Hanging basket transportation safety management method and system

Abstract

The invention discloses a hanging basket transportation safety management method and system, and relates to the field of damage detection. Collecting weight data of the hanging basket; establishing a steel wire rope health state prediction model by using a multi-layer perceptron (MLP) algorithm, predicting the damage degree of the steel wire rope, and obtaining a health state prediction result of the steel wire rope; according to the health state prediction result of the steel wire rope, a bidirectional long-short-term memory neural network BiLSTM is used for prediction, and a damage trend prediction result of the steel wire rope in a future period of time is obtained; comparing the collected weight data of the hanging basket with a preset safety weight range to obtain a judgment result of whether the weight of the hanging basket exceeds the limit or not; according to the health state prediction result of the steel wire rope, the damage trend prediction result and the judgment result of whether the weight of the hanging basket exceeds the limit or not, a decision tree algorithm is adopted to generate a hanging basket operation control scheme; aiming at low safety of a hanging basket in the prior art, the safety of the hanging basket is improved.

Description

Technical Field

[0001] The present application relates to the field of damage detection, and in particular to a hanging basket transportation safety management method and system. Background Art

[0002] In recent years, with the continuous increase in modern construction projects, complex structures such as high-rise buildings and long-span bridges have become increasingly more numerous. As an indispensable means of transportation for aerial operations, hanging baskets have been widely used in the fields of construction, bridge maintenance, and exterior wall cleaning. The hanging basket transportation operation has the characteristics of complex working environment, high safety risks, and high danger. The traditional manual experience-based management model is difficult to meet the increasingly stringent safety production requirements. How to ensure the safety and reliability of the hanging basket transportation process has become one of the key bottlenecks restricting the healthy development of the construction industry.

[0003] The safety risks during the operation of the hanging basket mainly come from two aspects: first, the steel wire rope used in the hanging basket is in a high stress state for a long time, with uneven force and severe wear, resulting in insufficient strength of the steel wire rope, which is prone to serious accidents such as breakage and failure; second, the dynamic load changes of the hanging basket's own structure are coupled with the wind-induced vibration of the main structure of the building. In extreme conditions such as bad weather, the probability of the hanging basket becoming unstable and overturning is greatly increased. If these risk factors are not monitored, warned and actively controlled in a timely and effective manner, it is very easy to cause safety accidents with casualties and huge economic losses.

[0004] At present, domestic and foreign scholars have conducted some research on the safety of hanging basket operation. The traditional method mainly uses strain gauges, accelerometers and other sensors to detect the tension, vibration and other physical quantities of the wire rope, and manually sets thresholds to determine whether the wire rope has a risk of damage. However, this single physical quantity monitoring method is difficult to fully reflect the health status of the wire rope, and lacks effective early warning capabilities for slow-changing damage such as wire rope fatigue, wear, and corrosion. Later, some scholars proposed to use signal processing methods to fuse multi-sensor information, extract the time-frequency characteristics of wire rope faults through algorithms such as wavelet analysis and Fourier transform, and combine machine learning models such as support vector machines and artificial neural networks to evaluate the degree of damage to the wire rope. However, these data-driven diagnostic methods still remain at the static diagnosis of the current moment, lacking predictive analysis of the future health status of the wire rope. In addition, for the control of the hanging basket power system, traditional methods mostly use classical control theories such as PID control and LQR control, and set fixed controller parameters in advance. This type of control scheme relies on an accurate hanging basket dynamics model, which is difficult to adapt to the real-time changes of parameters such as hanging basket mass and wind load, and the effect in actual engineering applications is not ideal. Summary of the invention

[0005] In response to the problem of low safety of hanging baskets in the prior art, the present application provides a method for managing the transportation safety of a hanging basket. By evaluating the current status of damage to the hanging basket wire rope and predicting future damage, a decision tree model is constructed to generate a hanging basket operation control strategy, and an online learning mechanism is introduced to dynamically optimize the control strategy, thereby improving the safety of the hanging basket.

[0006] The purpose of this application is achieved through the following technical solutions.

[0007] The present application provides a method for managing the safety of suspended basket transportation, comprising: S1, installing a flaw detection sensor on the steel wire rope of the suspended basket to collect the operation data of the steel wire rope; S2, collecting the weight data of the suspended basket through a three-pulley torque sensor; S3, establishing a steel wire rope health status prediction model based on the collected steel wire rope operation data using a multi-layer perceptron MLP algorithm to predict the degree of damage to the steel wire rope and obtain a prediction result of the health status of the steel wire rope; wherein the damage includes fatigue, rust, wear and wire breakage; S4, using a bidirectional long short-term memory neural network BiLSTM to perform prediction based on the prediction result of the health status of the steel wire rope to obtain a prediction result of the damage trend of the steel wire rope in the future; S5, comparing the collected suspended basket weight data with a preset safety weight range to obtain a judgment result of whether the suspended basket weight exceeds the limit; S6, using a decision tree algorithm to generate a control scheme for the operation of the suspended basket based on the prediction result of the health status of the steel wire rope, the prediction result of the damage trend, and the judgment result of whether the suspended basket weight exceeds the limit.

[0008] Among them, the flaw detection sensor is a non-destructive testing device used to detect internal defects and damage of the wire rope. Through certain physical principles, such as electromagnetic induction, ultrasound, etc., it can realize real-time monitoring of the health status of the wire rope. Common flaw detection sensors include electromagnetic flaw detection sensors and ultrasonic flaw detection sensors. Electromagnetic flaw detection sensors use the principle of electromagnetic induction to detect defects such as broken strands and corrosion of the wire rope by measuring the electromagnetic response of the wire rope under an alternating magnetic field; ultrasonic flaw detection sensors use the propagation and reflection characteristics of ultrasound in the wire rope to detect broken wires, wear and other damages of the wire rope by measuring the changes in ultrasonic signals. This application can realize real-time monitoring and early warning of the health status of the wire rope by installing flaw detection sensors on the wire rope of the hanging basket, timely discover potential risks and safety hazards of the wire rope, provide reliable data support for the safe operation of the hanging basket, and avoid accidents caused by wire rope failure.

[0009] The three-pulley torque sensor is a torque measuring device for measuring the weight of a hanging basket. By measuring the stress of the three pulleys in the lifting mechanism of the hanging basket, the weight of the hanging basket is indirectly calculated. The lifting mechanism of the hanging basket usually adopts the form of a three-pulley group, including a movable pulley and two fixed pulleys. The three-pulley torque sensor is respectively installed on the shafts of the three pulleys to measure the torque magnitude of each pulley. According to the size of the pulley and the arrangement of the rope, the lifting force acting on the movable pulley can be calculated through the torque balance equation, and then the weight of the hanging basket can be obtained. The present application can realize real-time monitoring of the weight of the hanging basket and grasp the load state of the hanging basket by installing a three-pulley torque sensor on the lifting mechanism of the hanging basket. Combined with the rated load of the hanging basket, it can be judged whether the hanging basket is overloaded, which provides an important basis for the safe operation of the hanging basket. At the same time, the weight data of the hanging basket is also a key parameter for controlling and optimizing the motion of the hanging basket, providing data support for realizing intelligent scheduling and efficient operation of the hanging basket.

[0010] Furthermore, the health status prediction result of the wire rope is obtained, including: S31, obtaining the historical operation data of the wire rope and constructing a training data set; the training data set includes the wire rope operation data and the corresponding damage type and damage degree labels; S32, constructing a wire rope health status prediction model based on a multi-layer perceptron MLP neural network, the model includes an input layer, multiple hidden layers and an output layer; wherein the hidden layer adopts a ReLU activation function, and the output layer adopts a Softmax function for multi-classification; S33, according to the training data set, the back propagation algorithm and the mini-batch gradient descent method are used to train the wire rope health status prediction model; S34, using the trained wire rope health status prediction model to predict the collected wire rope operation data, and selecting the damage type and damage degree combination with the largest prediction probability as the wire rope health status prediction result.

[0011] Furthermore, historical operation data of the wire rope is obtained to construct a training data set, including: collecting multi-sensor data of the hanging basket wire rope in different time periods and different operating conditions; wherein the multi-sensor data includes tension data measured by strain gauges and vibration data measured by accelerometers; preprocessing the tension data and extracting statistical features; the statistical features include mean, variance, peak value and root mean square; using wavelet packet decomposition algorithm to perform time-frequency decomposition of vibration data, and extracting energy features and wavelet coefficients of different frequency bands; using fast Fourier transform algorithm to perform frequency domain analysis on the vibration data after wavelet decomposition, and extracting spectral features and inverse spectrum features; the statistical features of tension data and energy features of vibration data are combined into a single frequency domain analysis. , wavelet coefficients, spectral features and inverse spectrum features are aligned according to timestamps to construct a multimodal fusion feature vector; the sliding window method is used to extract multi-scale data fragments on the multimodal fusion feature vector with a fixed window size and sliding step size, and each data fragment contains continuous multimodal fusion features and corresponding timestamps; according to the timestamp of each data fragment, the damage type and damage degree label of the corresponding fragment is obtained from the historical maintenance records and fault logs of the hanging basket wire rope, and the mapping relationship between multimodal fusion features and damage labels is established; the multi-scale data fragments are combined with the corresponding damage labels to construct a training data set, and each training sample contains a multimodal fusion feature fragment and a damage label.

[0012] Furthermore, the damage trend prediction result of the wire rope in the future is obtained, including: S41, according to the health status prediction result of the wire rope, the damage type and damage degree feature vector of the wire rope at the current moment are extracted as the input of the BiLSTM model; S42, the BiLSTM damage trend prediction model is constructed, and the model includes an input layer, a bidirectional LSTM layer, an attention layer and an output layer; S43, from the historical operation data of the wire rope obtained in step S3 and the corresponding historical damage label data, the loss feature sequence at each historical moment and the real damage label in the future period are extracted to construct a training data set for the BiLSTM model; S44, the cross entropy loss function and the Adam optimization algorithm are used to train the BiLSTM model using the training data set; S45, the extracted damage type and damage degree feature vector of the wire rope at the current moment are predicted using the trained BiLSTM model, and the damage type and damage degree prediction value of the wire rope in the future period are obtained as the loss trend prediction result of the wire rope.

