions that can be recognized by the underlying execution mechanism. This instruction set contains three layers of control logic:

[0091] Firstly, process path replanning: In the event of interruption of the original process due to equipment failure, new sample transfer path instructions are generated to guide AGV or robotic arm to transfer the sample to the alternative node.

[0092] Secondly, the backup instrument switching logic generates an instrument switching command when the main device is confirmed to be unavailable, activates the backup instrument in cold or hot backup state, and issues the corresponding initialization parameters.

[0093] Third, the parameter adjustment strategy generates control parameter correction instructions for minor process deviations, dynamically adjusting the current operating parameters of the equipment. For example, the stirring speed is adjusted to cope with the high viscosity of type B water, or the temperature control curve is adjusted to adapt to different heat capacity characteristics. These autonomous monitoring and decision instructions are encapsulated into standardized control messages and prepared to be sent to the execution control layer to achieve automated handling of abnormal operating conditions.

[0094] Optionally, the objective function of the multi-objective optimization strategy is a comprehensive utility function consisting of a weighted sum of three sub-objectives: timeliness of anomaly handling, economy of task completion, and safety of equipment operation. The timeliness sub-objective is defined as the estimated total time required from strategy initiation to recovery from the abnormal state to the steady-state region. A shorter time results in a higher score for this sub-objective. This estimated total time is obtained by performing a virtual rehearsal of the selected strategy in the digital twin space.

[0095] The economic sub-objective is defined as the sum of the cost of chemical reagents and the cost of equipment energy consumption during the execution of the strategy. The lower the cost, the higher the score of the sub-objective. The reagent consumption is calculated by the digital twin space based on the parameter adjustment in the strategy and the reaction kinetic model to estimate the volume and type of additional reagents required. The energy cost is obtained by multiplying the product of the rated power and the estimated running time of each actuator by the local industrial electricity price.

[0096] The safety sub-objective is defined as the ratio of the maximum instantaneous load of each actuator to its rated load limit during the execution of the strategy. The lower the ratio, the higher the score of the sub-objective. This ratio is obtained by extracting the peak torque, peak current and peak power parameters from the virtual dynamic model of each actuator in the digital twin space when performing virtual pre-simulation of the strategy, and then dividing them by the corresponding rated upper limit value.

[0097] The three sub-objectives each have different original dimensions and numerical ranges. To enable weighted summation, the system normalizes the original values ​​of each sub-objective. The maximum possible value for the timeliness sub-objective is set to three times the standard cycle time of the current process; the maximum possible value for the economic sub-objective is set to 1.5 times the current batch sample testing budget; and the maximum possible value for the safety sub-objective is set to 1.2 times the rated load of each mechanism (allowing for short-term overload limits). After normalization, all three sub-objectives are mapped to the range of 0 to 1, and all are minimized, meaning that smaller values ​​indicate better performance of the objective.

[0098] The comprehensive utility function is the weighted sum of the normalized values ​​of the three factors. The specific expression is: Comprehensive utility value = α × normalized timeliness value + β × normalized economy value + γ × normalized safety value, where α, β, and γ are the weight coefficients of the three sub-objectives. The values ​​of the three factors are all between 0 and 1, and the sum of the three factors is always equal to 1.

[0099] The default values ​​for the weighting coefficients are α=0.35, β=0.25, and γ=0.40. These default values ​​reflect the principle of prioritizing safety while considering efficiency and economy in unmanned laboratories. The weighting coefficients can be manually adjusted by the operator through the host computer interface before operation according to the urgency of the task or the importance of the sample. For urgent sample testing tasks, α can be increased to 0.50 and β can be decreased to 0.15. For high-value or precious sample testing tasks, γ can be increased to 0.55 and α can be decreased to 0.25.

[0100] In step S132, the specific steps are as follows:

[0101] At this point, a dynamic topology map of the internal logistics network of the unmanned laboratory is constructed. This map maps the passage status of each transmission channel, AGV (Automated Guided Vehicle) running path, and robotic arm working space in real time. When the inference engine determines that the current process path is blocked, such as detecting physical obstacles in the road ahead, physical failure of equipment nodes causing inability to receive samples, or congestion in the transmission network, the process path replanning mechanism is triggered.

[0102] The process path replanning mechanism locks the starting and target nodes of the obstructed path in the digital twin space and searches for alternative connected paths in the current dynamic topology map based on the shortest path algorithm in graph theory. During the search, the system comprehensively considers the physical distance, travel time cost, and real-time load of each node on the path to select an alternative flow path with the lowest overall cost. The system generates a path switching command and sends the coordinate sequence and control timing of the alternative path to the logistics scheduling system to drive the AGV or conveyor belt to change its running trajectory, thereby bypassing the obstructed area and ensuring that the sample can continue to flow to the next process node, avoiding the entire process from being stopped due to a single point of failure.

[0103] A first threshold, such as 90%, is set for decision-making on hard fault handling. This threshold defines the system's criteria for judging equipment failure or unrecoverable faults. When the confidence level of the abnormal operating condition input by the inference engine is higher than this first threshold, it means that the "abnormality" that has occurred in the current equipment, such as hardware damage or core sensor failure, cannot be repaired by parameter adjustment. The system then initiates the backup instrument switching logic. The specific execution process is as follows: the system retrieves redundant equipment or backup instruments that are functionally isomorphic to the faulty equipment and are in a ready state from the resource management list.

[0104] The system generates a device switching handshake command to interrupt control of the faulty device and seamlessly migrates the "task context information of the sample to be tested, including processed steps, current status parameters, and sample ID" to the backup instrument. It then issues a start command to the backup instrument, enabling it to take over subsequent testing tasks. This logic ensures that, under the premise of confirming that the fault is extremely severe, the continuity of the experimental process and the integrity of the data are guaranteed to the greatest extent by quickly isolating the fault source and activating redundant resources.

[0105] A second threshold, such as 80%, is set for decision-making on handling soft faults. This threshold is usually lower than the first threshold and is used to define the boundary where the equipment status deviates from the standard but still has the conditions to operate. When the confidence level of the abnormal operating condition input by the inference engine is between the second threshold and the first threshold, or is lower than the first threshold but higher than the minimum alarm line, it means that the current process is in a non-ideal operating state, such as slight blockage, environmental disturbance, or slight fluctuation of reagents. However, the detection quality can still be maintained on the original equipment by correcting the operating parameters.

