An assembled trolley control system

Abstract

The application provides an assembled trolley control system, and relates to the technical field of trolley intelligent control systems.The technical scheme points of the application are as follows: an environment sensing module, a posture detection module, an intelligent control module and an execution driving module are communicated;the environment sensing module comprises a laser radar submodule, a visual camera submodule, an ultrasonic sensor submodule and a data fusion submodule;the laser radar submodule is used for scanning the three-dimensional profile of a work area;the visual camera submodule is used for identifying component marker points;the ultrasonic sensor submodule is used for detecting close-range obstacles;and the data fusion submodule adopts a Kalman filtering algorithm to perform noise reduction processing on the multi-source sensor data.The assembled trolley control system provided by the application has the advantages of high positioning accuracy, strong environmental adaptability and good control coordination, and can meet the high-precision assembly requirements.

Description

Technical Field

[0001] This application relates to the field of intelligent control system technology for trolleys, and more specifically, to a control system for an assembly trolley. Background Technology

[0002] Existing prefabricated trolley operation relies heavily on manual experience, resulting in three core problems: First, low positioning accuracy, with component assembly deviations often exceeding 5mm, requiring repeated adjustments and leading to low construction efficiency; second, poor environmental adaptability, with dust and lighting changes in tunnels easily causing sensor failure, and reliance on manual judgment increasing the risk of safety accidents; and third, insufficient control coordination, with a lack of intelligent linkage between trolley movement and hoisting mechanism actions, easily leading to action conflicts. Existing control schemes mostly use single sensor feedback, failing to achieve multi-source data fusion and dynamic path optimization, making it difficult to meet the requirements of high-precision assembly.

[0003] The above problems urgently need to be addressed. Summary of the Invention

[0004] The purpose of this application is to provide an assembly trolley control system with advantages such as high positioning accuracy, strong environmental adaptability and good control coordination, which can meet the requirements of high-precision assembly.

[0005] This application provides a prefabricated trolley control system, the technical solution of which is as follows: It includes an environmental perception module, an attitude detection module, an intelligent control module, and an execution drive module, with each module communicating with each other; The environmental perception module includes a lidar submodule, a vision camera submodule, an ultrasonic sensor submodule, and a data fusion submodule. The lidar submodule is used to scan the three-dimensional contour of the work area, the vision camera submodule is used to identify component marker points, the ultrasonic sensor submodule is used to detect nearby obstacles, and the data fusion submodule uses a Kalman filter algorithm to perform noise reduction processing on the multi-source sensor data. The attitude detection module includes an inertial measurement submodule and a tilt sensor submodule. The inertial measurement submodule and the tilt sensor submodule synchronously collect data on the trolley travel angle, the boom pitch angle, and the component attitude. The intelligent control module includes an edge computing terminal submodule and a PLC submodule. The edge computing terminal submodule runs an improved A... The path planning algorithm and the improved fuzzy PID algorithm are used to output control commands. The PLC submodule receives the control commands output by the edge computing terminal submodule and sends them to the execution drive module. The execution drive module executes control commands and controls the movement of the control vehicle and the hoisting mechanism.

[0006] Furthermore, in this application, the visual camera submodule is equipped with an improved YOLOv8 algorithm, which improves the accuracy of marker point recognition by adding a component feature extraction layer; The improved YOLOv8 algorithm reduces computation by using depthwise convolutions and enhances the edge and texture features of the marked points. The formula is as follows: Where F base This represents the basic feature map (number of channels C) output by the C2f module. in =256); F base，i K represents the feature map of the i-th input channel; i F represents the depthwise convolution kernel of the i-th channel; dw This represents a local feature map of the output of a depthwise convolution, preserving the detailed features of the labeled points while ensuring algorithm speed. The improved YOLOv8 algorithm uses the ECA (Effective Channel Attention) mechanism to adaptively increase the weights of feature channels related to the marker points and suppress background interference. The formula is as follows: ω c =σ(AvgPool(F dw，c )·W); F mark，c =F dw，c ·ω c ; Where ω c F represents the attention weight of the c-th channel; dw，c The c-th channel of the deep convolutional feature map is represented by ; AugPool represents global average pooling, which extracts channel-level feature statistics; W represents a 1×1 convolution kernel; σ represents the Sigmoid activation function; F mark，c This represents the c-th channel feature map after attention weighting, with ω for the marked feature channel. c Approaching 1, the ω of the background channel c Approaching 0, thus achieving feature enhancement; F output by the component feature extraction layer mark (Number of channels C) out =256) is fused with the global features output by the SPPF module to obtain a fused feature map of local details and global context, as shown in the formula: F fusion =Concat(F mark SPPF(F) mark )); Where F fusion Represents the fused feature map; Concat represents the channel concatenation operation; F mark SPPF represents the feature map specific to the marker point; SPPF represents spatial pyramid pooling; SPPF(F)mark () represents the global feature map after SPPF processing; Improved YOLOv8 algorithm based on F fusion The marker features are used to generate candidate bounding boxes for markers, and the "class probability" and "coordinate position" of the markers are predicted by classification and regression branches, respectively: The classification branch is optimized using the cross-entropy loss function, with the following formula: ; Where L cls Represents the classification loss value; N represents the number of candidate boxes, y i =1 indicates that the i-th candidate box is a true marker point; S i y represents the confidence level of the i-th candidate box; i =0 indicates background.

[0007] The bounding box coordinates P of the regression branch prediction marker points pred =(x center y center (ω, h), the SloU (Scale-10U) loss function is used to optimize the positioning accuracy, and the formula is: ; Where x center y center L represents the pixel coordinates of the candidate box center point; ω and h represent the candidate box width and height, respectively; L reg Represents the regression loss value; SIoU represents the Scale-IoU similarity; P pred P represents the predicted bounding box of the i-th candidate box; gt The true bounding box represents the marker point; IoU represents the intersection-union ratio (measuring the degree of overlap) between the predicted box and the true box; Δ represents the distance between the center points of the predicted box and the true box; Λ represents the distance normalization factor; Θ represents the angle difference between the predicted box and the true box; Π represents the angle normalization factor, which is a fixed constant, Π=π / 2, i.e., 90°; Overlapping candidate boxes are removed using non-maximum suppression (NMS), retaining the markers with the highest confidence. The formula is as follows: NMS (P) pred IoU thresh =0.3); Where NMS represents the nonmaximum suppression function; P pred Represents the set of high-confidence candidate boxes; IoU thresh =0.3 represents the overlap threshold, and the final output is the precise coordinates of the marker point, which is used for subsequent assembly deviation calculation.

[0008] Furthermore, in this application, the improved A The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference within the work area; Improved A The path planning algorithm adds two penalty terms to ensure that the path simultaneously satisfies both "trolley passability" and "smooth hoisting," as shown in the formula: f(n) = g(n) + h(n) + C size (n)+C inertia (n); Where f(n) represents the total cost function of the node; n represents the current search node; g(n) represents the actual cost function; h(n) represents the heuristic function; C size (n) represents the penalty term for the trolley size constraint; C inertia (n) represents the hoisting inertia penalty term; G represents the path end node; Improved A The path planning algorithm calculates the cumulative path cost using the following formula: g(n) = g(n) prev )+dist(n prev (n); Where n prev Indicates the previous search node; dist(n prev (n) represents the distance between nodes, i.e., the distance between the previous node n. prev The Euclidean distance to the current node n; Improved A The path planning algorithm calculates the estimated remaining cost using the following formula: ; Where (x) n y n (x) represents the current node's raster coordinates; G y G ) represents the raster coordinates of the endpoint G; Improved A The path planning algorithm calculates the penalty term C for the trolley size constraint. size (n), the formula is: ; Improved A Path planning algorithm runtime space capacity determination logic: Taking node n as the center of the trolley, determine: [x n -W car / 2,x n +W car / 2]×[y n -L car / 2,y n +L carCheck if the rectangular area of ​​ / 2] contains "obstacle grid (1 or 2)". If it does, it is judged as "unacceptable" and a large penalty (10) is added. 6 The algorithm abandons the node; where W car Indicates the width of the trolley; L car Indicates the length of the trolley; Improved A Path planning algorithm calculates hoisting inertia penalty term C inertia (n), the formula is: C inertia (n)=k inertia ×θ(n)²; Where K inertia θ(n) represents the hoisting inertia coefficient; θ(n) represents the path turning angle: when the path is a straight line, θ(n) = 0, and the penalty term is 0; when the path turns, θ(n) increases, and the penalty term increases accordingly. Improved A Path planning algorithm execution constraint logic: θ(n)≤arcsin(W) car / R min (The maximum turning angle corresponding to the minimum turning radius of the trolley), if exceeded, the penalty term increases sharply (multiplied by 10), forcing the algorithm to choose a gentler turning path; where R min Indicates the minimum turning radius of the trolley; The edge computing terminal submodule performs path search according to the improved cost function f(n), and the process is as follows: 1) Initialize the open list (nodes to be explored) and the closed list (nodes already explored), and add the starting point S to the open list; 2) Select the node n with the smallest f(n) from the open list. If n is the destination G, then backtrack the path to generate the initial path. 3) Otherwise, expand the 8 neighboring nodes of n and calculate f(n) for each neighboring node; 4) Collision Detection: If the region corresponding to an adjacent node contains an obstacle grid (1 or 2), or C size (n)=10 6 If the node is not found, discard it; otherwise, add it to the open list. 5) Repeat steps 2)-4) until the destination G is found, generating the initial path P. init =[S, n1, n2, ..., G]; Improved A Path planning algorithm operation scenario adaptation logic: Dynamic interference grid 2) Real-time update. When construction workers are detected to have entered the current path, the algorithm immediately marks the area as impassable and searches for a new local path to achieve dynamic avoidance. When the initial path P initWhen a broken line exists (with excessively large angles), the improved A... The path planning algorithm performs a second-order smoothing optimization, as shown in the formula: ; Where P final (i) represents the coordinates of the i-th node after smoothing; α represents the smoothing coefficient, which balances path accuracy and smoothness; (1-α) represents the smoothing compensation coefficient. The edge computing terminal submodule will finalize the path P final The system is broken down into a "speed-position-time" instruction sequence and sent to the PLC submodule. The PLC submodule controls the execution drive module to drive the trolley along the final path. Simultaneously, the environmental perception module updates obstacle information every 100ms. If new dynamic disturbances are detected, the improved A... The path planning algorithm re-executes the local path search, optimizing only the disturbed road segments, without requiring global replanning, thus ensuring that the path is always collision-free.

[0009] Furthermore, in this application, the improved fuzzy PID algorithm dynamically adjusts the proportional, integral, and derivative parameters through attitude deviation and deviation change rate; The edge computing terminal submodule and the PLC submodule work together to execute the improved fuzzy PID algorithm. The execution logic includes data input, fuzzification, fuzzy processing, defuzzification, PID parameter update, and control output. Specifically: 1) Data Input: The attitude detection module collects θact(t) and θref(t) every 10ms and calculates the attitude deviation e(t) using the following formula: e(t) = θref(t) - θact(t); Where θref(t) represents the attitude reference value, that is, the target attitude parameters of the trolley at time t, such as the walking angle and the boom pitch angle; θact(t) represents the actual attitude value, that is, the actual attitude parameters of the trolley collected by the attitude detection module at time t; e(t) represents the attitude deviation, that is, the difference between the attitude reference value and the actual attitude value at time t. The rate of change of deviation ec(t) is calculated based on the deviation e(t-1) at the previous time step, using the following formula: ; Where ec(t) represents the rate of change of attitude deviation in period t; e(t) represents the real-time attitude deviation in period t; e(t-1) represents the attitude deviation in period (t-1); T s represents the discrete control period; t represents the control period index; Range normalization of e(t) and ec(t) (mapping to the interval [-1, +1]) facilitates fuzzy inference calculation; 2) Blurring: The normalized e(t) and ec(t) are transformed into fuzzy linguistic variables E and EC, and the membership degree of each fuzzy subset is calculated by the membership function; Where E and EC are fuzzy subsets, representing the linguistic variables after fuzzification of e(t) and ec(t), with values ​​{NB, NM, NS, ZO, PS, PM, PB}, corresponding to "negative large, negative medium, negative small, zero, positive small, positive medium, positive large" respectively; Taking an attitude deviation e(t) = 2° (normalized to 0.4) as an example, its fuzzification result is "PS (positive small)" and its membership degree μ PS =0.8, μ PM =0.2 (partially belonging to "the center"). Triangle membership function formula (taking e(t) as an example): ; Where x represents the normalized e(t); a, b, c represent the boundary values ​​of the fuzzy subset (e.g., the boundary of PS is [0.2, 0.4, 0.6]). 3) Fuzzy reasoning: The edge computing terminal submodule takes E and EC as inputs and infers U based on a preset fuzzy rule base. Kp U Ki U Kd ; Among them U Kp U Ki U Kd All are fuzzy output variables, corresponding to ΔK respectively. p , △K i , △K d The fuzzy linguistic variables take values ​​consistent with E and EC, namely {NB, NM, NS, ZO, PS, PM, PB}. The fuzzy inference synthesis formula (Mamdani) is as follows: ; Where μ R (x, y, u) represents the membership function of the fuzzy rule; This indicates the maximum operation; This indicates the operation of taking the smaller value; 4) Defuzzification: The edge computing terminal submodule uses the centroid method to convert the fuzzy output U Kp U Ki U Kd Transformed into precise parameter increment ΔK p , △K i , △K d The formula for the center of gravity method is (with △K) p (For example) ; in This indicates the fuzzy output variable U Kp The global integral, U is U Kp The range of values ​​for ; u represents the fuzzy output variable U Kp Continuous values ​​(replacing discretized u) k ); μ UKp (u) means U Kp Membership degree at continuous value u (instead of discretized μ) k ); The edge computing terminal submodule performs discretization calculations, dividing ΔK p The range of values ​​is divided into n discrete points, and the simplified formula is: ; Where △K p The proportional parameter increment (precise value) is the core result of the defuzzification output, used to update the PID proportional parameter; n is the number of fuzzy output discrete points, corresponding to the fuzzy output variable U. Kp The number of fuzzy subsets, where the fuzzy subsets are {NB, NM, NS, ZO, PS, PM, PB} (negative large, negative medium, negative small, zero, positive small, positive medium, positive large), each subset corresponds to one discrete point; k is the discrete point index, traversing all the discrete points corresponding to the fuzzy subsets, such as k=1 corresponding to NB, k=4 corresponding to ZO, k=7 corresponding to PB, its function is to extract the quantization value and membership degree of each fuzzy subset one by one, in preparation for weighted summation; u k The quantized value of the k-th discrete point belongs to the precise value corresponding to a pre-calibrated fuzzy subset (based on empirical calibration of the trolley PID parameters), such as ΔK. p u k Calibration value; μ k The membership degree of the k-th discrete point is μU, output by the Mamdani fuzzy inference synthesis formula. Kp (u k The physical meaning of this is that the greater the membership degree, the more the quantized value corresponding to the discrete point is adapted to the current working condition; This represents a weighted sum, which is the sum of the quantized values ​​of each discrete point according to their membership degree. The calculation logic is to highlight the contribution of discrete points with high membership degrees. This represents the sum of membership degrees, which is the sum of the membership degrees of all discrete points, and serves as a normalization function. 5) PID parameter update and control output: The edge computing terminal submodule superimposes the defuzzified parameter increments with the baseline parameters to update the real-time PID parameters. The parameter update formula is as follows: ; Where K p (t) / Ki (t) / K d (t) represent the updated real-time PID parameters (proportional / integral / derivative), which are all dynamically adjusted parameters at time t; K p0 / K i0 / K d0 These represent the PID reference parameters (proportional / integral / derivative), which are the initial reference values ​​from offline calibration and serve as the basis for parameter adjustments. Fuzzy increments are only fine-tuned to avoid control oscillations caused by sudden parameter changes; △K p / △K i / △K d These represent the increments of the PID parameters, which are the precise increments after fuzzy resolution using the centroid method. The improved fuzzy PID algorithm's constraint logic requires that the parameters be within a preset dynamic range; if they exceed this range, boundary values ​​are taken to prevent parameters from becoming uncontrollable. The edge computing terminal submodule implements PID control output, with the following formula: ; Where u(t) is the PID control output signal, representing the real-time control command output by the improved fuzzy PID algorithm to the execution drive module; t is the control time index (discrete-time variable), which is related to the control period T of the fuzzy PID. s Binding (t=0 is the initial control cycle, t=1 is the first control cycle, and so on); K p K(t) is a real-time proportional parameter that quickly responds to the current attitude deviation e(t). The larger the deviation, the more K... p The larger K is, the faster the correction can be achieved; i (t) represents the real-time integral parameter, which accumulates historical attitude deviations to eliminate steady-state errors and avoid the integral saturation problem of traditional PID controllers; K d (t) is the real-time differential parameter, reflecting the changing trend of attitude deviation and suppressing overshoot. The greater the rate of change of deviation, the better K d The larger e(t) is, the stronger the system damping; e(t) is the attitude deviation in the t-th period. This represents the cumulative integral term of attitude deviation, accumulating all attitude deviations from the initial period (k=0) to the current period (k=t). Each deviation is multiplied by the control period Ts, replacing continuous integration and eliminating steady-state errors that proportional control cannot resolve. `k` is the integral summation index, traversing all control periods from the initial period (k=0) to the current period (k=t). It is used only for the summation of the integral term and is distinct from the control time index `t` (k is the internal variable for summation, t is the current period index) to avoid confusion between the summation operation and the current period, ensuring the integrity of the integral term accumulation. `e(k)` represents the historical attitude deviation of the k-th period, belonging to cached historical attitude deviation data, providing historical data support for the integral term and ensuring the continuity of the integral accumulation. Its definition is completely consistent with `e(t)` (only the period is different). `T`s For discrete control cycles, discrete deviation data are transformed into integral / differential terms with a time dimension to ensure dimensional uniformity. The rate of change of attitude deviation (discrete differential term) is expressed as the difference between the deviation of the current cycle and the previous cycle, divided by the control cycle T. s The average rate of change of the deviation is obtained, which replaces the continuous derivative to avoid noise amplification and is used to predict the trend of deviation change and suppress overshoot in advance; e(t-1) is the attitude deviation in the (t-1)th cycle. After the calculation of the current cycle t is completed, e(t) is cached as e(t-1) for the next cycle to ensure the continuity of the derivative calculation and to be completely consistent with the acquisition source and preprocessing method of e(t). The PLC submodule receives the control output u(t) and converts it into control commands that the execution drive module can recognize, driving the trolley to adjust its attitude. At the same time, the attitude detection module collects the new θact(t+1) and enters the next control cycle, forming a closed-loop control.

