AI-based intelligent load prediction and deformation risk early warning system for laboratory bench base

Abstract

This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets, belonging to the field of AI early warning technology. It includes: an intelligent layout module for dynamically adjusting the initial layout based on the finite element mechanical model of the laboratory bench cabinet and clustering results based on historical load distribution; a preprocessing module for adaptively adjusting the weight coefficients of different data sources by combining the structural mechanics model of the laboratory bench cabinet with a particle swarm optimization algorithm, outputting a multi-source fusion standardized dataset; a model training module for training the model based on the dataset; and a prediction and early warning module for outputting predicted cabinet deformation values ​​within a preset time period, dynamically adjusting tiered early warning thresholds, comparing and triggering corresponding early warnings, and determining and outputting the deformation development trend. This ensures high accuracy in deformation prediction, effectively reducing damage to experimental equipment or experimental interruptions caused by cabinet deformation, and providing strong support for the safe operation of the laboratory.

Description

Technical Field

[0001] This invention relates to the field of AI early warning technology, and in particular to an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets. Background Technology

[0002] The base cabinet of the laboratory bench bears the load of experimental equipment and reagents for a long time, and is prone to deformation due to uneven load distribution, environmental factors (temperature and humidity), or frequent use (such as opening and closing drawers). If the deformation exceeds the limit, it may lead to damage to the cabinet, overturning of equipment, or even safety accidents.

[0003] In existing technologies, deformation monitoring of laboratory bench cabinets is mostly achieved through periodic manual inspections or real-time data collection using simple sensors.

[0004] Manual inspection is inefficient and lacks timeliness, making it impossible to detect potential risks in a timely manner;

[0005] Simple real-time data collection can only reflect the current deformation and lacks the ability to predict future deformation trends, making it difficult to provide early warnings.

[0006] Outdated data processing and analysis methods have not fully utilized multi-source data such as historical monitoring, simulated working conditions, and actual usage data for deep integration. There is also a lack of AI models that can simultaneously capture spatial characteristic load distribution and temporal dependence, resulting in insufficient deformation prediction accuracy.

[0007] Therefore, this invention proposes an AI-based intelligent load prediction and deformation risk early warning system for experimental bench cabinets. Summary of the Invention

[0008] This invention provides an AI-based intelligent load prediction and deformation risk warning system for laboratory bench cabinets, in order to solve the aforementioned technical problems.

[0009] This invention proposes an AI-based intelligent load prediction and deformation risk early warning system for experimental bench cabinets, comprising:

[0010] The intelligent layout module is used to dynamically adjust the initial layout based on the finite element mechanical model of the experimental platform cabinet and the clustering results of historical load distribution, and to collect the real-time parameters of the experimental platform cabinet, wherein the real-time parameters include: environmental parameters and load parameters, deformation parameters and equipment mechanical action parameters with location labels.

[0011] The preprocessing module is used to perform weighted fusion of historical detection data, simulated working condition data, actual usage data and real-time parameters with location tags under the experimental platform cabinet. It also adaptively adjusts the weight coefficients of different data sources by combining the structural mechanics model of the experimental platform cabinet and the particle swarm optimization algorithm, and outputs a multi-source fusion standardized dataset after preprocessing.

[0012] The model training module is used to construct a CNN-LSTM fusion model with attention mechanism that incorporates prior knowledge of structural mechanics, and to train the model based on the multi-source fusion standardized dataset.

[0013] The prediction and early warning module is used to input preprocessed real-time parameters into the trained model and output the predicted values ​​of cabinet deformation with location labels within a preset time period. At the same time, it standardizes the usage time, maintenance records and structural mechanical parameters of each monitoring position of the experimental bench cabinet, dynamically adjusts the graded early warning thresholds, and compares them with the predicted values ​​of cabinet deformation with location labels to trigger corresponding early warnings. Based on the structural mechanical model, it determines and outputs the deformation development trend of the experimental bench cabinet based on the predicted values ​​of cabinet deformation with location labels.

[0014] Preferably, the intelligent layout module includes:

[0015] The load acquisition unit is used to acquire historical load data of the experimental bench cabinet. The historical load data includes the load values ​​of different areas of the cabinet at each monitoring time and the corresponding timestamps.

[0016] Clustering unit is used to cluster the historical load data using density clustering algorithm to obtain several load distribution clusters. Each cluster corresponds to a typical load distribution pattern. Based on the finite element mechanical model of the experimental bench cabinet, the stress concentration area and deformation sensitive area of ​​the cabinet under the load distribution pattern corresponding to each cluster are calculated.

[0017] The function construction unit is used to construct a multi-objective optimization model with the objectives of maximizing the monitoring coverage of key areas corresponding to each cluster by the sensors and minimizing the redundancy between sensors. The objective function of the multi-objective optimization model is: ,in, Number of sensors; This represents the number of clusters. Let u0 be the weight coefficient of the u1th sensor for the key region of the u1th cluster. Let u0 be the coverage of the key region of the u1th cluster by the u0th sensor; The redundancy between the u0th sensor and the u2th sensor;

[0018] The scheme output unit is used to output the adjusted sensor layout scheme based on the solution results of the multi-objective optimization model.

[0019] Preferably, the CNN-LSTM fusion model includes:

[0020] The input layer receives location-labeled multi-source fusion normalized feature vectors based on the multi-source fusion normalized dataset; the attention layer receives location-labeled mechanical feature importance matrices based on the experimental bench base. Attention weights are assigned to the features of each dimension of the multi-source fusion standardized feature vector.

[0021] The CNN sublayer contains two convolutional layers and one pooling layer, used for convolutional operations. Extract spatial characteristics of load distribution, structural stress, and deformation with location labels, among which, This is a convolution operation; , For convolution kernel; , For bias terms; For activation functions; This is a pooling operation; Input features;

[0022] The feature fusion layer is used to perform multimodal fusion of spatial features with location labels output from CNN sublayers, weighted features output from the attention layer, and theoretical deformation features with location labels calculated based on the structural mechanics model.

[0023] The LSTM sublayer contains two hidden layers and is used to capture the temporal dependence of load-deformation relationships with location labels, the long-term effects of changes in equipment mechanical forces, and the dynamic laws of mechanical parameter decay.

[0024] The output layer is used to output the predicted values ​​of cabinet deformation with location tags within a preset future time period.

[0025] Preferably, the prediction and early warning module includes:

[0026] The data acquisition unit is used to standardize the usage time, maintenance records and structural mechanical parameters of each monitoring location to obtain a standardized parameter set for each monitoring location. Each monitoring location corresponds to a unique location label.

[0027] The threshold adjustment unit is used to construct a reinforcement learning model that integrates the structural mechanical constraints of the experimental platform cabinet. The reward function is to maximize the early warning accuracy and minimize the false alarm rate and false alarm rate. Based on the standardized parameter set corresponding to each location label, the unit dynamically adjusts the hierarchical early warning threshold under each location label. The adjusted threshold satisfies the allowable deformation range constraint calculated by the structural mechanical model of the corresponding monitoring location.