[0106] At this point, the system triggers a parameter adjustment strategy. The specific implementation process is as follows: based on the type of abnormal operating condition, the system determines the key control variables that need adjustment, such as temperature control duration, sample loading volume, or digestion power. Based on a digital twin space-based simulation model, it calculates the compensatory correction amount required to eliminate the current state deviation. For example, for uneven mixing caused by the high viscosity of type B water, it calculates the required increase in stirring time or digestion power, generates a parameter reconfiguration command, and sends it to the controller of the current equipment, dynamically modifying its operating setpoint. This strategy aims to eliminate potential abnormal effects and ensure the validity of experimental results through refined parameter compensation, without interrupting the process or switching equipment.

[0107] In step S14, the specific steps are as follows:

[0108] S141: The autonomous monitoring decision command is parsed into specific control timing signals through the edge computing gateway. The control timing signals are then sent to the execution control layer corresponding to each process node to drive the multi-degree-of-freedom robotic arm, injection pump and temperature control module to perform coordinated actions, thereby completing the abnormal handling.

[0109] S142: During the abnormal handling process, the current feedback, position deviation and action completion signal of the actuator are collected in real time. The current feedback, position deviation and action completion signal of the actuator are used as status feedback data. The status feedback data is transmitted back to the digital twin space, and virtual and real synchronous iteration is achieved during the parameter correction process.

[0110] S143: Continuously monitor the operating parameters of each process node. When the fluctuation variance of the key indicators is found to converge within the preset steady-state range within a preset number of consecutive sampling periods, it is determined that the abnormal handling is completed and the normal operation process is restored, thus forming an autonomous monitoring event in the unmanned laboratory site.

[0111] In step S141, the specific steps are as follows:

[0112] At this point, an edge computing gateway is constructed and deployed at the field end. This gateway serves as a protocol conversion hub between the upper-layer decision-making logic and the lower-layer execution mechanism. The edge computing gateway receives structured autonomous monitoring and decision-making instructions from the inference engine. These instructions contain high-level logical descriptions, such as "switch to backup instrument" and "adjust resolution parameters". The gateway is loaded with device driver models and communication protocol dictionaries for each process node, and is responsible for parsing the high-level logical descriptions into control timing signals that can be recognized by the lower-level hardware.

[0113] The specific analysis process includes: extracting the action type, target device ID, and target parameter value from the decision command; secondly, based on the physical characteristics of the target device, such as the kinematic model of the robotic arm, the flow response curve of the injection pump, and the thermal inertia of the temperature control module, calculating the precise time sequence required to achieve the action, determining the start time, duration, and action interval of each action stage, encapsulating the calculation results into a standard control message conforming to fieldbus protocols (such as Modbus, EtherCAT, PROFINET), and generating a control timing signal containing timestamps and address information to ensure that the command is executed precisely with millisecond-level time accuracy.

[0114] The parsed control timing signals are sent to the execution control layer corresponding to each process node in real time via industrial Ethernet or fieldbus network. This execution control layer is usually composed of programmable logic controllers (PLCs) or embedded motion controllers of each device. During the sending process, the edge computing gateway adopts a multi-channel concurrent transmission mechanism to ensure that "multiple nodes related to abnormal handling, such as robotic arm controllers, pump drivers, and temperature control instruments" can receive instructions synchronously.

[0115] To eliminate the asynchronous action problem caused by network transmission delay, the execution control layer does not immediately trigger the action after receiving the control timing signal. Instead, it performs local clock alignment based on the timestamp information in the signal to lock a unified action trigger time. The execution control layer performs integrity verification and legality checks on the signal. After confirming that there are no errors, it parses the control timing signal into a pulse sequence for driving the motor, an analog voltage signal, or a switch contact signal, and loads it into the corresponding control register or buffer to prepare for driving the physical device to take action.

[0116] Upon reaching the preset trigger time, the actuators at each process node begin to coordinate and work together based on control signals, completing the abnormal handling through the cooperation of physical actions. The specific linkage process involves the precise coordination of a multi-degree-of-freedom robotic arm, an injection pump, and a temperature control module: after receiving the motion command, the joint servo motors of the multi-degree-of-freedom robotic arm respond to the control signal, adjust the end posture according to the planned trajectory, and perform grasping, moving, or releasing actions to position the sample or reagent container to the target location; after receiving the fluid control command, the stepper motor drives the precision screw to rotate, controlling the precise displacement of the plunger, thereby extracting or discharging reagents at the set flow rate and volume, achieving the precise addition of chemical components.

[0117] After receiving the thermal control command, the solid-state relay or power regulator adjusts the output power of the heating element or controls the speed of the cooling fan according to the pulse width modulation (PWM) signal, so that the temperature in the reaction chamber rises and falls rapidly according to the new curve. The above-mentioned mechanisms work closely together under a unified time reference. For example, the robotic arm grabs and transfers the sample the moment the injection pump completes the injection action, or the temperature control module starts heating immediately after the sample is in place. Through this spatiotemporal coordination, automated physical intervention and handling of abnormal conditions in the B-type water quality testing process are realized.

[0118] In step S142, the specific steps are as follows:

[0119] At this time, during the execution of the abnormal handling command, the system activates a high-frequency status monitoring mechanism, using the sensing modules deployed inside each actuator to collect status feedback data in real time that characterizes the details of physical movements. The specific collection process includes: for multi-degree-of-freedom robotic arms, reading the real-time current feedback of the joint motors through the servo driver to reflect the actual load situation of the robotic arm overcoming gravity and friction during movement; reading the actual position coordinates of the end effector through the encoder and performing differential calculation between it and the command-planned position to obtain position deviation data in order to evaluate the following accuracy of the motion trajectory.

[0120] For the injection pump and fluid control unit, the phase signal or opening feedback of the stepper motor is collected to confirm the completion of the action signal, such as the digital level change of "injection completed" or "valve closed"; for the temperature control module, the current fluctuation of the heating element and the real-time jump data of the temperature sensor are collected; the current feedback, position deviation and action completion signal collected above constitute a multi-dimensional state feedback data set. This data set is cached in the register of the local controller at a microsecond sampling frequency to provide raw data support for subsequent virtual-real mapping.