[0010] Furthermore, this application also includes a safety warning module, which has a three-level warning mechanism, corresponding to audible and visual prompts, speed reduction operation, and emergency shutdown, with an emergency shutdown response time of ≤200ms.

[0011] Furthermore, this application also includes an electromagnetic interference suppression module, which integrates a shielding housing submodule and a signal filtering circuit submodule. The shielding housing submodule controls the activation of the copper-aluminum composite anti-interference shielding housing device, and the signal filtering circuit submodule suppresses industrial electromagnetic radiation interference through an LC filter network.

[0012] Furthermore, in this application, the environmental sensing module also includes a temperature sensor submodule, which collects the surface temperature of the component in real time; The edge computing terminal submodule is equipped with a component temperature deformation model unit. The component temperature deformation model unit combines the component surface temperature to correct the component assembly coordinates and compensate for the component size deviation caused by temperature changes. The calculation formula for the component temperature deformation model element is as follows: △L = α·L0·△T; X comp =X ref +△L; Where ΔL represents the temperature deformation increment of the component along the assembly direction, with positive values ​​indicating thermal expansion and negative values ​​indicating thermal contraction; α represents the linear expansion coefficient of the component material; L0 represents the original length of the component along the assembly direction, determined by the design drawings; ΔT represents the temperature change of the component, ΔT = T comp (t)-T0,T comp(t) represents the component surface temperature collected in real time by the temperature sensor; T0 represents the component calibration temperature; X comp This represents the corrected assembly target coordinates, which are the final control commands issued to the execution drive module; X ref Represents the original assembly reference coordinates, derived from improved A. The target location coordinates output by the path planning algorithm; The component temperature deformation model element will correct the assembly target coordinates X. comp It is converted into control commands and sent to the execution driver module.

[0013] Furthermore, in this application, the intelligent control module also includes a collaborative control submodule. This collaborative control submodule establishes communication with multiple trolleys via a 5G industrial module, enabling data sharing among the trolleys' operations and allocating lifting loads and receiving motion paths P based on a distributed collaborative algorithm. final (i); The distributed collaborative algorithm dynamically allocates the lifting load based on the load capacity of each vehicle, ensuring that the load on a single vehicle does not exceed the rated value. The formula is as follows: ; ; stF i ≤F rated,i (i=1,2,...,n); Where F i This represents the allocated load of the i-th trolley, which is the core control parameter output by the algorithm. The specific lifting load allocated to the i-th trolley is directly sent to the execution drive module of each trolley to ensure load balance among multiple trolleys; k i Let k be the load distribution coefficient for the i-th trolley. The distribution coefficient is positively correlated with the trolley's "remaining load capacity." The stronger the remaining capacity, the better the load distribution coefficient. i The larger the value, the more dynamic equilibrium is achieved where "the capable do more work"; F total For the total hoisting load of ultra-large components; F rated,i F is the rated lifting load of the i-th trolley; current,i This represents the actual current load of the i-th trolley. The total remaining load capacity of all trolleys is defined to ensure that the allocation coefficient is calculated based on the global remaining capacity, avoiding overloading of a single trolley while maximizing the utilization of the load potential of all trolleys; n is the number of trolleys in collaborative operation; j is the summation index number to ensure that the summation covers all collaborative trolleys without omissions or duplicate calculations; st is the constraint condition symbol, making F i ≤F rated,iThe load assigned to the i-th trolley must not exceed its rated load. This is to prevent safety accidents such as hydraulic system rupture and component falling caused by overloading of a single trolley, which is the bottom line for safety in collaborative operations. i is a trolley index variable that identifies a single trolley in collaborative operations. It is bound to the real-time position and attitude data of the trolley to ensure that load allocation and path coordination are executed synchronously. The collaborative control submodule allocates the load of the i-th trolley to the i-th trolley and converts it into control commands, which are then sent to the execution drive modules of each trolley. The intelligent control module also includes a collaborative control submodule, which establishes communication with multiple trolleys through a 5G industrial module to achieve data sharing of multiple trolley operations. It allocates the hoisting load and receives the motion path P based on a distributed collaborative algorithm. final (i); The distributed collaborative algorithm dynamically allocates the lifting load based on the load capacity of each vehicle, ensuring that the load on a single vehicle does not exceed the rated value. The formula is as follows: ; ; stF i ≤F rated,i (i=1,2,...,n); Where F i This represents the allocated load of the i-th trolley, which is the core control parameter output by the algorithm. The specific lifting load allocated to the i-th trolley is directly sent to the execution drive module of each trolley to ensure load balance among multiple trolleys; k i Let k be the load distribution coefficient for the i-th trolley. The distribution coefficient is positively correlated with the trolley's "remaining load capacity." The stronger the remaining capacity, the better the load distribution coefficient. i The larger the value, the more dynamic equilibrium is achieved where "the capable do more work"; F total For the total hoisting load of ultra-large components; F rated,i F is the rated lifting load of the i-th trolley; current,i This represents the actual current load of the i-th trolley. The total remaining load capacity of all trolleys is defined to ensure that the allocation coefficient is calculated based on the global remaining capacity, avoiding overloading of a single trolley while maximizing the utilization of the load potential of all trolleys; n is the number of trolleys in collaborative operation; j is the summation index number to ensure that the summation covers all collaborative trolleys without omissions or duplicate calculations; st is the constraint condition symbol, making F i ≤F rated,i The load assigned to the i-th trolley must not exceed its rated load. This is to prevent safety accidents such as hydraulic system rupture and component falling caused by overloading of a single trolley, which is the bottom line for safety in collaborative operations. i is a trolley index variable that identifies a single trolley in collaborative operations. It is bound to the real-time position and attitude data of the trolley to ensure that load allocation and path coordination are executed synchronously. The collaborative control submodule allocates the load of the i-th trolley to the i-th trolley and converts it into control commands, which are then sent to the execution drive modules of each trolley. The intelligent control module also includes a collaborative control submodule, which establishes communication with multiple trolleys through a 5G industrial module to achieve data sharing of multiple trolley operations. It allocates the hoisting load and receives the motion path P based on a distributed collaborative algorithm. final (i); The distributed collaborative algorithm dynamically allocates the lifting load based on the load capacity of each vehicle, ensuring that the load on a single vehicle does not exceed the rated value. The formula is as follows: ; ; stF i ≤F rated,i (i=1,2,...,n); Where F i This represents the allocated load of the i-th trolley, which is the core control parameter output by the algorithm. The specific lifting load allocated to the i-th trolley is directly sent to the execution drive module of each trolley to ensure load balance among multiple trolleys; k i Let k be the load distribution coefficient for the i-th trolley. The distribution coefficient is positively correlated with the trolley's "remaining load capacity." The stronger the remaining capacity, the better the load distribution coefficient. i The larger the value, the more dynamic equilibrium is achieved where "the capable do more work"; F total For the total hoisting load of ultra-large components; F rated,i F is the rated lifting load of the i-th trolley; current,i This represents the actual current load of the i-th trolley. The total remaining load capacity of all trolleys is defined to ensure that the allocation coefficient is calculated based on the global remaining capacity, avoiding overloading of a single trolley while maximizing the utilization of the load potential of all trolleys; n is the number of trolleys in collaborative operation; j is the summation index number to ensure that the summation covers all collaborative trolleys without omissions or duplicate calculations; st is the constraint condition symbol, making F i ≤F rated,i The load assigned to the i-th trolley must not exceed its rated load. This is to prevent safety accidents such as hydraulic system rupture and component falling caused by overloading of a single trolley, which is the bottom line for safety in collaborative operations. i is a trolley index variable that identifies a single trolley in collaborative operations. It is bound to the real-time position and attitude data of the trolley to ensure that load allocation and path coordination are executed synchronously. The collaborative control submodule allocates the load of the i-th trolley to the i-th trolley and converts it into control commands, which are then sent to the execution drive modules of each trolley.

[0014] Furthermore, in this application, the intelligent control module is also equipped with a remote control submodule, which connects to the remote control center through an encrypted communication protocol, supports AR visual interaction, and the remote control center can obtain data such as the trolley posture and working environment in real time, take over control when the local system fails, and implement unmanned emergency operations.

[0015] Furthermore, in this application, the execution drive module also includes a health diagnosis submodule, which collects operating data of key components of the actuator, analyzes the wear status of the components in combination with an improved BP neural network algorithm, predicts failure risks in advance, and generates maintenance prompts. The improved BP neural network algorithm is optimized through "momentum term + L2 regularization" to adapt to vibration data in industrial scenarios. The improved weight update formula (adding a momentum term to accelerate convergence) is as follows: w ij (k+1)=w ij (k)+η·δ j ·o i +α·[w ij (k)-w ij (k-1)]; b j (k+1)=b j (k)+η·δ j +α·[b j (k)-b j (k-1)]; Where w ij (k+1) represents the weight in the (k+1)th iteration, measuring the contribution of the output of the i-th neuron in the previous layer to the input of the j-th neuron in the next layer. This adapts to the high-dimensional coupling characteristics of vibration feature data, and accelerates weight convergence through the momentum term, thereby improving the accuracy of wear state identification. ij (k) represents the weight in the k-th iteration, serving as the benchmark for the next iteration's weight update. It inherits from previous iterations and is synchronized with the training process of the vibration feature data. With each iteration, the weights are adjusted based on sample errors, gradually optimizing the model's mapping relationship between "vibration features" and "wear status." k represents the iteration count, which is the training iteration count of the neural network. η represents the learning rate (step size factor), controlling the step size of the weight update. This avoids excessively large step sizes leading to training oscillations, and excessively small step sizes leading to slow convergence, while also preventing noise from causing disordered weight updates. δ j Let $\mathbf{j}$ represent the error term of the $j$-th neuron, which is the quantized value of the deviation between the neuron's output and the true label. Its calculation formula is: , where y j This represents the true label of the j-th neuron. Let σ′ represent the predicted output of the j-th neuron. j) represents the derivative of the activation function (when ReLU activation is first used, i.e. , where net j The net input to the j-th neuron is the weighted sum of the outputs and weights of the neurons in the previous layer, plus the bias term. , where w ij Let o be the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. i b is the output of the i-th neuron in the previous layer. j (the bias of the j-th neuron in the current layer); δ j Its function is to guide the direction of weight updates; the larger the error term, the greater the weight adjustment range, thus quickly reducing prediction bias. i This represents the output value of the i-th neuron. If i is an input layer neuron, then o i =x i (Input layer vibration feature data, such as normalized vibration amplitude); if i is a hidden layer neuron, then o i =σ(net i ), o i Used to convey effective information about vibration characteristics, providing feature contribution information for weight updates, and adapting to accurate gradient calculation in feature-coupled scenarios; α represents the momentum factor, introducing the inertia of the previous weight update, simulating momentum in physics, accelerating convergence and escaping local minima. Given the large number of local optima in industrial vibration data (spurious features caused by noise), α=0.9 can enhance the inertia of weight updates, preventing the model from getting trapped in local minima corresponding to spurious features, and improving generalization ability; W ij (k)-w ij (k-1) represents the weight increment between the k-th and (k-1)-th iterations. As input to the momentum term, it reflects the historical trend of weight updates. If the previous weight update was in the direction of decreasing error, the momentum term will help the current weight continue to adjust in that direction, accelerating convergence (e.g., if the weights are updated positively for three consecutive iterations, the momentum term will be superimposed with a positive increment); b j (k+1) represents the bias term in the (k+1)th iteration, i.e., the bias value of the j-th neuron in the neural network. It is used to adjust the activation threshold of the neuron, so that the model can better fit the nonlinear relationship between vibration characteristics and wear state, compensate for baseline drift of vibration data (such as feature offset caused by sensor zero drift), and improve the robustness of the model to noise; b j (k) represents the bias term in the k-th iteration, and is related to the weight w. ij(k) Synchronous updates jointly optimize the model's prediction accuracy. The momentum term of the bias term is optimized in sync with the weights to ensure an overall improvement in the model's convergence speed, adapting to the real-time analysis needs of the health diagnosis module; i represents the index of the previous layer's neurons, identifying the input neurons connected by the weights, clarifying the connection relationship of the weights, and avoiding hierarchical confusion; j represents the index of the next layer's neurons, identifying the output neurons connected by the weights, working in conjunction with the index i of the previous layer's neurons to uniquely determine a certain connection weight, ensuring accurate update logic; The formula for the error function (with L2 regularization to suppress overfitting) is: ; Where E represents the total error function (loss value), which measures the deviation between the model's predicted value and the true value, with the objective of convergence to E≤0.001; m is the number of training samples (engineering value m≥1000), and the samples are vibration characteristic data under different wear conditions (e.g., 200+ samples each for wear levels 0-4); y p The true label of the p-th sample (e.g., actual wear amount, wear level) is determined experimentally (e.g., by measuring wear amount using a laser diameter gauge); the model prediction value of the p-th sample (output from the output layer); λ represents the L2 regularization coefficient (engineering value λ=0.001), which suppresses excessively large weights and reduces the risk of model overfitting (avoiding overfitting to noisy data). The wear state mapping formula (from output layer to actual wear amount) is: ; Where Wear represents the actual wear amount of the component in the final output (unit: μm); k and c represent calibration coefficients (obtained from fitting experimental data, such as k=5.0 and c=0.1 for bearing components), realizing the mapping from model output value to actual wear amount; The health diagnosis submodule generates maintenance prompts based on the actual wear of the components.

