An AI-based mobile beam base installation method

Abstract

The application relates to the technical field of machine learning, in particular to an AI-based mobile beam base installation method, which comprises the following steps: embedded sensors acquire node mechanical parameters, AI predicted values and measured values are subjected to vector difference, a sliding window stores a difference value sequence, a derivative generates a direction sequence, K-means clustering outputs a state label, a state mapping table is established, an AI path frequency is adjusted, an LSTM compresses a calculation interval, a high-frequency node set is generated, a stress gradient matrix is constructed in combination with stress data, a random forest identifies an anomaly, outputs an abnormal coordinate, and is corrected through back propagation. In the application, embedded strain sensors are used to collect mechanical parameters in real time, AI prediction is used to construct a dynamic difference vector, a sliding window is used to capture state changes, a directional derivative and clustering are used for intelligent classification, state mapping is used for accurate identification, a calculation frequency interval is dynamically adjusted, high-frequency analysis is used for key nodes, stress gradient is used for auxiliary judgment, a random forest is used for anomaly identification and reverse correction, installation precision is improved, and system stability and coordination are enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of machine learning, in particular to an AI-based mobile beam base installation method. BACKGROUND

[0002] The technical field of machine learning relates to the development of intelligent algorithms for pattern recognition and decision optimization through data training of models, mainly including technical links such as feature extraction, parameter optimization, and prediction analysis, and is widely used in the field of industrial equipment automatic control. Among them, the traditional mobile beam base installation method refers to the spatial positioning and mechanical load matching problem of the mobile beam assembly in heavy machinery processing equipment. Laser range finder is used to collect base plane coordinate data, combined with manual calculation of mechanical distribution parameters, the height of the base support point is adjusted step by step through the hydraulic adjusting device, and the torque wrench is used to tighten the connecting bolts step by step.

[0003] The traditional method relies on laser range finder to obtain plane coordinate data, and its single-dimensional spatial positioning method is difficult to capture the dynamic coupling relationship of multi-dimensional mechanical parameters. The manual calculation of mechanical distribution parameters has subjective experience deviation and calculation efficiency bottleneck, which limits the mechanical matching accuracy to the professional level of the operator. The hydraulic step-by-step adjustment strategy lacks real-time feedback mechanism, and in the process of cooperation of the base support points, local overload or stress imbalance is easy to occur. The discrete operation mode of step-by-step tightening of the bolts is difficult to realize the synchronous optimization of the overall mechanical state, and the data fragmentation in each link of the installation process causes abnormal response lag, which easily leads to the decrease of the coordination of the base support system under the complex load working condition of heavy equipment, increases the cost and safety hazards of later debugging. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to provide an AI-based mobile beam base installation method.

[0005] In order to achieve the above purpose, the application adopts the following technical scheme: an AI-based mobile beam base installation method, comprising the following steps:

[0006] S1: Obtain the bending moment, shear force, deformation angle mechanical parameters of the base connection node through the embedded strain sensor, perform vector subtraction operation on the AI prediction value and the measured value, construct the node difference vector, and store the node difference vector set of continuous time sequence using a sliding window;

[0007] S2: Call the node difference vector set, perform first-order derivative operation on the adjacent period vectors to generate a direction derivative sequence, input it into the K-means clustering algorithm for trend classification, output the convergence state, oscillation state and divergence state labels, and establish a node trend state mapping table;

[0008] S3: Based on the oscillation state and divergence state labels in the node trend state mapping table, a frequency adjustment coefficient is set for the AI model path node, an LSTM network is used for dynamic compression calculation of the trigger interval, a high-frequency calculation node set is generated, and the calculation interval of the convergent state node is expanded to 1.5 times of the baseline value;

[0009] S4: The high-frequency calculation node set and stress sensor data are called to construct a spatial stress gradient vector matrix, the length difference of adjacent nodes is matched in space, input into a random forest model to identify abnormal amplitude, output an abnormal coordinate set, and perform a bias term back propagation correction.

[0010] As a further scheme of the application, the node difference vector set includes a timestamp index, a vector module length, and a direction angle component, the node trend state mapping table specifically includes node number, trend type, and direction derivative mean combination data, and the high-frequency calculation node set includes a node identifier, a trigger frequency coefficient, and an LSTM weight parameter.

[0011] As a further scheme of the application, the size of the sliding window is 5 consecutive timestamps, and the sliding step is 1 timestamp.

[0012] The time interval of the adjacent period vector is 2 times the sensor sampling period.

[0013] The baseline value is an initial calculation interval of 200 ms, and the expansion coefficient is based on a stress fluctuation stability threshold obtained through laboratory pressure testing.

[0014] The bias correction amount is the product of the abnormal amplitude value and the weight matrix, and the weight matrix elements reflect the spatial correlation strength of the nodes.

[0015] As a further scheme of the application, the node difference vector set acquisition step specifically includes:

[0016] S101: Collecting three-axis strain signals of embedded strain sensors, integrating cross-section stress distribution based on material elastic modulus parameters and cross-section inertia parameter to obtain a bending moment component, converting shear force components through a shear force transfer function, calculating deformation angle components based on node geometric size parameters, and generating node mechanical parameter measured values;

[0017] S102: Extracting bending moment prediction values, shear force prediction values, and deformation angle prediction values output by the AI model, respectively performing vector component subtraction operations with the node mechanical parameter measured values, calculating multiple component absolute difference values, constructing three-dimensional vector space coordinates in the order of bending moment-shear deformation angle, and generating node mechanical difference vectors.

