Building high-altitude operation risk prediction method

Abstract

The invention provides a building high-altitude operation risk prediction method, and relates to the field of data prediction. The MTS-MSCA prediction model provided by the invention comprises a time decomposition module, a channel decomposition module and a prediction module, specifically, the time decomposition module can extract a time pattern and model a time dependency relationship, and the channel decomposition module can perform channel pattern learning and model a time dependency relationship; the prediction module is used for combining the output of the time decomposition module and the output of the channel decomposition module and converting the combined output into an available operation risk prediction result. All the modules cooperate with one another to achieve accurate prediction of the high-altitude operation risk of the building.

Description

Technical Field

[0001] The invention belongs to the field of data prediction, and in particular relates to a method for predicting risks of high-altitude construction operations. Background Art

[0002] Working at height is one of the higher-risk jobs in the construction industry, involving scaffolding construction, steel structure construction, exterior wall maintenance and other links. Due to the high working height and complex environment, construction workers are easily affected by factors such as gravity, wind, and falling materials, resulting in serious threats to personal safety. According to global construction safety data statistics, high-altitude fall accidents account for more than 40% of casualties in the construction industry, among which accidents caused by poor safety management, inadequate personal protection measures and environmental changes dominate. In addition, traditional safety management methods rely more on manual inspections and experience-based judgments, which make it difficult to detect potential risks in a timely manner, resulting in remediation after the accident occurs, increasing economic losses and construction delays.

[0003] At present, the construction industry mainly relies on safety training, personal protective equipment, safety monitoring and on-site management to reduce the risk of accidents in high-altitude operations. However, these traditional methods have certain limitations. Manual inspections are easily affected by subjective factors of personnel, resulting in incomplete identification of hidden dangers. The use of PPE relies on the self-consciousness of workers and cannot completely eliminate risks. Although video surveillance can provide real-time images, it is difficult to provide early warning of potential dangers. In addition, bad weather has a greater impact on high-altitude operations, and the existing management system often lacks dynamic adjustment capabilities and is difficult to respond to environmental changes in real time.

[0004] The risk data of high-altitude construction operations are usually multivariate, high-dimensional and nonlinear. They are significantly affected by environmental conditions, wind speed, temperature and operation parameters, height, equipment status, and dynamic changes in operation time. They are often non-stationary and have complex time dependencies. This data characteristic results in highly nonlinear interactions and multi-scale feature patterns between variables. Models based on deep learning, especially the architecture that introduces a cross-variable attention mechanism, can automatically extract complex multi-dimensional features from the data, capture the dynamic interaction relationship between variables, and efficiently extract key features on multiple time scales. Compared with traditional linear models or manual feature selection, the cross-variable attention mechanism can better model the correlation between variables, effectively improve the prediction accuracy of complex risk data, and provide more reliable support for the risk management of high-altitude operations. Summary of the invention

[0005] The invention provides a method for predicting the risk of high-altitude construction operations. Aiming at the installation quality data of green environmentally friendly wall materials with non-stationary and multivariable factors, an MTS-MSCA model is proposed, which consists of a time decomposition module, a channel decomposition block and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above object specifically includes the following steps:

[0007] A method for predicting risks in high-altitude construction operations, characterized by including the following steps: S1. Collect risk data for high-altitude construction operations and preprocess the collected data related to risks in high-altitude construction operations; S2. Standardize the preprocessed risk data for high-altitude construction operations using the mean normalization method, and divide the risk data for operations into a training set and a test set; S3. Construct a time decomposition module, propose a multi-scale time decomposition strategy, introduce a multi-scale Mamba structure, divide the risk data for high-altitude operations into time blocks of different window sizes, and extract local and global time features respectively. The specific steps are as follows: S31. Input the risk data for high-altitude construction operations Decompose the data into time blocks of different scales S32. For time blocks of different scales Use the Mamba network to extract time patterns to obtain Mamba models of different scales S33. For Obtain For Fuse to obtain the final weight matrix S c ; S4. Construct a channel decomposition module, design a cross-variable attention mechanism to capture the interaction patterns between different variables in the time series, design an improved dynamic gated convolution, combine the historical hidden states, propose an adaptive feature adjustment factor, and dynamically adjust the importance of local features. The specific steps are as follows: S41. For the historical sequence at each time step, use SSM to extract the historical sequence X to obtain the forward time feature X f ; S42. For the historical sequence at each time step, use SSM to extract the historical sequence X to obtain the reverse time feature X b ; S43. For the forward time feature X f and the reverse time feature X b , use cross-variable attention to combine them to obtain the combined output X e ; S5. Construct a prediction module for combining the outputs of the time decomposition module and the channel decomposition module to obtain the final prediction result