[0028] The state space of the reinforcement learning model is the standardized parameter set for each location label, the action space is the adjustment amount of the hierarchical warning threshold for each location label, and the reward function is... ;

[0029] in, To improve the accuracy of early warnings; False alarm rate; This refers to the false negative rate. , , These are the weighting coefficients, and , 3.

[0030] Preferably, the preprocessing module includes:

[0031] The theoretical determination unit is used to calculate the theoretical response vector of load-deformation at each label location at each historical moment based on the structural mechanics model of the experimental bench cabinet.

[0032] The data determination unit is used to set the monitoring positions of the experimental bench base cabinet. Where M is the total number of monitoring locations; pj1 is the multi-source data from the j1-th monitoring location. , This refers to the historical detection data for the j1th monitoring location; This refers to the simulated operating condition data for the j1th monitoring location; This represents the actual usage data for the j1th monitoring location; This refers to the real-time parameter data for the j1th monitoring location;

[0033] Vector extraction unit, used for multi-source data at each monitoring location. Standardization processing is performed to obtain standardized data. And extract the multi-source standard vector of the j1th monitoring location at each historical moment;

[0034] The weighting construction unit is used to calculate the deviation between the multi-source standard vector and the corresponding theoretical response vector to determine the high-confidence weight of the j1-th monitoring location. With low reliability weight And construct the weight matrix Wj1=[ for the j1-th monitoring location. , ],in, ,in, Let be the element value of the k-th data source in the multi-source standard vector at the j1-th monitoring location at the t-th historical time. The number of historical moments; For all data sources under the k-th data source Standard deviation; It is a fluctuation penalty term, and ,and When k=1, it corresponds to h; when k=2, it corresponds to s; when k=3, it corresponds to h. Correspondingly, when k=4, and correspond;

[0035] Weight iteration unit, used to weigh all data sources involved The goal is to minimize the value in the data. The fusion weight coefficients of historical detection data, simulated working condition data, actual usage data and real-time parameters under each location label are iteratively optimized using the particle swarm optimization algorithm.

[0036] Tag fusion unit, used to import all tags from the k-th data source. The fusion error Ej1 and the weight vector constitute Matching degree function As a constraint, the optimized fusion weight coefficients of each location label are obtained, and the multi-source fusion data with location labels is output.

[0037] Preferably, the mechanical feature importance matrix ,in, This represents the theoretical deformation value of the structural mechanics model at the j1th monitoring location; for right The partial derivative matrix; Let J be the geometric weight matrix for the j1th monitoring location; For element-wise multiplication; This is the normalization function.

[0038] Preferably, the prediction and early warning module further includes:

[0039] Evolution determination units are used to construct a deformation evolution dynamics model with position labels based on the structural mechanics model of the experimental platform cabinet:

[0040] ,in, Let J be the vector of the rate of change of deformation at the j1th monitoring location; The structural inherent mechanical matrix of the j1th monitoring location; Let be the load-deformation transfer matrix for the j1th monitoring location; This represents the real-time load vector at the j1th monitoring location; Let J be the attenuation vector of the structural mechanical parameters at the j1-th monitoring location; Let J be the parameter attenuation-deformation influence matrix for the j1th monitoring location; This is the specific state vector of the j1th monitoring position;

[0041] The predicted deformation value of the base cabinet with location tags is used as the initial state. By combining real-time acquired load data and structural mechanical parameter attenuation model, the deformation evolution dynamic model is solved using an adaptive step-size numerical integration algorithm to obtain the deformation development trend curve of the j1th monitoring position within a preset time period in the future.

[0042] Preferred options also include:

[0043] The prediction unit is used to calculate the intersection of the deformation development trend curve and the safety threshold surface based on the safety threshold surface of the structural mechanics model, and to determine the prediction time window for deformation exceeding the limit.

[0044] Compared with the prior art, the beneficial effects of this application are as follows:

[0045] The intelligent layout module enables precise collection of multi-dimensional data from the cabinet, ensuring data coverage of key areas. The preprocessing module improves data quality and reliability through multi-source data fusion and optimization. The model training module constructs a fusion model incorporating mechanical laws, ensuring high accuracy in deformation prediction. The prediction and early warning module dynamically adjusts thresholds and provides precise early warnings, and can also predict deformation trends. Overall, it can identify the deformation risk of the experimental cabinet in advance, effectively reducing damage to experimental equipment or interruption of experiments caused by cabinet deformation, significantly improving the safety management level of the cabinet, extending its service life, and providing strong support for the safe operation of the laboratory.

[0046] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0047] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a structural diagram of an AI-based intelligent load prediction and deformation risk early warning system for an experimental bench cabinet, as described in an embodiment of the present invention. Detailed Implementation

[0050] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets, such as... Figure 1 As shown, it includes:

[0052] The intelligent layout module is used to dynamically adjust the initial layout based on the finite element mechanical model of the experimental platform cabinet and the clustering results of historical load distribution, and to collect the real-time parameters of the experimental platform cabinet, wherein the real-time parameters include: environmental parameters and load parameters, deformation parameters and equipment mechanical action parameters with location labels.

[0053] The preprocessing module is used to perform weighted fusion of historical detection data, simulated working condition data, actual usage data and real-time parameters with location tags under the experimental platform cabinet. It also adaptively adjusts the weight coefficients of different data sources by combining the structural mechanics model of the experimental platform cabinet and the particle swarm optimization algorithm, and outputs a multi-source fusion standardized dataset after preprocessing.

[0054] The model training module is used to construct a CNN-LSTM fusion model with attention mechanism that incorporates prior knowledge of structural mechanics, and to train the model based on the multi-source fusion standardized dataset.

[0055] The prediction and early warning module is used to input preprocessed real-time parameters into the trained model and output the predicted values ​​of cabinet deformation with location labels within a preset time period. At the same time, it standardizes the usage time, maintenance records and structural mechanical parameters of each monitoring position of the experimental bench cabinet, dynamically adjusts the graded early warning thresholds, and compares them with the predicted values ​​of cabinet deformation with location labels to trigger corresponding early warnings. Based on the structural mechanical model, it determines and outputs the deformation development trend of the experimental bench cabinet based on the predicted values ​​of cabinet deformation with location labels.

[0056] In this embodiment, the finite element mechanical model is a digital model constructed by disassembling the experimental platform cabinet into multiple non-overlapping small units according to the mechanical analysis requirements, and combining the elastic modulus and Poisson's ratio of the cabinet material. This model is used to simulate the stress and deformation state of the cabinet under different loads. Finite element analysis software such as ANSYS or ABAQUS is used to draw the three-dimensional structure of the cabinet, such as a cabinet with a length of 1200mm, a width of 600mm, and a height of 800mm, containing 4 columns and 3 beams. Material properties and constraints are defined, ultimately generating a model that can be used for mechanical calculations.