[0016] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0017] Beneficial effects: ① Positioning accuracy is significantly improved, meeting the requirements of high-precision assembly: Through the multi-source sensor data fusion system of the environmental perception module and the improvement of A A closed-loop control system is constructed using path planning and improved fuzzy PID control, combined with an improved YOLOv8 algorithm, providing a precise sensing foundation for positioning. A temperature deformation compensation model is used to offset dimensional deviations caused by thermal expansion and contraction, improving component assembly accuracy and eliminating the need for repeated adjustments, thus increasing construction efficiency. The improved fuzzy PID algorithm, through dynamic parameter tuning (response time ≤ 50ms), effectively suppresses attitude overshoot and component sway, ensuring that the control error of the boom pitch angle and trolley travel angle is ≤ 0.1°, further supporting high-precision assembly.

[0018] ② Strong environmental adaptability, suitable for complex construction scenarios: The environmental perception module, including the lidar submodule, vision camera submodule, and ultrasonic sensor submodule, achieves full-range environmental coverage from far to medium to near distances. The data fusion submodule uses a Kalman filter algorithm to reduce noise, solving the problem of sensor data distortion caused by tunnel dust and changes in lighting. The anti-electromagnetic interference module can suppress industrial electromagnetic radiation interference, ensuring stable and reliable transmission of sensor and control signals, and is suitable for strong electromagnetic environments such as substations and industrial plants. The temperature sensor submodule covers an operating temperature range of -10℃ to 50℃, and through deformation compensation, the system can maintain stable accuracy under extreme temperature conditions, breaking through the dependence of traditional systems on ambient temperature.

[0019] ③ Optimized control coordination, improving operational efficiency and safety: The architecture of the edge computing terminal submodule and PLC submodule enables high-speed data interaction between multiple modules, allowing intelligent linkage between the trolley's movement and the lifting mechanism's actions, avoiding conflicts inherent in traditional manual control. The collaborative control submodule, through 5G communication and distributed algorithms, achieves balanced load distribution and collision-free path coordination among multiple trolleys, breaking through the load limitations of a single trolley. This allows for efficient joint lifting of ultra-large components (such as thousand-ton bridge segments), improving operational efficiency. Improved A The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference (such as construction workers and temporary equipment), completely solving the collision risk caused by traditional path planning that "only focuses on the shortest path and ignores equipment characteristics".

[0020] ④ Ample safety redundancy reduces operational risks: A three-level early warning mechanism (audio-visual alerts → speed reduction → emergency shutdown) covers all scenarios from minor to severe risks, with an emergency shutdown response time of ≤200ms, quickly avoiding dangers such as collisions, overloads, and electrical leaks; the remote control submodule, through encrypted communication and AR visualization, enables unmanned operation in high-risk scenarios (deep tunnels, high altitudes), while also supporting remote takeover in case of local failure, preventing personnel from entering dangerous areas and reducing the accident rate. The health diagnosis submodule, through an improved BP neural network algorithm, predicts the wear status of key components of the actuator (bearings, hydraulic cylinders) 24-72 hours in advance, generating targeted maintenance prompts to avoid work interruptions or equipment damage caused by sudden failures, reducing maintenance costs.

[0021] ⑤ Strong functional scalability, adaptable to diverse operational needs: The collaborative control submodule supports dynamic networking of multiple vehicles, and the number of collaborative vehicles can be flexibly adjusted according to the weight and size of components; the remote control submodule is compatible with various interactive terminals such as AR glasses and large screens, supporting both daily visual monitoring and emergency takeover in case of local system failures, adapting to the control needs of different construction scenarios. The overall architecture adopts a modular design (such as safety early warning, anti-electromagnetic interference, and health diagnosis modules, which are all optional expansions), and can be flexibly configured according to project budget and operational complexity, balancing economy and practicality. Core algorithms (improved YOLOv8, A...) The path planning, improved fuzzy PID, and improved BP neural network all adopt lightweight design to adapt to the computing power of edge computing terminals and ensure long-term stable operation of the system. Attached Figure Description

[0022] Figure 1 An architecture diagram of an assembly trolley control system provided in this application; Figure 2 This is an architecture diagram of the environment perception module in this application; Figure 3 This is an architecture diagram of the attitude detection module in this application; Figure 4 This is an architecture diagram of the intelligent control module of this application; Figure 5 This is an architectural diagram of the electromagnetic interference suppression module of this application; Figure 6 This is an architecture diagram of the edge computing terminal submodule of this application; Figure 7 This is an architecture diagram of the execution driver module of this application. Detailed Implementation

[0023] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0024] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] Please refer to Figures 1 to 7 As shown, this application provides a prefabricated trolley control system, including an environmental perception module, an attitude detection module, an intelligent control module, and an execution drive module, all of which are communicatively connected. The environmental perception module includes a lidar submodule, a vision camera submodule, an ultrasonic sensor submodule, and a data fusion submodule. The lidar submodule scans the three-dimensional contour of the work area, the vision camera submodule identifies component marker points, the ultrasonic sensor submodule detects nearby obstacles, and the data fusion submodule uses a Kalman filter algorithm to reduce noise in the multi-source sensor data. The attitude detection module includes an inertial measurement submodule and a tilt sensor submodule. The inertial measurement submodule and the tilt sensor submodule synchronously collect the trolley's travel angle, the boom's pitch angle, and component attitude data. The intelligent control module includes a PLC submodule and an edge computing terminal submodule. The edge computing terminal submodule runs an improved AI... The path planning algorithm and the improved fuzzy PID algorithm are used. The PLC submodule receives the control commands output by the edge computing terminal submodule and sends them to the execution drive module. The execution drive module executes the control commands and controls the movement of the control vehicle and the hoisting mechanism.

[0026] Specifically, it includes an environmental perception module, an attitude detection module, an intelligent control module, and an execution drive module, with each module communicating with each other. This forms the core architecture of the system, breaking away from the drawbacks of traditional trolley systems where "modules are dispersed and manual coordination is required," and establishing an intelligent control mode of "centralized data hub + distributed functional modules." The combination of industrial Ethernet (transmission latency ≤10ms) and edge computing units enables high-speed data synchronization and local real-time processing between modules, replacing the simple architecture of the traditional "sensor-to-controller direct connection," and paving the way for subsequent algorithm execution (such as improving A...). Fuzzy PID controllers provide data interaction guarantees and are the foundation of system intelligence.

[0027] The environmental perception module includes a lidar submodule, a vision camera submodule, an ultrasonic sensor submodule, and a data fusion submodule. It constructs a multi-source perception system covering the entire scene without blind spots, solving the problem of traditional single sensors (such as those using only vision) being prone to failure due to environmental interference (dust, light). Through the complementary functions of different sensors, it achieves comprehensive perception of distances from far to medium to near, and of macroscopic environment to microscopic components.

[0028] The lidar submodule is used to scan the 3D contour of the work area: providing spatial reference data of the macroscopic environment for the system, solving the problems of low efficiency and poor accuracy of traditional manual surveying of work areas. The scanned 3D contour data is an improvement on A... The core input of the path planning algorithm, A The path planning algorithm identifies static obstacles (such as scaffolding and steel bar piles) based on this data, ensuring the environmental adaptability of the path planning; at the same time, it provides spatial boundary information of the work area for multiple vehicles to work together, avoiding boundary crossing conflicts.

[0029] The vision camera submodule is used to identify component marker points: achieving precise microscopic positioning of components is a key step in improving assembly accuracy and solves the pain point of "large deviations in visual judgment" in traditional manual alignment. Marker points (such as QR codes and reflective markers) serve as the "positioning reference" for components. The coordinate data of the marker points identified by the vision camera submodule are directly used to calculate the deviation between the actual position and the target position of the component, providing a core basis for improving the parameter adjustment of the fuzzy PID algorithm and supporting the achievement of assembly accuracy at the ±2mm level.

[0030] The ultrasonic sensor submodule is used to detect obstacles at close range: filling the "close-range blind spot" of lidar and visual sensors, building the "last line of defense" for collision protection, and solving the safety hazard of traditional sensors missing obstacles within 1 meter. For close-range operation scenarios such as trolley turning and crane arm lowering, it detects low / moving obstacles such as personnel and small tools in real time, triggering the safety warning module's gradient protection (such as deceleration and emergency stop) to improve operational safety.

[0031] The data fusion submodule employs a Kalman filter algorithm to denoise the multi-source sensor data, eliminating random noise and errors from the sensors (such as point cloud noise from lidar and illumination interference in visual recognition), outputting unified and reliable environmental data, and resolving control misjudgments caused by fluctuations in traditional single-sensor data. The Kalman filter algorithm formula is as follows: ①State prediction equation: ; Its function is to, based on the optimal state at the previous time step Predict the state at the current moment. A represents the state transition matrix, B represents the control input matrix, and u k-1 Indicates control input; ②State update equation: ; Combined with the current sensor observation value z k The residual between (multi-source sensor data) and the predicted value is expressed through the Kalman gain K. k Correct the predicted values ​​to obtain the optimal state at the current time. H is the observation matrix; the final output fused data retains both the spatial accuracy of the lidar and the positioning details of the vision, providing a "noise-free" decision-making basis for the subsequent control module.

[0032] The attitude detection module includes an inertial measurement submodule and a tilt sensor submodule: it constructs a "dual-sensor redundancy" attitude detection system to solve the problem of data distortion under vibration conditions caused by traditional single sensors (such as tilt sensors alone), and ensures the reliability of attitude data.

[0033] The inertial measurement submodule and the tilt sensor submodule synchronously collect data on the trolley's travel angle, the boom's pitch angle, and the component's attitude. This allows for real-time acquisition of the system's own state and the state of the controlled object, providing core feedback for closed-loop control and resolving the control instability problem caused by traditional control methods that "only focus on the component's position and ignore the trolley's attitude." Specifically, the inertial measurement submodule focuses on collecting data on the trolley's travel angle, boom's pitch angle, and component's attitude during motion changes, while the tilt sensor submodule focuses on collecting data on the trolley's travel angle, boom's pitch angle, and component's attitude when stationary or in uniform motion.

[0034] Among them, the trolley's traveling angle reflects whether the trolley is traveling along the planned path, and deviation data triggers improvement A. Path planning algorithms are used for path correction; the boom pitch angle is directly related to the lifting height of the component and is a key basis for improving the fuzzy PID algorithm to adjust the lifting speed, avoiding component swaying caused by rapid boom raising and lowering; component attitude data (such as tilt angle) is used to compensate for assembly coordinates, avoiding splicing deviations caused by component attitude offsets, and supporting high-precision assembly. "Synchronous acquisition" (timestamp error ≤1ms) ensures the spatiotemporal consistency of attitude data and environmental perception data, avoiding control logic confusion caused by data misalignment.

[0035] The intelligent control module includes a PLC submodule and an edge computing terminal submodule: it constructs a dual-core control architecture of "algorithm decision-making and hardware execution", which solves the bottleneck of insufficient computing power and inability to run complex algorithms in traditional "single PLC control", and realizes the separation and synergy of "intelligent decision-making and precise execution".

[0036] The edge computing terminal submodule runs an improved A Path planning algorithm and improved fuzzy PID algorithm: undertake the core function of "intelligent decision-making" of the system, and are the key to solving the problem of "fixed path and rigid parameters" of traditional trolleys, and realize adaptive control in dynamic environments.

[0037] Among them, the improved A The path planning algorithm dynamically generates collision-free paths based on the 3D contour data of the environmental perception module and the trolley size constraints, replacing traditional fixed path planning and adapting to the dynamic changes of the construction scenario. The improved fuzzy PID algorithm dynamically adjusts control parameters based on the deviation data of the attitude detection module and the positioning deviation of the environmental perception, replacing traditional fixed PID parameters and adapting to working conditions such as load changes and road bumps, ensuring control stability.

[0038] The PLC submodule receives control commands output by the edge computing terminal submodule and sends them to the execution drive module: it builds a bridge between "algorithm decision-making and mechanical action", solves the "adaptation gap" problem between complex algorithm output and actuator control signals, and ensures that decisions are implemented.

[0039] Among them, the PLC (Programmable Logic Controller) submodule has high anti-interference and fast response characteristics. It converts the "speed-position" commands (such as "traveling speed 0.3m / s, boom pitch angle +5°") output by the edge computing terminal into electrical signals (such as motor frequency conversion signals, hydraulic cylinder control signals) that can be recognized by the actuator, and at the same time provides real-time feedback on the execution status to form a control closed loop.

[0040] The execution drive module executes control commands and controls the movement of the vehicle and the hoisting mechanism: as the system's "mechanical execution terminal", it converts electrical signals into physical actions and is the ultimate carrier for achieving intelligent control goals, solving the problems of "uneven force and delayed action" in traditional manual operation of the execution mechanism.

[0041] Specifically, the execution targets include the trolley's travel motor (controlling forward, backward, and steering) and the hydraulic / motor system of the hoisting mechanism (controlling the hoisting arm's pitch, rotation, and component lifting). Based on the instructions of the PLC submodule, precise control of the movements is achieved through technologies such as frequency conversion drive and servo control (e.g., travel speed error ≤ 0.01m / s, hoisting arm angle error ≤ 0.1°). Ultimately, the components are assembled according to the planned path and target posture, achieving the core functional objectives of the system.

[0042] In some preferred embodiments, the visual camera submodule is equipped with an improved YOLOv8 algorithm, which improves the accuracy of marker point recognition by adding a component feature extraction layer.