[0008] Preferably, in the step S1, collect the risk data of building high-altitude operations, including environmental parameter data, the status data of operators, and the operating status data of equipment. Among them, the environmental parameter data includes wind speed data, temperature data, humidity data, air pressure data, and light intensity data in the high-altitude operation area. The status data of operators includes the heart rate data, standing stability data, gait data, body surface temperature data, and posture angle data of construction workers. The operating status data of equipment includes vibration data, stress data, energy consumption data, displacement data, and connection stability data of the operation platform. Use Z-Socre to process the outliers in the data. The specific formula is: In the formula, X i is the value of the i-th data point in the risk data of building high-altitude operations, Z i is the standard score of the i-th data point, μ is the mean of this variable, σ is the standard deviation of this variable, |Z i | > Z thresh The data point X i is regarded as an outlier. Among them, Z thresh takes the value of 3.

[0009] Preferably, in the step S2, use the mean normalization method to standardize the preprocessed risk data of building high-altitude operations. The specific formula is: In the formula, X is the preprocessed risk data of building high-altitude operations, μ is the mean of the risk-related data of building high-altitude operations, σ is the standard deviation, and then divide the risk data of building high-altitude operations into a training set and a test set according to a proportion.

[0010] Preferably, in the step S3, in S31, input the risk data of building high-altitude operations Among them, T is the time step, N is the number of variables. Divide the input data into time blocks of different scales. The specific formula is: In the formula, T s is the length of the time window of different scales, Patch is the chunking operation, N is the number of channels of each Patch segment, P is the number of Patch segments divided by each time series, T p is the length of each Patch segment.

[0011] Preferably, by decomposing the aerial work risk data into time blocks of different time scales, according to the dynamic change characteristics of the working environment conditions, working parameters, and equipment status variables, short-term local risk patterns and long-term global risk trends are respectively extracted. By deeply learning and capturing risk patterns at different time scales, the model can more comprehensively identify risk change characteristics, thereby improving the prediction accuracy of potential aerial work risks.

[0012] Preferably, in S3 and S32, the Use the Mamba network to extract time patterns, and use the state space model SSM(·) to capture time-dependent information and non-linear dynamic characteristics. The specific formula is: In the formula, s is different time scales, σ is the non-linear activation function ReLU, and Linear is a linear transformation. The specific formula is: In the formula, W L is the linear projection matrix, b L is the bias term, and Conv is the convolution operation used to extract local time patterns. The specific formula is: In the formula, W c (k) is the convolution kernel, K is the convolution kernel size, t is the current moment, and b c is the bias; SSM(·) adopts an explicit state update equation. The specific formula is: h′(t) = Ah(t) + Bx(t); y(t) = Ch(t); In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, h(t) is the hidden state at the current time step, x(t) is the observed value at the current time step, and y(t) is the predicted value of the state update; Use the first-order hold method to process discrete data. The specific formula is: h t = Ah t-1 + B 0 x t + B 1 x t-1 ; B 0 =(I - e AΔt )A -1 B; B 1 = ΔtB - B 0 ; Among them, x t-1 is the input value at the previous moment, xt is the input value at the current moment, B 0 is the contribution of the previous time step to the state update, B 1 is the contribution of the current time step to the state update, and Δt is the time step size.

[0013] Preferably, the Mamba network structure is used to deeply analyze the high-altitude operation risk data at different scales. A linear layer is used to model the global trend of the risk data, and a convolutional layer captures the local time pattern features. The information extracted from each layer is integrated through a multi-scale fusion technology, enhancing the model's ability to extract complex patterns in time series, enabling it to identify the key time features affecting risk prediction, and significantly improving the model's response speed and prediction accuracy.