[0018] S103: The node mechanical difference value vector in the timestamp sequence is aligned, data is intercepted by using a fixed time window, a window range is updated by sliding in a preset step length, continuous vector data points in the window are stored, and a node difference value vector set is generated.

[0019] As a further scheme of the present application, the material elastic modulus parameter is determined by the ASTM E111 standard.

[0020] The shear transfer function is , wherein, represents a shear transfer function value calculated by integration, represents a shear stress function distributed along a cross section, represents an integral variable along a cross section position coordinate, represents a shear stress value at a cross section position coordinate , represents a differential unit along a cross section position coordinate, and an integral symbol represents an operation of accumulating and summing shear stresses along the entire cross section;

[0021] The fixed time window has a duration of 10 seconds, and the preset step length is 2 seconds.

[0022] As a further scheme of the present application, the step of obtaining the node trend state mapping table specifically comprises:

[0023] S201: Adjacent time window vector data in the node difference value vector set is called, a change rate of a plurality of vector components on a time axis is calculated, a first derivative value is calculated by using a central difference method, and a directional derivative sequence is generated;

[0024] S202: A cluster algorithm category number parameter is set, the directional derivative sequence is taken as a feature input, an Euclidean distance between a sample point and an initial cluster center is calculated, a centroid coordinate is iteratively updated until a distance square sum converges, a class number to which the sample point belongs is output, and a trend classification label is generated;

[0025] S203: A correspondence between a timestamp sequence and the trend classification label is established, a class number and a state type mapping rule are defined, a two-dimensional relationship table of a time interval field and a state type field is constructed, and a node trend state mapping table is generated;

[0026] The category number parameter is set to 3, and respectively corresponds to a convergence state, an oscillation state, and a divergence state.

[0027] As a further scheme of the present application, the step of S3 specifically comprises:

[0028] S301: Extract state duration values and amplitude values corresponding to oscillation state and divergence state labels based on the node trend state mapping table, assign weight coefficients to the two types of labels respectively, and calculate a node frequency adjustment coefficient through a linear superposition formula;

[0029] S302: Call the node frequency adjustment coefficient input LSTM network to build a time series prediction model, capture node state change trends using a sliding window mechanism, dynamically adjust the forgetting gate weight through a gating unit, output a plurality of node trigger interval compression ratios, and generate a dynamic trigger interval sequence in combination with a reference interval value;

[0030] S303: According to the dynamic trigger interval sequence, filter nodes with a compression ratio higher than a set threshold to form a high-frequency computing node set, synchronously extract a convergent state node set, perform scalar multiplication operation on the corresponding trigger interval and the reference value, update the calculation interval parameter using an expansion coefficient, and generate a node calculation interval configuration set.

[0031] As a further scheme of the application, the weight coefficient value range is 0.5-1.2, which is determined by gradient descent method optimization;

[0032] The reference interval value is 200ms, which is derived from the clock cycle configuration of the embedded system;

[0033] The set threshold is a compression ratio exceeding 30%, which is dynamically adjusted according to real-time calculation resource load.

[0034] As a further scheme of the application, the abnormal coordinate set acquisition step is specifically:

[0035] S401: Call the high-frequency computing node set and stress sensor data to build a three-dimensional space coordinate system, calculate the stress gradient vector module length difference between adjacent nodes, and establish a spatial stress gradient difference matrix;

[0036] S402: Input the spatial stress gradient difference matrix into the random forest model, set the decision tree split threshold, evaluate feature importance through the Gini coefficient, filter coordinate points with an abnormal amplitude exceeding a set threshold, and generate an abnormal amplitude feature set;

[0037] S403: According to the abnormal amplitude feature set, execute the back propagation algorithm, calculate the gradient change of the bias term, adjust the weight update step size using the momentum factor, iteratively correct the network parameters, and generate an abnormal coordinate set.

[0038] As a further scheme of the application, the decision tree split threshold is 1.2 times the abnormal amplitude set threshold;

[0039] The momentum factor value range is 0.8-0.95, which is linearly attenuated according to the number of iterations.

[0040] Compared with the prior art, the application has the advantages and positive effects that:

[0041] In the application, multi-dimensional mechanical parameters are collected in real time by the embedded strain sensor, and vector operation is performed on the AI prediction value to construct a dynamic difference vector set, the sliding window mechanism is combined to capture the mechanical state changes in the continuous time sequence, the directional derivative sequence and the clustering algorithm are used to intelligently classify the node trend, the state mapping table is established to realize accurate identification of different mechanical states, the calculation frequency and interval are dynamically adjusted based on the state label, high-frequency calculation and spatial stress gradient analysis are implemented at key nodes, the random forest model is used to identify abnormal amplitude and perform reverse correction, the mechanical parameter matching accuracy during the base installation process is effectively improved, dynamic adaptive adjustment of the installation process is realized, response lag caused by manual intervention is reduced, error accumulation phenomenon in step-by-step adjustment is avoided, and the stability and coordination of the base support system under complex working conditions are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 It is a total flowchart of the AI-based mobile beam base installation method of the application.

[0043] Figure 2 It is a schematic diagram of the node difference vector set acquisition process of the application.

[0044] Figure 3 It is a schematic diagram of the node trend state mapping table establishment process of the application.