[0043] Specifically, in the scenario of component marker recognition, the traditional YOLOv8 algorithm has the problem of insufficient adaptation to industrial scenarios: the component markers (such as QR codes and reflective markers) are small in size (usually 2-5cm), easily affected by changes in dust / lighting in the construction environment, and the background (concrete surface, steel reinforcement texture) is complex, which leads to insufficient feature capture of the markers by the original algorithm, and easy to miss or misdetect.

[0044] The core optimization of the improved YOLOv8 algorithm is the addition of a "component-specific feature extraction layer" to the backbone feature extraction network. This layer enhances key features such as texture and edges of the markers through "lightweight convolution + attention mechanism" while maintaining real-time performance (e.g., recognition speed ≥ 30 frames / second), adapting to the dynamic assembly requirements of prefabricated trolleys. In other words, the vision camera submodule incorporates the improved YOLOv8 algorithm, enhancing marker recognition accuracy by adding a component feature extraction layer, achieving a recognition speed ≥ 30 frames / second.

[0045] It should be noted that the marker points are visual recognition feature reference points pre-placed at key locations on components (such as bridge segments, precast concrete beams, steel structure components, etc.) and the trolley body. They are the core objective of the visual camera submodule to achieve accurate positioning and dynamic tracking by improving the YOLOv8 algorithm, and also the spatial coordinate reference for trolley posture control and component hoisting alignment.

[0046] For example, the application of the improved YOLOv8 algorithm in the visual camera submodule consists of four steps: image preprocessing → improved feature extraction → marker detection → post-processing. The entire process revolves around "improving the recognizability of marker features and reducing environmental interference," specifically: 1) Image preprocessing: adapting to construction environment interference.

[0047] The raw image I captured by the visual camera needs to be preprocessed to remove noise caused by dust and changes in lighting: Adaptive illumination equalization: CLAHE (contrast-limited adaptive histogram equalization) is used to adjust the image brightness distribution, avoiding blurring of marker features caused by overexposure in strong light or underexposure in weak light. The formula is: I clahe =CLAHE(I, clipLimit=2.0, tileGridSize=(8, 8)); Where I represents the original input image; I clahe This represents the image after CLAHE processing; CLAHE represents the algorithm name; clipLimit=2.0 represents the contrast limit threshold (to avoid local overexposure); tileGridSize=(8,8) is the image block size to ensure local lighting balance.

[0048] Gaussian blur noise reduction: High-frequency noise caused by filtering dust, the formula is: ; Among them I pre Indicates the final preprocessed image; G σ σ represents the Gaussian kernel function; σ=1.0 represents the standard deviation of the Gaussian kernel (balancing noise reduction and feature preservation). The normalization coefficient is represented by exp(·); the natural exponential function is represented by x and y; the kernel relative coordinates are represented by x and y. This represents the convolution operation. The preprocessed data I... pre The input algorithm backbone network provides a clean input foundation for feature extraction.

[0049] 2) Improved feature extraction: Add a dedicated feature layer for component marker points.

[0050] The original YOLOv8 algorithm relies on the C2f module (which fuses multi-scale features) and SPPF (spatial pyramid pooling to enhance global features) for feature extraction, but it is insufficient in capturing the "local features" of component markers (such as the black and white block texture of a QR code and the bright edges of a reflective cursor). The improved YOLOv8 algorithm adds a "MarkFeature-specific feature extraction layer" between the output of the C2f module and the input of the SPPF module. The architecture is "lightweight deep convolution (DWConv) + ECA attention mechanism", and the specific process is as follows: ① Lightweight depthwise convolution: Extracts local features.

[0051] The computational cost is reduced by using depthwise convolution (DWConv) while enhancing the edge and texture features of the marker points. The formula is as follows: ; Where F base This represents the basic feature map (number of channels C) output by the C2f module. in =256); F base，i K represents the feature map of the i-th input channel; i This represents the depthwise convolution kernel for the i-th channel (size 3×3, one kernel per channel group, computational cost is only 1 / C of that of a regular convolution). in ); F dw This represents the local feature map output by the depthwise convolution, preserving the detailed features of the marked points while ensuring algorithm speed (avoiding frame rate drops caused by adding new layers).

[0052] ②ECA attention mechanism: reinforces key features.

[0053] The improved YOLOv8 algorithm uses the ECA (Effective Channel Attention) mechanism to adaptively increase the weights of feature channels related to marker points, suppressing background interference (such as concrete textures). The formula is as follows: ω c =σ(AvgPool(F dw，c )·W); F mark，c =F dw，c ·ω c ; Where ωc F represents the attention weight of the c-th channel; dw，c F represents the c-th channel of the deep convolutional feature map; AugPool represents global average pooling, which extracts channel-level feature statistics; W represents a 1×1 convolution kernel (adaptively adjusts channel weights with minimal computation); σ represents the Sigmoid activation function (output weights ∈ [0, 1]); mark，c This represents the c-th channel feature map after attention weighting, with ω for the marked feature channel. c Approaching 1, the ω of the background channel c Approaching 0, feature enhancement is achieved.

[0054] ③ Feature fusion and global modeling.

[0055] F output by the component feature extraction layer mark (Number of channels C) out =256) is fused with the global features output by the SPPF module to obtain a fused feature map of "local details + global context", providing comprehensive feature support for subsequent detection. The formula is: F fusion =Concat(F mark SPPF(F) mark )); Where F fusion Represents the fused feature map; Concat represents the channel concatenation operation; F mark SPPF represents the feature map specific to the marker point; SPPF represents spatial pyramid pooling, which enhances the multi-scale adaptability of features; SPPF(F) mark ) represents the global feature map after SPPF processing.

[0056] 3) Marker detection head: precise positioning.

[0057] Fusion Feature Map F fusion Inputting a YOLOv8 detector head, the decoupled head (classification branch + regression branch) predicts the "class probability" (whether it is a marker point) and "coordinate position" of the marker point, respectively, where F fusion N is the source of features generated by F, determining the effectiveness and scale of candidate boxes; that is, N is the source of features generated by F. fusion The quantification results of feature representation ability guide F in reverse. fusion The optimization direction, and the synergistic optimization of the two, is the core logic for improving the accuracy of marker recognition in the YOLOv8 algorithm: Classification branch: The confidence score S∈[0,1] of the output labeled point (threshold set to 0.5, S≥0.5 is considered a valid labeled point), optimized using the cross-entropy loss function, the formula is: ; Where Lcls Represents the classification loss value; N represents the number of candidate boxes, y i =1 indicates that the i-th candidate box is a true marker point; S i y represents the confidence level of the i-th candidate box; i =0 indicates background.

[0058] Regression branch: Predict the bounding box coordinates P of the marker points pred =(x center y center The SloU (Scale-LoU) loss function is used to optimize the localization accuracy (adapted to small target location marking), and the formula is: ; Where x center y center L represents the pixel coordinates of the candidate box center point; ω and h represent the candidate box width and height, respectively; L reg Represents the regression loss value; SIoU represents the Scale-IoU similarity; P pred P represents the predicted bounding box of the i-th candidate box; gt SIoU represents the true bounding box of the marker point; IoU represents the intersection-over-union ratio (measuring the degree of overlap) between the predicted box and the true box; Δ represents the distance between the center points of the predicted box and the true box; Λ represents the distance normalization factor; Θ represents the angle difference between the predicted box and the true box; Π represents the angle normalization factor, which is a fixed constant, Π=π / 2, i.e., 90°; SIoU loss considers overlap, distance and angle simultaneously, solving the problem of large localization deviation of small target marker points, and the convergence speed is faster than the original CIoU.

[0059] 4) Post-processing: Filter valid marker points.

[0060] Overlapping candidate boxes are removed using non-maximum suppression (NMS), retaining the markers with the highest confidence. The formula is as follows: NMS (P) pred IoU thresh =0.3); Where NMS represents the nonmaximum suppression function; P pred Represents the set of high-confidence candidate boxes; IoU thresh =0.3 represents the overlap threshold (marker points are usually independently distributed, and setting the threshold to 0.3 can avoid missed detections), and finally outputs accurate marker point coordinates for subsequent assembly deviation calculations.

[0061] The improved YOLOv8 algorithm described above serves several purposes in practical applications: First, it enhances the accuracy of marker recognition. The newly added MarkFeature layer, through "local feature extraction + attention weighting," strengthens the feature recognition of markers under dust and lighting interference, further improving the accuracy compared to the original YOLOv8 and preventing assembly deviations caused by missed or false detections of markers. Second, it ensures recognition speed (e.g., ≥30 frames / second). By employing lightweight deep convolution (DWConv) and ECA attention (without fully connected layers, resulting in low computational cost), the additional computational cost of the new layer is less than the original algorithm. Combined with GPU acceleration (edge ​​computing terminals equipped with NVIDIA Jetson Xavier NX), the inference speed is further improved (reaching 35-40 frames / second), meeting the "real-time positioning" requirements in dynamic assembly of the trolley (e.g., processing time per frame ≤33ms). Supporting high-precision assembly: reducing and ensuring the positioning error of the marker point coordinates (e.g., ≤0.5mm), providing accurate position feedback for subsequent parameter adjustment of the improved fuzzy PID algorithm and assembly coordinate correction, is the key sensing foundation for the system to achieve high assembly accuracy (e.g., ±1mm-±2mm).

[0062] In some preferred embodiments, the improved A The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference within the work area.

[0063] Specifically, traditional A The path planning algorithm only searches for the optimal path based on the "point model" without considering the actual volume of the trolley (which can easily lead to collisions due to narrow paths) and the hoisting inertia (which can easily lead to component swaying). Furthermore, it cannot adapt to dynamic disturbances in the construction scenario (such as moving construction personnel and temporary equipment).

[0064] Improvement A The core optimization of the path planning algorithm is to incorporate "trolley size constraints" and "lifting inertia coefficient" into the path search process, and dynamically update obstacle information based on real-time data from the environmental perception module. This ultimately generates a smooth path that is "collision-free, low-inertia, and adaptable to dynamic scenarios," addressing the pain points of traditional algorithms such as "the path is feasible but the trolley cannot pass" and "sharp turns on the path cause component swaying," thus adapting to the lifting operations requirements of prefabricated trolleys. This is the improved A... The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference within the work area.

[0065] For example, improve A The application of path planning algorithms in the prefabricated trolley control system consists of four steps: environment modeling, cost function optimization, path search and collision detection, and dynamic adjustment and smoothing. The entire process is integrated with the environment perception module and the intelligent control module. 1) Environment modeling: Constructing dynamic grid maps (the basis of path search).

[0066] The environmental perception module uses lidar to scan the 3D contour of the work area and extract static obstacles (OBS). static Ultrasonic sensors and vision cameras detect dynamic interference in real time. dynamic And transmit it to the edge computing terminal.

[0067] The edge computing terminal converts the above data into a GridMap. The core logic is as follows: Static obstacle grid: Obs dynamic The corresponding spatial location is marked as "1", indicating that passage is not possible; Dynamic interference grid: Obs dynamic The corresponding spatial location is marked as "2" and is updated in real time (e.g., the grid status changes synchronously when construction workers move). Free grid: Unoccupied areas are marked as "0" and can be used as path candidate areas.

[0068] Function: It transforms complex working environments into discrete grids that can be processed by algorithms, providing "environmental boundary" data for path search and avoiding the collision risk caused by traditional algorithms "ignoring obstacle details".

[0069] 2) Cost function optimization: incorporating size constraints and inertia penalties.

[0070] Traditional A The cost function of the path planning algorithm is f(n) = g(n) + h(n). The improved algorithm adds two penalty terms to ensure that the path simultaneously satisfies both "trolley passability" and "smooth hoisting," as shown in the formula: f(n) = g(n) + h(n) + C size (n)+C inertia (n); Where f(n) represents the total cost function of the node; n represents the current search node; g(n) represents the actual cost function; h(n) represents the heuristic function (estimated cost); C size (n) represents the penalty term for the trolley size constraint; C inertia (n) represents the hoisting inertia penalty term; G represents the path end node; ① Actual cost g(n): Cumulative path cost, its formula is: g(n) = g(n) prev )+dist(n prev (n); Where n prev Indicates the previous search node; dist(n prev (n) represents the distance between nodes, i.e., the distance between the previous node n. prevEuclidean distance to the current node n (inter-grid distance, 0.1m for 1 grid); Function: To record the actual length of the route traveled, ensuring the core objective of minimizing the route length.

[0071] ② Heuristic function h(n): estimates the remaining cost, and its formula is: ; Where (x) n y n (x) represents the current node's raster coordinates; G y G ) represents the raster coordinates of the endpoint G; multiplying by 0.1 converts the raster index distance to the actual distance (m), ensuring that the dimensions of h(n) and g(n) are consistent; Function: To guide the algorithm to search towards the destination, reduce the exploration of invalid paths, and improve search efficiency.

[0072] ③Trolley size constraint penalty item C size (n), the formula is: ; Space capacity determination logic: Taking node n as the center of the trolley, determine: [x n -W car / 2,x n +W car / 2]×[y n -L car / 2,y n +L car Check if the rectangular area of ​​ / 2] contains "obstacle grid (1 or 2)". If it does, it is judged as "unacceptable" and a large penalty (10) is added. 6 The algorithm abandons the node; where W car Indicates the width of the trolley; L car Indicates the length of the trolley; Function: To solve the problem of traditional A The path planning algorithm, based on a "point model," ignores the issue of trolley volume, ensuring that the planned path width and turning radius meet the actual passage requirements of the trolley (e.g., path width ≥ W). car +0.5m safety margin).

[0073] ④ Lifting inertia penalty item C inertia (n), its formula is: C inertia (n)=k inertia ×θ(n)²; Where K inertiaθ(n) represents the hoisting inertia coefficient; θ(n) represents the path turning angle: when the path is a straight line, θ(n) = 0, and the penalty term is 0; when the path turns, θ(n) increases, and the penalty term increases accordingly. Constraint logic: θ(n)≤arcsin(W) car / R min (The maximum turning angle corresponding to the minimum turning radius of the trolley), if exceeded, the penalty term increases sharply (multiplied by 10), forcing the algorithm to choose a gentler turning path; where R min Indicates the minimum turning radius of the trolley; Function: To prevent the hoisted components from swaying due to inertia when the path turns sharply (for example, if the angle is too large, the centrifugal force of the components will increase, which may cause them to expand into obstacles or deviate from the target position), and to ensure a smooth hoisting process.

[0074] 3) Path search and collision detection: Generate initial feasible paths.

[0075] Edge computing terminals perform path search according to the improved cost function f(n). The core process is as follows: ① Initialize the open list (nodes to be explored) and the closed list (nodes already explored), and add the starting point S to the open list; ② Select the node n with the smallest f(n) from the open list. If n is the destination G, then backtrack the path to generate the initial path. ③ Otherwise, expand the 8 adjacent nodes of n (4 adjacent nodes in the up, down, left, and right directions and 4 adjacent nodes in the diagonal direction in the grid map) and calculate f(n) for each adjacent node. ④ Collision detection: If the region corresponding to an adjacent node contains an obstacle grid (1 or 2), or C size (n)=10 6 If the node is not found, discard it; otherwise, add it to the open list. ⑤ Repeat steps 2-4 until the destination G is found, generating the initial path P. init =[S, n1, n2, ..., G].