[0057] In this embodiment, the clustering results of historical load distribution refer to the load data of different areas of the cabinet over the past year, which includes the load value and corresponding timestamp at each monitoring time. The density clustering algorithm DBSCAN is used to group data with similar load distribution patterns into one class, forming several clusters. Daily load data of five monitoring areas (front left, front right, rear left, rear right, and center) of the cabinet are collected by sensors. The DBSCAN algorithm is then called using the scikit-learn library of Python, and a reasonable neighborhood radius and minimum sample number are set for clustering. For example, three clusters may be obtained, corresponding to three typical distribution patterns: high load in the center (600-800N), low load at the edge (200-300N), high load on the left (500-700N), low load on the right (100-200N), and uniform load (300-400N).

[0058] In this embodiment, dynamically adjusting the initial layout refers to optimizing the initial installation positions of the sensors based on the clustering results of historical load distribution to improve the monitoring accuracy of key areas of the cabinet. Specifically, an initial layout is set after uniformly installing the sensors. When the proportion of newly added clusters is ≥15% and the time since the last adjustment is ≥1 month, the adjustment is triggered. The optimal layout is calculated using a multi-objective optimization model (aiming at maximizing the coverage of key areas of the clusters by the sensors and minimizing the redundancy between sensors). For example, if the clustering results show that the middle of the cabinet is a stress concentration area and there are no sensors in the middle of the initial layout, then two sensors are added to the middle during the adjustment, while one redundant sensor at the edge is reduced.

[0059] In this embodiment, the real-time parameters include four categories. The first is environmental parameters, which are data on environmental factors affecting the mechanical performance of the cabinet. Specifically, this involves collecting temperature and humidity data using an SHT30 temperature and humidity sensor, and collecting vibration frequencies generated by nearby equipment using an ADXL345 vibration sensor. The second is load parameters with location tags. The location tag is a unique identifier for each monitoring location, and the load parameter is the force value at that location, including vertical and horizontal loads. Specifically, a ZEMICH3F load cell is installed at each monitoring location, and each sensor is bound to a location tag. After data collection, the location tag, vertical load, and horizontal load are recorded. For example, the load parameters for the front left area are: front left area - 500N (vertical) - 80N (horizontal). Thirdly, deformation parameters with location tags, i.e., deformation data at each monitoring location, such as deflection and rotation angle. KeyenceIL-300 displacement sensors are installed at each location, and location tags are attached to collect deflection data in the range of 0-5mm and rotation angle data in the range of 0-1°. Fourthly, equipment mechanical action parameters, which refer to the mechanical effects of experimental equipment on the cabinet during operation, such as impact force and continuous pressure. Specifically, force sensors are installed at the contact points between the equipment and the cabinet to collect mechanical data during equipment operation. For example, the instantaneous impact force of 200N generated by the centrifuge on the cabinet during operation is a mechanical action parameter of the equipment.

[0060] In this embodiment, the historical detection data with location tags refers to the load, deformation, and other data collected by sensors at various monitoring locations of the cabinet over a past period, with each data point associated with a corresponding location tag and timestamp. The data format of the monitoring data is location tag-timestamp-load value-deformation value, for example, rear right area-timestamp-420N-0.3mm.

[0061] In this embodiment, the simulated operating condition data is obtained by simulating the load, deformation, and other data at various monitoring locations of the cabinet under different extreme or typical operating conditions using finite element software, and each data point is associated with a location tag. Specifically, the finite element mechanical model of the cabinet is loaded into the ABAQUS software, and different operating conditions are set, such as full load condition: each area bears an 800N vertical load; off-center load condition: the left area bears a 1000N vertical load, and the right area bears a 200N vertical load. After running the simulation, the load and deformation data at each location are output and associated with location tags.

[0062] In this embodiment, the actual usage data refers to the relevant operational data during the daily use of the cabinet, excluding real-time sensor monitoring. This includes the weight of the equipment placed by the experimenter, the equipment placement location, and the usage duration, all linked to location tags. Through the laboratory web log system, experimenters record the equipment name, equipment weight, placement location tag, and usage duration after each use, and export the data periodically. For example, a single usage data entry might be: Microscope - 30kg - Front Left Area - 4 hours.

[0063] In this embodiment, the structural mechanics model is a model of the relationship between the force and deformation of the cabinet based on classical mechanics theory. It is used to calculate the theoretical deformation value and provide a reference for data fusion.

[0064] In this embodiment, the adaptive adjustment of the weight coefficients of different data sources is based on the deviation between the data from each data source and the theoretical values ​​of the structural mechanics model, as well as the data fluctuation, and the weight coefficients are dynamically updated to ensure the reliability of the fused data.

[0065] Preprocessing includes data cleaning and data standardization. Data cleaning removes outliers from multi-source data, such as data with loads exceeding the cabinet's maximum load capacity of 1500N or data with negative deformation values. This is generally achieved using the 3σ principle. Data standardization scales the data to the 0-1 range to eliminate the influence of dimensions. The standardized value is obtained using the Min-Max standardization formula: Standardized value = (Original value - Minimum value) / (Maximum value - Minimum value).

[0066] In this embodiment, the multi-source fusion standardized dataset is a dataset with a uniform format and location labels after weighted fusion and preprocessing, which is used for subsequent model training.

[0067] In this embodiment, the preset duration is set according to needs, such as 24 hours or 72 hours. The deformation prediction value is the model's prediction result of the deformation at different time points in the future for each location, and is associated with the location label.

[0068] In this embodiment, the usage time of each monitoring location is the cumulative time from the first time it is put into use to the current time, which is used to assess the aging degree of the cabinet.

[0069] In this embodiment, the maintenance record is a record of the maintenance status of each monitoring location, including the number of maintenance operations, the most recent maintenance interval, and the maintenance level, which is used to determine the reliability of the location.

[0070] In this embodiment, the structural mechanical parameters are the cabinet structure and material parameters corresponding to each monitoring position, including cross-sectional dimensions, elastic modulus, yield strength, etc., and material aging and degradation are taken into account.

[0071] In this embodiment, the dynamic adjustment of the graded early warning threshold refers to the critical value that divides the deformation risk into three levels: low, medium, and high. For example, the initial threshold is: low risk ≤ 0.2, medium risk 0.2-0.5, and high risk > 0.5. The dynamic adjustment is to adjust the threshold according to the standardized parameters of each location, such as usage time, maintenance records, and structural mechanical parameters, so as to make the early warning more accurate.

[0072] In this embodiment, the comparison triggering of corresponding early warnings involves comparing the predicted deformation value at each location with the adjusted graded early warning threshold to determine the risk level and trigger an early warning. The system compares in real time. If the predicted deformation value at a certain location is 0.48, and the adjusted high-risk threshold is 0.45, it is determined to be high-risk. An early warning message containing the location tag, risk level, and predicted deformation value is pushed to the management personnel via the laboratory management APP. Simultaneously, an early warning pop-up is displayed on the laboratory monitoring screen. For example, if the predicted deformation in the front right area is 0.3, and the adjusted medium-risk threshold is 0.2-0.4, a medium-risk early warning is triggered, indicating that the front right area, for example, with the location tag: front right area - coordinates (900, 200) mm, has a predicted deformation value of 0.3, is at a medium-risk level, and attention should be paid to the equipment placement.