[0045] Figure 4 It is a schematic diagram of the calculation node dynamic adjustment process of the application.

[0046] Figure 5 It is a schematic diagram of the abnormal coordinate identification and correction process of the application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0048] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0049] Embodiment one: please refer to Figure 1 and Figure 2 The present application provides a technical solution: an AI-based mobile beam base installation method, comprising the following steps:

[0050] S1: Obtain the bending moment, shear force and deformation angle mechanical parameters of the base connection node through the embedded strain sensor, perform vector subtraction operation on the AI prediction value and the measured value, construct the node difference vector, and store the node difference vector set of continuous time sequence using a sliding window;

[0051] The node difference vector set includes a time stamp index, a vector module length and a direction angle component;

[0052] The size of the sliding window is 5 continuous time stamps, and the sliding step is 1 time stamp;

[0053] The step of obtaining the node difference vector set is specifically:

[0054] S101: Collect the three-axis strain signals of the embedded strain sensor, integrate the cross-section stress distribution to obtain the bending moment component in combination with the material elastic modulus parameter and the cross-section moment of inertia parameter, convert the shear force component through the shear force transfer function, calculate the deformation angle component according to the node geometric size parameter, and generate the measured value of the node mechanical parameter;

[0055] The material elastic modulus parameter is determined by the ASTM E111 standard;

[0056] The shear force transfer function is , wherein, represents the shear force transfer function value obtained by integration, represents the shear stress function along the cross-section distribution, represents the integral variable along the cross-section position coordinate, represents the shear stress value at the cross-section position coordinate , represents the differential unit along the cross-section position coordinate, and the integral symbol This represents the operation of summing the shear stress over the entire cross section;

[0057] The fixed time window lasts for 10 seconds, with a preset step size of 2 seconds.

[0058] S101: The moving beam base is the connection base between the moving crossbeam of a large gantry crane and the ground support column. This base node is numbered ZN-03. First, an embedded fiber Bragg grating (FBG) strain sensor array, pre-arranged inside and on the surface of node ZN-03 along the three principal stress directions (x, y, z axes), is used to collect triaxial strain signals during the gantry crane's moving operation. At time stamp T1 (e.g., 10:00:00 AM), the collected measured triaxial strain values ​​are as follows: , , .

[0059] The determination of the material's elastic modulus parameter strictly followed the ASTM E111 standard test method. Standard specimens were prepared using the same material as the moving beam base (e.g., Q355B steel) and subjected to uniaxial tensile tests on a universal testing machine. During the test, the deformation of the specimen under different loads was precisely measured using an extensometer, and multiple sets of stress-strain data points were recorded. For example, when a stress of 100 MPa was applied, the measured strain was 488 μm. When a stress of 150 MPa is applied, the measured strain is 731. When a stress of 200 MPa is applied, the measured strain is 975. Plotting these data points on a stress-strain coordinate system, a linear regression is performed on the data from the initial linear elastic stage. The slope of the resulting straight line is the elastic modulus E of the material. Based on multiple sets of measurement data, the average elastic modulus is calculated as follows: MPa. This value is a fixed constant that is used continuously in subsequent calculations.

[0060] The moment of inertia parameters of the cross section are determined based on the specific geometric design drawings of node ZN-03. The cross section of this node is I-shaped, and its moment of inertia about the neutral axis (x-axis) is... It is calculated using standard geometric formulas. For example, for an I-beam section with a web height of 400mm, a flange width of 200mm, and both the web and flange thicknesses of 10mm, its moment of inertia is... The calculated value is This value is also used as a definite geometric constant in the calculation.

[0061] Using the acquired strain signal Using the aforementioned material and geometric parameters, an integral calculation is performed on the stress distribution across the cross-section. First, the normal stress at a point on the outer edge of the cross-section is calculated according to Hooke's Law. MPa. Subsequently, based on beam bending theory, the bending moment components... pass Calculation, where This is the distance (200mm) from the outer edge of the cross-section to the neutral axis. Substitute the values... N·mm, which is 84.7 kN·m. Similarly, calculate the bending moment components in the other two directions. and Finally, a vector of measured bending moment values ​​is generated.

[0062] The shear force components are obtained through a shear force transfer function. The shear force transfer function is described here. This refers to integrating the shear stress distribution across the cross section. In practice, the shear force value is first calculated from the rate of change of the bending moment along the beam length. For simplicity, the formula is presented here. Explained as shear stress In cross-sectional area Points on Shear stress The distribution is based on Calculation, where Let be the area moment of the region above the point in question about the neutral axis. Let be the cross-sectional width at that point. We convert the shear force to bending moment using a pre-established shear force-bending moment relationship model (shear force transfer function) in finite element analysis software, inputting a bending moment value of 84.7 kN·m, and obtain the corresponding shear force components. It is 55.2 kN.

[0063] The deformation angle component is calculated based on the node's geometric dimensions. The node's geometric dimensions, such as the length L (e.g., 500 mm) and thickness H (e.g., 80 mm) of the connecting plate, are pre-stored in the system. Deformation angle It is approximately equal to the ratio of the deflection to the length of the node connection plate under bending moment. The deformation angular component is obtained using material mechanics formulas. Depend on Calculate. Substitute the values. radian.

[0064] Through the above steps, at timestamp T1, a vector of measured values ​​of nodal mechanical parameters was generated. To simplify the subsequent explanation, we will only take one principal component of bending moment, shear force, and deformation angle as an example, that is, the measured value vector is (84.7kN·m, 55.2kN, 0.0125rad).