[0013] Furthermore, a BiLSTM damage trend prediction model is constructed, including: converting the extracted damage type and damage degree feature vectors of the wire rope at the current moment into a fixed-length one-hot encoding form as the input feature of the BiLSTM damage trend prediction model; mapping the input features to a low-dimensional vector space through an embedding layer to obtain the distribution of the damage type and damage degree vectors; concatenating the damage type and damage degree vectors output by the embedding layer according to the feature dimension to obtain the damage feature vector at the current moment; using two parallel forward and reverse LSTMs to perform forward and reverse transmission on the damage feature vector respectively, to obtain the forward and backward long-distance temporal dependencies in the development process of the wire rope damage respectively; at each time step of the forward LSTM network, concatenating the forward hidden state of the current time step with the damage feature vector at the current moment according to the feature dimension as the input of the next time step; at each time step of the reverse LSTM network, concatenating the reverse hidden state of the current time step with the damage feature vector at the current moment according to the feature dimension as the input of the previous time step. The input of one time step; at each time step, the forward hidden state output by the forward LSTM network and the reverse hidden state output by the reverse LSTM network are spliced ​​according to the feature dimension to obtain the current moment bidirectional LSTM hidden state vector that integrates the bidirectional time series information; using the Luong attention mechanism, the attention weights of the current moment and each moment are calculated by the similarity between the current moment bidirectional LSTM hidden state vector and the bidirectional LSTM hidden state vector at each moment, and the bidirectional LSTM hidden state vectors at each moment are weighted summed based on the attention weights to obtain the current moment attention feature vector; the current moment attention feature vector and the current moment bidirectional LSTM hidden state vector are spliced ​​according to the feature dimension to obtain the current moment damage prediction feature vector; the current moment damage prediction feature vector is mapped to the predicted probability distribution of the damage degree of each damage type of the wire rope in the future period of time through the fully connected layer and the Softmax function, and the damage type with the largest probability value and the corresponding damage degree are used as the damage trend prediction result at the current moment.

[0014] Furthermore, a decision tree algorithm is used to generate a control scheme for the operation of the hanging basket, including: S61, constructing an input feature vector based on the prediction results of the health status of the wire rope at the current moment, the prediction results of the damage trend of the wire rope in the future, and the judgment result of whether the weight of the hanging basket exceeds the limit; S62, obtaining the control scheme and control effect of the hanging basket under different working conditions based on the historical operation data of the hanging basket, and performing a multi-index quantitative evaluation on the control effect to obtain the optimal control scheme under each working condition and form a control knowledge base; S63, matching the constructed input feature vector with the control knowledge base, and constructing a training data set for the decision tree algorithm based on the matching results, wherein each training sample in the training data set contains the input feature vector and the corresponding optimal control scheme; S64, using the CART decision tree algorithm, based on the control knowledge base, Taking minimization of the Ni index as the partitioning criterion, binary tree modeling is performed on the training data set constructed in the training step S63 to generate an initial decision tree; S65, the initial decision trees are combined through an integrated learning strategy to obtain a suspended basket operation control decision model; S66, a gradient boosting decision tree algorithm is adopted to generate an online learning model of suspended basket operation control according to the training data set constructed in step S63; S67, the input feature vector constructed in S61 is input into the suspended basket operation control decision model obtained in step S65 and the online learning model obtained in step S66, respectively, to obtain a control scheme based on historical experience and a control scheme based on online learning, respectively; S68, according to the control scheme based on historical experience and the control scheme based on online learning, a suspended basket operation control decision scheme is obtained through weight fusion.

[0015] Furthermore, an online learning model for the operation control of the hanging basket is generated, including: using the training data set constructed in step S63, according to the preset number of decision trees and sampling ratio, using the Boosting strategy to generate a decision tree sequence; setting the maximum depth of each decision tree in the decision tree sequence, the minimum number of leaf node samples, and the maximum number of partition features; using the grid search method to traverse different combinations of hyperparameters; using the cross-validation method to divide the training data set constructed in step S63 into K mutually exclusive subsets, each time selecting K-1 subsets as training sets and the remaining 1 subset as a validation set; for each group Hyperparameter combination, using K different training sets and validation sets, respectively train and evaluate the decision tree sequence to obtain K groups of model evaluation results; among which, the model evaluation results include mean absolute error MAE, root mean square error RMSE and determination coefficient R2; the K groups of model evaluation results under each group of hyperparameter combinations are fused using the weighted average soft voting method to obtain the comprehensive performance evaluation results of the decision number sequence under the current hyperparameter combination; all hyperparameter combinations are traversed, and according to the comprehensive performance evaluation results, the optimal hyperparameter combination and the corresponding decision tree sequence are selected as the online learning model for hanging basket operation control.

[0016] Furthermore, a suspension basket operation control decision plan is obtained, including: respectively obtaining a control plan based on historical experience and a control plan based on online learning; performing linear weighted fusion of the control plan based on historical experience and the control plan based on online learning according to a preset fusion weight to obtain a suspension basket operation control decision plan; obtaining current state parameters during the operation of the suspension basket, and calculating the rate of change of the suspension basket operation condition according to the difference between the current state parameters and the historical state parameters; wherein the state parameters include wire rope tension, suspension basket displacement and suspension basket speed; comparing the calculated rate of change of the suspension basket operation condition with a preset threshold range; if the rate of change of the suspension basket operation condition exceeds the preset threshold range, adjusting the fusion weight, otherwise maintaining the fusion weight unchanged;

[0017] Furthermore, if the rate of change of the operating condition of the suspended basket exceeds the preset threshold range, the fusion weight is adjusted, including: setting the weight adjustment coefficient α of the historical experience control scheme and the weight adjustment coefficient β of the online learning control scheme, both of which have initial values ​​of 1; obtaining the calculated rate of change of the operating condition of the suspended basket, and calculating the change value of the weight adjustment coefficient according to the degree of deviation of the rate of change from the upper and lower limits of the preset threshold range. and ; The degree of deviation is expressed by the absolute value of the difference between the rate of change and the upper and lower limits of the threshold range; the weight adjustment coefficient is updated, and the update formula is: , ,in, and are the weight adjustment coefficients of the updated historical experience control scheme and the online learning control scheme, and is a preset update rate parameter used to control the speed of change of the weight adjustment coefficient; and Perform normalization: , , using the normalized weight adjustment coefficient and , calculate the weights of the historical experience control scheme and the online learning control scheme and : , ,in, and They are the weights of the historical experience control scheme and the online learning control scheme before fusion weight adjustment.

[0018] Another aspect of the present application also provides a hanging basket transportation safety management system for executing a hanging basket transportation safety management method of the present application.

[0019] Compared with the prior art, the advantages of this application are:

[0020] According to the historical operation data of the suspended basket, the control scheme and control effect under different working conditions are quantitatively evaluated, and the optimal control parameter combination under each working condition is obtained in a multi-objective weighted manner. The prior scheduling experience can be fully utilized to form a global scheduling knowledge base, and the complex and changeable suspended basket working conditions are mapped with the optimal control strategy. Through knowledge-driven, blind attempts are avoided every time a new working condition is encountered, providing explainable and reliable strategy guidance for real-time scheduling. The CART classification regression tree is used to construct the suspended basket state and the optimal control scheme into a nonlinear decision tree, and the state space is recursively divided based on the Gini index as the basis for tree node splitting. The complex multi-level logical relationship between the state and the control strategy can be adaptively fitted. Compared with the linear model and the rule table method, it is more suitable for dealing with the combinatorial explosion problem of high-dimensional states. The model has a natural feature combination ability, can automatically mine the cross terms between state factors, and improve the refinement of the strategy. By integrating multiple complementary decision trees, the tree model is combined and optimized using Bootstrap sampling and majority voting principles, which further reduces the variance of a single decision tree and enhances the generalization performance and anti-interference ability of the control strategy for unknown working conditions.

[0021] On the basis of the decision tree integration model, the online learning paradigm is further used to introduce real-time feedback information to dynamically correct the control strategy. The gradient boosted decision tree (GBDT) is used to iteratively train a series of weak learners in the form of residual fitting, and the gradient descent method is used to optimize the forward step-by-step algorithm to approach the global optimal strategy in the direction of the steepest descent. Compared with the offline batch trained decision tree, GBDT can efficiently process massive online data streams, strengthen the strategy defects through error-driven incremental learning, and quickly adapt to new dynamically changing working conditions. It not only improves the real-time performance of the strategy, but also takes into account the convergence stability of the strategy. The cross-validation and grid search mechanisms are introduced, and the hyperparameters of GBDT such as the number of trees, maximum depth, minimum number of leaf node samples, and number of partition features are automatically optimized with reference to multivariate evaluation indicators such as mean absolute error, root mean square error, and determination coefficient. The accuracy and robustness of the online strategy are further improved through parameter integration.

[0022] Aiming at the dynamic changes of the operating conditions of the suspended basket, an adaptive weight adjustment mechanism based on the state parameter change rate is designed. The stability of the suspended basket state is used as the fusion coefficient to realize the dynamic weighting of the global scheduling strategy based on historical experience and the local fine-tuning strategy based on online feedback. When the operating conditions of the suspended basket fluctuate violently, the weight of the online learning branch is adaptively increased according to the degree of deviation between the state parameter change rate and the preset threshold, so that the control decision is more dependent on the real-time working conditions, thereby achieving rapid response and precise optimization of specific working conditions; when the state of the suspended basket is stable, the weight adjustment mechanism automatically restores the dominant position of the global strategy branch, and smoothly transitions the online learning scheme to the overall scheduling framework, on the one hand, making full use of the guiding role of prior experience, and on the other hand, avoiding the random fluctuations of local feedback data affecting the stability of the long-term strategy. Finally, by weighted fusion of the policy outputs of the two branches, the online optimization and global scheduling are balanced in the dynamic trade-off, overcoming the shortcomings of a single model in a complex environment, and forming a suspended basket control strategy that takes into account both generalization and adaptability. At the same time, the output effect of the fusion strategy is used as new sample feedback to continuously optimize the decision-making model and online learning model, further enhancing the closed-loop and self-learning ability of the hanging basket operation control. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present application will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0024] Figure 1 is an exemplary flow chart of a method for managing the safety of suspended basket transportation according to some embodiments of the present application;

[0025] Figure 2 is a schematic diagram of a three-pulley torque sensor according to some embodiments of the present application;

[0026] Figure 3 is an exemplary flow chart of generating a prediction result of a wire rope health status according to some embodiments of the present application;

[0027] Figure 4 is an exemplary flow chart of constructing a training data set for a wire rope health status prediction model according to some embodiments of the present application;

[0028] Figure 5 It is an exemplary flow chart for generating a hanging basket operation control decision plan according to some embodiments of the present application. DETAILED DESCRIPTION

[0029] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, a flaw detection sensor is installed on the wire rope of the hanging basket to collect the running data of the wire rope; the weight data of the hanging basket is collected through a three-pulley torque sensor; according to the collected running data of the wire rope, a wire rope health status prediction model is established by using a multi-layer perceptron MLP algorithm to predict the damage degree of the wire rope and obtain the prediction result of the health status of the wire rope; wherein the damage includes fatigue, rust, wear and wire breakage; according to the prediction result of the health status of the wire rope, a bidirectional long short-term memory neural network BiLSTM is used to predict and obtain the prediction result of the damage trend of the wire rope in the future; the collected hanging basket weight data is compared with the preset safety weight range to obtain the judgment result of whether the hanging basket weight exceeds the limit; according to the prediction result of the health status of the wire rope, the prediction result of the damage trend, and the judgment result of whether the hanging basket weight exceeds the limit, a decision tree algorithm is used to generate a control scheme for the hanging basket operation.