[0073] In this embodiment, the deformation development trend is used to predict changes in risk.

[0074] The beneficial effects of the above technical solution are as follows: the intelligent layout module enables accurate collection of multi-dimensional data from the cabinet, ensuring data coverage of key areas; the preprocessing module improves data quality and reliability through multi-source data fusion and optimization; the model training module constructs a fusion model incorporating mechanical laws, ensuring high accuracy in deformation prediction; the prediction and early warning module dynamically adjusts thresholds and provides accurate early warnings, and can also predict deformation trends. Overall, it can identify the deformation risk of the experimental cabinet in advance, effectively reducing damage to experimental equipment or interruption of experiments caused by cabinet deformation, significantly improving the safety management level of the cabinet, extending its service life, and providing strong support for the safe operation of the laboratory.

[0075] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets. The intelligent layout module includes:

[0076] The load acquisition unit is used to acquire historical load data of the experimental bench cabinet. The historical load data includes the load values ​​of different areas of the cabinet at each monitoring time and the corresponding timestamps.

[0077] Clustering unit is used to cluster the historical load data using density clustering algorithm to obtain several load distribution clusters. Each cluster corresponds to a typical load distribution pattern. Based on the finite element mechanical model of the experimental bench cabinet, the stress concentration area and deformation sensitive area of ​​the cabinet under the load distribution pattern corresponding to each cluster are calculated.

[0078] The function construction unit is used to construct a multi-objective optimization model with the objectives of maximizing the monitoring coverage of key areas corresponding to each cluster by the sensors and minimizing the redundancy between sensors. The objective function of the multi-objective optimization model is: ,in, Number of sensors; This represents the number of clusters. Let u0 be the weight coefficient of the u1th sensor for the key region of the u1th cluster. Let u0 be the coverage of the key region of the u1th cluster by the u0th sensor; The redundancy between the u0th sensor and the u2th sensor;

[0079] The scheme output unit is used to output the adjusted sensor layout scheme based on the solution results of the multi-objective optimization model.

[0080] In this embodiment, ,in, The maximum load for the u1th cluster; This is the sum of the maximum loads across all clusters; Let u0 be the monitoring accuracy of the 0th sensor; This is the sum of the accuracies of all sensors.

[0081] In this embodiment, the monitoring coverage rate of the key area corresponding to each cluster by the sensor is the proportion of the range of the key area of ​​the cluster monitored by the sensor. For example, the key area of ​​cluster 1 is the central 300mm×300mm area. Sensor A is installed in the center, and the monitoring range covers 200mm×200mm of this area. The coverage rate is the area covered by the monitoring range divided by the area of ​​the corresponding key area.

[0082] In this embodiment, the redundancy between sensors is the degree of overlap of the monitoring areas of multiple sensors. For example, sensor A monitors a 200mm×200mm area in the middle, and sensor B monitors a 200mm×200mm area slightly to the right of the middle. The overlapping area is 100mm×200mm, and the redundancy is: the area of ​​the overlapping area / (the area of ​​the monitoring range corresponding to sensor A + the area of ​​the monitoring range corresponding to sensor B).

[0083] The multi-objective optimization model is a mathematical model that simultaneously optimizes two objectives: maximizing the monitoring coverage of key areas by sensors and minimizing the redundancy between sensors. The objective function finds the optimal solution by quantifying these two objectives.

[0084] In this embodiment, Python's DEAP library is used for programming. Input parameters such as the number of sensors, cluster information, sensor accuracy, and key areas are used to run the algorithm and obtain the Pareto optimal solution, which is the set of solutions balancing coverage and redundancy. A suitable scheme is then selected from this set. For example, after optimization, a layout with one sensor each in the center, front left, and rear right achieves a total coverage of 85% and a redundancy of ≤15% in the key areas. The sensor installation coordinates are then output, such as a layout where the central sensor is installed at the center of the cabinet at (600, 300) mm.

[0085] The beneficial effects of the above technical solution are as follows: the load acquisition unit accurately collects historical load data, the clustering unit combines the finite element model to analyze the key areas under typical load modes, and the function construction and solution output unit optimizes the sensor layout. Finally, the sensor achieves high coverage and low redundancy monitoring of the key areas of the cabinet, providing a precise data acquisition basis for subsequent load prediction and deformation early warning, and improving the comprehensiveness and efficiency of the mechanical state monitoring of the cabinet.

[0086] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for experimental platform cabinets. The CNN-LSTM fusion model includes:

[0087] The input layer receives location-labeled multi-source fusion normalized feature vectors based on the multi-source fusion normalized dataset; the attention layer receives location-labeled mechanical feature importance matrices based on the experimental bench base. Attention weights are assigned to the features of each dimension of the multi-source fusion standardized feature vector.

[0088] The CNN sublayer contains two convolutional layers and one pooling layer, used for convolutional operations. Extract spatial characteristics of load distribution, structural stress, and deformation with location labels, among which, This is a convolution operation; , For convolution kernel; , For bias terms; For activation functions; This is a pooling operation; Input features;

[0089] The feature fusion layer is used to perform multimodal fusion of spatial features with location labels output from CNN sublayers, weighted features output from the attention layer, and theoretical deformation features with location labels calculated based on the structural mechanics model.

[0090] The LSTM sublayer contains two hidden layers and is used to capture the temporal dependence of load-deformation relationships with location labels, the long-term effects of changes in equipment mechanical forces, and the dynamic laws of mechanical parameter decay.

[0091] The output layer is used to output the predicted values ​​of cabinet deformation with location tags within a preset future time period.

[0092] Preferably, the mechanical feature importance matrix ,in, This represents the theoretical deformation value of the structural mechanics model at the j1th monitoring location; for right The partial derivative matrix; Let J be the geometric weight matrix for the j1th monitoring location; For element-wise multiplication; This is the normalization function.

[0093] In this embodiment, X is the input feature, and is Two-dimensional feature tensor, where, Here, C represents the spatial grid dimension of the cabinet (i.e., the location label), C represents the number of feature channels, and represents the load / deformation / environment. It should be noted that the input feature X and the theoretical deformation value... All features are standardized before computation; K1 is a 3×3 convolution kernel used to extract local spatial features, and K2 is a 2×2 convolution kernel used to compress dimensions to ensure that the output features correspond one-to-one with the spatial grid of the location labels.

[0094] In this embodiment, the experimental platform cabinet is divided into five regions with unique location labels: front left, front right, rear left, rear right, and center. For each region, parameters such as load (vertical / horizontal load), deformation (deflection / rotation), and environmental factors (temperature / humidity) are standardized to form feature vectors. For example, the feature vector for the front left region includes standardized dimensions such as vertical load 0.5, horizontal load 0.3, temperature 0.4, and humidity 0.2. These feature vectors are read from the preprocessed CSV data file and input into the model input layer.