[0121] A high-bandwidth, low-latency data feedback channel is obtained, and industrial fieldbus (such as EtherCAT) or time-sensitive network (TSN) is used to transmit cached state feedback data from the lower execution control layer to the upper digital twin space in real time. During the transmission process, the system accurately timestamps the data packets to ensure that the feedback data is strictly synchronized with the simulation and deduction clock in the digital twin space. After receiving the data, the digital twin space loads it as an external input variable into the corresponding virtual entity model.

[0122] Specifically, the current feedback data of the robotic arm is input into the virtual dynamics model to update the load parameters of the virtual joints in real time; the position deviation data is mapped to the virtual kinematics model to correct the display coordinates of the virtual end effector in the three-dimensional scene; and the action completion signal is used as a state trigger to drive the work order process state in the virtual model to switch. Through the above feedback mechanism, the digital twin space is no longer just a one-way deduction based on the preset model, but becomes a dynamic mirror that can perceive the execution details of the physical scene in real time, ensuring the consistency between the virtual state and the physical state.

[0123] Based on the returned state feedback data, a closed-loop control logic for parameter correction is initiated within the digital twin space to achieve synchronous iteration between the virtual and real worlds. The solver in the digital twin space calculates the residual between the current feedback state and the desired target state in real time. For example, when the current feedback value of the robotic arm is detected to be significantly higher than the predicted value of the theoretical model, and the position deviation shows a continuous increasing trend, the system determines that there may be unmodeled interference forces in the physical field, such as pipe adhesion caused by type B water quality.

[0124] At this point, the digital twin space uses Kalman filtering or gradient descent algorithms to perform online correction and calibration of the physical parameters inside the virtual model, such as friction coefficient, load mass, and fluid viscosity coefficient. The corrected parameter model will generate a new predictive control law and send it back to the execution control layer of the physical site to adjust the output torque or speed of the actuator. Through the iterative process of "physical execution feedback - virtual parameter correction - new instruction issuance", the system continuously reduces the error between the model prediction and the actual response, so that the digital twin can gradually approach the real physical characteristics when dealing with special working conditions such as type B water quality, and achieve precise control and adaptive optimization of the abnormal handling process.

[0125] In step S143, the specific steps are as follows:

[0126] At this time, after the abnormal handling response is initiated, the system maintains high-frequency continuous monitoring of the operating parameters of each affected process node, and does not terminate data acquisition due to the execution of the handling command; the system selects key indicator parameters that can characterize whether the current process has returned to stability from the real-time aggregated multi-source sensing data stream. These parameters are usually the physical quantities most sensitive to disturbances, such as the temperature and pressure change rate in the reactor, the motion smoothness parameters of the robotic arm, the pressure fluctuation value of the fluid pipeline, and the spectral absorption rate of the chemical reaction.

[0127] To quantitatively assess the stability of the system state, the system constructs a sliding time window and extracts statistical features from the collected key indicator sequences within this window, focusing on calculating the variance of the fluctuation of the indicator. This variance reflects the degree of dispersion of the physical quantity in the time dimension. The smaller the variance value, the weaker the fluctuation of the physical quantity, and the more the system is approaching a stable state; conversely, it indicates that the system is still in a transitional process of oscillation or adjustment. The system uses the calculated variance as the core metric for judging whether the abnormal handling is effective.

[0128] A steady-state determination model based on time accumulation logic is constructed to determine whether process nodes have truly returned to the preset normal operating track. The specific determination logic is as follows: the system pre-sets a steady-state interval corresponding to the key indicators. This interval is determined based on the tolerance range of the equipment during normal operation and the accuracy requirements of the Type B water quality testing process. For example, the temperature fluctuation range is controlled within ±0.5℃. At the same time, a threshold for the number of continuous sampling cycles used to confirm stability is set, for example, N=10 cycles, covering a time span of about 30 seconds to 1 minute.

[0129] The system compares the variance of key indicators calculated in each sampling period with the preset steady-state range. Only when the monitored variance value falls within the steady-state range for a preset number of sampling periods without any instantaneous jumps outside the range, does the system determine that the dynamic behavior of the current process has converged to a steady state. This mechanism prevents misjudgments caused by accidental instantaneous disturbances or hysteresis effects, and ensures that the anomaly is confirmed to have been effectively suppressed or eliminated only when the system shows a continuous and stable recovery trend.

[0130] After completing the steady-state convergence determination, the system officially confirms the end of the anomaly handling process and triggers the state machine to switch from "anomaly handling mode" back to "normal operation mode". The system sends a process recovery instruction to the dispatch center to release the path blockage or resource lock previously triggered by the anomaly, restore the normal flow of the sample queue to be tested, and continue to execute the subsequent testing steps of type B water quality according to the established process. At the same time, the system automatically records the entire process data from anomaly occurrence, identification, decision-making, execution to recovery confirmation, including timestamps, anomaly type, handling strategy and feedback parameters, and encapsulates it into a complete "autonomous monitoring event" log and stores it in the database. This log not only serves as the verification basis for the success of this type B water quality test, but also serves as new prior knowledge to update the historical fault database and knowledge graph, which is used to optimize the parameters of the digital twin model and improve the reasoning ability to deal with similar complex working conditions in the future, thereby realizing the self-evolution and closed loop of the autonomous monitoring capability of the unmanned laboratory.

[0131] Please see Figure 2 The specific application scenarios and corresponding embodiments of this application are as follows:

[0132] In step S11, the unmanned water quality analysis laboratory is a standardized automated testing system that has been built. Its routine testing targets are "Type A water quality, such as conventional surface water". The current research and development task requires to use the unmanned water quality analysis laboratory on-site to verify the autonomous monitoring method for "Type B water quality, such as industrial wastewater with high salinity and high organic matter concentration".

[0133] In the physical space of the unmanned water quality analysis laboratory, the researchers confirmed the locations of the "sample scanning position" and the "Type B water quality-specific reagent rack" in the sample flow path and marked them on the digital map as key flow nodes for this test. Considering the strong corrosiveness and precipitation characteristics of Type B water quality, the "high-temperature digestion station" and "centrifugation station" in the pretreatment area were marked as high-risk monitoring nodes, the "heavy metal analyzer" and "organic matter analyzer" in the analysis and testing area were marked as core data acquisition nodes, and the "acidic waste liquid tank" in the waste liquid disposal area was marked as a key focus node.