[0076] Scene adaptation: The dynamic interference grid ② is updated in real time. When a construction worker is detected to have entered the current path, the algorithm immediately marks the area as impassable and searches for a new local path to achieve dynamic avoidance.

[0077] 4) Path smoothing: Optimize path continuity (adapt to trolley execution).

[0078] Initial path P init There may be broken lines (excessive angles), requiring secondary smoothing optimization. The core formula is: ; Where P final(i) represents the coordinates of the i-th node after smoothing; α=0.7 represents the smoothing coefficient, balancing path accuracy and smoothness; (1-α) represents the smoothing compensation coefficient; Function: To transform a broken line path into a continuous and smooth curve, ensuring that the trolley steering mechanism and hoisting mechanism can operate smoothly (such as gradually changing speed when turning, avoiding sudden braking and sharp turns).

[0079] 5) Path distribution and dynamic adjustment.

[0080] The edge computing terminal will ultimately path P final The system is broken down into a "speed-position-time" command sequence (e.g., "0-5s: travel to node n1 at a speed of 0.3m / s, turning angle 0°"), which is then sent to the PLC submodule via industrial Ethernet. The PLC submodule controls the execution module to drive the trolley along the final path. Simultaneously, the environmental perception module updates obstacle information every 100ms. If new dynamic disturbances (such as temporarily stacked materials) are detected, the improved A... The path planning algorithm re-executes the local path search (optimizing only the disturbed road segments, without global replanning) to ensure that the path is always collision-free.

[0081] A In practical applications, the path planning algorithm serves several purposes: Dynamic obstacle and interference avoidance: Through real-time grid map updates and collision detection, it achieves full-scene avoidance of static obstacles (scaffolding, rebar piles) and dynamic interference (construction workers, mobile equipment), reducing the conflict rate to 0%; Adapting to the actual passage requirements of the trolley: Size constraint penalties ensure that the path width and turning radius meet the trolley's volume and minimum turning radius requirements, avoiding the problem of "the path is feasible, but the trolley cannot pass"; Reducing the impact of hoisting inertia: Inertia penalties make the path smoother, reducing component sway amplitude (e.g., controlled within ±2cm), avoiding assembly deviations caused by swaying; Improving work efficiency: Heuristic functions guide the algorithm to quickly search for the optimal path, reducing not only path planning time but also local dynamic adjustment time, adapting to the real-time operation requirements of the trolley, and improving the assembly efficiency of large components.

[0082] In some preferred embodiments, the improved fuzzy PID algorithm dynamically adjusts the proportional, integral, and derivative parameters based on the attitude deviation and the rate of change of deviation.

[0083] Specifically, the traditional PID algorithm uses a fixed ratio (K). p ), integral (K) i ), differential (K) d The parameters are not suitable for dynamic working conditions such as load fluctuations (e.g., changes in component weight), road bumps, and hoisting inertial interference in prefabricated trolley operations, and are prone to problems such as large overshoot, delayed response, and large steady-state error.

[0084] The core optimization of the improved fuzzy PID algorithm is the introduction of a fuzzy inference module. Taking "attitude deviation" and "deviation change rate" as inputs, it dynamically adjusts the PID parameters according to preset fuzzy rules, enabling the parameters to adapt to changes in operating conditions in real time. Simultaneously, a lightweight inference design ensures a parameter response time ≤50ms, addressing the pain point of "parameter rigidity" in traditional PID algorithms and achieving precise and stable control of the trolley's attitude (travel angle, boom pitch angle). In other words, the improved fuzzy PID algorithm dynamically adjusts the proportional, integral, and derivative parameters based on attitude deviation and deviation change rate, with a parameter response time ≤50ms.

[0085] For example, the improved fuzzy PID algorithm is executed collaboratively by the edge computing terminal submodule and the PLC submodule. The process is divided into "data input → fuzzification → fuzzy inference → defuzzification → PID parameter update and control output", with a total time of ≤50ms, adapting to real-time control requirements. Specifically: 1) Data input and preprocessing (10ms).

[0086] The attitude detection module collects θact(t) and θref(t) every 10ms and calculates the attitude deviation e(t) using the following formula: e(t) = θref(t) - θact(t); Where θref(t) represents the attitude reference value, that is, the target attitude parameters of the trolley at time t, such as the travel angle and the boom pitch angle, derived from the improved A The path nodes output by the path planning algorithm are transformed; θact(t) represents the actual attitude value, that is, the actual attitude parameters of the trolley collected by the attitude detection module at time t; e(t) represents the attitude deviation, that is, the difference between the attitude reference value and the actual attitude value at time t. The rate of change of deviation ec(t) is calculated based on the deviation e(t-1) at the previous time step, using the following formula: ; Where ec(t) represents the rate of change of attitude deviation in period t; e(t) represents the real-time attitude deviation in period t; e(t-1) represents the attitude deviation in period (t-1); T s represents the discrete control period; t represents the control period index; Range normalization of e(t) and ec(t) (mapping to the interval [-1, +1]) facilitates fuzzy inference calculation.

[0087] 2) Fuzzification: Convert precise quantities to fuzzy quantities (5ms) The normalized e(t) and ec(t) are transformed into fuzzy linguistic variables E and EC. The core is to calculate the membership degree of each fuzzy subset through the membership function.

[0088] Where E and EC are fuzzy subsets, representing the linguistic variables after fuzzification of e(t) and ec(t), with values ​​{NB, NM, NS, ZO, PS, PM, PB}, corresponding to "negative large, negative medium, negative small, zero, positive small, positive medium, positive large" respectively; Taking an attitude deviation e(t) = 2° (normalized to 0.4) as an example, its fuzzification result is "PS (positive small)" and its membership degree μ PS =0.8, μ PM =0.2 (partially belonging to "the center"). Triangle membership function formula (taking e(t) as an example): ; Where x represents the normalized e(t); a, b, c represent the boundary values ​​of the fuzzy subset (e.g., the boundary of PS is [0.2, 0.4, 0.6]).

[0089] Function: To convert precise deviation values ​​into fuzzy language descriptions that are compatible with human control experience (such as "small positive deviation" and "medium negative rate of change").

[0090] 3) Fuzzy reasoning: Based on the rule base, output parameter increments (15ms).

[0091] The edge computing terminal submodule takes E and EC as inputs and infers U based on a preset fuzzy rule base. Kp U Ki U Kd (The fuzzy quantity of parameter increment).

[0092] Among them U Kp U Ki U Kd All are fuzzy output variables, corresponding to ΔK respectively. p , △K i , △K d The fuzzy linguistic variables take values ​​consistent with E and EC, namely {NB, NM, NS, ZO, PS, PM, PB}. The fuzzy inference synthesis formula (Mamdani) is as follows: ; Where μ U (u) represents the membership degree of the output fuzzy variable, indicating the degree of membership of the output fuzzy variable U (e.g., U0). Kp Corresponding △K p U Ki Corresponding △K i ), u is the specific fuzzy level of U, the membership degree at the discrete point u. Its engineering significance is that the higher the membership degree, the more the parameter increment corresponding to the discrete point u is adapted to the current working condition. This indicates the maximum value operation, which merges the output membership degrees of all matching rules and retains the maximum value. Its engineering logic is as follows: if multiple fuzzy rules match the current input at the same time (such as "E=PS and EC=ZO" and "E=PM and EC=NS" both match), the maximum value of the output membership degree of each rule is taken as the final fusion result to ensure that the weight of the best-fitting rule is maximized. This indicates the minimum value operation, which is used to calculate the trigger strength of a single rule. Its engineering logic is as follows: for a single rule (e.g., "If E=PB and EC=PB, then U..."),... Kp =PB”, taking the minimum value among the membership degrees of input E, input EC, and the rule itself as the trigger strength of the rule (the rule is effective only if all three satisfy a certain membership degree); x is the normalized precise quantity of attitude deviation, corresponding to the normalized precise value of attitude deviation e(t), which serves to connect precise input and fuzzy linguistic variables and is the intermediate carrier of fuzzification; y is the normalized precise quantity of deviation change rate, corresponding to the normalized precise value of deviation change rate ec(t), which, together with x, constitutes the precise input pair for fuzzy inference, ensuring the uniformity of the dimensions of the input data; μ E (x) represents the membership degree of the fuzzy subset of attitude deviation, indicating the degree to which the precise quantity x belongs to the fuzzy linguistic variable E (such as NB, NM, ZO, etc.). Its purpose is to quantify the degree of fuzziness in "the current attitude deviation belongs to positive small / positive medium / positive large," providing a basis for rule matching; μ EC (y) represents the membership degree of the fuzzy subset of the deviation change rate, indicating the degree to which the precise quantity y belongs to the fuzzy linguistic variable EC; μ R (x, y, u) represents the membership function of a fuzzy rule. The engineering setting is that the default membership degree of the preset rules in the rule base is 1 (e.g., "If E=PB and EC=PB, then U..."). Kp =PB” of μ R =1), indicating that the rule itself is completely valid; U is the output fuzzy linguistic variable, representing the output linguistic variable of fuzzy inference, corresponding to the fuzzy description of the PID parameter increment, with values ​​of {NB, NM, NS, ZO, PS, PM, PB} (negative large, negative medium, negative small, zero, positive small, positive medium, positive large), and its engineering correspondence is U. Kp Corresponding △K p U Ki Corresponding to K i U Kd Corresponding △K dIts function is to transform precise parameter increment requirements into fuzzy language descriptions, adapting to human control experience (e.g., "when the deviation is large, the proportional parameter increment is large"); u represents the discrete points of the output fuzzy variable, and the specific value of the output fuzzy linguistic variable U (i.e., one of {NB, NM, NS, ZO, PS, PM, PB}) is represented by a fuzzy level corresponding to the parameter increment for each discrete point. By traversing all discrete points, the complete membership distribution of the output variable can be obtained (e.g., U...). Kp The membership degrees at each point NB, NM, ..., PB provide data support for defuzzification (centroid method); Function: To fuse the outputs of all matching rules and obtain the fuzzy membership distribution of parameter increments.

[0093] 4) Defuzzification: Fuzzy quantity → Precise quantity (10ms).

[0094] The edge computing terminal submodule uses the centroid method (low computational cost, fast response) to convert the fuzzy output U... Kp U Ki U Kd Transformed into precise parameter increment ΔK p , △K i , △K d The formula for the center of gravity method is (with △K) p (For example) ; in This indicates the fuzzy output variable U Kp The global integral, U is U Kp The range of values ​​(e.g., △K) p [-3.0, 3.0]); u represents the fuzzy output variable U. Kp Continuous values ​​(replacing discretized u) k ); μ UKp (u) means U Kp Membership degree at continuous value u (instead of discretized μ) k ); Engineering Implementation: Discretization Calculation (discretizing ΔK) p The range of values ​​is divided into n discrete points), and the simplified formula is: ; Where △K p The proportional parameter increment (precise value) is the core result of the defuzzification output, used to update the PID proportional parameter; n is the number of fuzzy output discrete points, corresponding to the fuzzy output variable U. KpThe number of fuzzy subsets, where the fuzzy subsets are {NB, NM, NS, ZO, PS, PM, PB} (negative large, negative medium, negative small, zero, positive small, positive medium, positive large), each subset corresponds to one discrete point; k is the discrete point index, traversing all the discrete points corresponding to the fuzzy subsets, such as k=1 corresponding to NB, k=4 corresponding to ZO, k=7 corresponding to PB, its function is to extract the quantization value and membership degree of each fuzzy subset one by one, in preparation for weighted summation; u k The quantized value of the k-th discrete point belongs to the precise value corresponding to a pre-calibrated fuzzy subset (based on empirical calibration of the trolley PID parameters), such as ΔK. p u k Calibration value; μ k The membership degree of the k-th discrete point is μU, output by the Mamdani fuzzy inference synthesis formula. Kp (u k The physical meaning of this is that the greater the membership degree, the more the quantized value corresponding to the discrete point is adapted to the current working condition; This represents a weighted sum, which is the sum of the quantized values ​​of each discrete point according to their membership degree. The calculation logic is to highlight the contribution of discrete points with high membership degrees. This represents the sum of membership degrees, which is the sum of the membership degrees of all discrete points (serving a normalization function). Where u k μ represents the parameter increment value at discrete points. k Indicates the membership degree of the corresponding discrete point; output accuracy ΔK p ≤0.01, to ensure fine-grained parameter adjustment.

[0095] 5) PID parameter update and control output (8ms).

[0096] The defuzzy parameter increments are superimposed with the baseline parameters to update the real-time PID parameters. The parameter update formula is as follows: ; Where K p (t) / K i (t) / K d (t) represent the real-time PID parameters (proportional / integral / derivative), which are all dynamically adjusted parameters at time t. These are the core inputs of PID control, adapting to the trolley's attitude deviation / rate of change of deviation. For example, when the deviation is large, K... p (t) increases, accelerating the correction speed; K p0 / K i0 / K d0 This represents the PID reference parameters (proportional / integral / derivative), which are the initial reference values ​​from offline calibration (adapted to the normal operating conditions of the trolley). These serve as the basis for parameter adjustments, with fuzzy increments only fine-tuned to avoid control oscillations caused by sudden parameter changes; △K p / △K i / △Kd This represents the increment of the PID parameters (defuzzy output), which is the precise increment after defuzzification using the centroid method. The improved fuzzy PID algorithm's constraint logic requires that the parameters be within a preset dynamic range (e.g., K). p (∈[2.0, 8.0]), if it exceeds the limit, the boundary value is taken to avoid parameter runaway.