[0031] Specifically, S1, install flaw detection sensors on the wire rope of the hanging basket to collect the running data of the wire rope; Specifically, sensor selection and arrangement: According to the diameter, length, material and other parameters of the wire rope of the hanging basket, select suitable non-destructive flaw detection sensors such as magnetostrictive, eddy current, and electromagnetic ultrasonic. Comprehensively consider the detection sensitivity, resolution, dynamic range and other indicators to select the sensor model with the best performance. The flaw detection sensors are arranged at equal intervals along the axial direction of the wire rope, and the sensor spacing is determined according to the signal attenuation characteristics of the wire rope and the required detection accuracy. At the same time, the sensor array is arranged at different cross sections of the wire rope to scan the entire cross section of the wire rope. The voltage signal output by the flaw detection sensor is collected in real time using a high-speed data acquisition card, and the analog signal is converted into a digital signal. According to the Nyquist sampling theorem, the sampling frequency should be at least twice the highest frequency of the wire rope damage signal. The collected original signal is denoised, and a bandpass filter is used to filter out low-frequency trend items and high-frequency noise to improve the signal-to-noise ratio. In order to eliminate the gain difference between sensors, the amplitude of each channel signal is normalized. Since the sampling start time of each sensor may be inconsistent, the signals of different channels need to be synchronized. Taking the speed pulse signal of the wire rope as the reference, the time delay of each channel signal is determined by cross-correlation analysis, and time shift compensation is performed to align the signals of each channel on the time axis. In addition, the spatial position of the wire rope needs to be aligned. The running displacement of the wire rope is recorded by the encoder, and the signals collected by each sensor are mapped to the spatial coordinate system of the wire rope, and the corresponding relationship between the signal amplitude and the spatial position of the wire rope is established.

[0032] like Figure 2As shown, S2 collects the weight data of the hanging basket through a three-pulley torque sensor; a spoke-type torque sensor suitable for the pulley shaft diameter and load range of the hanging basket is selected, and the sensor range should cover the empty weight and full load weight of the hanging basket. The torque sensor is connected to the pulley shaft through a coupling, and the output end of the sensor is connected to the data acquisition module. The power supply voltage and grounding of the sensor must meet the explosion-proof and waterproof requirements. A torque sensor is installed on each of the three pulleys of the hanging basket, and the installation position of the sensor should be as close as possible to the pulley bearing to reduce the influence of the pulley's deadweight on the measurement accuracy. The analog voltage signal output by the torque sensor is converted into a digital signal through a data acquisition card. According to the sensitivity and nonlinear error of the torque sensor, the collected raw signal is calibrated and corrected. Since the hanging basket will generate vibration and impact during operation, causing high-frequency noise interference to the torque signal, the collected signal needs to be low-pass filtered. Digital filtering algorithms such as Butterworth filter and Kalman filter can be used to adaptively smooth the torque signal. Since the stress states of the three pulleys may be different, the signal amplitudes and phases of the three torque sensors are inconsistent, and the three signals need to be time synchronized. Taking the operating status signal of the hanging basket (such as the speed signal of the lifting motor) as the benchmark, the time delay of the three torque signals is determined by cross-correlation analysis, and time shift alignment is performed. The synchronized torque signal is divided into time windows, and the length of the time window and the sliding step are determined according to the operating cycle and sampling frequency of the hanging basket. The torque signal in each time window represents a complete hanging basket weight measurement cycle.

[0033] S3, based on the collected wire rope operation data, a wire rope health status prediction model is established using a multi-layer perceptron MLP neural network to predict the damage degree of the wire rope and obtain the wire rope health status prediction result. Figure 3 As shown, S31, the historical operation data of the wire rope is obtained to construct a training data set. In this embodiment, 5 hanging basket wire ropes with the same model and specifications and different usage time are selected as data collection objects, which are marked as R1~R5 respectively. 10 strain gauges and 10 accelerometers are arranged at equal intervals on each wire rope, and the sampling frequency is set to 1kHz. For 3 consecutive months, 2 hours of data are collected every day, covering typical working conditions such as lifting, lowering, translation, and rotation of the hanging basket. The collected data is denoised, and a 4th-order Butterworth low-pass filter is used with a cutoff frequency set to 500Hz to remove high-frequency noise. A method combining the 3σ criterion and the Chauvenet criterion is used to eliminate outliers in the collected data. In the end, 2160 hours of multi-sensor data were collected for each wire rope.

[0034] Extract features. For the tension data collected by each strain gauge, extract four statistical features such as mean, variance, peak value and root mean square. The calculation window length of each feature is 1 minute, and the sliding step length is 10 seconds. The tension data of each wire rope is converted into a 10×4×(2160×60 / 10)=518400-dimensional statistical feature matrix. For the vibration data collected by each accelerometer, use 4-layer wavelet packet decomposition, select db4 wavelet as the mother wavelet, and get 16 frequency bands after decomposition. Extract 2 energy features (band energy mean and band energy standard deviation) and 2 wavelet coefficient features (band wavelet coefficient mean and band wavelet coefficient standard deviation) in each frequency band. The calculation window length of each feature is 1 minute, and the sliding step length is 10 seconds. The vibration data of each wire rope is converted into a 10×16×4×(2160×60 / 10)=2073600-dimensional time-frequency domain feature matrix. For the vibration data of each frequency band, the fast Fourier transform is used to calculate the spectrum and inverse spectrum. Two features are extracted from the spectrum (frequency center of gravity and frequency standard deviation), and two features are extracted from the inverse spectrum (inverse frequency center of gravity and inverse frequency standard deviation). The calculation window length of each feature is 1 minute, and the sliding step length is 10 seconds. The vibration data of each wire rope is converted into a frequency domain feature matrix of 10×16×4×(2160×60 / 10)=2073600 dimensions.

[0035] Data fusion, align the statistical feature matrix of tension data, the time-frequency domain feature matrix of vibration data, and the frequency domain feature matrix according to the timestamp, and splice them into a high-dimensional feature matrix by column. The dimension of this matrix is ​​(10×4+10×16×4×2)×(2160×60 / 10)=4665600×12960. Each column represents the multimodal fusion feature vector of the wire rope at a timestamp, which contains 120 components and fully describes the health status of the wire rope at that moment.

[0036] Data partitioning, on the multimodal fusion feature matrix, a sliding window method is used to extract multi-scale data fragments. By setting three different window sizes (10 minutes, 30 minutes, and 60 minutes) and corresponding sliding steps (1 minute, 2 minutes, and 5 minutes), 3×(2160-10+1)=6453 10-minute fragments, 3×(2160-30+1)=6393 30-minute fragments, and 3×(2160-60+1)=6303 60-minute fragments can be obtained. Each fragment contains a continuous multimodal fusion feature vector and a corresponding timestamp range.

[0037] Obtain labels. For each data segment extracted from the wire rope, query the historical maintenance records and fault logs of the hanging basket according to its timestamp range to obtain whether typical damages such as broken strands, rust, stranding, and flattening of the wire rope have occurred in the corresponding segment, as well as the severity of the damage (such as the number of broken strands, the proportion of rusted area, the proportion of stranding length, and the flattening deformation rate). Based on expert experience, set a quantitative standard for the severity of the damage, and discretize the continuous degree of damage into four levels: normal, mild damage, moderate damage, and severe damage. Finally, obtain the damage type and damage degree labels for each data segment, and construct a mapping relationship between multimodal fusion features and damage labels.

[0038] like Figure 4 As shown in the figure, a training data set is constructed, and the data segments of all wire ropes and their corresponding damage labels are summarized and randomly divided into a training set and a test set in a ratio of 8:2. In the training set, the data segments are paired with the damage labels to construct training samples. Each training sample contains a 120×L multimodal fusion feature segment (L is the segment length) and a damage type + damage degree label. In the test set, the test samples are also constructed, but the damage labels are temporarily hidden to evaluate the generalization performance of the model. Finally, the training set contains 100084 samples and the test set contains 25021 samples. The sample ratio of the four categories of normal, mild damage, moderate damage and severe damage in the data set is about 7:2:1:1, which is consistent with the actual law of wire rope health degradation. The data set is moderate in size and reasonable in category distribution. It integrates multi-source heterogeneous sensor data and expert knowledge, and can effectively support the training and testing of deep learning prediction models.

[0039] S32: Construct a wire rope health status prediction model based on a multi-layer perceptron MLP neural network, where: Input layer: Input the 120×L-dimensional multimodal fusion feature fragment of each training sample into the input layer of the MLP neural network, where L is the fragment length, which takes values ​​of 10, 30 or 60. Therefore, the number of nodes in the input layer is set to 120×L, which is consistent with the dimension of the feature fragment. Hidden layer: Set 3 fully connected hidden layers, containing 256, 128 and 64 neurons respectively. The first hidden layer maps the 120×L-dimensional features of the input layer to 256 dimensions, the second hidden layer maps the 256-dimensional features to 128 dimensions, and the third hidden layer maps the 128-dimensional features to 64 dimensions. Each hidden layer uses the ReLU activation function to perform nonlinear transformation on the output of the previous layer to improve the nonlinear fitting ability of the network. The network structure is a "wide to narrow" pyramid shape. By reducing the dimension layer by layer, high-level semantic information in the feature fragment is extracted while avoiding the risk of overfitting. Output layer: According to the damage type and damage degree labels in the data set, 12 output nodes are set, corresponding to the combination of 3 damage types (broken strands, rusted strands, and detached strands) and 4 damage degrees (normal, mild, moderate, and severe). The output layer uses the Softmax activation function to map the 64-dimensional features of the previous hidden layer into 12 probability values, each of which represents the possibility that the sample belongs to the corresponding combination of damage type and degree. The Softmax function ensures that the sum of the 12 probability values ​​is 1, which conforms to the properties of probability distribution.