[0014] Preferably, in step S3 and S33, the Mamba models at different scales are passed through a multi-layer perceptron and a Softmax operation to obtain weight matrices at different scales The specific formula is: In the formula, Softmax(·) is the activation function, and MLP is the multi-layer perceptron. The specific formula is: MLP(X) = σ(W 2 (σ(W 1 X + b 1 )) + b 2 ); In the formula, W 1 , W 2 are two-layer fully connected weight matrices, b 1 , b 2 are bias terms, and σ is the non-linear activation function ReLU; The weight matrices at different scales are fused to obtain the final weight matrix S c The specific formula is: In the formula, w s is the learnable weight.

[0015] Preferably, the output of the Mamba model is processed through a multi-layer perceptron, and the Softmax operation is used to calculate the weight matrices at each scale. Then, these weight matrices are fused into the final weight matrix. The dynamic weight adjustment method allows the model to effectively balance information between different time scales, optimizes the overall structure of the prediction model, and improves the adaptability of the model in the changing environment of high-altitude operation risks and the accuracy of prediction.

[0016] Preferably, in the steps S4 and S41, gated convolution is used to extract local features, and SSM is used to model time dependence to obtain the forward time feature X f , and the specific formula is: X f = X gated + X ssm ; In the formula, X gated is the local feature extracted by gated convolution, and the specific formula is: X gated = G dynamic ⊙ X conv ; G dynamic = σ(Conv g (X forward ) ⊙ Φ(h t-1 )); X conv = Conv(X forward ); In the formula, X forward is the forward input of the historical input X, Conv g (·) is the gated convolution layer, σ is the non-linear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element-wise multiplication, Φ(h t-1 ) is the non-linear transformation of the fused hidden state h t-1 , and the specific formula is: Φ(h t-1 ) = ReLU(W Φ h t-1 + b Φ ); In the formula, W Φ is the learnable weight parameter, b Φ is the learnable bias parameter; X ssm is the time dependence feature extracted by SSM, and the specific formula is: X ssm = Ch t ; h t = Ah t-1 + BX forward ; In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, h t is the hidden state at the current time step, and h t-1 is the hidden state at the previous time step.

[0017] Preferably, the forward branch of the bidirectional Mamba structure is used to process the historical sequence in the high-altitude operation risk data. By extracting the forward time features related to the dynamic changes of the operation environment conditions, operation parameters, and equipment status, the antecedent trends and patterns of risk variables are captured, providing the model with information on the development trend of risk data in the time dimension, strengthening the model's understanding of risk behaviors in the initial and intermediate stages of the operation process, and significantly improving the accuracy of risk prediction.

[0018] Preferably, in steps S4 and S42, an improved dynamic gating convolution is used to extract local features, and SSM is used to model time dependence to obtain the reverse time feature X b , and the specific formula is: X b = X gated + X ssm ; In the formula, X gated is the local feature extracted by the gating convolution, and the specific formula is: X gated = G dynamic ⊙ X conv ; G dynamic = σ(Conv g (X backward ) ⊙ Φ(h t-1 )); X conv = Conv(X backward ); In the formula, X backward is the reverse input of the historical input X, Conv g (·) is the gating convolution layer, σ is the non-linear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element-wise multiplication, Φ(h t-1 ) is the non-linear transformation of the fused hidden state h t-1 , and the specific formula is: Φ(h t-1 ) = ReLU(W Φ h t-1 + b Φ ); In the formula, W Φ is the learnable weight parameter, b Φ is the learnable bias parameter; X ssm is the time dependence feature extracted by SSM, and the specific formula is: X ssm = Ch t ; h t = Ah t-1+BX backward ; In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, h t is the hidden state at the current time step, and h t-1 is the hidden state at the previous time step.

[0019] Preferably, the historical sequence in the high-altitude operation risk data is processed through the reverse branch of the bidirectional Mamba structure to capture the dynamic features from the end of the risk data forward. By modeling the gradual change process of environmental conditions, the cumulative effect of operation parameters, and the gradual decay trend of equipment status, the reverse dynamic information related to risk changes is extracted, the end stage of the risk sequence and its development trajectory are comprehensively analyzed, the potential risk patterns in the later stage of the sequence are effectively captured, and the risk prediction accuracy of the model is significantly improved.