[0095] In this embodiment, taking the central region as an example, the theoretical deformation value of the region under different loads is first calculated using a structural mechanics model (e.g., using beam theory from mechanics of materials, combined with the cross-sectional dimensions of the cabinet beam and the elastic modulus of Q235 steel; assuming the theoretical deflection is 0.5mm when the load is 400N). Then, the partial derivative matrix of the theoretical deformation with respect to input features (such as load and temperature) is calculated (e.g., for every 1N increase in load, the theoretical deformation increases by 0.001mm, corresponding to a partial derivative element of 0.001; for every 1℃ change in temperature, the theoretical deformation changes by 0.0005mm, corresponding to a partial derivative element of 0.0005). Simultaneously, based on the fact that the central region is a geometrically critical area of ​​stress concentration, a geometric structure weight matrix Gj1 is set (the load-related weight element for the central region is set to 0.9, and the environmental-related weight element is set to 0.1). The partial derivative matrix and the geometric structure weight matrix are multiplied element-by-element, and then normalized using the Softmax function to obtain the mechanical feature importance matrix. This gives the load features of the central region higher attention weights, allowing the model to focus more on key features that significantly affect deformation.

[0096] In this embodiment, the input feature X is The two-dimensional feature tensor is defined as follows: H=5 (the cabinet is divided into 5 grids vertically), W=5 (divided into 5 grids horizontally), and C=3 (feature channels are load, deformation, and environment). The first convolutional kernel K1 is 3×3 in size and is used to extract local spatial features, such as how the loads in adjacent grid regions jointly affect the deformation. The second convolutional kernel K2 is 2×2 in size and is used to compress the feature dimension. b1 and b2 are bias terms and activation functions. ReLU is used to set negative outputs to 0 while retaining positive features. The pooling operation uses max pooling, which takes the maximum value from a 2×2 feature block to reduce computation. For example, the input tensor undergoes K1 convolution (generating 64 feature channels), ReLU activation, K2 convolution (generating 32 feature channels), and ReLU activation again, followed by max pooling. This captures the spatial pattern of high load in the front-left region and synchronous deformation changes in the surrounding regions. The CNN sublayer outputs a 32-dimensional spatial feature vector of the central region. The attention layer weights the central load features to obtain a 16-dimensional vector. The structural mechanics model calculates the theoretical deformation of the central region to be 0.4 mm (normalized to 0.2 mm, forming a 1-dimensional feature vector). These three vectors are concatenated into a 49-dimensional fused feature vector, allowing the model to utilize information from spatial distribution, feature importance, and theoretical mechanics simultaneously.

[0097] In this embodiment, the input to the LSTM sublayer is the temporal feature sequence output by the feature fusion layer, such as the fused features of 24 time points each day over the past 7 days. The LSTM learns, through memory units such as cell state, input gate, forget gate, and output gate, the temporal pattern of increased load due to morning equipment loading, followed by a gradual increase in deformation several hours later, and the long-term pattern of material elastic modulus decreasing year by year after long-term equipment use, resulting in a gradual increase in deformation under the same load. For example, a centrifuge operating at fixed times each day generates impact loads, and the LSTM can remember the long-term cumulative effect of such periodic loads on deformation.

[0098] In this embodiment, the input to the output layer is the temporal features processed by the LSTM sublayer. Through a fully connected layer, it outputs the deformation value corresponding to each location label and each future time point. For example, when predicting the next 24 hours, it outputs results every 6 hours, such as the deformation of the front left region after 6 hours being 0.3mm and after 12 hours being 0.35mm, and the deformation of the middle region after 6 hours being 0.4mm and after 12 hours being 0.42mm, providing data support for subsequent early warning.

[0099] The beneficial effects of the above technical solution are as follows: by receiving multi-source standardized features through the input layer, highlighting key mechanical features through the attention layer, extracting spatial distribution patterns through the CNN sub-layer, integrating multimodal information through the feature fusion layer, capturing temporal and long-term changes through the LSTM sub-layer, and accurately predicting future deformation through the output layer, the solution fully integrates prior knowledge of structural mechanics with a data-driven approach. This ensures both theoretical rationality and the accurate capture of the spatiotemporal patterns of complex loads and deformations, providing high-precision prediction results for early warning of deformation risks in experimental platform cabinets.

[0100] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets. The prediction and early warning module includes:

[0101] The data acquisition unit is used to standardize the usage time, maintenance records and structural mechanical parameters of each monitoring location to obtain a standardized parameter set for each monitoring location. Each monitoring location corresponds to a unique location label.

[0102] The threshold adjustment unit is used to construct a reinforcement learning model that integrates the structural mechanical constraints of the experimental platform cabinet. The reward function is to maximize the early warning accuracy and minimize the false alarm rate and false alarm rate. Based on the standardized parameter set corresponding to each location label, the unit dynamically adjusts the hierarchical early warning threshold under each location label. The adjusted threshold satisfies the allowable deformation range constraint calculated by the structural mechanical model of the corresponding monitoring location.

[0103] The state space of the reinforcement learning model is the standardized parameter set for each location label, the action space is the adjustment amount of the hierarchical warning threshold for each location label, and the reward function is... ;

[0104] in, To improve the accuracy of early warnings; False alarm rate; This refers to the false negative rate. , , These are the weighting coefficients, and , 3.

[0105] In this embodiment, maintenance records are quantified into three parameters: maintenance frequency normalized to 0-1; most recent maintenance interval normalized to 0-1; maintenance level: full maintenance = 1, partial maintenance = 0.5, no maintenance = 0;

[0106] In this embodiment, the usage duration is the time from when the monitoring location was activated to the present.

[0107] In this embodiment, maintenance logs from the laboratory management system are extracted and normalized using the pandas library in Python. Structural mechanical parameters, such as the material's elastic modulus, are used. For example, if the initial elastic modulus at a certain location is 206 GPa, and due to aging it is currently 180 GPa, and the elastic modulus decay range is 100-206 GPa, the normalized value is (180-100) / (206-100) ≈ 0.755. This value is then normalized after calculation using the material aging model and test data. It should be noted that the normalized parameter set for each monitoring location is a collection of these processed parameters. For example, the parameter set for a certain location might include usage time 0.275, maintenance frequency 0.6, most recent maintenance interval 0.5, maintenance level 1, and a normalized elastic modulus value of 0.755.

[0108] In this embodiment, 3=1, and The value of 3 is determined according to the actual needs of the scenario. For example, it can be 0.6, 0.2, or 0.2, with priority given to ensuring the accuracy of the early warning.

[0109] In this embodiment, the reinforcement learning model that integrates structural mechanics constraints incorporates allowable deformation limits for the cabinet structure into the reinforcement learning process. For example, the maximum allowable deformation at a certain location, calculated by the structural mechanics model, is 2mm, which corresponds to ≤0.5 after standardization. The model is constructed using Python's StableBaselines library. In the reward calculation, if the threshold adjustment causes the predicted deformation to exceed the allowable range, a penalty is applied.