[0040] S33: The back propagation algorithm and mini-batch gradient descent method are used to train the wire rope health status prediction model. Specifically, parameter initialization: Initialize the weight matrix and bias vector of each layer of the MLP neural network. The hidden layer weight matrix uses the Xavier uniform initialization method to keep the variance of each layer output unchanged. The initialization formula is: ,in and are the input and output dimensions of the weight matrix, respectively. The hidden layer bias vector is initialized to 0. The output layer weight matrix and bias vector are initialized using a normal distribution with a mean of 0 and a standard deviation of 0.01. Reasonable parameter initialization helps accelerate model convergence and avoid gradient vanishing or exploding problems.

[0041] Mini-batch division: randomly shuffle the training set and divide it into mini-batches of size 128. Each mini-batch contains 128 training samples, and gradients can be calculated in parallel to improve training efficiency. If the number of iterations is set to 100, each epoch needs to traverse 100084 / 128≈784 mini-batches. Loss function calculation: the cross-entropy loss function is used to measure the difference between the probability distribution of damage type and degree predicted by the MLP model and the true label of the sample. For each of the 128 samples in the mini-batch, the cross-entropy loss of the predicted probability distribution and the true label is calculated in parallel, and then the average is taken as the loss function value of the mini-batch. The formula for the cross-entropy loss function is: , where is the one-hot encoding value of the kth class in the true label (one-hot encoding form) of the sample; The probability distribution that the i-th sample belongs to the k-th class predicted by the model.

[0042] Gradient back propagation: Use the back propagation algorithm to calculate the gradient of the loss function with respect to the parameters of each layer of the MLP model. First, calculate the gradient of the loss function with respect to the Softmax function of the output layer, and then back propagate the gradient layer by layer according to the chain rule until the input layer. The weight matrix of the layer and the bias vector , its gradient formula is and ,in For the The output of the layer, For the The error term of the layer is given by ,in is the weighted input of the lth layer, is the activation function, is the Hadamard product.

[0043] Parameter update: Adopt Adam optimization algorithm to update weight matrix and bias vector of each layer of MLP model according to the gradient calculated by back propagation. Adam algorithm comprehensively considers the first-order moment estimation and second-order moment estimation of gradient, adaptively adjusts the learning rate of each parameter, and accelerates convergence. Set the initial learning rate to 0.001, the exponential decay rate of first-order moment estimation to 0.9, and the exponential decay rate of second-order moment estimation to 0.999. , the weight decay coefficient is 0.0005. According to the Adam algorithm formula, update the Layer weight matrix and the bias vector The formula is and ,in and are the correction values ​​of the first-order moment estimate and the second-order moment estimate, respectively. is the learning rate, is the smoothing term.

[0044] Iterative training: Repeat until 100 epochs are completed or the accuracy of the model on the validation set has not improved for 5 consecutive epochs. After each epoch, the cross entropy loss and classification accuracy of the model are evaluated on the training set and validation set to monitor the training progress of the model. If the performance of the model on the validation set is found to have significantly decreased, the training is stopped in time to prevent overfitting. Finally, the model with the highest classification accuracy on the validation set is selected as the optimal model for the subsequent wire rope health status prediction task.

[0045] S34: Use the trained wire rope health status prediction model to predict the collected wire rope operation data. For the newly collected wire rope operation data, extract the statistical features of the tension data, the time-frequency domain features and the frequency domain features of the vibration data according to the method in step S31, and align them according to the timestamp to construct a 120×L-dimensional multimodal fusion feature fragment, where L is the length of the time window to be predicted. Input the extracted feature fragment into the trained MLP neural network model, forward propagate and calculate the 12 node values ​​of the output layer, and convert them into the probability distribution of damage type and damage degree through the Softmax function. Select the combination of damage type and damage degree with the largest probability value as the prediction result of the health status of the wire rope in the current time window. For example, if the probability value of the third node of the output layer (corresponding to "broken strand + mild") is 0.8, which is higher than other nodes, it is predicted that the current wire rope has a mild broken strand damage. Compare the predicted damage type and damage degree with the preset warning threshold. For example, if the warning threshold of "broken strands + mild" is set to 0.5, then when the predicted probability of this damage combination exceeds 0.5, it is considered that the wire rope is at risk of broken strands and needs to be taken seriously. According to the prediction results and risk levels, corresponding maintenance decision suggestions are given, such as "it is recommended to focus on checking the broken strands of the wire rope during the next routine maintenance", "it is recommended to increase the frequency of wire rope inspections and pay close attention to the development of strand damage", etc.

[0046] S4, based on the prediction results of the health status of the wire rope, the bidirectional long short-term memory neural network BiLSTM is used to predict the damage trend prediction results of the wire rope in the future, including: S41: Based on the prediction results of the health status of the wire rope, the damage type and damage level feature vector of the wire rope at the current moment are extracted as the input of the BiLSTM model: For the prediction results of the health status of the wire rope at each moment, the damage type and damage level information are extracted. The damage types include broken strands, rusted strands and broken strands, which are represented by 1, 2, and 3 respectively; the damage level includes normal, mild, moderate and severe levels, which are represented by 0, 1, 2, and 3 respectively. The extracted damage type and damage level information are combined into a binary group (damage_type, damage_level) to represent the damage status of the wire rope at the current moment. For example, (2, 1) means that the wire rope has mild rust at the current moment. Collect the damage state tuples of the wire rope at N consecutive moments to form a damage state sequence [(damage_type_1, damage_level_1), (damage_type_2, damage_level_2), ..., (damage_type_N, damage_level_N)] as the input feature of the BiLSTM model. The value of N is determined according to the time step setting of the BiLSTM model, usually 10~50.

[0047] S42: Construct a BiLSTM damage trend prediction model, which includes an input layer, a bidirectional LSTM layer, an attention layer, and an output layer. Specifically, the input layer: convert each binary group (damage_type, damage_level) in the damage state sequence into a 7-dimensional damage state feature vector. For damage_type, convert it into a 3-dimensional one-hot vector [1, 0, 0], [0, 1, 0], or [0, 0, 1], corresponding to the three types of damage: broken strands, rusted strands, and raised strands. For damage_level, convert it into a 4-dimensional one-hot vector [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], or [0, 0, 0, 1], corresponding to the four levels of damage: normal, mild, moderate, and severe. The 3D one-hot vector of damage type and the 4D one-hot vector of damage degree are concatenated according to the feature dimension to obtain a 7D damage state feature vector. For example, [0, 1, 0, 0, 0, 1, 0] indicates that corrosion damage has occurred and the damage degree is moderate. The damage state feature vectors of N time steps are arranged in chronological order to form a 7×N 2D tensor as the input of the BiLSTM model.

[0048] Embedding layer: Set an embedding matrix of size 7×D, where D is the dimension of the embedding vector, usually 32 to 128. Multiply the 7×N damage state feature tensor of the input layer by the embedding matrix to obtain a D×N embedding vector sequence. The embedding layer can learn the potential association between damage type and damage degree, converting high-dimensional sparse one-hot encoding into low-dimensional dense distributed representation, which helps improve the generalization ability of the model.

[0049] Bidirectional LSTM layer: Set up a forward LSTM network and a reverse LSTM network. The hidden layer dimensions of both networks are H, which is usually 64 to 256. Input the D×N embedding vector sequence output by the embedding layer into the forward LSTM network and the reverse LSTM network at the same time. In the forward LSTM network, the hidden layer dimensions are H, which is usually 64 to 256. For each time step t, the D-dimensional embedding vector at time t is processed in the order of and the H-dimensional forward hidden state at time t-1 As the input of the LSTM unit, the forward hidden state at time t is calculated through the input gate, forget gate, output gate and memory unit and memory cell status In the reverse LSTM network, according to the time step For each time step t, the D-dimensional embedding vector at time t is processed in the order of and the H-dimensional reverse hidden state at time t+1 As the input of the LSTM unit, through the input gate, forget gate, output gate and memory unit, calculate the reverse hidden state at time t and memory cell status At each time step t, the H-dimensional forward hidden state output by the forward LSTM network is and the H-dimensional reverse hidden state output by the reverse LSTM network According to the feature dimension splicing, a 2H-dimensional bidirectional hidden state vector is obtained h t = [ h t f , h t b ] , as the feature representation of the fusion of bidirectional temporal information at time t.

[0050] Attention layer: set an H×2H attention weight matrix , which is used to calculate the similarity between the bidirectional hidden state vector at the current moment and the bidirectional hidden state vector at each moment. For the current moment t, calculate its bidirectional hidden state vector And the bidirectional hidden state vector at each moment The attention weight . Luong attention mechanism is used, and the similarity is calculated using a bilinear function, that is, . For attention weight Normalize and get the normalized attention weight Based on normalized attention weights Bidirectional hidden state vector at each moment Perform weighted summation to obtain the attention feature vector at the current time t . The attention feature vector at the current time t is With the bidirectional hidden state vector According to the feature dimension splicing, a 3H-dimensional damage prediction feature vector is obtained p t = [ h t , c t ] , as a high-level feature representation that comprehensively considers bidirectional temporal dependencies and attention mechanisms at time t.