[0134] Based on the existing sensing facilities of the unmanned water quality analysis laboratory, targeted deployments and enhancements were made for the characteristics of Type B water quality. During the sample transfer process, a high-definition industrial camera deployed on the top of the AGV was used to collect image streams of "Type B water quality sample bottles: usually wide-mouth bottles of special colors", focusing on capturing the visual characteristics of whether there are residual crystals or turbidity on the bottle body. An RFID reader was used to scan the sample label to identify the identity information of "Type B-Wastewater-01".

[0135] In the pretreatment stage, due to the violent reaction of type B water, a high-temperature resistant vibration sensor is additionally deployed on the outer wall of the digester to collect vibration signals of a specific frequency caused by the violent reaction during the digestion process; at the end of the reagent addition arm, a high-precision weight sensor is deployed to collect the trace weight change signals of the special strong acid reagent required for type B water.

[0136] During the analysis and testing phase, the instrument status codes of the heavy metal analyzer are collected in real time through the Modbus protocol gateway. The status codes such as "light chamber temperature" and "argon pressure" are monitored in particular, because the high salinity of type B water can easily cause nebulizer blockage, and the status codes will reflect this potential risk in real time.

[0137] After the test started, the sensing unit began working in real time; the vision unit continuously acquired posture image streams of the Type B water sample during the robotic arm's grasping process, ensuring that the sample bottle did not slip; the weight sensor transmitted the weight difference signal of the sample before and after the addition of the digestion reagent in real time, accurate to 0.1 mg; the laser positioning system continuously output the three-dimensional position coordinates of the robotic arm during the movement of the Type B water sample, ensuring that it accurately avoids obstacles and is precisely placed in the digestion hole; the environmental sensor collected the temperature and humidity in the laboratory, because some chemical reactions of Type B water are sensitive to environmental temperature; the protocol gateway high-frequency parsed the status codes sent by the analyzer, and once it identified a specific hexadecimal code representing "abnormal negative pressure" or "ignition failure", it was input as key abnormal data into the subsequent processing flow. Through the above acquisition, the system obtained a complete set of multi-source heterogeneous sensing data for the characteristics of Type B water, laying the data foundation for subsequent spatiotemporal alignment and anomaly identification.

[0138] The vision system of the unmanned water quality analysis laboratory acquires an image stream at a rate of 30 frames per second, while the temperature sensor of the digester reports data once per second. When testing a type B water sample, the vision system captured the image feature of an untightened sample cap at timestamp T1 the instant the robotic arm placed the sample into the digester, while the temperature sensor recorded the temperature change at timestamp T1+50ms. To accurately correlate this abnormal state, the system uses a global clock source to align the two times. Through an interpolation algorithm, the system calculates that the estimated temperature inside the digester at time T1 is 25.3 degrees Celsius. This precisely binds the image event of the "untightened cap" to the temperature state "before placement" on the same time slice, ensuring that subsequent analysis can accurately determine the impact of this state on the subsequent heating process.

[0139] In the pretreatment stage of type B water quality, AGV transports the sample to the centrifuge station; the lidar on the AGV collects the local coordinates centered on the AGV, showing that the distance of the obstacle from the center of the AGV (x_local, y_local) is (0.5m, 0m); at the same time, the global monitoring camera installed on the laboratory ceiling detects waste liquid dripping in front of the centrifuge through image recognition.

[0140] To determine whether the AGV will crush the waste liquid, the system calls a pre-calibrated transformation matrix between the AGV and the global coordinate system to convert the AGV's local coordinates into the laboratory's global coordinates (X_global, Y_global). Simultaneously, the pixel coordinates of the waste liquid in the camera image are mapped to the global ground coordinates. After spatial mapping, the system confirms that the AGV's planned path coincides with the waste liquid's location in the global coordinate system, thereby triggering an obstacle avoidance decision. This process ensures the spatial consistency between mobile carrier data and fixed environment monitoring data.

[0141] During the acid digestion of type B water, the liquid level sensor may produce false high values ​​due to the violent reaction after the addition of strong acid and the instantaneous jump caused by bubbles. For example, the liquid level may jump from 10cm to 15cm and then fall back. At this time, the data cleaning mechanism is activated. The statistical module based on the 3σ principle detects that the jump value deviates from three standard deviations of "mean in the current time window: 10cm" and marks it as a potential outlier.

[0142] Meanwhile, the Kalman filter predicts the next liquid level to be 10.1 cm based on the physical trend of the rising liquid level. Because the residual between the "actual measured value: 15 cm" and the predicted value is too large, the system determines that the data is "environmental noise: bubble interference" and removes it. The smoothed data accurately reflects the steady rise of the liquid level. In addition, in response to the high-frequency jitter noise of the conductivity sensor caused by the high salinity of the type B water, the system uses a low-pass filter algorithm to smooth it out, eliminating electromagnetic interference and outputting a stable and reliable target data stream, ensuring the accuracy of subsequent digital twin inference.

[0143] The system retrieves general water quality analysis process maps from the knowledge base and identifies the current task as "Type B water quality detection". The map automatically loads specific branches related to Type B water quality: since Type B water quality is high-salt industrial wastewater, the map instantiates specific process logics of "salt interference suppression" and "high temperature and high pressure digestion". The system clarifies that at the "digestion" process node, not only should the conventional temperature be considered, but also the "pressure" parameter should be given special attention, because high-salt wastewater is prone to boiling over at high temperatures. The map also sets a safe pressure threshold range for Type B water quality, such as not exceeding 2.5 MPa. This rule guides the focus of subsequent feature extraction.

[0144] After the type B water sample enters the digestion station, the system extracts features from the target data stream based on the spectral guidance. For the signals collected by the pressure sensor, the system extracts the key temporal feature of "pressure change rate," because the type B water reacts violently, and the pressure change rate can predict the boiling trend in advance. For the monitoring image stream inside the digester, the system extracts the feature of "gradient change in solution color depth," because the degradation process of high organic matter concentration is accompanied by obvious color changes, from dark brown to light color, and this visual feature vector directly reflects the progress of the digestion reaction. At the same time, the system extracts the feature of "ripple coefficient" of heating power to determine whether the heating system is experiencing abnormal fluctuations due to increased load caused by high salt crystallization. These features together constitute a high-dimensional vector characterizing the current operating status of the digestion process.