[0065] S102: Extract the bending moment prediction value, shear force prediction value, and deformation angle prediction value output by the AI model, and perform vector component subtraction operation with the measured values of the node mechanical parameters respectively, calculate multiple component absolute difference values, construct a three-dimensional vector space coordinate point in the order of bending moment-shear force-deformation angle, and generate a node mechanical difference vector;

[0066] S102: At the same timestamp T1, the AI model outputs its predicted mechanical parameters based on historical data and current working conditions. For example, the bending moment prediction value is 82.5 kN·m, the shear force prediction value is 56.0 kN, and the deformation angle prediction value is 0.0120 rad. Then, perform vector component subtraction operation to calculate the absolute difference value of each component. The bending moment difference is kN·m. The shear force difference is kN. The deformation angle difference is rad. According to the order of bending moment-shear force-deformation angle, construct a three-dimensional vector space coordinate point, and thus generate the node mechanical difference vector at T1 .

[0067] S103: Align the node mechanical difference vectors in the timestamp sequence, use a fixed time window to cut data, update the window range by a preset step size, store the continuous vector data points in the window, and generate a node difference vector set.

[0068] S103: The duration of the fixed time window is set to 10 seconds, and the preset step size is 2 seconds. This means that the system will cut continuous data from T1 to T1+10 seconds. The size of the sliding window is 5 consecutive timestamps, and the sliding step size is 1 timestamp. Assuming the sensor sampling frequency is 1 Hz, 10 node mechanical difference vectors will be collected in the 10-second fixed window. The sliding window will process these 10 vectors. For example, the first window contains vectors from T1 to T5 When T6 data comes in, the window slides and contains vectors from T2 to T6 . Store the continuous vector data points in each sliding window, and these stored data sets constitute the node difference vector set. Each element in the set contains a timestamp index, a difference vector itself (containing three components), a vector magnitude calculated from the vector components (for example, the magnitude at T1 is ), and a direction angle component (for example, the angle with the bending moment axis ).

[0069] Please refer to Figure 1 and Figure 3 , S2: Call the node difference vector set, perform first-order derivative operation on adjacent period vectors to generate a direction derivative sequence, input to the K-means clustering algorithm for trend classification, output convergence state, oscillation state, and divergence state labels, and establish a node trend state mapping table;

[0070] The node trend state mapping table is specifically a combination of node number, trend type, and average value of directional derivative;

[0071] The time interval of the adjacent time period vector is twice the sensor sampling period;

[0072] The obtaining step of the node trend state mapping table is specifically:

[0073] S201: Call the adjacent time window vector data in the node difference vector set, calculate the change rate of multiple vector components on the time axis, calculate the first derivative value by using the central difference method, and generate a directional derivative sequence;

[0074] S201: Accept the node difference vector set generated by S1, which stores the difference vectors of ZN-03 node at consecutive time stamps. For example, call the data in a sliding window from the set, which contains five consecutive time stamp (T1 to T5) difference vectors, respectively , , , , . The time interval of the adjacent time period vector is set to twice the sensor sampling period. If the sensor sampling period is 1 second, the time interval is 2 seconds.

[0075] The system calls the vector data of the adjacent time window in the node difference vector set. Here, the change of the vector in the window over time is concerned. To calculate the directional derivative, the central difference method is used. This method uses the information of two points before and after a point to calculate the derivative of the point, which can provide more accurate approximation than forward or backward difference. Taking the calculation of the directional derivative at time stamp T3 as an example, the vector at the previous time stamp T2 and the vector at the next time stamp T4 are needed. The time interval is two sampling periods, i.e. 2 seconds. For the bending moment component, its change rate at T3 is kN·m / s. For the shear force component, its change rate at T3 is kN / s. For the deformation angle component, its change rate at T3 is rad / s. Thus, the directional derivative vector at T3 is generated . Repeat this process to calculate the directional derivatives of all points that can be calculated in the window, for example , , , forming a directional derivative sequence.

[0076] S202: Set the number of cluster algorithm category parameters, input the directional derivative sequence as the feature, calculate the Euclidean distance between the sample point and the initial cluster center, iteratively update the centroid coordinates until the distance square converges, output the sample point category number, and generate the trend classification label;

[0077] The number of categories is set to 3, corresponding to convergence state, oscillation state, and divergence state respectively;

[0078] S202: The directional derivative sequence generated by S201, such as , is input as a feature for trend classification. The number of categories parameter K is preset to 3, which is mapped to “convergence state”, “oscillation state” and “divergence state” respectively. The execution action of the clustering process is as follows: First step: initial centroid selection. Randomly select 3 vectors from the directional derivative sequence as the initial cluster centers. For example, assume the initial centroid is , , , representing the initial divergence, oscillation, and convergence trends respectively. Second step: sample assignment. Calculate the Euclidean distance between each directional derivative vector (sample point) in the sequence and the 3 initial centroids. Take as an example: the distance square sum from to

[0079] .

[0080] to :

[0081] .

[0082] to :

[0083] .