[0051] Output layer: Set a 3H×12M fully connected weight matrix and a 12M-dimensional bias vector , where M is the number of time steps to be predicted in the future. For the damage prediction feature vector at the current time t , mapped to a 12M-dimensional prediction vector through a fully connected layer . The prediction vector According to the time step dimension, it is divided into M 12-dimensional sub-vectors, and each sub-vector corresponds to the predicted value of the damage state in the future time step. The Softmax function is applied to each 12-dimensional sub-vector to map it into a joint probability distribution of damage type and damage degree. The first 3 elements represent the probability of the damage type (broken strands, rusted strands, and raised strands), and the last 9 elements represent the probability of the damage degree under each damage type (normal, mild, moderate, and severe). For each future time step, the damage type with the largest probability value and the corresponding damage degree are selected as the damage trend prediction result of the time step. The BiLSTM damage trend prediction model constructed in this embodiment can make full use of the bidirectional temporal dependency in the historical damage state sequence of the wire rope, and adaptively pay attention to the damage characteristics at different times through the attention mechanism, thereby improving the accuracy of the damage trend prediction. At the same time, through the end-to-end modeling method, the prediction tasks of the damage type and the damage degree are jointly optimized, and the intrinsic relationship between the two can be captured to obtain a more reliable and complete damage trend prediction result.

[0052] S43: From the historical operation data of the wire rope obtained in step S3 and the corresponding historical damage label data, extract the damage feature sequence at each historical moment and the real damage label in the future period, and construct a training data set for the BiLSTM model. Specifically, preprocess the historical operation data of the wire rope obtained in step S3, and group and synchronize the data according to the sensor type and sampling frequency. For the strain gauge data, resample at a frequency of 100 sampling points per minute to obtain the strain time series; for the accelerometer data, resample at a frequency of 1000 sampling points per minute to obtain the acceleration time series. Time-align the strain time series and the acceleration time series to ensure that the strain data and acceleration data at each moment have the same timestamp. Preprocess the historical damage label data, convert the damage records in text format into structured timestamp-damage type-damage degree triples, and sort them by timestamp.

[0053] Extract features, use a sliding time window method, with a step size of 1 minute, and extract multimodal fusion feature vectors for each historical moment on the historical operation data. For each historical moment t, select strain data and acceleration data in a 5-minute time window before and after t as the center, totaling 10 minutes of data. For the strain data in the time window, extract four statistical features such as mean, variance, peak-to-peak value and root mean square; for the acceleration data, extract five statistical features such as mean, standard deviation, root mean square, kurtosis and skewness. Concatenate the four statistical features of strain data and the five statistical features of acceleration data according to the feature dimensions to obtain a 9-dimensional multimodal fusion feature vector, which represents the vibration and tension characteristics of the wire rope at time t. Using the MLP neural network model trained in step S32, predict the 9-dimensional multimodal fusion feature vector extracted at each historical moment t to obtain the damage type of the wire rope at time t. and degree of injury . The predicted and Combine into a tuple , represents the damage state of the wire rope at time t. The sliding time window is taken as a step of 10 minutes, and the damage state binary sequence in each window is extracted at the historical moment. Each sequence contains the damage state of N consecutive moments, which represents the development process of wire rope damage over a period of time.

[0054] Construct a damage label sequence. For each damage state sequence, find the damage label in the historical damage label data according to the timestamp t_N of the last moment. The actual damage record within the next M time steps. If there is a damage record within M time steps, the damage type and damage degree in the damage record are extracted to form a damage label binary tuple If there is no damage record within M time steps, the damage label binary is set to (normal, normal), indicating that the wire rope is in a normal state during this time period. The damage label binary of M time steps is arranged in chronological order to obtain a damage label sequence, which corresponds to the damage state sequence and indicates the actual damage development of the wire rope in the future period.

[0055] Construct a dataset and arrange all the extracted damage state sequences and corresponding damage label sequences in chronological order to obtain a dataset of size L, where L is the total number of damage state sequences. Each data sample contains a damage state sequence and a damage label sequence. The former is used as the input feature of the BiLSTM model, and the latter is used as the training target of the model. The dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1, which are used for model training, hyperparameter tuning, and performance evaluation, respectively.

[0056] S44: Use the cross entropy loss function and Adam optimization algorithm to train the BiLSTM model using the training data set. According to the method in S42, a BiLSTM damage trend prediction model is constructed, including an input layer, an embedding layer, a bidirectional LSTM layer, an attention layer, and an output layer. In the input layer, each binary in the damage state sequence is Converted into a 7-dimensional one-hot encoding vector. In the embedding layer, the 7-dimensional damage state feature vector is mapped to a 64-dimensional embedding space. In the bidirectional LSTM layer, the hidden layer dimension is set to 128, and the forward and reverse LSTMs encode the damage state sequence respectively. In the attention layer, the Luong attention mechanism is used to calculate the attention weights of the current moment and each moment to obtain the damage prediction feature vector at the current moment. In the output layer, the damage prediction feature vector is mapped to the probability distribution of damage type and damage degree in the next M time steps through the fully connected layer.

[0057] The cross entropy loss function is used as the loss function of the BiLSTM model to measure the difference between the damage probability distribution predicted by the model and the actual damage label. For each training sample, the damage probability distribution predicted by the BiLSTM model for the next M time steps is compared with the damage label sequence, and the cross entropy loss of each time step is calculated. The cross entropy losses of M time steps are averaged to obtain the overall loss value of the training sample. The loss values ​​of all training samples are averaged to obtain the average cross entropy loss of the current training batch as the optimization target for model training. The Adam optimization algorithm is used to optimize the parameters of the BiLSTM model, setting the initial learning rate to 0.001 and the momentum factor to is 0.9, is 0.999. In each training batch, the gradient of each parameter of the model is calculated according to the average cross entropy loss of the current batch. The Adam optimization algorithm is used to adaptively adjust the learning rate of each parameter and update the model parameters according to the first-order moment estimation and second-order moment estimation of the gradient. The gradient clipping technique is used to limit the gradient norm within a preset threshold to avoid the gradient explosion problem. The training set data is divided into several small batches, each batch contains 32 training samples, which are sequentially input into the BiLSTM model for forward propagation and back propagation. In each training batch, the average cross entropy loss is calculated, and the model parameters are updated using the Adam optimization algorithm to complete a parameter iteration. The number of training rounds is set to 50. In each training round, all data in the training set are completely traversed and parameter updated. At the end of each training round, the performance of the model is evaluated on the validation set, and the average cross entropy loss and damage prediction accuracy are calculated. The early stopping strategy is used. If the performance on the validation set does not improve after 5 consecutive training rounds, the training is stopped and the current optimal model parameters are saved.

[0058] S45: Use the trained BiLSTM model to predict the extracted damage type and damage degree feature vector of the wire rope at the current moment, and obtain the predicted value of the damage type and damage degree of the wire rope in the future period of time as the damage trend prediction result of the wire rope, and obtain the multi-sensor operation data of the wire rope at the current moment t, including strain data and acceleration data. In the same way as the training data set, with t as the center, select the sensor data in the time window of 5 minutes before and after. Extract the 4 statistical features of the strain data and the 5 statistical features of the acceleration data in the time window to construct a 9-dimensional multimodal fusion feature vector. Input the 9-dimensional multimodal fusion feature vector at the current moment t into the MLP neural network model trained in step S32. Through the forward propagation calculation of the MLP model, the predicted damage type of the wire rope at moment t is obtained. and predict the extent of injury The predicted damage type and damage degree are combined into a damage state binary The predicted damage state binary at the current moment is Append to the end of the historical damage feature sequence. If the length of the historical damage feature sequence exceeds the preset sequence length N, delete the first damage state tuple at the beginning of the sequence and keep the sequence length N. The updated historical damage feature sequence represents the damage state of the wire rope at the most recent N moments including the current moment. Input the updated historical damage feature sequence into the trained BiLSTM damage trend prediction model. Through the forward propagation calculation of the BiLSTM model, the probability distribution of damage type and damage degree for the next M time steps is obtained. For each future time step, the combination of damage type and damage degree with the highest probability is selected as the predicted damage state for that time step. Arrange the predicted damage states of the M time steps in chronological order to obtain the predicted sequence of damage type and damage degree of the wire rope in the future period of time.

[0059] S5, compare the collected basket weight data with the preset safety weight range to obtain a judgment result on whether the basket weight exceeds the limit; specifically, according to the design parameters and safety specifications of the basket, the preset safety weight range of the basket is [ W min , W max ] . It indicates the minimum safe weight of the hanging basket, which is usually set to 20% to 30% of the rated load, taking into account the weight of the hanging basket, the weight of permanent equipment and the weight of necessary tools and materials. Indicates the maximum safe weight of the hanging basket, that is, the rated load of the hanging basket, which is generally calculated based on design parameters such as wire rope strength, motor power, and braking performance. The preset safe weight range is used as the configuration parameter of the hanging basket control system and stored in the non-volatile memory of the system to ensure data persistence and consistency. In the main loop of the hanging basket control system, the hanging basket weight measurement value W transmitted by the data acquisition module is read in real time. The measured value W is compared with the preset safe weight range. [ W min , W max ] Compare and determine whether W is within the safe range. , indicating that the weight of the hanging basket is too light, there may be safety hazards such as missing tools and materials, equipment falling off, etc., and an underweight alarm signal is generated. , indicating that the weight of the hanging basket exceeds the rated load, there is an overload risk, which may lead to serious safety accidents such as wire rope breakage and brake failure, and an overload alarm signal is generated. , indicating that the weight of the hanging basket is within the safe range, meets the normal operating conditions, and no alarm signal is generated. According to the results of the weight data comparison, a judgment result is generated as to whether the weight of the hanging basket exceeds the limit. If an underweight alarm signal is generated, a prompt message of "the weight of the hanging basket is underweight and there is a safety hazard" is output, and the yellow warning light on the hanging basket control panel is turned on. If an overweight alarm signal is generated, a prompt message of "the weight of the hanging basket exceeds the limit and there is a safety risk" is output, and the red warning light on the hanging basket control panel is turned on, and the sound and light alarm device is triggered at the same time. If the weight of the hanging basket is within the safe range, a prompt message of "the weight of the hanging basket is normal and meets the operating conditions" is output, and the green indicator light on the hanging basket control panel is turned on. The judgment result is uploaded to the hanging basket safety monitoring system via wired or wireless communication, and the history of changes in the weight status of the hanging basket is recorded to provide data support for accident analysis and responsibility tracing.