[0110] The accuracy rate (Acc) of early warnings is the proportion of correct early warnings out of the total number of early warnings. For example, if 80 out of 100 early warnings are correct, then Acc = 80 / 100 = 0.8. The false alarm rate (FPR) is the proportion of cases where there is no risk but are mistakenly identified as risky. For example, if 10 out of 50 cases where there is no risk, then FPR = 10 / 50 = 0.2. The false negative rate (FNR) is the proportion of cases where there is risk but no early warning is issued. For example, if 5 out of 30 cases where there is risk, then FNR ≈ 5 / 30 ≈ 0.167.

[0111] In this embodiment, the state space is the environmental state perceived by reinforcement learning, i.e., the standardized parameter set for each location, such as the parameter set for the front left region; the action space is the adjustment amount of the threshold, such as increasing the low-risk threshold by 0.05. The graded warning threshold is the critical value for classifying risk levels (e.g., initial low risk ≤ 0.5, medium risk 0.5-0.8, high risk > 0.8). The model will dynamically adjust based on the standardized parameter set. For example, if a location has been used for a long time and has been maintained less, the low-risk threshold will be adjusted to ≤ 0.4. After adjustment, the threshold must meet the allowable deformation range calculated by the structural mechanics model. For example, if the maximum allowable deformation for a location is ≤ 0.5 after standardization, the predicted deformation after threshold adjustment cannot exceed this value.

[0112] The beneficial effects of the above technical solution are as follows: by standardizing the use, maintenance, and mechanical parameters of each monitoring location through the acquisition unit, and then by dynamically optimizing the early warning threshold through reinforcement learning combined with structural mechanical constraints by the threshold adjustment unit, the accuracy of early warning is improved by utilizing historical and real-time data, and the rationality of the threshold is ensured by mechanical constraints. Ultimately, the accuracy of early warning is maximized and the false alarm and missed alarm rates are minimized, providing a more intelligent and reliable threshold determination basis for early warning of deformation risk of experimental bench cabinet.

[0113] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets. The preprocessing module includes:

[0114] The theoretical determination unit is used to calculate the theoretical response vector of load-deformation at each label location at each historical moment based on the structural mechanics model of the experimental bench cabinet.

[0115] The data determination unit is used to set the monitoring positions of the experimental bench base cabinet. Where M is the total number of monitoring locations; pj1 is the multi-source data from the j1-th monitoring location. , This refers to the historical detection data for the j1th monitoring location; This refers to the simulated operating condition data for the j1th monitoring location; This represents the actual usage data for the j1th monitoring location; This refers to the real-time parameter data for the j1th monitoring location;

[0116] Vector extraction unit, used for multi-source data at each monitoring location. Standardization processing is performed to obtain standardized data. And extract the multi-source standard vector of the j1th monitoring location at each historical moment;

[0117] The weighting construction unit is used to calculate the deviation between the multi-source standard vector and the corresponding theoretical response vector to determine the high-confidence weight of the j1-th monitoring location. With low reliability weight And construct the weight matrix Wj1=[ for the j1-th monitoring location. , ],in, ,in, Let be the element value of the k-th data source in the multi-source standard vector at the j1-th monitoring location at the t-th historical time. The number of historical moments; For all data sources under the k-th data source Standard deviation; It is a fluctuation penalty term, and ,and When k=1, it corresponds to h; when k=2, it corresponds to s; when k=3, it corresponds to h. Correspondingly, when k=4, and correspond;

[0118] Weight iteration unit, used to weigh all data sources involved The goal is to minimize the value in the data. The fusion weight coefficients of historical detection data, simulated working condition data, actual usage data and real-time parameters under each location label are iteratively optimized using the particle swarm optimization algorithm.

[0119] Tag fusion unit, used to import all tags from the k-th data source. The fusion error Ej1 and the weight vector constitute Matching degree function As a constraint, the optimized fusion weight coefficients of each location label are obtained, and the multi-source fusion data with location labels is output.

[0120] In this embodiment, the fusion error Ej1 is an L×1 column vector composed of the errors of the j1-th monitoring location at L historical moments under the k-th data source. It is an L×1 column vector composed of the exclusive weights of the j1th monitoring position at L historical time points.

[0121] In this embodiment, the structural mechanics model can adopt the beam theory of mechanics of materials, treating the cabinet crossbeam as a simply supported beam. Given the material (e.g., Q235 steel, elastic modulus E=206GPa) and cross-sectional dimensions (e.g., crossbeam cross-section moment of inertia I=105mm4, span L=1000mm), when the central area of ​​the cabinet is subjected to a vertical load of 500N at a certain historical moment, the theoretical deformation is calculated to be 0.3mm using the deflection formula. This value, along with load and other parameters, is standardized to form a theoretical response vector. Using Python combined with structural mechanics formulas, the load and structural parameters at each historical moment are input to calculate the theoretical response vector in batches.

[0122] In this embodiment, the multi-source data for each location includes historical detection data, simulated operating condition data, actual usage data, and real-time parameter data for the corresponding location.

[0123] In this embodiment, the multi-source data for each monitoring location is standardized. For example, in the historical monitoring data of the front left area, the load range is 100-1000N. At a certain moment, the load of 500N is standardized to (500-100) / (1000-100)=0.444 after Min-Max standardization. In the simulated working condition data, the deformation range is 0-1mm. The simulated 0.5mm is standardized to 0.5. In the actual usage data, the equipment weight range is 20-100kg. The standardized value corresponding to a 50kg equipment is (50-20) / (100-20)=0.375. In the real-time parameters, the temperature range is 20-30℃. The current 25℃ is standardized to 0.5. After processing, the multi-source standard vector for each historical moment is extracted, such as [0.444,0.5,0.375,0.5] at a certain moment.

[0124] In this embodiment, the deviation between the multi-source standard vector and the theoretical response vector is calculated, high-confidence weights and low-confidence weights are determined, and a weight matrix is ​​constructed. Assuming the number of historical time points L=10, and the average deviation over the 10 time points is 0.12, then max(0.1,1-0.12)=0.88; if the standard deviation of the data source deviation is 0.05, and the fluctuation penalty term is 0.951, this part contributes 0.88×0.951=0.837.

[0125] In this embodiment, weights [0.3, 0.2, 0.2, 0.3] are assigned to historical detection, simulated operating conditions, actual usage, and real-time parameters. These weights are then iteratively adjusted to optimize the data source. Minimize as much as possible. Implement PSO using Python's pyswarms library, setting the objective function to each data source. Find the minimum value and iteratively optimize the weights.

[0126] In this embodiment, a fusion error Ej1 is introduced as belonging to a certain data source. The bias vector and the matching degree function Mp of the weight matrix Wj1 are used as constraints to optimize the fusion weight coefficients and output multi-source fused data with position labels. Matrix operations are performed using Python to adjust the weights to a reasonable matching degree, ultimately obtaining the multi-source fused data.

[0127] In this embodiment, when hour, ,otherwise, .