[0097] The PID control output formula is: ; Where u(t) is the PID control output signal, the real-time control command output by the algorithm to the execution drive module; t is the control time index (discrete-time variable), and is related to the control period T of the fuzzy PID. s Binding (t=0 is the initial control cycle, t=1 is the first control cycle, and so on); K p K(t) is a real-time proportional parameter that quickly responds to the current attitude deviation e(t). The larger the deviation, the more K... p The larger K is, the faster the correction can be achieved; i (t) represents the real-time integral parameter, which accumulates historical attitude deviations to eliminate steady-state errors (such as small attitude shifts after the trolley stops), thus avoiding the integral saturation problem of traditional PID controllers; K d (t) is the real-time differential parameter, reflecting the changing trend of attitude deviation and suppressing overshoot (such as oscillations during rapid boom movements). The greater the rate of change of deviation, the better K... d The larger the value of θref(t), the stronger the system damping; e(t) is the attitude deviation in the t-th period, e(t) = θref(t) - θact(t), where θref(t) is the target attitude value in the t-th period, derived from the improved A The path planning algorithm outputs parameters such as the trolley travel angle and the boom pitch angle; θact(t) is the actual attitude value in the t-th cycle, which is synchronously collected by the inertial measurement submodule and the tilt sensor submodule. A positive deviation indicates that the actual attitude is lower than the target, and a negative deviation indicates that the actual attitude is higher than the target. This represents the cumulative integral term for attitude deviation, accumulating all attitude deviations from the initial period (k=0) to the current period (k=t). Each deviation is multiplied by the control period Ts, replacing continuous integration and eliminating steady-state errors that proportional control cannot address, such as small attitude deviations caused by hydraulic system leaks. `k` is the integral summation index, traversing all control periods from the initial period (k=0) to the current period (k=t). It is used only for the summation of the integral term and is distinct from the control time index `t` (k is the internal variable for summation, t is the current period index), avoiding confusion between the summation operation and the current period, and ensuring the integrity of the integral term accumulation. `e(k)` represents the historical attitude deviation of the k-th period, belonging to cached historical attitude deviation data, providing historical data support for the integral term, ensuring the continuity of the integral accumulation, and is completely consistent with the definition of `e(t)` (only the period is different). `T` s For discrete control cycles, discrete deviation data are transformed into integral / differential terms with a time dimension to ensure dimensional uniformity. The rate of change of attitude deviation (discrete differential term) is expressed as the difference between the deviation of the current cycle and the previous cycle, divided by the control cycle T. s The average rate of change of the deviation is obtained, replacing the continuous derivative to avoid noise amplification. It is used to predict the trend of deviation change and suppress overshoot in advance. For example, when the deviation increases from 1° to 2°, the derivative term outputs a positive signal to increase the control damping. e(t-1) is the attitude deviation in the (t-1)th cycle. After the calculation of the current cycle t is completed, e(t) is cached as e(t-1) for the next cycle to ensure the continuity of the derivative term calculation. It is completely consistent with the acquisition source and preprocessing method of e(t) (both come from the attitude detection module and Kalman filter). 6) Control signal issuance and feedback (2ms).

[0098] The PLC submodule receives the control output u(t) and converts it into control commands (such as motor frequency conversion commands and hydraulic cylinder solenoid valve control signals) that can be recognized by the execution drive module to drive the trolley attitude adjustment; at the same time, the attitude detection module collects the new θ. act (t+1), enter the next control cycle, forming a closed-loop control.

[0099] The improved PID fuzzy algorithm's practical application benefits include: dynamic adaptation to operating conditions: Real-time adjustment of PID parameters through fuzzy inference reduces attitude overshoot (from 15% in traditional PID to below 3%) and steady-state error (≤0.1°) under conditions such as load changes (e.g., component weight increasing from 10 tons to 50 tons) and road bumps; reduced parameter response time (e.g., ≤50ms): The fuzzy inference module employs a lightweight design (discrete membership function, simplified rule base search), combined with parallel processing of the edge computing terminal submodule, allowing for a total control cycle time of ≤30ms with a 20ms redundancy to meet real-time control requirements; ensures stable attitude control: Dynamic adjustment of differential parameters enhances system damping, reducing sway amplitude during boom pitch angle adjustments (e.g., ≤±2cm), preventing component collisions, and supporting high assembly accuracy (e.g., ±1mm); and strong engineering practicality: the reference parameter K... p0 K i0 K d0 It can be calibrated according to the trolley model and component weight. The fuzzy rule base is optimized based on on-site construction experience and can be implemented without complicated debugging.

[0100] In some preferred embodiments, a safety warning module is also included. The safety warning module has a three-level warning mechanism, corresponding to audible and visual prompts, speed reduction operation, and emergency shutdown, with an emergency shutdown response time of ≤200ms.

[0101] For example, considering the "high-risk, high-precision, and dynamic working conditions" of prefabricated trolley hoisting operations, the triggering conditions of the three-level early warning mechanism need to be based on data from multiple modules such as environmental perception, attitude detection, load detection, health diagnosis, and leakage current monitoring, and designed according to a "progressive risk gradient" (from minor alerts to emergency avoidance) to ensure the accuracy and timeliness of the early warning (e.g., emergency shutdown response time ≤ 200ms). Specific triggering conditions are as follows: A Level 1 warning (audio-visual alert) indicates a minor risk and only requires attention. Its triggering conditions (any one of which needs to be met) are: Environmental perception: Ultrasonic sensors detect nearby obstacles at a distance ≥1.5m and <2.0m (non-emergency collision risk, only avoidance is required); visual cameras identify construction personnel entering the work area warning range (≥5m, not within the dangerous work radius); Attitude deviation: The attitude detection module collects the deviation of the trolley travel angle / lifting boom pitch angle e(t)∈[±0.5°, ±1.0°] (slight deviation, within the dynamic adjustment range of the PID algorithm, does not affect assembly accuracy); Load / leakage: The load sensor detects a lifting load ≥80% of the rated load. Load <90% (approaching heavy load, warning against further load increase); Leakage detection module detects leakage current ≥10mA and <30mA (minor leakage, not reaching the safety threshold); Health diagnosis: The health diagnosis submodule detects minor wear on key components of the actuator (bearings, hydraulic cylinders) through an improved BP neural network (wear Ra=0.8~1.5μm, not affecting operation in the short term); Data interference: The anti-electromagnetic interference module detects a sensor data transmission error rate ≥0.005% and <0.01% (increased interference, warning to pay attention to data reliability).

[0102] Warning measures: The yellow indicator light on the control box flashes and the buzzer sounds an intermittent alarm (frequency 1Hz). This does not affect the normal operating speed of the trolley, but only reminds the operator to pay attention to the risk source and prepare to avoid or adjust it.

[0103] Level 2 warning (reduced speed operation) indicates moderate risk, requiring speed reduction for avoidance. The triggering conditions (any one of which is sufficient) are: Environmental perception: Ultrasonic sensors detect an obstacle at a distance ≥1.0m and <1.5m (increased collision risk, requiring speed reduction for avoidance); Visual cameras / LiDAR detect dynamic interference (such as mobile devices, construction personnel) entering the danger radius of the work area (≥3m and <5m); Attitude deviation: Attitude deviation e(t)∈[±1.0°, ±2.0°] (PID algorithm can still be adjusted, but normal speed easily leads to increased overshoot); Deviation The rate of change e(t) ∈ [±0.5° / s, ±1.0° / s] (the deviation shows an increasing trend); Load / Leakage Category: Lifting load ≥90% of rated load and <100% (close to the rated value, avoid overload); Leakage current ≥30mA and <50mA (increased leakage risk, not reaching the electric shock danger threshold); Health Diagnosis Category: Moderate wear of actuator components (Ra=1.5~3.2μm), vibration amplitude collected by vibration sensor ≥1.5g and <2.0g (g is the acceleration due to gravity, the stability of equipment operation decreases); Path / Environment Category: Improvement A The path planning algorithm detects new static obstacles (such as temporarily piled steel bars) ahead of the path, and needs to replan the local path at low speed; the operating environment temperature is ≥45℃ and <50℃ (high temperature can easily lead to a decrease in the efficiency of the hydraulic system).

[0104] Warning measures: The buzzer will sound a continuous alarm (frequency 2Hz) and the red indicator light will flash. The PLC submodule will automatically reduce the trolley travel speed and the lifting arm movement speed to 50% of the rated speed. Operators can adjust the path or avoid risks while the speed is reduced. If the risk is not eliminated within 10 seconds, it will automatically upgrade to a level three warning.

[0105] A Level 3 warning (emergency shutdown) indicates severe risk and requires immediate evacuation. Its triggering conditions (any one of which must be met, with a response time ≤ 200ms) are: Collision Risk Category: Ultrasonic sensor detects obstacle distance <1.0m (imminent collision); LiDAR / visual camera detects dynamic interference (construction personnel, mobile equipment) entering the hoisting danger radius (<3m); Attitude Loss of Control Category: Attitude deviation e(t) ≥ ±2.0° (exceeding the adjustment capability of PID algorithm, which may cause component swaying and falling); Hoisting boom pitch angle change rate > ±1.0° / s (attitude loss of control); Overload / Leakage Category: Hoisting load ≥ 105% of rated load (severe overload, which may cause hydraulic system rupture); Leakage current ≥ 50mA (reaching the human electric shock danger threshold, G (B / T3811-2008 standard); Equipment failure type: The health diagnosis submodule detects severe wear (Ra≥3.2μm) or failure of key components of the actuator (such as bearing jamming, hydraulic cylinder leakage ≥5mL / min); vibration amplitude ≥2.0g (equipment is on the verge of failure); Data failure type: sensor data transmission error rate ≥0.01% (data is severely distorted and cannot support precise control); failure of core sensors such as inertial measurement submodule (IMU) and vision camera (no data output); Human intervention type: operator presses emergency stop button (hardware redundancy trigger, highest priority).

[0106] Warning measures: The buzzer sounds a high-decibel continuous alarm and the red and yellow lights are constantly on at the same time. The PLC submodule immediately cuts off the power supply to the execution drive module (the motor and hydraulic cylinder stop moving). The hoisting mechanism activates the mechanical locking device (to prevent the component from falling). Ensure that all actions stop within 100ms. The remaining 100ms is used for risk status locking and alarm information uploading, meeting the response requirement of "≤200ms".

[0107] In some preferred embodiments, an electromagnetic interference suppression module is also included. The electromagnetic interference suppression module integrates a shielding housing submodule and a signal filtering circuit submodule. The shielding housing submodule controls the activation of the copper-aluminum composite anti-interference shielding housing device, and the signal filtering circuit submodule suppresses industrial electromagnetic radiation interference through an LC filter network.

[0108] Specifically, the anti-electromagnetic interference module integrates a shielding housing submodule and a signal filtering circuit submodule: constructing a complete "shielding-filtering" two-level protection system, forming a multi-layered electromagnetic interference protection barrier. Specifically, it provides system-level protection by combining shielding and filtering technologies to block external electromagnetic radiation intrusion and purify internal signal transmission, covering both conducted and radiated interference, the two main forms of EMI; modular design facilitates system integration and maintenance, with the anti-interference shielding housing providing basic protection and the filtering circuit providing fine suppression of interference in specific frequency bands; and enhanced reliability: the dual protection mechanism enhances the system's stability in strong electromagnetic environments (such as industrial sites), ensuring accurate transmission of control signals and sensor data.

[0109] The shielding shell submodule controls the activation of the copper-aluminum composite anti-interference shielding shell device: utilizing the complementary properties of copper and aluminum, it achieves wide-bandwidth, high-efficiency electromagnetic shielding. The working mechanism of the copper-aluminum composite material has the following advantages: High-frequency shielding: the copper layer (inner layer) possesses high conductivity (approximately 5.96 × 10⁻⁶). 7 S / m), generating strong reflection loss for high-frequency electromagnetic waves (100kHz~1GHz) (about 3dB higher than aluminum), reflecting most of the interference back to the source; Low-frequency absorption: The aluminum layer (outer layer) provides good permeability and weight advantages, effectively absorbing low-frequency magnetic field interference (10Hz~100kHz) and reducing overall weight. Composite effect: The shielding effectiveness (SE) formed by the combination of the two materials is better than that of a single material, and can cover the full spectrum of interference from power supply interference to radio frequency radiation in industrial environments.

[0110] The signal filtering circuit submodule suppresses industrial electromagnetic radiation interference through an LC filter network: Residual electromagnetic interference that has penetrated the shielding enclosure is precisely filtered out by the LC filter network, ensuring signal purity. The working principle of the LC filter network is as follows: Differential mode interference suppression: The LC circuit, composed of an inductor (L) and a capacitor (C), forms an impedance barrier on the signal line, generating high impedance for specific frequency interference and low impedance for useful signals; Common mode interference elimination: Through the combination of common mode inductors and capacitors, common mode noise is bypassed to ground, purifying the power supply and signal reference potential; Frequency selectivity: A specific cutoff frequency (e.g., 100kHz) can be designed to precisely filter out common industrial environment interferences such as switching power supply noise and inverter harmonics (typical frequency range 50Hz~1MHz). The implementation methods are specifically divided into: π-type filter structure: composed of two capacitors and one inductor (such as C1-L-C2), which forms multi-stage attenuation for high-frequency interference, improving noise suppression capability by 3-5 times; T-type filter structure: suitable for analog circuits with high signal integrity requirements, reducing signal distortion; multi-stage LC cascade: for complex industrial electromagnetic environments, a 2-3 stage filter network is constructed, with a total attenuation of over 60dB. This ensures accurate data from precision measuring equipment such as IMUs and tilt sensors (error <0.05°); prevents false triggering of control signals (such as emergency stop malfunction rate <0.001%); and guarantees A... The path planning and fuzzy PID control algorithm operates stably (control cycle fluctuation < ±5%).

[0111] In some preferred embodiments, the environmental sensing module further includes a temperature sensor submodule, which collects the surface temperature of the component in real time; the edge computing terminal submodule is equipped with a component temperature deformation model unit, which combines the component surface temperature to correct the component assembly coordinates and compensate for the component size deviation caused by temperature changes.

[0112] Specifically, the environmental perception module also includes a temperature sensor submodule, which collects the surface temperature of the components in real time: providing accurate input data for deformation compensation and solving the problem of "unable to quantify deformation due to lack of temperature data". Real-time acquisition ensures dynamic tracking of temperature changes, ensuring timely compensation and adapting to temperature fluctuation scenarios during construction (such as sudden temperature changes caused by tunnel ventilation).

[0113] The edge computing terminal submodule is equipped with a component temperature deformation model unit. The component temperature deformation model unit combines temperature data to correct the component assembly coordinates and compensate for the component size deviation caused by temperature changes. By quantifying temperature deformation, the deviation is converted into the assembly coordinate correction amount, realizing precise linkage compensation of "temperature-deformation-coordinates", which solves the technical pain point of "assembly misalignment caused by temperature deformation".

[0114] For example, the calculation formula for the component temperature deformation model element is as follows: △L = α·L0·△T; X comp =X ref +△L; Where △L represents the temperature deformation increment of the component along the assembly direction (unit: mm), with positive values ​​indicating thermal expansion and negative values ​​indicating thermal contraction; α represents the linear expansion coefficient of the component material (unit: 1 / ℃); L0 represents the original length of the component along the assembly direction (unit: mm), determined by the design drawings (e.g., the assembly direction length of a bridge segment L0 = 5000 mm); △T represents the temperature change of the component (unit: ℃), △T = T comp (t)-T0,T comp (t) represents the real-time surface temperature of the component (°C) collected by the temperature sensor; T0 represents the component's calibration temperature (usually 20°C, i.e., the reference temperature corresponding to the dimensions marked on the design drawings); X comp This represents the corrected assembly target coordinates (unit: mm), which are the final control commands issued to the execution drive module; X ref Indicates the original assembly reference coordinates (unit: mm), derived from Improvement A. The target location coordinates output by the path planning algorithm.

[0115] The compensation logic is as follows: the temperature sensor collects T... comp (t) → Calculate △T → Obtain △L through deformation model → Correct the original coordinates X ref → Output X comp ; Accuracy assurance: Taking a 5m long concrete component as an example, if the temperature change ΔT = 10℃, then ΔL = 1.0 × 10 -5 ×5000×10=0.5mm. If this deviation is not compensated, it will cause misalignment during assembly. This effect can be completely offset by coordinate correction. Scene adaptation: Adaptable to construction temperature range of -10~50℃, covering extreme working conditions such as outdoor winter and tunnel summer, ensuring that assembly accuracy is not affected by temperature fluctuations.