[0060] like Figure 5 As shown, S6, according to the prediction results of the health status of the wire rope, the prediction results of the damage trend, and the judgment results of whether the weight of the hanging basket exceeds the limit, a decision tree algorithm is used to generate a control scheme for the operation of the hanging basket, including: S61: The obtained prediction results of the health status of the wire rope at the current moment, the prediction results of the damage trend of the wire rope in the future period, and the judgment results of whether the weight of the hanging basket exceeds the limit are used to construct an input feature vector. First, the prediction results of the health status of the wire rope at the current moment are extracted, and the health status of the multi-sensor data of the wire rope at the current moment is predicted using the trained MLP neural network model. The output of the MLP model is the damage type (broken strand, rusted, and stranded) and the damage degree (normal, mild, moderate, and severe) of the wire rope at the current moment. The damage type is encoded as an integer from 1 to 3, corresponding to broken strand, rusted, and stranded respectively; the damage degree is encoded as an integer from 0 to 3, corresponding to normal, mild, moderate, and severe respectively. The encoded damage type and damage degree are combined into a 2-dimensional feature vector to represent the health status of the wire rope at the current moment. For example, (2, 1) indicates that the wire rope is slightly corroded at the current moment.

[0061] Then, the damage trend prediction results of the wire rope in the future are extracted, and the trained BiLSTM model is used to predict the damage trend of the health status sequence of the wire rope at the current moment and the previous N moments. The output of the BiLSTM model is the damage type and damage degree sequence of the wire rope in the future M time steps, and each time step contains a damage type and a damage degree. For each future time step, the predicted damage type and damage degree are encoded as integers from 1 to 3 and 0 to 3, respectively, which is consistent with the encoding method of the health status characteristics at the current moment. The damage type and damage degree encodings of the future M time steps are concatenated in sequence to obtain a 2M-dimensional feature vector, which represents the damage development trend of the wire rope in the future. For example, (2, 1, 2, 2, 3, 2, .....) indicates that the wire rope will have mild corrosion, moderate corrosion, severe corrosion and other damages in the future.

[0062] Extract the result of whether the weight of the hanging basket exceeds the limit, collect the weight data of the hanging basket in real time through the weighing sensor, and compare it with the preset safe weight range [ W min , W max ] For comparison. If the basket weight , then the basket is judged to be underweight and coded as 1; if , then the basket is judged to be in an overweight state and coded as 3; if , then the weight of the hanging basket is judged to be normal and encoded as 2. The weight state of the hanging basket is encoded as a 1-dimensional feature, indicating the current load condition of the hanging basket. For example, 2 means that the weight of the hanging basket is within a safe range.

[0063] Construct an input feature vector, and sequentially concatenate the current health status feature of the wire rope (2-dimensional), the future damage trend feature (2M-dimensional), and the weight status feature (1-dimensional) of the basket to obtain a 2+2M+1-dimensional feature vector. This feature vector combines the current health status of the wire rope, the future damage trend, and the load condition of the basket, and comprehensively describes the current operating status of the basket. For example, if the wire rope is currently slightly corroded (2, 1), and moderate and severe corrosion will occur in the future (2, 2, 3, 2), and the weight of the basket is normal (2), then the constructed input feature vector is (2, 1, 2, 2, 3, 2, ......., 2), with a dimension of 2+2M+1. The three types of heterogeneous information, namely the current health status of the wire rope, the future damage trend, and the weight status of the basket, are extracted as a unified feature representation, and an input feature vector that comprehensively describes the operating status of the basket is constructed. This feature vector can be used as the input of the subsequent decision tree algorithm to generate a control scheme for the operation of the basket.

[0064] S62: According to the historical operation data of the suspended basket, the control scheme and control effect of the suspended basket under different working conditions are obtained, and the control effect is quantitatively evaluated by multiple indicators to obtain the optimal control scheme under each working condition, form a control knowledge base, and collect the operation record data of the suspended basket over a period of time, including operating condition parameters, control instructions and operating effect evaluation indicators. The operating condition parameters include the height, speed, acceleration, load, etc. of the suspended basket, which can be collected in real time by various sensors on the suspended basket. The control instructions include lifting, descending, braking, etc., which can be extracted from the control system log of the suspended basket. The operating effect evaluation indicators include operating smoothness, energy consumption, response time, etc., which can be calculated based on data such as acceleration, current, and timestamp. The collected historical data is cleaned and preprocessed to remove outliers and invalid data, and the data is converted into a structured table form. The operating condition parameters are extracted from the historical data to form a multidimensional feature vector to represent the operating status of the suspended basket. The K-means clustering algorithm is used to divide the operating condition feature vectors in the historical data into K clusters, each cluster representing a typical operating condition. For each cluster, the corresponding control instruction sequence under the working condition is extracted as the candidate control scheme for the working condition. The candidate control scheme can be a complete instruction sequence or a set of key control parameters, such as lifting speed, lowering speed, braking torque, etc.

[0065] Perform a multi-index evaluation of the control effect, and extract the corresponding operating effect evaluation index for the candidate control scheme under each working condition. Set different weight coefficients according to the importance of the index, and perform weighted summation on each index to obtain the comprehensive performance score of the control scheme. For example, the weight of operating smoothness can be set to 0.4, the weight of energy consumption can be set to 0.3, and the weight of response time can be set to 0.3, indicating that smoothness is the most important consideration. The calculation formula for the comprehensive performance score is: ,in are the weight coefficients of the three indicators. stability, energy, and response represent the quantitative evaluation values ​​of operation smoothness, energy consumption, and response time, respectively, which can be mapped to the same scale through normalization or standardization.

[0066] Select the optimal control scheme. For each operating condition, compare the comprehensive performance scores of the candidate control schemes and select the control scheme with the highest score as the optimal control scheme under that operating condition. The optimal control scheme means the instruction sequence or parameter combination that can achieve the best control effect under that operating condition, taking into account factors such as running smoothness, energy consumption and response time. If the comprehensive performance scores of multiple control schemes are the same, they can be selected based on other auxiliary indicators or expert experience.

[0067] Construct a control knowledge base, combine the operating parameters of each working condition and the corresponding optimal control scheme into a control knowledge, and form a rule form of "if the working condition then the control scheme". For example, "if the height = 50m and the speed = 1m / s and the load = 1000kg then the lifting speed = 0.5m / s and the braking torque = 500N·m". Collect the control knowledge of all working conditions to form a control knowledge base as training material for the subsequent decision tree algorithm. Each piece of knowledge in the control knowledge base represents the optimal control strategy obtained through optimization evaluation under specific working conditions, which reflects the empirical law of hanging basket control.

[0068] S63: Match the constructed input feature vector with the control knowledge base, and construct a training data set for the decision tree algorithm based on the matching results. Each training sample in the training data set contains the input feature vector and the corresponding optimal control scheme. First, match the feature vector with the control knowledge. For each input feature vector, extract the feature subvector representing the operating condition of the hanging basket as the query condition. In the control knowledge base, search for the K control knowledge most similar to the query condition as the candidate matching results. The similarity can be measured using commonly used distance measurement methods such as Euclidean distance and cosine similarity. The calculation formula for Euclidean distance is: , where x and y are two eigenvectors, are their values ​​on the i-th feature. The smaller the Euclidean distance, the closer the two feature vectors are in the feature space, and the more similar the corresponding operating conditions are.

[0069] Construct training samples, and for each input feature vector, use it as the input feature X of the decision tree. Extract the optimal control scheme from the matched K control knowledge and count their occurrence frequencies. Select the optimal control scheme with the highest occurrence frequency as the output label Y corresponding to the input feature vector. Combine the input feature X and the output label Y into a training sample (X, Y), which represents the optimal control strategy to be adopted under a given operating state. If there are multiple optimal control schemes with the same and highest occurrence frequency among the matched K control knowledge, you can randomly select one as the output label, or make a decision based on other auxiliary information.

[0070] Construct a training data set and repeat for all input feature vectors to obtain N training samples . Combine these N training samples into a training data set , as the input of the decision tree algorithm. Each sample of the training data set consists of two parts: the input feature X represents the current operating status of the hanging basket, including the health status of the wire rope, damage trend, weight of the hanging basket, etc.; the output label Y represents the optimal control scheme to be taken in this state, including specific operations such as lifting, lowering, and braking. The number of samples N in the training data set depends on the number of input feature vectors. It is usually necessary to cover as many operating conditions and control strategies as possible to improve the generalization performance of the decision tree model. In this application, the historical operation data is converted into a control knowledge base, and the real-time operation status is matched with the control knowledge base to construct a training data set for the decision tree algorithm. The training data set takes the operating status characteristics as input and the optimal control scheme as output, reflecting the empirical laws and optimization strategies of the hanging basket operation control.

[0071] S64: Using the CART decision tree algorithm, with the minimization of the Gini index as the partitioning criterion, a binary tree modeling is performed on the training data set constructed in the training step S63 to generate an initial decision tree. First, the CART (Classification and Regression Tree) algorithm is used to generate the initial decision tree from the training data set. Starting from the root node, binary tree modeling is performed recursively. The root node contains all training samples, each of which consists of an input feature vector X and an output control solution label Y. For each non-leaf node, an optimal feature and partitioning threshold are selected to divide the sample into two left and right child nodes. The left and right child nodes are recursively divided until the preset stopping condition is met, such as the number of samples contained in the node is less than the threshold, the depth of the tree reaches the upper limit, etc. The setting of the stopping condition requires a trade-off between the complexity and fitting ability of the decision tree, and is usually tuned through methods such as cross-validation.

[0072] For each non-leaf node, traverse all input features, and for all possible values ​​of each feature, try to divide the sample into two left and right child nodes. Assume that the current node is t, and the sample set is , there are K features in total, and the kth feature is possible values. For the vth value of the kth feature, Divide into two subsets and , respectively contain samples with feature k values ​​less than or equal to v and greater than v. Calculate the Gini index after partitioning Select the feature k with the smallest Gini index and the value v as the optimal partitioning scheme, and divide the sample set Divide into left and right child nodes.

[0073] Calculate the Gini index. For each node t, calculate its Gini index Gini(t) to measure the impurity of the node. Assume that node t contains samples, there are K categories in total, and the number of samples in the kth category is The Gini index of node t is: . The value range of the Gini index is [0, 1]. The smaller the value, the higher the purity of the node, that is, the more concentrated the category distribution of the samples in the node. When the node contains only samples of one category, the Gini index reaches the minimum value of 0; when the number of samples of each category in the node is equal, the Gini index reaches the maximum value of 1-1 / K. For each leaf node, the output control scheme label Y of all samples in the node is counted. The control scheme with the highest frequency of occurrence is selected as the prediction result of the leaf node. If there are multiple control schemes with the same and highest frequency of occurrence, one can be randomly selected as the prediction result, or a decision can be made based on prior knowledge. The prediction result of the leaf node represents the optimal control strategy that should be adopted when the input feature vector X falls into the node.