[0084] Since the distance from to is the smallest, assign to category 1. Repeat this calculation and assignment process for all other sample points in the sequence. Third step: update the centroid. After all sample points are assigned, recalculate the centroid of each category. The new centroid is the average of all vectors in that category. For example, if category 1 is finally assigned to and , the new centroid is . Fourth step: iteration. Repeat the second and third steps, i.e. constantly reassign sample points and update centroid coordinates. When the iteration is over, the category assignment of all sample points no longer changes, or the change in all centroid coordinates is less than a small convergence threshold (e.g. ) is reached, the iteration process stops, and the distance square sum reaches convergence. At this time, the final class number (1, 2, or 3) to which each sample point belongs is output, generating a trend classification label.

[0085] S203: Establish the correspondence between the timestamp sequence and the trend classification label, define the class number and state type mapping rule, construct the two-dimensional relationship table of the time interval field and the state type field, and generate the node trend state mapping table.

[0086] S203: Establish the correspondence between the timestamp sequence and the trend classification label. For example, in the above process, the directional derivative at T3 is classified into class 1. We predefine the mapping rule: class 1 is mapped to "divergent state", class 2 is mapped to "oscillatory state", and class 3 is mapped to "convergent state". Therefore, the trend state at T3 is labeled as "divergent state". All labels within a time interval (e.g. T2-T4) are integrated, and the mean value of the directional derivative vector within the interval is calculated. For example, if the state of T2-T4 is "divergent state", the mean vector of is calculated. Finally, a two-dimensional relationship table, i.e. the node trend state mapping table, is constructed, as shown in Table 1 below.

[0087] Table 1 Node Trend State Mapping Table

[0088]

[0089] As shown in Table 1, the table clearly combines the node number, the time interval of trend occurrence, the trend type output by the clustering algorithm, and the specific quantitative indicator (mean value of directional derivative) of the trend, generating a structured node trend state mapping table.

[0090] Please refer to Figure 1 and Figure 4 , S3: Based on the oscillatory state and divergent state labels in the node trend state mapping table, set the frequency adjustment coefficient for the AI model path nodes, dynamically compress the calculation trigger interval using LSTM network, generate a high-frequency calculation node set, and expand the calculation interval of the convergent state nodes to 1.5 times the baseline value;

[0091] The high-frequency calculation node set includes node identifier, trigger frequency coefficient, and LSTM weight parameter;

[0092] The baseline value is the initial calculation interval of 200ms, and the expansion coefficient is based on the stress fluctuation stability threshold obtained through laboratory stress test;

[0093] The steps of S3 are as follows:

[0094] ​S301: Extract the state duration value and amplitude value corresponding to the oscillation state and divergence state label based on the node trend state mapping table, assign weight coefficients to the two types of labels respectively, and calculate the node frequency adjustment coefficient through a linear superposition formula;

[0095] The weight coefficient value range is 0.5-1.2, which is determined by gradient descent method optimization;

[0096] S301: The system operates based on the node trend state mapping table generated by S2 (see Table 1). Extract the node information labeled as "oscillation state" and "divergence state" from the table. For example, node ZN-03 is labeled as "divergence state" in time interval T2-T4 and as "oscillation state" in T5-T7.

[0097] The system extracts the corresponding duration and amplitude values of these states. The duration is directly obtained from the time interval, for example, the duration of T2-T4 is 3 time stamps. The amplitude value is extracted from the node difference vector set of S1, specifically the maximum value of the difference vector norm in this time interval. Assuming that the maximum value of the difference vector norm in the T2-T4 interval is 3.1, and the maximum value in the T5-T7 interval is 1.5. The system assigns weight coefficients to the two types of labels "divergence state" and "oscillation state" respectively.

[0098] The weight coefficient value range is 0.5 to 1.2. This range is verified through a series of offline experiments. The experimental process is as follows: construct a simulated moving beam base finite element model, apply loads of different properties (stable, periodic, and impact). Run the technical solution and manually adjust the weight combination, record the prediction accuracy and computational resource consumption of the AI model under different weights. The goal is to find a weight range that allows the system to quickly respond when the model deviation increases (corresponding to "oscillation state" and "divergence state"), while not excessively consuming computational resources when the model is stable. The weight is optimized and determined by gradient descent method. Define a cost function, for example where is the prediction error, is the computational cost, is the weight, is the hyperparameter. Calculate the partial derivative of the cost function with respect to the weight by iteration, and update the weight along the negative gradient direction until the cost function converges. Experimental results show that when the weight coefficient is in the range of 0.5 to 1.2, the overall performance of the system is optimal.

[0099] Assign a weight coefficient to "divergence state" , and assign a weight coefficient to "oscillation state" . Then, calculate the frequency adjustment coefficient of node ZN-03 in the T2-T4 time period through a linear superposition formula. The formula is: where is the normalized duration, is the normalized amplitude value, is the hyperparameter (e.g. , indicating that the amplitude has a greater impact). Assuming the maximum duration is 10 timestamps and the maximum amplitude is 5.0, the normalized duration of T2-T4 , the normalized amplitude . The frequency adjustment coefficient of this node is: . Similarly, calculate the frequency adjustment coefficient of T5-T7 "oscillation state" .