[0145] The system integrates the extracted features with environmental background data, equipment operating parameters, and sample attribute labels. It incorporates the following data: "Environmental background data: Current laboratory humidity is 80%, this high humidity environment may make type B waste liquid more hygroscopic and deliquescent." Secondly, it integrates equipment operating parameters, noting that the digester's sealing ring has been used 900 times, approaching the wear threshold—a crucial parameter for high-pressure digestion. It also integrates sample attribute labels, clearly identifying the current sample as "Type B - High-salt - Industrial wastewater." Through feature fusion, a key data combination is constructed, encompassing "high-humidity environmental background + high-load equipment status + high-salt sample attributes + violent reaction characteristics." This combination is input into the digital twin space, enabling subsequent anomaly identification logic to accurately deduce based on the complex context of "this is a high-humidity environment, near-wear equipment processing high-risk type B waste liquid," rather than simply relying on a single pressure value.

[0146] In step S12, the sensing system of the unmanned water quality analysis laboratory collects key data combinations after the type B water sample enters the digester, including ambient temperature of 25℃, sample attribute "type B - high organic matter", digester set temperature of 180℃, and real-time pressure of 1.8MPa. These data are mapped to the digital twin space in real time. In the virtual scene, the three-dimensional model representing the digester receives the pressure parameter of 1.8MPa, the flow field mesh inside the virtual space is loaded with the fluid attribute parameters of "high organic matter", and the virtual sensor is synchronized to the current operating state. At this time, the virtual digester in the digital twin space is not only a reproduction of the geometry, but also has the same fluid attributes and boundary conditions as the physical site.

[0147] The system initiates a fluid dynamics and thermodynamics engine within the digital twin space to perform simulations. Due to the high organic matter concentration in type B water, the simulation calculates that under the current heating power, a severe local exothermic effect occurs within the reaction system, resulting in a local temperature field higher than the set value. Furthermore, the high-viscosity fluid forms a non-uniform shear force distribution during stirring. Based on the simulation results, the system extracts deep physical parameters from the core region of the virtual flow field, including "maximum local temperature rise, +5℃", "bottom velocity vector, 0.2m / s", "reaction exothermic rate, 150J / s", and "turbulent dissipation rate". These parameters are encoded into a state feature vector Vcurrent=[5.0,0.2,150,...], which accurately characterizes the complex physical reaction state of type B water during the digestion process, which cannot be directly measured by external sensors alone.

[0148] The system retrieves the state baseline vector Vbaseline for the "high-temperature and high-pressure digestion" process from the standard process library for type B water. Considering the special characteristics of type B water, the ideal reaction exothermic rate of this baseline vector is set to 100 J / s, and the ideal bottom flow velocity is 0.5 m / s. It is assumed that high viscosity causing flow obstruction is abnormal. The system performs multidimensional Euclidean distance calculation between Vcurrent and Vbaseline. During the calculation, since the "reaction exothermic rate" is directly related to safety, it is given a higher weight of 0.6, while the "flow velocity" weight is 0.4. The calculation results show that there is a significant distance between the two in Euclidean space. The deviation metric matrix shows that the deviation in the "reaction exothermic rate" dimension is +50, indicating that the reaction is too violent. The deviation in the "bottom flow velocity" dimension is -0.3, indicating that there are dead zones in the stirring. The final generated deviation metric matrix clearly points out that although the current digestion process meets the macroscopic temperature standard, the microscopic kinetic state has deviated from the standard process range, and there is a risk of boiling over.

[0149] The deviation measurement matrix output by the digital twin space shows that the deviation value in the dimension of "reaction exothermic rate" reached +60J / s, while the ideal value is 100J / s, and the preset safety threshold range for this dimension is ±20J / s. Although the deviations of "temperature" and "pressure" are still within the threshold range, the system judges that there is a potential risk of thermal runaway in the digestion process because the key indicator of "reaction exothermic rate" is seriously exceeded. It then triggers the abnormal mode matching logic and starts the fault diagnosis process.

[0150] The system activates a Bayesian network containing historical fault knowledge for the specific scenario of "processing type B water in an unmanned laboratory for water quality analysis." Root nodes in the network include: "reagent concentration too high," "stirring blade scaling leading to artificially high rotation speed," "cooling circulation blockage," and "abnormal sample properties, high organic matter content." The input observational evidence is "extremely high reaction exothermic rate" and "low bottom flow velocity vector." The Bayesian network, combined with prior data from the historical fault database, shows that for conventional type A water, "reagent concentration too high" has a higher prior probability of causing this phenomenon. However, for type B water with high salinity, the historical database shows a higher correlation between "stirring blade scaling" and "abnormal sample properties." Through reverse engineering, the network finds that "stirring blade scaling" can well explain "low flow velocity," as scaling increases resistance and "abnormal exothermic rate," resulting in uneven mixing and localized overheating.

[0151] The inference results show that, under the current observational evidence, the posterior probability of the abnormal condition type "scaling of the agitator leading to a decrease in mixing efficiency" is 92.4%, while the posterior probability of "abnormal sample properties" is 5.1%, and the posterior probability of "excessively high reagent concentration" is 2.5%. The system locks the type with the highest probability as the identification result, that is, determines the abnormal condition as "scaling of the agitator". At the same time, the 92.4% posterior probability is used as the confidence level of this identification. Since this confidence level is far higher than the minimum judgment standard of 85%, the system confirms that the diagnosis result is credible and outputs it as the final abnormal identification result, indicating that the high salinity of type B water causes rapid scaling of the agitator, which in turn affects the mixing effect.

[0152] In step S13, the inference engine receives an abnormal signal of "Digester 01 - Stirring Blade Scaling"; the engine immediately scans the laboratory's resource status:

[0153] Instrument load rate: The query found that "Digester 02" is currently processing a batch of conventional type A water samples with a load rate of 85%, but will be idle in 10 minutes; the standby "Digester 03" is completely idle with a load rate of 0%.

[0154] Reagent remaining: It was detected that only 150 mL of the special strong acid reagent required for the digestion of type B water quality remained, which is only enough to support the processing of 3 samples. It should be used sparingly.

[0155] Queue length: The queue of samples to be tested shows that there are still 5 type B water quality samples waiting to be processed, indicating a high degree of urgency.

[0156] The inference engine performs strategy deduction in the digital twin space for the anomaly of "stirring blade fouling"; it evaluates three strategies:

[0157] Strategy A: Shut down and repair "Digester 01". Scenario: The long repair time resulted in a severe backlog of 5 Type B water samples, violating the "task efficiency" objective.