[0074] Generate an initial decision tree and recursively generate CART decision trees repeatedly until all nodes meet the stopping condition. The generated decision tree aims to minimize the Gini index and achieves recursive partitioning of the sample space by greedily selecting the optimal partitioning features and thresholds. The structure of the decision tree consists of the partitioning features, thresholds, and left and right subtrees of the nodes. The leaf nodes store the corresponding control scheme prediction results. The initial decision tree may have overfitting problems, and the generalization performance needs to be further improved. It needs to be optimized through pruning, integration, and other methods.

[0075] S65: The initial decision trees are combined through the integrated learning strategy to obtain the hanging basket operation control decision model, and the initial decision trees are integrated using the Bagging (Bootstrap Aggregating) strategy. Randomly select N samples (repeatable) from the dataset to form a sub-dataset Repeat random sampling M times to obtain M sub-datasets , the number of samples in each sub-dataset is the same as that in the original training set. Bagging introduces random perturbations in each sub-dataset through random sampling with replacement, thus enhancing the diversity of decision trees. , use the CART algorithm in step S64 to train a decision tree Each decision tree Independently train on different data subsets to obtain M decision trees with different structures and parameters Parallel training can make full use of computing resources and accelerate the construction process of the decision tree integration model. During the training process, the generation process of each decision tree is the same as step S64, including recursive partitioning, Gini index calculation, control scheme prediction, etc.

[0076] For a new input feature vector , input them into M decision trees respectively, and get M prediction results Each prediction result Represents the i-th decision tree for the feature vector The M prediction results are integrated by voting method, and the control scheme with the highest number of votes is selected as the final integrated prediction result. If there are multiple control schemes with the same number of votes and the highest number of votes, one can be randomly selected as the final prediction result, or weighed based on prior knowledge. Integrated prediction results The prediction results of M decision trees are combined, and the robustness and accuracy of prediction are improved through the voting mechanism. The input of the decision model is the eigenvector X of the operation state of the suspended basket, and the output is the corresponding optimal control scheme Y. When , it is input into M decision trees, and the final control solution is obtained through integrated prediction. The decision model reduces the variance of a single decision tree and improves the stability and generalization performance of the prediction through the Bagging integration strategy. The complexity of the decision model is related to the number of decision trees M, the depth of each decision tree and the number of nodes, and it is necessary to weigh the performance and efficiency of the model. This application adopts the CART decision tree algorithm and the Bagging integration learning strategy to construct an efficient and robust hanging basket operation control decision model. The model can predict the optimal control scheme according to the real-time operating status characteristics of the hanging basket, realize the intelligent control and optimization of the hanging basket operation process, and improve the safety, reliability and efficiency of the hanging basket operation. At the same time, by integrating multiple decision trees, the risk of overfitting of the model is effectively reduced.

[0077] S66: Using the gradient boosted decision tree (GBDT) algorithm, based on the training data set constructed in step S63, an online learning model for the hanging basket operation control is generated: Decision tree sequence generation: Using the Boosting strategy, based on the training data set , set the number of decision trees M and the sampling ratio η, initialize the sample weight distribution to uniform distribution, and iteratively train M CART decision trees. In each round of iteration m, according to the current sample weight distribution, Sampling is performed to obtain a sub-dataset Using the method of step S64, based on Training a CART decision tree , and calculate its The prediction error on According to the prediction error , calculate the decision tree The weight coefficient , used to measure The importance of the final integrated model. Update the sample weight distribution, increase the weight of the samples that are predicted incorrectly, and reduce the weight of the samples that are predicted correctly, so that the next round of iteration focuses on difficult samples. Hyperparameter optimization: Set the hyperparameter range of the decision tree sequence, including the maximum depth of a single decision tree, the minimum number of samples in the leaf node, and the maximum number of partition features. Use the grid search method to traverse all hyperparameter combinations. Use K-fold cross validation to Divide into K mutually exclusive subsets, select K-1 subsets as training sets each time, and the remaining 1 subset as validation set. For each set of hyperparameter combinations, based on different training sets and validation sets, train and evaluate the decision tree sequence to obtain K sets of model evaluation results, including mean absolute error MAE, root mean square error RMSE and determination coefficient . The K groups of evaluation results are integrated using a weighted average soft voting method to obtain the comprehensive performance of the decision tree sequence under the current hyperparameter combination. The hyperparameter combination with the best comprehensive performance is selected as the optimal hyperparameter for the decision tree sequence. Online learning model: Using the optimal hyperparameters, based on the complete training data set , train and obtain M decision trees , the weight coefficient of each tree is , constituting the GBDT online learning model for hanging basket operation control.

[0078] S67: Input the input feature vector constructed in S61 into the basket operation control decision model (Bagging) obtained in step S65 and the online learning model (GBDT) obtained in step S66, respectively, to obtain a control scheme based on historical experience and a control scheme based on online learning: Historical experience control scheme: Input feature vector Input into the Bagging decision model, and after parallel prediction and voting of M decision trees, a control plan based on historical experience is obtained. . Online learning control scheme: The input feature vector Input into the GBDT online learning model, and after weighted prediction by M decision trees, a control scheme based on online learning is obtained. .

[0079] S68: According to the control scheme based on historical experience and the control scheme based on online learning, the suspension basket operation control decision scheme is obtained through weight fusion. First, the fusion weight is initialized and the initial weight of the historical experience control scheme is set. and the initial weights of the online learning control scheme , indicating the importance of the two control schemes in the fusion decision. and The value range of is [0, 1] and satisfies , the initial value can be set based on prior knowledge or experience, e.g. , . Introduce weight adjustment coefficients α and β for dynamic adjustment and The initial values ​​of the weights are all 1, indicating that the weights of the two control schemes remain unchanged at the beginning.

[0080] Calculate the rate of change of the operating condition of the hanging basket, obtain the current state parameters during the operation of the hanging basket, including wire rope tension, hanging basket displacement and speed, etc., which can be collected in real time through sensors. Calculate the difference between the current state parameters and the historical state parameters to obtain the change in the state parameters. Divide the change in the state parameters by a time window (such as 1 second) to obtain the rate of change of the state parameters, which reflects the speed of change of the operating condition of the hanging basket. For example, the rate of change of wire rope tension = |current tension value-tension value at the previous moment| / time window, and the calculation method of displacement change rate and speed change rate is similar.

[0081] Adjust the fusion weight and set the preset threshold range to represent the normal range of the rate of change of the suspended basket operating condition. If it exceeds this range, it is considered that the condition has changed significantly. Compare the calculated rate of change of the suspended basket operating condition with the preset threshold range to determine whether it exceeds the normal range. If it exceeds the threshold range, calculate the change in the weight adjustment coefficients α and β according to the degree to which the rate of change deviates from the threshold. and . and The calculation formula is: Δα=k1×(rate of change-upper threshold), Δβ=k2×(lower threshold-rate of change), where and is the proportional coefficient, which controls the adjustment range. Update the weight adjustment coefficient: , ,in and is the weight adjustment rate parameter, which controls the adjustment speed and has a value range of (0, 1). and Normalize them to ensure that their sum is 1: , .

[0082] Use the normalized weight adjustment coefficient and , calculate the final weights of the historical experience control scheme and the online learning control scheme and . , represents the final weight of the historical experience control scheme; , represents the final weight of the online learning control scheme. and control schemes based on online learning Multiply by the corresponding final weights and , and then add them together to get the fused hanging basket operation control decision plan . , which represents the control decision result that comprehensively considers historical experience and online learning. The integrated control decision plan It takes into account the stability of historical experience and the adaptability of online learning, and can dynamically adjust the control strategy according to changes in the operating conditions of the hanging basket.

[0083] The fused hanging basket operation control decision plan Output to the hanging basket control system as the optimal control instruction at the current moment. The specific control parameters contained in , such as lifting speed, descending speed, braking torque, etc., are used to perform corresponding control operations. While executing the control operation, the operating status of the suspended basket is continuously monitored, the real-time state parameters are obtained, and the operating condition change rate is calculated and the weight is adjusted to realize the dynamic update and optimization of the control decision. The present application combines the control scheme based on historical experience and the control scheme based on online learning to obtain a suspended basket operation control decision scheme that comprehensively considers multiple factors. The scheme can dynamically adjust the weight ratio of historical experience and online learning according to the changes in the operating conditions of the suspended basket, which not only ensures the stability and reliability of the control decision, but also improves the control system's adaptability to environmental changes and real-time optimization capabilities. At the same time, through real-time monitoring and feedback, the closed-loop optimization of the control decision is realized, and the safety, efficiency and intelligence level of the suspended basket operation control are continuously improved.

Claims

1. A method for managing the safety of suspended basket transportation, characterized in that: include: S1, install flaw detection sensors on the wire rope of the hanging basket to collect wire rope operation data; S2, collects the basket weight data through a three-pulley torque sensor; S3, based on the collected wire rope operation data, a wire rope health status prediction model is established using the multi-layer perceptron MLP algorithm to predict the damage degree of the wire rope and obtain the wire rope health status prediction result; wherein the damage includes fatigue, rust, wear and wire breakage; S4, based on the prediction results of the health status of the wire rope, the bidirectional long short-term memory neural network BiLSTM is used to predict the damage trend of the wire rope in the future; S5, comparing the collected hanging basket weight data with a preset safe weight range to obtain a judgment result of whether the hanging basket weight exceeds the limit; S6, based on the prediction results of the health status of the wire rope, the prediction results of the damage trend, and the judgment result of whether the weight of the hanging basket exceeds the limit, a decision tree algorithm is used to generate a control plan for the operation of the hanging basket.

2. The method for managing the safety of suspended basket transportation according to claim 1, characterized in that: S3, based on the collected wire rope operation data, a wire rope health status prediction model is established using a multi-layer perceptron MLP neural network to predict the damage degree of the wire rope and obtain the wire rope health status prediction results, including: S31, obtaining historical operation data of the steel wire rope and constructing a training data set; the training data set includes the steel wire rope operation data and corresponding damage type and damage degree labels; S32, constructing a wire rope health status prediction model based on a multi-layer perceptron MLP neural network, the model includes an input layer, multiple hidden layers and an output layer; wherein the hidden layer uses a ReLU activation function, and the output layer uses a Softmax function for multi-classification; S33, training the wire rope health status prediction model based on the training data set using the back propagation algorithm and the mini-batch gradient descent algorithm; S34, using the trained wire rope health status prediction model to predict the collected wire rope operation data, and selecting the damage type and damage degree combination with the largest prediction probability as the wire rope health status prediction result.