[0100] S302: Call the node frequency adjustment coefficient input LSTM network, build a time series prediction model, use the sliding window mechanism to capture the node state change trend, dynamically adjust the forgetting gate weight through the gate unit, output multiple node trigger interval compression ratios, and combine the reference interval value to generate a dynamic trigger interval sequence;

[0101] The reference interval value is 200ms, derived from the clock cycle configuration of the embedded system;

[0102] S302: Input a series of node frequency adjustment coefficients generated by S301, such as , into a trained long short-term memory (LSTM) network in chronological order. The LSTM network is used to predict the short-term trend of node state changes. The reference calculation interval value is set to 200ms, which is derived from the clock cycle configuration of the embedded hardware system and is a stable and reliable hardware parameter. The gating units (input gate, forgetting gate, output gate) inside the LSTM network dynamically adjust their weights using the input frequency adjustment coefficient sequence. For example, when a higher adjustment coefficient value of 0.5412 is input, it indicates that the node state is unstable, and the forgetting gate will reduce its weight to reduce the memory of past stable state information; at the same time, the input gate will increase its weight, allowing the current unstable state information to have a greater impact on the network state. Through this mechanism, the LSTM network can capture the state change trend from stable to divergent or oscillation. The network finally outputs a trigger interval compression ratio. For example, for an input of 0.5412, the LSTM network predicts that the next state will continue to be unstable, and the output compression ratio is 0.6 (i.e. compress 60%). Combine this compression ratio with the reference interval value to generate a dynamic trigger interval: . This forms a dynamic trigger interval sequence.

[0103] S303: According to the dynamic trigger interval sequence, filter out the nodes with compression ratio higher than the set threshold to form a high-frequency computing node set, synchronously extract the convergent state node set, perform scalar multiplication operation on the corresponding trigger interval and the reference value, update the computing interval parameter using the expansion coefficient, and generate a node computing interval configuration set;

[0104] The set threshold is set to be more than 30%, and is dynamically adjusted according to the real-time computing resource load.

[0105] S303: Set a compression ratio threshold, which is dynamically adjusted according to the real-time computing resource load. For example, when the system load is 70%, the threshold is set to 30%. If the computing interval compression ratio of any node is higher than 30%, the node is filtered out to form a high-frequency computing node set. In the example of S302, the compression ratio of node ZN-03 is 60%, which is higher than 30%, so ZN-03 is added to the high-frequency computing node set. The set records the node identifier (ZN-03), the trigger frequency coefficient (i.e. the dynamic trigger interval 80ms), and the LSTM weight parameter for subsequent fine-tuning. At the same time, the system synchronously extracts the node set marked as "convergent state". For example, node ZN-05 is marked as "convergent state" in the interval T8-T10. Its corresponding computing trigger interval will be subjected to scalar multiplication operation with an expansion coefficient. The expansion coefficient is set to 1.5, which is based on laboratory stress test. In the test, a constant load is applied to the base, and the stress fluctuation is monitored. When the stress fluctuation amplitude is continuously below a certain stable threshold (e.g. less than 0.1% of the full scale) for 1 minute, the system state at that time is recorded, and the proportion of computing resources that can be saved is back calculated, and finally it is determined that the expansion coefficient of 1.5 is safe and efficient. Therefore, the computing interval of node ZN-05 is updated to: . Finally, the system generates a complete node computing interval configuration set, which contains the latest computing interval parameters of all nodes (high-frequency, regular, and low-frequency).

[0106] Please refer to Figure 1 and Figure 5 , S4: Call the high-frequency computing node set and the stress sensor data to construct a spatial stress gradient vector matrix, spatially match the adjacent node length difference, input into the random forest model to identify the abnormal amplitude, output the abnormal coordinate set, and perform bias term back propagation correction;

[0107] The abnormal coordinate set specifically refers to the combined data of three-dimensional coordinates, time stamp, and abnormal amplitude;

[0108] The bias correction amount is the product of the abnormal amplitude value and the weight matrix, and the weight matrix elements reflect the spatial correlation strength of the nodes;

[0109] The steps of obtaining the abnormal coordinate set are as follows:

[0110] S401: Call the high-frequency computing node set and the stress sensor data, build a three-dimensional spatial coordinate system, calculate the stress gradient vector length difference between adjacent nodes, and establish a spatial stress gradient difference matrix;

[0111] S401: The system calls the high-frequency computing node set generated by S3. This set contains node ZN-03, whose calculation interval is compressed to 80ms. At the same time, the system obtains the stress sensor raw data of ZN-03 and its adjacent nodes (e.g. ZN-02, ZN-04) at the latest timestamp (e.g. T11, 10:00:11 AM) from the sensor data stream.

[0112] The system first builds a three-dimensional spatial coordinate system with the geometric center of the moving beam base as the origin. According to the design drawings, the coordinates of ZN-02, ZN-03, and ZN-04 are (-0.5, 1.0, 0.0) meters, (0.0, 1.0, 0.0) meters, and (0.5, 1.0, 0.0) meters, respectively. At T11, the principal stress values collected at each node are MPa, MPa, MPa. The system calculates the stress gradient vector between adjacent nodes. For example, the component of the stress gradient between ZN-03 and ZN-02 in the x direction is MPa / m. Similarly, the gradient components in other directions are calculated to form the stress gradient vector. Then the vector length is calculated, for example, the stress gradient vector length between ZN-03 and ZN-02 is MPa / m. The stress gradient vector length between ZN-03 and ZN-04 is MPa / m. Next, the difference between the gradient lengths of these adjacent nodes is calculated. For example, the spatial stress gradient difference at high-frequency node ZN-03 is MPa / m. Repeat this calculation for all high-frequency nodes and their neighborhoods to build a spatial stress gradient difference matrix.