[0116] In some preferred embodiments, the intelligent control module also includes a collaborative control submodule, which establishes communication with multiple trolleys through a 5G industrial module to achieve data sharing of operations among the multiple trolleys and allocates lifting loads and receives motion paths based on a distributed collaborative algorithm.

[0117] Specifically, the intelligent control module also includes a collaborative control submodule. This submodule establishes communication with multiple vehicles via a 5G industrial module, enabling data sharing among them. It builds a highly reliable, low-latency collaborative communication link, resolving collaborative failures caused by "data asynchrony among multiple devices." Specifically, the 5G industrial module offers advantages such as: communication latency ≤20ms (meeting real-time collaborative requirements), bandwidth ≥100Mbps (supporting multi-dimensional data transmission), and strong anti-interference capabilities (suitable for industrial construction scenarios). Data sharing includes: real-time location, load status, attitude data, obstacle detection results, and remaining lifting capacity for each vehicle, providing comprehensive input data for the distributed collaborative algorithm to ensure decision consistency.

[0118] Distributed collaborative algorithm-based allocation of lifting load and movement path: Through decentralized algorithmic logic, dynamic optimization of load and path is achieved, avoiding the problems of "single point of failure in centralized control" and "low efficiency of manual coordination." Decentralization refers to avoiding all sub-modules being controlled by the same module, but instead optimizing to a collaborative processing approach. Specifically, load allocation: Based on the rated load and real-time load margin of each trolley, the total lifting load is evenly distributed to avoid overloading of a single trolley; path coordination: Based on shared obstacle data, the movement path of each trolley is dynamically planned to ensure no collisions and no intersections, adapting to the joint lifting needs of ultra-large components (such as long-span bridge segments and tunnel segment assemblies).

[0119] For example, the distributed collaborative algorithm dynamically allocates the lifting load based on the load capacity of each vehicle, ensuring that the load of a single vehicle does not exceed the rated value and the allocation error is ≤5%. The core formula is as follows: ; ; stF i ≤F rated,i (i=1,2,...,n) Where F i This represents the allocated load of the i-th trolley, which is the core control parameter output by the algorithm. The specific lifting load allocated to the i-th trolley is directly sent to the execution drive module of each trolley to ensure load balance among multiple trolleys; k i Let k be the load distribution coefficient for the i-th trolley. The distribution coefficient is positively correlated with the trolley's "remaining load capacity." The stronger the remaining capacity, the better the load distribution coefficient. i The larger the value, the more dynamic equilibrium is achieved where "the capable do more work"; F total For the total hoisting load of ultra-large components; F rated,i F is the rated lifting load of the i-th trolley; current,i This represents the actual current load of the i-th trolley. The total remaining load capacity of all trolleys is defined to ensure that the allocation coefficient is calculated based on the global remaining capacity, avoiding overloading of a single trolley while maximizing the utilization of the load potential of all trolleys; n is the number of trolleys in collaborative operation; j is the summation index number to ensure that the summation covers all collaborative trolleys without omissions or duplicate calculations; st is the constraint condition symbol, making F i ≤F rated,i The load assigned to the i-th trolley must not exceed its rated load. This is to prevent safety accidents such as hydraulic system rupture and component falling caused by overloading of a single trolley, which is the bottom line for safety in collaborative operations. i is a trolley index variable that identifies a single trolley in collaborative operations. It is bound to the real-time position and attitude data of the trolley to ensure that load allocation and path coordination are executed synchronously. The path coordination assistance logic is as follows: each vehicle adjusts its path based on a shared global obstacle map and its own position through a distributed negotiation mechanism. Using the "collaborative operation area center" as a reference, each vehicle is assigned an independent movement trajectory zone (trajectory zone width ≥ vehicle width + 0.5m safety margin). If a vehicle detects a new obstacle on its path, it synchronizes this information to other vehicles in real time via 5G communication. Each vehicle then uses a local path replanning algorithm (improved A...) The path planning algorithm is dynamically adjusted to ensure that overall coordination is not affected.

[0120] In some preferred embodiments, the intelligent control module is also equipped with a remote control submodule, which connects to the remote control center through an encrypted communication protocol, supports AR visual interaction, and the remote control center can obtain data such as the trolley's posture and working environment in real time, and take over control in the event of a local system failure to carry out unmanned emergency operations.

[0121] Specifically, the intelligent control module also includes a remote control submodule, which connects to the remote control center via an encrypted communication protocol: establishing a secure and interference-resistant remote communication link to solve the problems of "remote data transmission leakage and tampering" and "communication interruption in industrial environments." It employs the AES-256 encryption algorithm (symmetric encryption and asymmetric key exchange) to ensure data transmission security (cracking difficulty ≥10). 25 (Times); communication latency ≤30ms (adapting to real-time control requirements), compatible with 5G industrial modules / fiber optic communication, electromagnetic interference resistance in synergy with the aforementioned electromagnetic interference resistance module; by establishing a reliable connection between local and remote systems, it provides secure link support for subsequent data sharing and remote takeover.

[0122] Supporting AR visualization interaction, the remote control center can acquire real-time data such as trolley posture and working environment, breaking the "information blind spot" of remote control and solving the operational obstacles of "not being able to see or see accurately remotely." The 3D working environment data collected by the environmental perception module (LiDAR, visual camera) and the trolley posture data from the posture detection module are overlaid on the AR terminal (such as AR glasses, large screen) to generate a visualization interface that blends "virtual annotation with real-world scene" (such as component target location annotation, obstacle highlighting, and posture deviation value overlay). The transmission frequency is synchronized with the local system (≥10Hz) to ensure that the data acquired by the remote control center is consistent with the on-site status, providing accurate basis for remote decision-making.

[0123] Taking over control in case of local system failure and implementing unmanned emergency operations: Providing fault redundancy to avoid operation interruptions or safety accidents caused by local system failure is the core emergency function. Among them, the remote control submodule monitors the status of the local system in real time (such as the validity of sensor data, the rationality of algorithm output, and the response status of actuators). When a "local system fault signal" is detected (such as no effective output for three consecutive control cycles or sensor data distortion), the takeover process is automatically triggered (e.g., the control can ensure that the takeover response time is ≤50ms). No manual on-site inspection is required; operations such as component reset, trolley evacuation, and fault alarm can be completed remotely, reducing personnel risks in high-risk scenarios and minimizing operation interruption time.

[0124] For example, the core of AR visualization is to achieve "precise overlay of virtual annotations and real-world scenes," and the core mapping relationship is as follows: P AR (x, y, z) = Calib(P) scene (x,y,z))+Offset(x,y,z); Where P AR (x, y, z) represents the virtual coordinates displayed on the AR terminal (e.g., the target assembly position of a component); P scene (x, y, z) represents the real-world 3D coordinates collected by the environmental perception module; Calib(·) represents the calibration function (to eliminate installation deviations of the camera / LiDAR, with a calibration error ≤ 0.5 mm); Offset(x, y, z) represents the virtual-real fusion offset (dynamically adjusted according to the AR terminal calibration parameters to ensure superposition accuracy ≤ 1 mm).

[0125] Function: To ensure that the "virtual guidance" seen by remote operators is completely aligned with the actual scene, and to avoid remote operation errors caused by visual discrepancies.

[0126] In some preferred embodiments, the execution drive module also includes a health diagnosis submodule. The health diagnosis submodule collects operating data of key components of the actuator through vibration sensors, analyzes the wear status of the components in combination with an improved BP neural network algorithm, predicts failure risks in advance, and generates maintenance prompts.

[0127] Specifically, the execution drive module also includes a health diagnosis submodule. This submodule collects operational data from key components of the actuator using vibration sensors: acquiring "physical characteristic data" of the component's operating status to provide accurate input for fault diagnosis and solve the problem of "inability to determine component health status due to lack of effective data support." The data collection targets are: core vulnerable components of the actuator (such as the hydraulic pump bearing of the lifting arm, the travel motor bearing, and the hydraulic cylinder piston rod); the collection parameters include: vibration amplitude (unit: g), root mean square value (RMS), peak factor, kurtosis, and characteristic frequency (unit: Hz), totaling 5-8 key vibration characteristics (wear will lead to increased vibration amplitude and kurtosis); the collection frequency is ≥100Hz, with one set of data collected every 10ms to track the component's operating status in real time and avoid missed diagnoses due to data lag.

[0128] By combining an improved BP neural network algorithm to analyze component wear status, the system can predict failure risks in advance and generate maintenance tips. The optimized BP neural network transforms "vibration characteristic data" into "quantified wear status," enabling accurate fault prediction and targeted maintenance. This addresses the core issues of "low efficiency and high false alarm rate in traditional manual inspection." The algorithm's core is an improved BP neural network (optimizing convergence speed and generalization ability) that replaces the traditional "threshold judgment method" (which can only determine whether a fault exists, but cannot quantify the degree of wear). Output results include: ① Component wear level (0-4 levels: Level 0 no wear, Level 1 slight wear, Level 2 moderate wear, Level 3 severe wear, Level 4 fault); ② Specific wear amount (unit: μm); ③ Maintenance tips (e.g., "Hydraulic pump bearing wear 1.2μm, Level 3 wear, recommended replacement within 72 hours"). The early warning value lies in: proactively avoiding sudden failures (e.g., bearing seizure causing sudden boom stoppage) and reducing operation downtime (e.g., reducing from the traditional 4-8 hours / time to 1-2 hours / time).

[0129] To address the issue of "poor performance," the improved BP neural network algorithm is optimized through "momentum term + L2 regularization" to adapt to the complexity of vibration data in industrial scenarios (high noise levels and coupled features). The improved weight update formula (adding a momentum term to accelerate convergence) is as follows: w ij (k+1)=w ij (k)+η·δ j ·o i +α·[w ij (k)-w ij(k-1)]; b j (k+1)=b j (k)+η·δ j +α·[b j (k)-b j (k-1)]; Where w ij (k+1) represents the weight in the (k+1)th iteration, measuring the contribution of the output of the i-th neuron in the previous layer to the input of the j-th neuron in the next layer. It adapts to the high-dimensional coupling characteristics of vibration feature data (such as vibration amplitude and kurtosis), and accelerates weight convergence through the momentum term, thereby improving the accuracy of wear state identification; w ij (k) represents the weight in the k-th iteration, serving as the benchmark for the next iteration's weight update. It inherits from previous iterations and is synchronized with the training process of the vibration feature data. With each iteration, the weights are adjusted based on sample errors, gradually optimizing the model's mapping relationship between "vibration features" and "wear status." k represents the iteration number index, which is the training iteration count of the neural network. η represents the learning rate (step size factor), controlling the step size of the weight update. This avoids excessively large step sizes leading to training oscillations, and excessively small step sizes leading to slow convergence. Given the high noise level in vibration data, η=0.01 balances the error correction speed with noise immunity, preventing noise-induced weight update disorder. δ j Let $\mathbf{j}$ represent the error term of the $j$-th neuron, which is the quantized value of the deviation between the neuron's output and the true label. Its calculation formula is: , where y j This represents the true label of the j-th neuron (e.g., the amount of wear corresponding to the vibration sample). Let σ′ represent the predicted output of the j-th neuron. j ) represents the derivative of the activation function (when ReLU activation is first used, i.e. Where netj represents the net input of the j-th neuron, which is the weighted sum of the outputs and weights of the previous layer's neurons, plus the bias term. , where w ij Let o be the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. i b is the output of the i-th neuron in the previous layer. j (the bias of the j-th neuron in the current layer); δ j Its function is to guide the direction of weight updates; the larger the error term, the greater the weight adjustment range, thus quickly reducing prediction bias. i This represents the output value of the i-th neuron. If i is an input layer neuron, then o i =x i (Input layer vibration feature data, such as normalized vibration amplitude); if i is a hidden layer neuron, then o i =σ(neti ), o i Used to convey effective information about vibration characteristics, providing feature contribution information for weight updates, and adapting to accurate gradient calculation in feature-coupled scenarios; α represents the momentum factor, introducing the inertia of the previous weight update, simulating momentum in physics, accelerating convergence and escaping local minima. Given the large number of local optima in industrial vibration data (spurious features caused by noise), α=0.9 can enhance the inertia of weight updates, preventing the model from getting trapped in local minima corresponding to spurious features, and improving generalization ability; W ij (k)-w ij (k-1) represents the weight increment between the k-th and (k-1)-th iterations. As input to the momentum term, it reflects the historical trend of weight updates. If the previous weight update was in the direction of decreasing error, the momentum term will help the current weight continue to adjust in that direction, accelerating convergence (e.g., if the weights are updated positively for three consecutive iterations, the momentum term will be superimposed with a positive increment); b j (k+1) represents the bias term in the (k+1)th iteration, i.e., the bias value of the j-th neuron in the neural network. It is used to adjust the activation threshold of the neuron, so that the model can better fit the nonlinear relationship between vibration characteristics and wear state, compensate for baseline drift of vibration data (such as feature offset caused by sensor zero drift), and improve the robustness of the model to noise; b j (k) represents the bias term in the k-th iteration, and is related to the weight w. ij (k) Synchronous updates jointly optimize the model's prediction accuracy. The momentum term of the bias term is optimized in sync with the weights to ensure an overall improvement in the model's convergence speed, adapting to the real-time analysis requirements of the health diagnosis submodule; i represents the index of the previous layer's neurons, identifying the input neurons connected by the weights, clarifying the connection relationship of the weights, and avoiding hierarchical confusion; j represents the index of the next layer's neurons, identifying the output neurons connected by the weights, working in conjunction with the index i of the previous layer's neurons to uniquely determine a certain connection weight, ensuring accurate update logic; The formula for the error function (with L2 regularization to suppress overfitting) is: ; Where E represents the total error function (loss value), which measures the deviation between the model's predicted value and the true value, and the goal is to converge to E≤0.001; m represents the number of training samples (engineering value m≥1000), the samples are vibration feature data under different wear conditions (e.g., 200+ samples each for wear levels 0-4), 1 / m is used to normalize the error and avoid fluctuations in the magnitude of the error value caused by differences in the number of samples. This represents the training sample summation operator, ensuring that error calculation covers all training samples, reflecting the model's fit to the full dataset, and avoiding insufficient training caused by local sample bias; y pThe true label of the p-th sample (such as actual wear amount or wear level) is determined by experiment (e.g., by measuring wear amount using a laser diameter gauge). This represents the model prediction value (output of the output layer) for the p-th sample. λ represents the squared error of the p-th sample. The squared error amplifies the weight of larger deviations, so that the model prioritizes correcting samples with large prediction errors (such as vibration samples in a severely worn state), avoiding the cancellation of positive and negative deviations and ensuring the effectiveness of error calculation; λ represents the L2 regularization coefficient (engineering value λ=0.001), which suppresses excessively large weights and reduces the risk of model overfitting (avoiding overfitting to noisy data). This represents the L2 regularization term (weight penalty term), which penalizes the sum of squares of all connection weights in the neural network. It penalizes weights with large absolute values, forcing the model to use more small weights to fit the data, simplifying the model structure, improving generalization ability, and avoiding overfitting to vibration noise; w ij The term represents the connection weights of the neural network, which are the strength of the connections between neurons and reflect the contribution of the output of the previous layer of neurons to the input of the next layer of neurons. The regularization term acts directly on the weights and works in conjunction with the momentum term to optimize the weight updates (the momentum term accelerates convergence, and the regularization term suppresses overfitting), adapting to the complexity of industrial vibration data; i and j are neuron index variables, where i is the index of the previous layer of neurons (input layer / hidden layer) and j is the index of the next layer of neurons (hidden layer / output layer). Their function is to traverse the connection weights between all neurons in the neural network, ensuring that the regularization term covers all weights without any omissions or penalties.