[0158] Strategy B: Transfer the sample to "Digester 02". Expected result: A 10-minute wait is required, and "Digester 02" lacks special protection against high-salinity wastewater, posing a risk of cross-contamination and violating the "safety" objective.

[0159] Strategy C: Transfer the sample to the standby "Digester 03" and adjust the digestion parameters. Analytical results: "Digester 03" is a corrosion-resistant device suitable for type B water quality; although it consumes some energy to start, it can process the water immediately, and the remaining strong acid reagent is just sufficient. This strategy achieves the best balance between "efficiency," "safety," and "resource constraints," resulting in Pareto optimality.

[0160] Therefore, engine-locked strategy C is the optimal response strategy. Based on strategy C, the inference engine generates the following specific autonomous monitoring decision instructions:

[0161] Process path replanning: Generate new AGV path instructions, cancel the route to "Digester 01", and plan a new path to the workstation where "Digester 03" is located.

[0162] Backup instrument switching logic: Generate equipment activation command, remotely wake up "Digester 03", and issue preheating command to make its temperature quickly reach 180℃ required for type B water quality.

[0163] Parameter adjustment strategy: Considering the high viscosity of type B water, specific control parameters are generated for "Digester 03". The stirring speed is increased from the default 800 rpm to 1200 rpm, and the digestion and pressure holding time is extended by 5 minutes to ensure complete reaction.

[0164] During the transfer of the Type B water sample to the "optical testing station," the AGV navigation system detected that the planned route was blocked due to a reagent leak at the previous Type A water testing equipment. The ground maintenance robot was physically blocking and cleaning the area, causing the current path to be obstructed. At this point, the process path replanning mechanism was triggered. The system recalculated in the digital twin map and found an alternative path that bypassed the leak area to the back door of the optical testing station. Although this path was 5 meters longer, it was still passable. The system immediately issued a path replanning command, instructing the AGV carrying the Type B water sample to change its route, avoid the leak area, and successfully reach the testing station.

[0165] During the "heavy metal digestion" of another batch of type B water samples, the inference engine detected an abnormality of "sealing ring breakage" in "digester 01" with a calculated confidence level as high as 98%, exceeding the preset first threshold of 90%. This indicated that digester 01 had completely lost its high-pressure sealing capability, could not continue to work, and could not be repaired by software. The system immediately activated the backup instrument switching logic, calling "digester 04" which was in cold backup state from the equipment pool. The system synchronously transmitted the "acidification status" and "reaction time" information of the current type B water sample to digester 04, and controlled the robotic arm to transfer the sample to digester 04, which then took over the remaining high-temperature digestion task to ensure uninterrupted detection.

[0166] When performing "UV digestion" on type B water, the inference engine identified a slight anomaly of "incomplete organic matter digestion" with a confidence level of 75%, which is below the second threshold of 80% but above the minimum monitoring line. This indicates that the instrument hardware itself is normal, but the high organic matter concentration of type B water exceeds the preset digestion intensity for conventional type A water. The system determines that there is no need to switch instruments, but instead triggers a parameter adjustment strategy. After calculation, the system generates a compensatory correction instruction: increasing the current "UV digestion power" from the default 800W to 950W and extending the "digestion time" from 30 minutes to 38 minutes. This instruction is directly sent to the current UV digester, compensating for the difference in sample properties through enhanced physical force, so that type B water samples can achieve the same digestion effect as type A samples.

[0167] In step S14, the edge computing gateway receives the aforementioned high-level instruction; the gateway queries the device model library and learns that the solenoid valve ID corresponding to the "backup pipeline" is "Valve_B", the motor ID corresponding to "increase stirring speed" is "Motor_Mix", and the target speed is set to "1200rpm"; based on the "time: approximately 2 seconds" for the robotic arm to reach the backup pipeline and the motor acceleration characteristics, the gateway generates precise control timing signals, which take 1 second to accelerate from 0 to 1200rpm: the instruction is issued at time T0; "Valve_B" is triggered to open at time T+2; the syringe pump is triggered to discharge liquid at a rate of "5mL / s" at time T+2.5; "Motor_Mix" is triggered to accelerate linearly from time T+1 to T+2, and finally a high-precision control message conforming to the EtherCAT protocol is generated.

[0168] Control messages are synchronously sent via industrial Ethernet to the PLC controlling the robotic arm, the PLC controlling the fluid system, and the temperature control / stirring module controlling the digester. After receiving the message, each execution control layer reads its internal high-precision clock, locks T0 time as the reference, and calibrates the local time deviation. The PLC controlling the fluid system stores the received timing signal into the I / O buffer, preparing to drive the relay to activate at T+2 seconds. The motor driver of the digester loads the acceleration ramp parameters to the servo driver, preparing to execute the speed-up action.

[0169] The process begins with the multi-degree-of-freedom robotic arm moving, its joint motors rotating in tandem. The end effector grips the Type B water sample bottle, smoothly moving it to the bottom of the backup tubing in 2 seconds. As soon as the robotic arm is in position and stable, time T+2 arrives, and the stepper motor of the syringe pump rotates, driving the plunger to quickly inject the special digesting agent required for Type B water into the sample bottle through the backup tubing. Simultaneously, the stirring motor inside the digester responds to the acceleration command at time T+1, driving the stirring paddle to rotate at high speed, creating a strong vortex during reagent injection to prevent localized precipitation or crystallization of the high-salt wastewater upon injection. The temperature control module synchronously increases the heating power according to the command to cope with the heat capacity changes after the injection of Type B water. Through the close coordination of these actions, the system successfully bypasses the blocked main tubing, completing reagent addition and sample mixing, and the anomaly is successfully resolved.

[0170] As the stirring motor accelerates to 1200 rpm, the servo driver in the execution control layer monitors the rapid increase in motor current in real time. Due to the high viscosity of type B water, the current stabilizes at 4.5A, while the current for conventional type A water is only 2.8A at the same speed. At the same time, the encoder feedback shows that although the set speed is 1200 rpm, due to the heavy load, the actual speed fluctuates slightly between 1175 rpm and 1185 rpm, and the "position deviation: speed deviation" persists. When the stirring operation reaches the preset 5-minute time node, the timer generates a high-level action completion signal. These include the 4.5A current, the -25 rpm speed deviation, and the "complete" status word, which are packaged into status feedback data by the system.