3. The method for managing the safety of suspended basket transportation according to claim 2, characterized in that: S31, obtaining historical operation data of the wire rope and constructing a training data set; The training data set contains wire rope operation data and corresponding damage type and damage degree labels, including: Collect multi-sensor data of the hanging basket wire rope in different time periods and different operating conditions; the multi-sensor data includes tension data measured by strain gauges and vibration data measured by accelerometers; Preprocessing the tension data and extracting statistical features; the statistical features include mean, variance, peak value and root mean square; The wavelet packet decomposition algorithm is used to perform time-frequency decomposition on the vibration data, and the energy characteristics and wavelet coefficients of different frequency bands are extracted; The fast Fourier transform algorithm is used to perform frequency domain analysis on the vibration data after wavelet decomposition to extract the spectrum features and inverse spectrum features; The statistical features of the tension data, the energy features, wavelet coefficients, spectrum features, and inverse spectrum features of the vibration data are aligned according to the timestamps to construct a multimodal fusion feature vector; Using the sliding window method, with a fixed window size and sliding step size, multi-scale data segments are extracted on the multimodal fusion feature vector. Each data segment contains continuous multimodal fusion features and corresponding timestamps. According to the timestamp of each data segment, the damage type and damage degree label of the corresponding segment is obtained from the historical maintenance records and fault logs of the hanging basket wire rope, and the mapping relationship between multimodal fusion features and damage labels is established; The multi-scale data segments are combined with the corresponding damage labels to construct a training dataset, where each training sample contains a multimodal fusion feature segment and a damage label.

4. The method for managing the safety of suspended basket transportation according to claim 2, characterized in that: S4, based on the prediction results of the health status of the wire rope, uses the bidirectional long short-term memory neural network BiLSTM to predict the damage trend of the wire rope in the future, including: S41, based on the prediction result of the health status of the steel wire rope, extract the damage type and damage degree feature vector of the steel wire rope at the current moment as the input of the BiLSTM model; S42, construct a BiLSTM damage trend prediction model, which includes an input layer, a bidirectional LSTM layer, an attention layer, and an output layer; S43, extracting the loss feature sequence at each historical moment and the real damage label in the future from the historical operation data of the wire rope obtained in step S3 and the corresponding historical damage label data, and constructing a training data set for the BiLSTM model; S44, using the cross entropy loss function and Adam optimization algorithm, the BiLSTM model is trained using the training dataset; S45, using the trained BiLSTM model to predict the extracted damage type and damage degree feature vector of the wire rope at the current moment, to obtain the damage type and damage degree prediction value of the wire rope in the future period of time as the loss trend prediction result of the wire rope.

5. The method for managing the safety of suspended basket transportation according to claim 4, characterized in that: S42, build a BiLSTM damage trend prediction model, which includes an input layer, a bidirectional LSTM layer, an attention layer, and an output layer, including: The extracted damage type and damage degree feature vector of the wire rope at the current moment is converted into a fixed-length one-hot encoding form as the input feature of the BiLSTM damage trend prediction model; The input features are mapped to a low-dimensional vector space through an embedding layer to obtain the distribution of damage type and damage degree vectors; The damage type and damage degree vectors output by the embedding layer are concatenated according to the feature dimension to obtain the damage feature vector at the current moment; Two parallel forward and backward LSTMs are used to forward and backward propagate the damage feature vector respectively, so as to obtain the forward and backward long-distance temporal dependency relationship in the development process of wire rope damage. At each time step of the forward LSTM network, the forward hidden state of the current time step and the damage feature vector at the current moment are concatenated according to the feature dimension as the input of the next time step; At each time step of the reverse LSTM network, the reverse hidden state of the current time step is concatenated with the damage feature vector at the current moment according to the feature dimension as the input of the previous time step; At each time step, the forward hidden state output by the forward LSTM network and the reverse hidden state output by the reverse LSTM network are concatenated according to the feature dimension to obtain the current bidirectional LSTM hidden state vector that integrates the bidirectional time series information. Using the Luong attention mechanism, the attention weights at the current moment and at each moment are calculated by the similarity between the current moment's bidirectional LSTM hidden state vector and the bidirectional LSTM hidden state vector at each moment. The bidirectional LSTM hidden state vectors at each moment are weighted summed based on the attention weights to obtain the attention feature vector at the current moment. The current moment attention feature vector is concatenated with the current moment bidirectional LSTM hidden state vector according to the feature dimension to obtain the current moment damage prediction feature vector; The damage prediction feature vector at the current moment is mapped to the predicted probability distribution of the damage degree of each damage type of the wire rope in the future through the fully connected layer and the Softmax function. The damage type with the largest probability value and the corresponding damage degree are taken as the damage trend prediction result at the current moment.

6. The method for managing the safety of suspended basket transportation according to any one of claims 1 to 5, characterized in that: S6, based on the prediction results of the health status of the wire rope, the prediction results of the damage trend, and the judgment result of whether the weight of the hanging basket exceeds the limit, a decision tree algorithm is used to generate a control plan for the operation of the hanging basket, including: S61, constructing an input feature vector using the obtained prediction result of the current health status of the wire rope, the prediction result of the damage trend of the wire rope in the future, and the judgment result of whether the weight of the hanging basket exceeds the limit; S62, according to the historical operation data of the hanging basket, obtain the control scheme and control effect of the hanging basket under different working conditions, and perform multi-index quantitative evaluation on the control effect, obtain the optimal control scheme under each working condition, and form a control knowledge base; S63, matching the constructed input feature vector with the control knowledge base, and constructing a training data set for the decision tree algorithm according to the matching result, wherein each training sample in the training data set includes the input feature vector and the corresponding optimal control solution; S64, using the CART decision tree algorithm, taking the minimization of the Gini index as the partitioning criterion, performing binary tree modeling on the training data set constructed in the training step S63, and generating an initial decision tree; S65, combining the initial decision trees through an integrated learning strategy to obtain a hanging basket operation control decision model; S66, using a gradient boosting decision tree algorithm to generate an online learning model for hanging basket operation control according to the training data set constructed in step S63; S67, inputting the input feature vector constructed in S61 into the hanging basket operation control decision model obtained in step S65 and the online learning model obtained in step S66, respectively, to obtain a control scheme based on historical experience and a control scheme based on online learning, respectively; S68, according to the control scheme based on historical experience and the control scheme based on online learning, a hanging basket operation control decision scheme is obtained through weight fusion.

7. The method for managing the safety of suspended basket transportation according to claim 6, characterized in that: S66, using a gradient boosting decision tree algorithm, based on the training data set constructed in step S63, generating an online learning model for the operation control of the hanging basket, including: Using the training data set constructed in step S63, a decision tree sequence is generated using a Boosting strategy according to a preset number of decision trees and a sampling ratio; Set the maximum depth of each decision tree in the decision tree sequence, the minimum number of samples for leaf nodes, and the maximum number of partition features; use the grid search method to traverse different combinations of hyperparameters; The training data set constructed in step S63 is divided into K mutually exclusive subsets by using a cross-validation method, and K-1 subsets are selected as training sets each time, and the remaining 1 subset is used as a validation set; For each set of hyperparameter combinations, K different training sets and validation sets are used to train and evaluate the decision tree sequence, respectively, to obtain K sets of model evaluation results; the model evaluation results include mean absolute error MAE, root mean square error RMSE and determination coefficient R2; The K groups of model evaluation results under each hyperparameter combination are fused using a weighted average soft voting method to obtain the comprehensive performance evaluation results of the decision number sequence under the current hyperparameter combination; All hyperparameter combinations are traversed, and based on the comprehensive performance evaluation results, the optimal hyperparameter combination and the corresponding decision tree sequence are selected as the online learning model for the gondola operation control.

8. The method for managing the safety of suspended basket transportation according to claim 7, characterized in that: S68, according to the control scheme based on historical experience and the control scheme based on online learning, a hanging basket operation control decision scheme is obtained through weight fusion, including: Obtain control schemes based on historical experience and control schemes based on online learning respectively; According to the preset fusion weights, the control scheme based on historical experience and the control scheme based on online learning are linearly weighted fused to obtain the hanging basket operation control decision scheme; Obtain the current state parameters during the operation of the hanging basket, and calculate the rate of change of the operating condition of the hanging basket according to the difference between the current state parameters and the historical state parameters; wherein the state parameters include the wire rope tension, the hanging basket displacement and the hanging basket speed; The calculated rate of change of the operating condition of the hanging basket is compared with a preset threshold range; If the change rate of the operating condition of the hanging basket exceeds the preset threshold range, the fusion weight is adjusted, otherwise the fusion weight is maintained unchanged.

9. The method for managing the safety of suspended basket transportation according to claim 8, characterized in that: If the change rate of the operating condition of the hanging basket exceeds the preset threshold range, the fusion weight is adjusted, including: Set the weight adjustment coefficient α of the historical experience control scheme and the weight adjustment coefficient β of the online learning control scheme, both of which have an initial value of 1; Obtain the calculated change rate of the operating condition of the suspended platform, and calculate the change value of the weight adjustment coefficient according to the degree of deviation of the change rate from the upper and lower limits of the preset threshold range and ; The degree of deviation is expressed by the absolute value of the difference between the rate of change and the upper and lower limits of the threshold range; Update the weight adjustment coefficient, and the update formula is: ; ; in, and are the weight adjustment coefficients of the updated historical experience control scheme and the online learning control scheme, and is a preset update rate parameter used to control the changing speed of the weight adjustment coefficient; Adjust the updated weight coefficient and Perform normalization: ; ; Use the normalized weight adjustment coefficient and , calculate the weights of the historical experience control scheme and the online learning control scheme and : ; ; in, and They are the weights of the historical experience control scheme and the online learning control scheme before fusion weight adjustment.

10. A hanging basket transportation safety management system, characterized in that: include: At least one processing unit; used to execute instructions to implement the hanging basket transportation safety management method described in any one of claims 1 to 9.

Images