[0128] In this embodiment, .

[0129] The beneficial effects of the above technical solution are as follows: by providing theoretical reference through structural mechanics model, standardization and deviation analysis are performed on multi-source data from historical testing, simulated working conditions, actual use, and real-time monitoring. Combined with particle swarm optimization and intelligent optimization of fusion weights with matching degree constraints, the consistency between data and mechanical laws is ensured, while reducing data fluctuation interference. This provides a high-quality and highly reliable standardized dataset for subsequent model training, effectively supporting the accuracy of deformation prediction and risk warning.

[0130] This invention provides an AI-based intelligent load prediction and deformation risk early warning system for laboratory bench cabinets. The prediction and early warning module further includes:

[0131] Evolution determination units are used to construct a deformation evolution dynamics model with position labels based on the structural mechanics model of the experimental platform cabinet:

[0132] ,in, Let J be the vector of the rate of change of deformation at the j1th monitoring location; The structural inherent mechanical matrix of the j1th monitoring location; Let be the load-deformation transfer matrix for the j1th monitoring location; This represents the real-time load vector at the j1th monitoring location; Let J be the attenuation vector of the structural mechanical parameters at the j1-th monitoring location; Let J be the parameter attenuation-deformation influence matrix for the j1th monitoring location; This is the specific state vector of the j1th monitoring position;

[0133] The predicted deformation value of the base cabinet with location tags is used as the initial state. By combining real-time acquired load data and structural mechanical parameter attenuation model, the deformation evolution dynamic model is solved using an adaptive step-size numerical integration algorithm to obtain the deformation development trend curve of the j1th monitoring position within a preset time period in the future.

[0134] Preferred options also include:

[0135] The prediction unit is used to calculate the intersection of the deformation development trend curve and the safety threshold surface based on the safety threshold surface of the structural mechanics model, and to determine the prediction time window for deformation exceeding the limit.

[0136] In this embodiment, It is a three-dimensional vector, including: deflection, rotation angle, and strain. It is a 3×3 structure matrix. The load vector is a 2D vector, including: vertical load and horizontal load. It is a 3×2 transfer matrix. ,in, This is the initial elastic modulus; The attenuation coefficient is determined by material aging experiments, where t is the service life in years; This is the initial Poisson's ratio.

[0137] The safety threshold surface is a two-dimensional surface defined by the location xu and the usage duration t, where the threshold is the threshold value at a specific point (xu,t). ,in, For position xu, based on the material yield strength under usage time t, the attenuation model is the same. , The moment of inertia of the cross section; The maximum bending moment; This is the distance from the centroid of the cross section to the edge.

[0138] In this embodiment, It is the rate of change of deformation at a certain monitoring location, such as a three-dimensional vector composed of the rate of change of deflection, rotational speed, and strain in the front left region of the cabinet. The 3×3 matrix that reflects the mechanical properties of the cabinet structure itself is determined by the structural geometry and material properties. For example, the stiffness matrix of the front left region can be extracted through finite element analysis. Describe the transmission relationship of vertical and horizontal loads to deformation. For example, for every 1 N increase in vertical load, the deflection increases by 0.001 mm, the rotation angle increases by 0.0001°, and the strain increases by [missing information]. For every 1N increase in horizontal load, the deflection increases by 0.0005mm, the rotation angle increases by 0.0002°, and the strain increases by [missing information]. Based on this, a matrix is ​​constructed. A two-dimensional vector containing vertical and horizontal loads, such as the front left region currently bearing a vertical load of 500N and a horizontal load of 80N, the vector is [500, 80].

[0139] In this embodiment, E0 is the initial elastic modulus (e.g., 206 GPa for Q235 steel initially), and k0 is the decay coefficient (obtained from material aging experiments, e.g., k0 = 0.05 per year).

[0140] In this embodiment, A matrix describing the effect of material parameter decay on deformation, such as how a 1 GPa decrease in elastic modulus results in a 0.0002 mm increase in deflection, a 0.00001° increase in rotation angle, and a 0.00001° increase in strain. Construct a 3×1 matrix.

[0141] In this embodiment, the numerical integration algorithm with adaptive step size adopts the RK45 algorithm.

[0142] In this embodiment, the intersection point of the deformation development trend curve and the safety threshold surface is determined by comparing the deformation trend curve obtained from the evolution determination unit (e.g., the curve showing the deflection increasing from 0.4 mm to 0.8 mm in the next 72 hours) with the safety threshold surface (e.g., the safety deflection threshold curve changing over time at that location) to find the point of intersection. For example, if the trend curve shows a deflection of 0.72 mm after 48 hours, this time is the starting point of the predicted time window for deformation exceeding the limit. Using tools such as Python's scipy.optimize.root, the functions of the deformation trend curve and the safety threshold surface are numerically solved to determine the time of exceeding the limit.

[0143] The beneficial effects of the above technical solution are: by combining structural mechanics and material attenuation law, a dynamic model is constructed to accurately predict the future deformation trend of each position of the cabinet; based on the safety threshold surface, the time window for deformation exceeding the limit is determined in advance. The two work together to achieve a forward-looking early warning of the cabinet deformation risk, providing a basis for timely maintenance or adjustment measures, and effectively improving the safety, reliability and service life of the experimental bench cabinet.

[0144] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based experimental bench base cabinet intelligent load prediction and deformation risk early warning system, characterized in that, The method comprises the following steps: An intelligent layout module is used to dynamically adjust an initial layout based on a finite element mechanics model of a bench cabinet of an experiment table and a clustering result according to historical load distribution, and to collect real-time parameters of the bench cabinet of the experiment table, wherein the real-time parameters comprise environmental parameters and load parameters, deformation parameters and equipment mechanics action parameters with position tags; A preprocessing module is used to perform weighted fusion on historical detection data with position tags, simulated working condition data, actual use data and real-time parameters of the bench cabinet of the experiment table, to adaptively adjust weight coefficients of different data sources in combination with a structural mechanics model of the bench cabinet of the experiment table and a particle swarm optimization algorithm, and to perform preprocessing to output a multi-source fusion standardized data set; A model training module is used to construct a CNN-LSTM fusion model with attention mechanism integrated with structural mechanics prior, and to perform model training based on the multi-source fusion standardized data set; A prediction and early warning module is used to input preprocessed real-time parameters into the trained model to output bench cabinet deformation prediction values with position tags within a future preset time length, to perform standardized processing on use time lengths, maintenance records and structural mechanics parameters of each monitoring position of the bench cabinet of the experiment table, to dynamically adjust graded early warning thresholds, to compare the graded early warning thresholds with the bench cabinet deformation prediction values with position tags to trigger corresponding early warnings, and to determine a deformation development trend of the bench cabinet of the experiment table based on the bench cabinet deformation prediction values with position tags based on the structural mechanics model and to output the deformation development trend. The CNN-LSTM fusion model comprises: An input layer is used to receive a multi-source fusion standardized feature vector with position tags based on the multi-source fusion standardized data set; An attention layer is used to determine the importance of the labeled force characteristics matrix of the bottom cabinet of the experimental bench based on the labeled force characteristics matrix An attention weight is assigned to each dimension of the multi-source fusion standardized feature vector. CNN sublayer, comprising 2 convolution layers and 1 pooling layer, is used for convolution operation The spatial features of the load distribution with position labels, structural stress and deformation are extracted, for convolution operation; 、 for convolution kernel; 、 for bias term; for activation function; for pooling operation; for input feature; A feature fusion layer is used to perform multi-modal fusion on spatial features with position tags output by a CNN sublayer, weighted features output by an attention layer and theoretical deformation features with position tags calculated based on a structural mechanics model; An LSTM sublayer comprises two hidden layers and is used to capture time sequence dependence of load-deformation relationship with position tags, long-term influence of equipment mechanics action changes and dynamic laws of mechanics parameter attenuation; An output layer is used to output bench cabinet deformation prediction values with position tags within a future preset time length.