[0113] S402: Input the spatial stress gradient difference matrix into the random forest model, set the decision tree split threshold, evaluate feature importance through the Gini coefficient, filter coordinate points with abnormal amplitude exceeding the set threshold, and generate an abnormal amplitude feature set;

[0114] The decision tree split threshold is 1.2 times the threshold value of the abnormal amplitude;

[0115] S402: input the spatial stress gradient difference matrix generated in S401 into a pre-trained random forest model. The model is composed of multiple decision trees (e.g. 100). The split threshold of the decision tree is set to 1.2 times the abnormal amplitude threshold. The abnormal amplitude threshold itself is set according to the maximum spatial stress gradient difference under normal working conditions, for example 20 MPa / m. Then, the split threshold of the internal node of the decision tree is MPa / m. The model evaluates the importance of the features by Gini Impurity and decides how to split. The calculation method of Gini Impurity is , where is the proportion of samples belonging to class k in the data set. When splitting the node, the algorithm traverses all possible features and split points, and selects the one that minimizes the weighted average of the Gini coefficients of the two child nodes after splitting. For example, a parent node contains 10 gradient difference samples , of which 5 are abnormal (greater than 24) and 5 are normal (less than or equal to 24). The Gini coefficient of this parent node is . If 24 is taken as the split point, the left child node contains (5 normal, 0 abnormal), and its Gini coefficient is . The right child node contains (0 normal, 5 abnormal), and its Gini coefficient is also 0. The weighted Gini coefficient after splitting is . Since the Gini impurity has decreased from 0.5 to 0, this is a good split. Each tree in the random forest makes a judgment on the input gradient difference (e.g. 10 MPa / m). Since 10 MPa / m is much smaller than the split threshold 24 MPa / m, most trees will classify it as "normal". But if the gradient difference of a node is 35 MPa / m, most trees will classify it as "abnormal". The model aggregates the voting results of all trees to finally filter out the coordinate points with abnormal amplitude exceeding the set threshold (20 MPa / m), generating the abnormal amplitude feature set.

[0116] S403: perform back propagation algorithm according to the abnormal amplitude feature set, calculate the gradient change of the bias term, adjust the weight update step size with the momentum factor, iteratively correct the network parameters, and generate the abnormal coordinate set;

[0117] The momentum factor ranges from 0.8 to 0.95 and decays linearly according to the number of iterations.

[0118] S403: Assuming that the gradient difference value of ZN-03 node at T12 time is 35 MPa / m, it is identified as an anomaly by the random forest model. The system performs a back propagation algorithm according to this anomaly amplitude feature set to correct the bias term of the AI prediction model. The calculation of the gradient change of the bias term is directly related to the anomaly amplitude value. The value range of the momentum factor (Momentum Factor) is 0.8 to 0.95. The setting basis is: a larger momentum (such as 0.95) is used in the early stage of iteration to accelerate convergence, and as the iteration number increases, a linear decay strategy is adopted to use a smaller momentum (such as 0.8) in the later stage of iteration to prevent overshooting the optimal point. The specific correction process is: first, calculate the gradient of the bias term wherein is the learning rate. Then, update the weight in combination with the momentum: wherein is the momentum factor, is the current speed term, is the speed term at the last time. For example, assuming that the momentum factor is 0.9, the speed term at the last time is 0.1, the anomaly amplitude is 35, the relevant weight is 0.1, and the learning rate is 0.01, then the gradient is 0.215. The current speed term is 0.1. The new bias correction amount is -0.215. This correction amount is back propagated to the AI prediction network to adjust the bias parameter of the corresponding node. After multiple iterations of correction, a combined data including three-dimensional coordinates, time stamps, and anomaly amplitudes is finally generated, that is, an anomaly coordinate set, for example .

[0119] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments still belong to the protection scope of the present application technical solution.​

Claims

1. An AI-based mobile girder pedestal installation method, characterized by, The method comprises the following steps: S1: Obtain the bending moment, shear force and deformation angle mechanical parameters of the base connecting node through the embedded strain sensor, perform vector subtraction operation on the AI prediction value and the measured value, construct a node difference vector, and store the node difference vector set of continuous time sequence by using a sliding window; S2: Call the node difference vector set, perform first-order derivative operation on the adjacent time period vectors to generate a direction derivative sequence, input it into a K-means clustering algorithm for trend classification, output convergence state, oscillation state and divergence state labels, and establish a node trend state mapping table; S3: Based on the oscillation state and divergence state labels in the node trend state mapping table, set a frequency adjustment coefficient for the AI model path node, dynamically compress the calculation trigger interval by using an LSTM network, generate a high-frequency calculation node set, and expand the convergence state node calculation interval to 1.5 times of the reference value at the same time; S4: Call the high-frequency calculation node set and stress sensor data, construct a spatial stress gradient vector matrix, perform spatial matching on the adjacent node module length difference, input it into a random forest model to identify abnormal amplitude, output an abnormal coordinate set, and perform bias term back propagation correction.

2. The AI-based mobile girder pedestal installation method of claim 1, wherein, The node difference vector set comprises a time stamp index, a vector module length and a direction angle component, the node trend state mapping table specifically comprises combined data of a node number, a trend type and a direction derivative mean value, the high-frequency calculation node set comprises a node identifier, a trigger frequency coefficient and an LSTM weight parameter, and the abnormal coordinate set specifically comprises combined data of a three-dimensional coordinate, a time stamp and an abnormal amplitude.