[0171] The aforementioned status feedback data was transmitted back to the digital twin space in milliseconds without jitter via the TSN network. In the virtual digester model, the system was loaded with a 4.5A current load, and the speed of the virtual agitator was corrected to 1175 rpm in real time. On the digital twin interface, the original green "ideal operation" indicator turned into a yellow "high load" state, and the vortex morphology in the virtual flow field appeared slightly weaker than expected due to the reduced speed. This virtual-real synchronization ensured that the R&D personnel saw not only "command sent" at the remote monitoring end, but also the real physical state of "command execution blocked".

[0172] The fluid dynamics engine in the digital twin space compared the virtual flow field with the expected flow field and found that the current mixing intensity was insufficient to break up the local crystal clusters that might form in type B water. Based on the returned high current and speed deviation data, the system performed parameter correction iterations: the algorithm adjusted the fluid's "dynamic viscosity" parameter in the model from the default 1.5 mPa·s to 2.4 mPa·s to match the high viscosity characteristics of type B water. Based on the corrected model, the system recalculated the required theoretical torque and concluded that the current of 4.5A was close to the motor's rated load, and simply increasing the voltage might lead to overload. Therefore, the system generated a new optimization instruction, modifying the stirring strategy to "intermittent high-intensity stirring"—that is, running for 10 seconds and pausing for 2 seconds, while simultaneously adjusting the temperature control curve to match the stirring intervals. The new instruction was issued again, and after the physical motor executed this strategy, the current dropped back to a safe range, and the mixing effect gradually reached the requirements for type B water treatment during the iteration, achieving adaptive optimization that synchronized virtual and real systems.

[0173] The system continuously monitors the operating status of the standby digester 03, selecting "reaction system temperature" and "stirring motor current" as key indicators. Due to the intense chemical reaction of type B water, temperature and current usually exhibit normal physical fluctuations. The system collects data at a frequency of 10 times per second and calculates the fluctuation variance of these two indicators within a 5-second sliding time window. For example, at time T1, the system calculates the temperature fluctuation variance to be 0.15 (°C²) and the motor current fluctuation variance to be 0.05 (A²).

[0174] The system is preset to have a steady-state temperature range of less than 0.2 °C² for the high-temperature digestion process of type B water, and a preset number of continuous sampling cycles of 10, which means it needs to be stable for 50 seconds. After the monitoring started, from the first cycle to the tenth cycle, although there were occasional small bubbles bursting and causing disturbances during the digestion of type B water, the calculated temperature fluctuation variance remained between 0.12 and 0.18, and the motor current fluctuation variance also remained low. The system confirmed that the variance values ​​in these 10 consecutive cycles did not exceed the preset range and there was no sudden increase, thus determining that the digestion process had overcome the previous mixing unevenness problem and the kinetic state had converged to a stable reaction equilibrium state.

[0175] Based on the above steady-state determination, the system confirmed that the anomaly caused by "scaling of the agitator" had been successfully handled. The process controller automatically canceled the red alarm label for the sample, updated the status to "detecting", and continued to execute the subsequent cooling and titration steps. The system generated a complete autonomous monitoring event record, which included: "Anomaly type: decreased stirring efficiency; Handling strategy: switch to backup instrument digester 03 and parameter correction; Recovery time: 15 minutes; Effect evaluation: index convergence successful". This event was archived to the knowledge base, and the equipment selection logic for type B water quality treatment was updated, that is, the corrosion-resistant digester 03 was selected first, marking the successful closure of this autonomous monitoring task.

[0176] Please see Figure 3 The autonomous monitoring system for the unmanned laboratory site is applied to the aforementioned autonomous monitoring method for the unmanned laboratory site; the autonomous monitoring system for the unmanned laboratory site includes:

[0177] Data module 21 is used in the unmanned laboratory to collect multi-source heterogeneous sensing data in real time based on sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste liquid disposal, perform spatiotemporal alignment of multi-source heterogeneous sensing data, and determine the corresponding key data combination in combination with data preprocessing.

[0178] The digital twin module 22 is used to input various key data combinations into a preset digital twin space, and to perform dynamic simulation in the digital twin space to generate state feature vectors for each process node. By performing deviation processing with the state reference vector, potential abnormal working condition types and their corresponding confidence levels are identified.

[0179] The autonomous monitoring module 23 is used to input various abnormal working conditions and their confidence levels into the inference engine of the unmanned laboratory, and dynamically generate autonomous monitoring decision instructions in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategy.

[0180] The anomaly handling module 24 is used to send autonomous monitoring decision instructions to the execution control layer of each process node, drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling, and at the same time transmit the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range.

[0181] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0182] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An autonomous monitoring method for an unmanned laboratory site, characterized in that, include: In the unmanned laboratory, based on the sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste liquid disposal, multi-source heterogeneous sensing data is collected in real time, the multi-source heterogeneous sensing data is spatiotemporally aligned, and the corresponding key data combination is determined by combining data preprocessing. Each key data combination is input into a preset digital twin space, and dynamic simulation is performed in the digital twin space to generate state feature vectors for each process node. By performing deviation processing with the state baseline vector, potential abnormal working condition types and their corresponding confidence levels are identified. Each abnormal operating condition type and its confidence level are input into the inference engine of the unmanned laboratory, and autonomous monitoring decision instructions are dynamically generated in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategy. The autonomous monitoring and decision-making instructions are sent to the execution control layer of each process node to drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling. At the same time, the status feedback data during the handling process is transmitted back to the digital twin space in real time until each process node returns to the steady-state operation range.

2. The autonomous monitoring method for unmanned laboratory sites according to claim 1, characterized in that, In the unmanned laboratory setting, sensing units deployed at various stages of sample transfer, pretreatment, analysis and detection, and waste disposal collect multi-source heterogeneous sensing data in real time. This data is then spatiotemporally aligned, and combined with data preprocessing to determine corresponding key data combinations, including: In the unmanned laboratory, each process node of sample transfer, pretreatment, analysis and detection and waste disposal is marked. Corresponding sensing units are deployed for each process node. Through the sensing units, corresponding multi-source heterogeneous sensing data is collected in real time. The multi-source heterogeneous sensing data covers image stream, weight signal, position coordinates, environmental parameters and instrument status codes.