2. The AI-based experimental bench base cabinet intelligent load prediction and deformation risk early warning system according to claim 1, characterized in that, The intelligent layout module comprises: A load acquisition unit is used to acquire historical load data of the bench cabinet of the experiment table, wherein the historical load data comprises load values of different regions of the bench cabinet at each monitoring time and corresponding time stamps; A clustering unit is used to perform clustering on the historical load data by using a density clustering algorithm to obtain a plurality of load distribution clustering clusters, each clustering cluster corresponding to a typical load distribution mode, and to calculate stress concentration regions and deformation sensitive regions of the bench cabinet under the load distribution mode corresponding to each clustering cluster based on a finite element mechanics model of the bench cabinet of the experiment table; The function construction unit is configured to construct a multi-objective optimization model with the maximum monitoring coverage of the key areas corresponding to the cluster groups by the sensor pairs and the minimum redundancy among the sensors as the target, wherein a target function of the multi-objective optimization model is: wherein, is the number of sensors; is the number of cluster groups; is a weight coefficient of a key area of a u1th cluster group by a u0th sensor, is a coverage of a key area of a u1th cluster group by a u0th sensor; is a redundancy of a u2th sensor to the u0th sensor. A scheme output unit is used to output an adjusted sensor layout scheme according to a solution result of the multi-objective optimization model. 3.The AI-based intelligent load prediction and deformation risk early warning system for lab bench base cabinets according to claim 1, characterized in that, The prediction and early warning module comprises: An acquisition unit is used to perform standardized processing on use time lengths, maintenance records and structural mechanics parameters of each monitoring position to obtain a standardized parameter set of each monitoring position, wherein one monitoring position corresponds to one unique position tag. A threshold adjusting unit is configured to construct a reinforcement learning model fused with structural mechanics constraints of a test bench bottom cabinet, to maximize a warning accuracy and minimize false and missed alarm rates as a reward function, to dynamically adjust a hierarchical warning threshold under each position label based on a standardized parameter set corresponding to each position label, and to satisfy an allowable deformation range constraint calculated by a structural mechanics model of a corresponding monitoring position. The state space of the reinforcement learning model is a set of normalized parameters of each position label, the action space is an adjustment amount of the graded early warning threshold of each position label, and the reward function is ; wherein, is the early warning accuracy rate; is the false alarm rate; is the missed alarm rate; , , is a weight coefficient, and , 3. 4.The AI-based intelligent load prediction and deformation risk early warning system for lab bench base cabinets according to claim 1, characterized in that, The preprocessing module comprises: A theoretical determination unit is configured to calculate a theoretical response vector of a load-deformation at each historical time based on a structural mechanics model of the test bench bottom cabinet. A data determination unit is configured to set a monitoring position set of a bench bottom cabinet , wherein M is the total number of monitoring positions; pj1 is multi-source data of the j1th monitoring position , is historical detection data of the j1th monitoring position; is simulation working condition data of the j1th monitoring position; is actual use data of the j1th monitoring position; is real-time parameter data of the j1th monitoring position; a vector extraction unit configured to extract multi-source data of each monitoring position performing standardization processing to obtain standardized data and extracting the multi-source standardized vector of the j1th monitoring position at each historical time The weighting construction unit is used to calculate the deviation between the multi-source standard vector and the corresponding theoretical response vector to determine the high-confidence weight of the j1-th monitoring location. With low confidence weight And construct the weight matrix Wj1=[ for the j1-th monitoring location. , ],in, ,in, Let be the element value of the k-th data source in the multi-source standard vector at the j1-th monitoring location at the t-th historical time. The number of historical moments; For all data sources under the k-th data source Standard deviation; It is a fluctuation penalty term, and ,and When k=1, it corresponds to h; when k=2, it corresponds to s; when k=3, it corresponds to h. Correspondingly, when k=4, and correspond; a weight iteration unit configured to iteratively optimize the fusion weight coefficients of the historical detection data, the simulated working condition data, the actual use data and the real-time parameters under each position label in combination with a particle swarm optimization algorithm, with the minimum value in the following formula as the target ​ a label fusion unit for introducing a matching degree function of the fusion error Ej1 composed of the fusion error Ej1 of all the data sources under the kth data source and the weight vector As a constraint condition, the fusion weight coefficient of each position label after optimization is obtained, and the multi-source fusion data with position labels is output.​​ 5.The AI-based intelligent load prediction and deformation risk early warning system for lab bench base cabinets according to claim 1, characterized in that, The mechanical characteristic importance degree matrix wherein, is a theoretical deformation value of the structural mechanical model at the j1th monitoring position; is is a partial derivative matrix of ; is a geometric structure weight matrix of the j1th monitoring position; is an element-by-element multiplication; is a normalization function. 6.The AI-based intelligent load prediction and deformation risk early warning system for lab bench base cabinets according to claim 1, characterized in that, The prediction and warning model further comprises: An evolution determination unit is configured to construct a deformation evolution dynamics model with position labels based on a structural mechanics model of the test bench bottom cabinet. wherein, is the rate of change of deformation vector at the j1th monitoring location; is the structural inherent mechanical matrix at the j1th monitoring location; is the load-deformation transfer matrix at the j1th monitoring location; is the real-time load vector at the j1th monitoring location; is the structural mechanical parameter decay vector at the j1th monitoring location; is the parameter decay-deformation influence matrix at the j1th monitoring location; is the configuration-specific state vector at the j1th monitoring location; Using the position-labeled base cabinet shape change prediction value as an initial state , combined with real-time collected load data and structural mechanics parameter attenuation model, using adaptive step numerical integration algorithm to solve the shape evolution dynamic model, get the deformation trend curve of the j1 monitoring position in the future preset length. 7.The AI-based intelligent load prediction and deformation risk early warning system for lab bench base cabinets according to claim 6, characterized in that, Further comprising: A pre-judgment unit is configured to calculate an intersection of the deformation development trend curve and a safety threshold surface based on a safety threshold surface of the structural mechanics model, and to determine a pre-judgment time window of deformation overrun.

Images