[0130] The wear state mapping formula (from output layer to actual wear amount) is: ; Wear represents the actual wear amount of the component in the final output (unit: μm), which is the core indicator used to generate maintenance tips; k and c represent calibration coefficients (obtained by fitting experimental data, such as k=5.0 and c=0.1 for bearing components), which realize the mapping from the model output value to the actual wear amount.

[0131] The improvements to this algorithm are as follows: the momentum term α increases the convergence speed, avoiding the problem that traditional backpropagation (BP) still fails to converge after more than 1,000 iterations on industrial data; the L2 regularization coefficient λ reduces the fault misjudgment rate and adapts to the noise interference of vibration data in construction scenarios; and the ReLU activation function solves the gradient vanishing problem and improves the feature extraction capability of deep networks (adapting to the complex mapping of 5-8 dimensional vibration features).

[0132] Through the above technical solutions, the prefabricated trolley control system provided in this application has the following advantages: ① Positioning accuracy is significantly improved, meeting the requirements of high-precision assembly: Through the multi-source sensor data fusion system of the environmental perception module and the improvement of A A closed-loop control system is constructed using path planning and improved fuzzy PID control, combined with an improved YOLOv8 algorithm, providing a precise sensing foundation for positioning. A temperature deformation compensation model is used to offset dimensional deviations caused by thermal expansion and contraction, improving component assembly accuracy and eliminating the need for repeated adjustments, thus increasing construction efficiency. The improved fuzzy PID algorithm, through dynamic parameter tuning (response time ≤ 50ms), effectively suppresses attitude overshoot and component sway, ensuring that the control error of the boom pitch angle and trolley travel angle is ≤ 0.1°, further supporting high-precision assembly.

[0133] ② Strong environmental adaptability, suitable for complex construction scenarios: The environmental perception module, including the lidar submodule, vision camera submodule, and ultrasonic sensor submodule, achieves full-range environmental coverage from far to medium to near distances. The data fusion submodule uses a Kalman filter algorithm to reduce noise, solving the problem of sensor data distortion caused by tunnel dust and changes in lighting. The anti-electromagnetic interference module can suppress industrial electromagnetic radiation interference, ensuring stable and reliable transmission of sensor and control signals, and is suitable for strong electromagnetic environments such as substations and industrial plants. The temperature sensor submodule covers an operating temperature range of -10℃ to 50℃, and through deformation compensation, the system can maintain stable accuracy under extreme temperature conditions, breaking through the dependence of traditional systems on ambient temperature.

[0134] ③ Optimized control coordination, improving operational efficiency and safety: The architecture of the edge computing terminal submodule and PLC submodule enables high-speed data interaction between multiple modules, allowing intelligent linkage between the trolley's movement and the lifting mechanism's actions, avoiding conflicts inherent in traditional manual control. The collaborative control submodule, through 5G communication and distributed algorithms, achieves balanced load distribution and collision-free path coordination among multiple trolleys, breaking through the load limitations of a single trolley. This allows for efficient joint lifting of ultra-large components (such as thousand-ton bridge segments), improving operational efficiency. Improved A The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference (such as construction workers and temporary equipment), completely solving the collision risk caused by traditional path planning that "only focuses on the shortest path and ignores equipment characteristics".

[0135] ④ Ample safety redundancy reduces operational risks: A three-level early warning mechanism (audio-visual alerts → speed reduction → emergency shutdown) covers all scenarios from minor to severe risks, with an emergency shutdown response time of ≤200ms, quickly avoiding dangers such as collisions, overloads, and electrical leaks; the remote control submodule, through encrypted communication and AR visualization, enables unmanned operation in high-risk scenarios (deep tunnels, high altitudes), while also supporting remote takeover in case of local failure, preventing personnel from entering dangerous areas and reducing the accident rate. The health diagnosis submodule, through an improved BP neural network algorithm, predicts the wear status of key components of the actuator (bearings, hydraulic cylinders) 24-72 hours in advance, generating targeted maintenance prompts to avoid work interruptions or equipment damage caused by sudden failures, reducing maintenance costs.

[0136] ⑤ Strong functional scalability, adaptable to diverse operational needs: The collaborative control submodule supports dynamic networking of multiple vehicles, and the number of collaborative vehicles can be flexibly adjusted according to the weight and size of components; the remote control submodule is compatible with various interactive terminals such as AR glasses and large screens, supporting both daily visual monitoring and emergency takeover in case of local system failures, adapting to the control needs of different construction scenarios. The overall architecture adopts a modular design (such as safety early warning, anti-electromagnetic interference, and health diagnosis modules, which are all optional expansions), and can be flexibly configured according to project budget and operational complexity, balancing economy and practicality. Core algorithms (improved YOLOv8, A...) The path planning, improved fuzzy PID, and improved BP neural network all adopt lightweight design to adapt to the computing power of edge computing terminals and ensure long-term stable operation of the system.

[0137] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A prefabricated trolley control system, characterized in that, It includes an environmental perception module, an attitude detection module, an intelligent control module, and an execution drive module, and the modules are interconnected. The environmental perception module includes a lidar submodule, a vision camera submodule, an ultrasonic sensor submodule, and a data fusion submodule. The lidar submodule is used to scan the three-dimensional contour of the work area, the vision camera submodule is used to identify component marker points, the ultrasonic sensor submodule is used to detect nearby obstacles, and the data fusion submodule uses a Kalman filter algorithm to perform noise reduction processing on the multi-source sensor data. The attitude detection module includes an inertial measurement submodule and a tilt sensor submodule. The inertial measurement submodule and the tilt sensor submodule synchronously collect data on the trolley travel angle, the boom pitch angle, and the component attitude. The intelligent control module includes an edge computing terminal submodule and a PLC submodule. The edge computing terminal submodule runs an improved A... The path planning algorithm and the improved fuzzy PID algorithm are used to output control commands. The PLC submodule receives the control commands output by the edge computing terminal submodule and sends them to the execution drive module. The execution drive module executes control commands and controls the movement of the control vehicle and the hoisting mechanism.

2. The prefabricated trolley control system according to claim 1, characterized in that, The visual camera submodule is equipped with an improved YOLOv8 algorithm, which improves the accuracy of marker point recognition by adding a component feature extraction layer.

3. The prefabricated trolley control system according to claim 2, characterized in that, The improved A The path planning algorithm incorporates trolley size constraints and hoisting inertia coefficients to dynamically avoid static obstacles and dynamic interference within the work area; Improved A The path planning algorithm adds two penalty terms to ensure that the path simultaneously satisfies "trolley passability" and "smooth hoisting," as shown in the formula: f(n)=g(n)+h(n)+C size (n)+C inertia (n); Where f(n) represents the total cost function of the node; n represents the current search node; g(n) represents the actual cost function; h(n) represents the heuristic function; C size (n) represents the penalty term for the trolley size constraint; C inertia (n) represents the hoisting inertia penalty term; G represents the path end node; Improved A The path planning algorithm calculates the cumulative path cost using the following formula: g(n)=g(n prev )+dist(n prev ,n); Where n prev Indicates the previous search node; dist(n prev (n) represents the distance between nodes, i.e., the distance between the previous node n. prev The Euclidean distance to the current node n; Improved A The path planning algorithm calculates the estimated remaining cost using the following formula: ； Where (x) n y n (x) represents the current node's raster coordinates; G y G ) represents the raster coordinates of the endpoint G; Improved A The path planning algorithm calculates the penalty term C for the trolley size constraint. size (n), the formula is: ； Improved A Path planning algorithm runtime space capacity determination logic: Taking node n as the center of the trolley, determine: [x n -W car / 2,x n +W car / 2]×[y n -L car / 2,y n +L car If the rectangular area of ​​ / 2] contains "obstacle grid (1 or 2)", it is determined as "unacceptable", and a maximum penalty (106) is added to make the algorithm abandon the node; where W car Indicates the width of the trolley; L car Indicates the length of the trolley; Improved A Path planning algorithm calculates hoisting inertia penalty term C inertia (n), the formula is: C inertia (n)=k inertia ×θ(n)²; Where K inertia θ(n) represents the hoisting inertia coefficient; θ(n) represents the path turning angle: when the path is a straight line, θ(n) = 0, and the penalty term is 0; when the path turns, θ(n) increases, and the penalty term increases accordingly. Improved A Path planning algorithm execution constraint logic: θ(n)≤arcsin(W) car / R min (The maximum turning angle corresponding to the minimum turning radius of the trolley), if exceeded, the penalty term increases sharply (multiplied by 10), forcing the algorithm to choose a gentler turning path; where R min Indicates the minimum turning radius of the trolley; When the initial path P init When a broken line exists (with excessively large angles), the improved A... The path planning algorithm performs a second-order smoothing optimization, as shown in the formula: ； Where P final (i) represents the coordinates of the i-th node after smoothing; α represents the smoothing coefficient, which balances path accuracy and smoothness; (1-α) represents the smoothing compensation coefficient; The edge computing terminal submodule will finalize the path P final The instruction sequence is broken down into "speed-position-time" and sent to the PLC submodule. The PLC submodule controls the execution drive module to drive the trolley to travel along the final path.

4. The prefabricated trolley control system according to claim 3, characterized in that, The improved fuzzy PID algorithm dynamically adjusts the proportional, integral, and derivative parameters based on the attitude deviation and the rate of change of deviation.

5. The prefabricated trolley control system according to claim 4, characterized in that, It also includes a safety warning module, which has a three-level warning mechanism, corresponding to audible and visual prompts, speed reduction operation, and emergency shutdown, with an emergency shutdown response time of ≤200ms.

6. The prefabricated trolley control system according to claim 5, characterized in that, It also includes an electromagnetic interference suppression module, which integrates a shielding housing submodule and a signal filtering circuit submodule. The shielding housing submodule controls the activation of the copper-aluminum composite anti-interference shielding housing device, and the signal filtering circuit submodule suppresses industrial electromagnetic radiation interference through an LC filter network.

7. The prefabricated trolley control system according to claim 6, characterized in that, The environmental sensing module also includes a temperature sensor submodule, which collects the surface temperature of the component in real time. The edge computing terminal submodule is equipped with a component temperature deformation model unit. The component temperature deformation model unit combines the component surface temperature to correct the component assembly coordinates and compensate for the component size deviation caused by temperature changes. The calculation formula for the component temperature deformation model element is as follows: △L = α·L0·△T; X comp =X ref +△L; Where ΔL represents the temperature deformation increment of the component along the assembly direction, with positive values ​​indicating thermal expansion and negative values ​​indicating thermal contraction; α represents the linear expansion coefficient of the component material; L0 represents the original length of the component along the assembly direction; and ΔT represents the temperature change of the component, ΔT = T comp (t)-T0,T comp (t) represents the component surface temperature collected in real time by the temperature sensor; T0 represents the component calibration temperature; X comp This represents the corrected assembly target coordinates, which are the final control commands issued to the execution drive module; X ref Represents the original assembly reference coordinates, derived from improved A. The target location coordinates output by the path planning algorithm; The component temperature deformation model element will correct the assembly target coordinates X. comp It is converted into control commands and sent to the execution driver module.

8. The prefabricated trolley control system according to claim 7, characterized in that, The intelligent control module also includes a collaborative control submodule, which establishes communication with multiple trolleys via a 5G industrial module to achieve data sharing among the trolleys' operations. This submodule allocates the lifting load and receives the motion path P based on a distributed collaborative algorithm. final (i); The distributed collaborative algorithm dynamically allocates the lifting load based on the load capacity of each vehicle, ensuring that the load on a single vehicle does not exceed the rated value. The formula is as follows: ； ； s.t.F i ≤F rated,i (i=1,2,...,n); Where F i This represents the distributed load of the i-th trolley; k i F is the load distribution factor for the i-th trolley; total For the total hoisting load of ultra-large components; F rated,i F is the rated lifting load of the i-th trolley; current,i This represents the actual current load of the i-th trolley. is the total remaining load capacity of all trolleys; n is the number of trolleys working in collaboration; j is the number of summation indices; st is the constraint symbol; i is the trolley index variable; The collaborative control submodule allocates the load of the i-th trolley to the i-th trolley and converts it into control commands, which are then sent to the execution drive modules of each trolley.

9. The prefabricated trolley control system according to claim 8, characterized in that, The intelligent control module is also equipped with a remote control submodule, which connects to the remote control center through an encrypted communication protocol, supports AR visual interaction, and allows the remote control center to obtain trolley posture and working environment data in real time. In the event of a local system failure, the remote control center can take over control and implement unmanned emergency operations.

10. The prefabricated trolley control system according to claim 9, characterized in that, The execution drive module also includes a health diagnosis submodule, which collects operating data of key components of the actuator, analyzes the wear status of the components in combination with an improved BP neural network algorithm, predicts failure risks in advance, and generates maintenance prompts. The improved BP neural network algorithm is optimized through "momentum term + L2 regularization" to adapt to vibration data in industrial scenarios. The improved weight update formula (adding a momentum term to accelerate convergence) is as follows: w ij (k+1)=w ij (k)+η·δ j ·o i +α·[w ij (k)-w ij (k-1)]; b j (k+1)=b j (k)+η·δ j +a·[b j (k)-b j (k-1)]; Where w ij (k+1) represents the weight in the (k+1)th iteration; w ij (k) represents the weight of the k-th iteration; k represents the iteration number index; η represents the learning rate (step size factor); δ j δ represents the error term of the j-th neuron; j Its function is to guide the direction of weight updates; i Represents the output value of the i-th neuron; α represents the momentum factor; W ij (k)-w ij (k-1) represents the weight increment between the k-th and (k-1)-th iterations; b j (k+1) represents the bias term in the (k+1)th iteration; b j (k) represents the bias term in the k-th iteration; i represents the index of the previous neuron; j represents the index of the next neuron. The formula for the error function (with L2 regularization to suppress overfitting) is: ； Where E represents the total error function (loss value); m represents the number of training samples; 1 / m is used to normalize the error; Represents the training sample summation operator; y p This represents the true label of the p-th sample; This represents the model prediction value for the p-th sample; Let represent the squared error of the p-th sample; λ represents the L2 regularization coefficient. w represents the L2 regularization term; ij Represents the connection weights of the neural network; i and j are neuron index variables, where i is the index of the neuron in the previous layer and j is the index of the neuron in the next layer; The wear state mapping formula (from output layer to actual wear amount) is: ； Wear represents the actual wear amount of the component in the final output; This represents the predicted value of the neural network output layer; k and c represent the calibration coefficients. The health diagnosis submodule generates maintenance tips based on the actual wear of the components.

Images