3. The autonomous monitoring method for unmanned laboratory sites according to claim 2, characterized in that, In the unmanned laboratory setting, based on sensing units deployed at various stages of sample transfer, pretreatment, analysis and detection, and waste disposal, multi-source heterogeneous sensing data is collected in real time. This data is then spatiotemporally aligned, and combined with data preprocessing to determine corresponding key data combinations. The process also includes: A spatiotemporal registration mechanism based on timestamp indexing and spatial coordinate transformation is adopted to map multi-source heterogeneous sensing data to a unified global spatiotemporal reference. An outlier removal mechanism is used to clean the data, thereby eliminating environmental noise interference and outputting the corresponding target data stream. Obtain a pre-defined knowledge graph of water quality analysis processes, extract time-series feature vectors representing the operational status of the processes from the target data stream, and construct corresponding key data combinations through feature fusion. These key data combinations include environmental background data, equipment operating parameters, and sample attribute labels.

4. The autonomous monitoring method for unmanned laboratory sites according to claim 1, characterized in that, The various key data combinations are input into a preset digital twin space, where dynamic simulations are performed to generate state feature vectors for each process node. By processing the deviations from the state baseline vectors, potential abnormal operating condition types and their corresponding confidence levels are identified, including: The key data combinations are mapped in real time to a preset digital twin space, and dynamic simulations are performed based on the physics engine and fluid dynamics rules. During the simulation, a state feature vector representing the physical state of the current process is generated. At the same time, a state reference vector is obtained, and the state feature vector is compared with the state reference vector. Multidimensional Euclidean distance is introduced in the comparison process to construct the corresponding deviation metric matrix.

5. The autonomous monitoring method for unmanned laboratory sites according to claim 4, characterized in that, The various key data combinations are input into a preset digital twin space, where dynamic simulations are performed to generate state feature vectors for each process node. By processing the deviations from the state baseline vectors, potential abnormal operating condition types and their corresponding confidence levels are identified. This also includes: If any dimension of the deviation measurement matrix exceeds the preset threshold range, the abnormal pattern matching logic is triggered. By combining a Bayesian network with the historical fault database, reverse tracing is performed to identify the specific abnormal working condition type and calculate the posterior probability of the abnormal working condition type as the confidence level.

6. The autonomous monitoring method for unmanned laboratory sites according to claim 1, characterized in that, The process involves inputting various abnormal operating condition types and their confidence levels into the inference engine of the unmanned laboratory, and dynamically generating autonomous monitoring decision instructions based on the resource constraints of the unmanned laboratory. These autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies, including: Each abnormal operating condition type and its confidence level are input into the inference engine of the unmanned laboratory to simultaneously obtain the resource constraints of the unmanned laboratory. These resource constraints include the current instrument load rate, reagent balance, and sample queue length. Based on a multi-objective optimization strategy, the inference engine searches for the optimal response strategy in the digital twin space, thereby dynamically generating autonomous monitoring decision instructions. These autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies.

7. The autonomous monitoring method for unmanned laboratory sites according to claim 6, characterized in that, The process involves inputting various abnormal operating condition types and their confidence levels into the inference engine of the unmanned laboratory, and dynamically generating autonomous monitoring and decision-making instructions based on the resource constraints of the unmanned laboratory. These instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies. Process path replanning is used to generate alternative flow paths when the current path is blocked. Backup instrument switching logic is used to initiate redundant equipment takeover tasks when the confidence level is higher than the first threshold. Parameter adjustment strategy is used to compensate for temperature control time, sample volume or digestion power when the confidence level is lower than the second threshold.

8. The autonomous monitoring method for unmanned laboratory sites according to claim 1, characterized in that, The process of issuing autonomous monitoring and decision-making instructions to the execution control layer of each process node to drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling, and simultaneously transmitting the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range, includes: The autonomous monitoring and decision-making instructions are parsed into specific control timing signals through the edge computing gateway. The control timing signals are then sent to the execution control layer corresponding to each process node to drive the multi-degree-of-freedom robotic arm, injection pump and temperature control module to perform coordinated actions, thereby completing the abnormal handling. During the abnormal handling process, the current feedback, position deviation and action completion signal of the actuator are collected in real time. The current feedback, position deviation and action completion signal of the actuator are used as status feedback data and transmitted back to the digital twin space. The virtual and real synchronous iteration is realized during the parameter correction process.

9. The autonomous monitoring method for an unmanned laboratory site according to claim 8, characterized in that, The process of issuing autonomous monitoring and decision-making instructions to the execution control layer of each process node to drive the corresponding actuators to coordinate actions to complete the anomaly handling, and simultaneously transmitting the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range, also includes: The system continuously monitors the operating parameters of each process node. When the fluctuation variance of key indicators converges to the preset steady-state range within a preset number of consecutive sampling periods, it determines that the anomaly handling is completed and the normal operation process is restored, thus forming an autonomous monitoring event at the unmanned laboratory site.

10. An autonomous monitoring system for an unmanned laboratory site, characterized in that, The autonomous monitoring system at the unmanned laboratory site is applied to the autonomous monitoring method at the unmanned laboratory site as described in any one of claims 1-9; The autonomous monitoring system at the unmanned laboratory site includes: The data module is used in the unmanned laboratory to collect multi-source heterogeneous sensing data in real time based on sensing units deployed at each process node of sample transfer, pretreatment, analysis and detection and waste disposal. It performs spatiotemporal alignment of the multi-source heterogeneous sensing data and determines the corresponding key data combination in combination with data preprocessing. The digital twin module is used to input various key data combinations into a preset digital twin space, and to perform dynamic simulation in the digital twin space to generate state feature vectors for each process node. By processing the deviation from the state baseline vector, potential abnormal working condition types and their corresponding confidence levels are identified. The autonomous monitoring module is used to input various abnormal operating conditions and their confidence levels into the inference engine of the unmanned laboratory, and dynamically generate autonomous monitoring decision instructions in combination with the resource constraints of the unmanned laboratory. The autonomous monitoring decision instructions include process path replanning, backup instrument switching logic, and parameter adjustment strategies. The anomaly handling module is used to issue autonomous monitoring and decision-making instructions to the execution control layer of each process node, drive the corresponding execution mechanism to coordinate actions to complete the anomaly handling, and at the same time transmit the status feedback data during the handling process back to the digital twin space in real time until each process node returns to the steady-state operating range.

Images