3. The AI-based mobile girder pedestal installation method of claim 2, wherein, The size of the sliding window is 5 continuous time stamps, and the sliding step is 1 time stamp; The time interval of the adjacent time period vectors is 2 times of the sensor sampling period; The reference value is an initial calculation interval of 200 ms, and the expansion coefficient is based on the stress fluctuation stability threshold obtained through laboratory pressure test; The bias term correction amount is the product of the abnormal amplitude value and the weight matrix, and the weight matrix elements reflect the node spatial correlation strength.

4. The AI-based mobile girder pedestal installation method of claim 3, wherein, The acquisition step of the node difference vector set specifically comprises: S101: Collect three-axis strain signals of the embedded strain sensor, integrate the cross-section stress distribution to obtain the bending moment component by combining the material elastic modulus parameter and the cross-section moment of inertia parameter, convert the shear force component through a shear force transfer function, calculate the deformation angle component according to the node geometric size parameter, and generate the measured value of the node mechanical parameter; S102: Extract the bending moment prediction value, the shear force prediction value and the deformation angle prediction value output by the AI model, respectively perform vector component subtraction operation on the measured value of the node mechanical parameter, calculate a plurality of component absolute difference values, construct three-dimensional vector space coordinate points in the order of bending moment-shear force-deformation angle, and generate a node mechanical difference vector; S103: Align the node mechanical difference vector in the time stamp sequence, perform data interception by using a fixed time window, slide and update the window range by a preset step, store the continuous vector data points in the window, and generate a node difference vector set.

5. The AI-based mobile girder pedestal installation method of claim 4, wherein, The material elastic modulus parameter is determined by the ASTM E111 standard; the shear transfer function is wherein represents the value of the shear transfer function calculated by integration, represents the shear stress function distributed along the cross section, represents the integral variable along the cross section position coordinate, represents the shear stress value at the cross section position coordinate , represents the differential unit along the cross section position coordinate, the integral sign represents the operation of cumulatively summing the shear stress along the entire cross section; The fixed time window lasts for 10 seconds, and the preset step is 2 seconds.

6. The AI-based mobile girder pedestal installation method of claim 5, wherein, The obtaining step of the node trend state mapping table is specifically as follows: S201: adjacent time window vector data in the node difference vector set is called, the change rate of a plurality of vector components on the time axis is calculated, the first derivative value is calculated by using the central difference method, and a direction derivative sequence is generated; S202: the number of cluster algorithm categories is set, the direction derivative sequence is taken as a feature input, the Euclidean distance between the sample point and the initial cluster center is calculated, the centroid coordinates are iteratively updated until the distance square sum converges, the category number to which the sample point belongs is output, and a trend classification label is generated; S203: a corresponding relationship between a timestamp sequence and the trend classification label is established, a category number and a state type mapping rule are defined, a two-dimensional relationship table of a time interval field and a state type field is constructed, and a node trend state mapping table is generated; The number of categories is set to 3, corresponding to a convergent state, an oscillation state, and a divergent state.

7. The AI-based mobile girder pedestal installation method of claim 6, wherein, The step S3 is specifically as follows: S301: based on the node trend state mapping table, state duration values and amplitude values corresponding to oscillation state and divergent state labels are extracted, weight coefficients are assigned to the two types of labels respectively, and a node frequency adjustment coefficient is calculated and generated by using a linear superposition formula; S302: the node frequency adjustment coefficient is called to input an LSTM network, a time series prediction model is constructed, a sliding window mechanism is used to capture the node state change trend, a forgetting gate weight is dynamically adjusted through a gate unit, a plurality of node trigger interval compression ratios are output, and a dynamic trigger interval sequence is generated in combination with a reference interval value; S303: according to the dynamic trigger interval sequence, nodes with a compression ratio higher than a set threshold value are screened to form a high-frequency calculation node set, a convergent state node set is synchronously extracted, scalar multiplication operation is performed on the corresponding trigger interval and the reference value, an expansion coefficient is used to update the calculation interval parameter, and a node calculation interval configuration set is generated.

8. The AI-based mobile girder pedestal installation method of claim 7, wherein, The weight coefficient is in the range of 0.5-1.2 and is determined by gradient descent optimization; The reference interval value is 200 ms, which is derived from the clock cycle configuration of an embedded system; The set threshold value is that the compression ratio exceeds 30%, and is dynamically adjusted according to the real-time calculation resource load.

9. The AI-based mobile girder pedestal installation method of claim 8, wherein, The obtaining step of the abnormal coordinate set is specifically as follows: S401: the high-frequency calculation node set and stress sensor data are called to construct a three-dimensional space coordinate system, the stress gradient vector module length difference between adjacent nodes is calculated, and a spatial stress gradient difference matrix is established; S402: the spatial stress gradient difference matrix is input into a random forest model, a decision tree splitting threshold value is set, the feature importance is evaluated by using a Gini coefficient, coordinate points with an abnormal amplitude exceeding a set threshold value are screened, and an abnormal amplitude feature set is generated; S403: the back propagation algorithm is executed according to the abnormal amplitude feature set, the gradient change amount of the bias term is calculated, the weight update step is adjusted by using a momentum factor, the network parameters are iteratively corrected, and an abnormal coordinate set is generated.

10. The AI-based mobile girder pedestal installation method of claim 9, wherein, The decision tree splitting threshold value is 1.2 times the set threshold value of the abnormal amplitude; The momentum factor is in the range of 0.8-0.95 and is linearly attenuated according to the number of iterations.

Images