[0020] Preferably, in steps S4 and S43, cross-variable attention is used to combine the forward input X forward and the reverse input X backward to obtain the combined output X e , and the specific formula is: X e =SelfAttention(X f +X b ); Among them, X f is the forward time feature, X b is the reverse time feature, and SelfAttention is the attention mechanism. The specific formula is: X e =A c V; Among them, V is the feature content, and A c is the attention weight. The specific formula is: In the formula, d is the feature dimension, and Q, K, and V are the feature representations after linear transformation. The specific formula is: Q = W Q (X f +X b ); K = W K (X f +X b ); V = W V (X f +X b ); In the formula, W Q , W K , and W V are weight matrices.

[0021] Preferably, through the cross-variable attention mechanism, the bidirectional time features are deeply analyzed. In view of the dynamic changes in the environmental conditions, operation parameters, and equipment status time features in the high-altitude operation risk data, the interaction and correlation between these features are strengthened. By capturing the interaction patterns and changing trends of different risk variables in the time dimension, the focusing ability of the model on key risk features is improved, and the analysis efficiency and prediction accuracy of the model for high-dimensional and multi-variable risk data under complex environmental conditions are enhanced.

[0022] Preferably, in step S5, the historical and future sequences are regarded as a unified sequence, and through the shared weight matrix S c consistency is achieved to reconstruct historical data and predict future data for the channels. The specific formula is: X′ c =A c S c ; In the formula, X′ c is the reconstructed channel feature, is the predicted channel feature, A c is the learnable basis signal matrix for reconstruction, A p is the learnable basis signal matrix for prediction, S c is the shared weight matrix, representing the mixing ratio of variables in the feature pattern. In the time dimension, the historical and future sequences are uniformly represented to reconstruct historical data and predict future data for time. The specific formula is: X′ t =A t S t ; In the formula, X′ t is the time reconstruction of the historical sequence, is the time prediction of the future sequence, A t is the shared time basis signal matrix, S t is the time coefficient matrix of the historical sequence, S p is the time coefficient matrix of the future sequence. The features on the two axes of time and channel are concatenated together to obtain the final prediction result The specific formula is: In the formula, is the prediction result on the channel, is the prediction result on time, Concat(·) is the concatenation operation of time and channel features, f MLP is the MLP mapping process. The specific formula is: f MLP (x) = W 2 ·σ(W 1 ·x + b 1 ) + b 2 ; In the formula, W 1 is the weight of the first layer, W 2 is the weight of the second layer, b 1 is the bias of the first layer, b 2 is the bias of the second layer, and σ(·) is the non - linear activation function ReLU.

[0023] Preferably, by processing time and channel features in a unified framework and splicing them together, the consistent representation of historical and future sequences and the comprehensive utilization of features are achieved. It not only fully integrates multi - dimensional information inputs such as environmental conditions, operation parameters, and equipment status, but also effectively improves the model's global perception ability of risk features, thus significantly enhancing the comprehensiveness and accuracy of prediction results.

[0024] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows: The present invention proposes an MTS - MSCA prediction model, which is applied to the scenario of building high - altitude operation risk prediction, including a time decomposition module, a channel decomposition module, and a prediction module. Specifically, the time decomposition module can extract time patterns and model time - dependent relationships, the channel decomposition module can perform channel pattern learning, and the prediction module is used to combine the outputs of the time decomposition module and the channel decomposition module and convert them into available operation risk prediction results. The modules cooperate with each other to achieve accurate prediction of building high - altitude operation risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flowchart of the steps of a method for predicting building high - altitude operation risks.

[0026] Figure 2 is a structural diagram of the MTS - MSCA prediction model.

[0027] Figure 3 is a structural diagram of the time decomposition module.

[0028] Figure 4 is a structural diagram of the channel decomposition module.

[0029] Figure 5 is a fitting effect diagram of the MTS - MSCA prediction model for realizing building high - altitude operation risk prediction. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0031] Please refer to Figures 1 - 5 , the present invention provides a technical solution: a method for predicting risks in high-altitude construction operations. By constructing a time decomposition module, time patterns can be extracted and time dependencies can be modeled. The channel decomposition module can perform channel pattern learning. The prediction module is used to combine the outputs of the time decomposition module and the channel decomposition module and convert them into available operation risk prediction results. The specific steps are as Figure 1 shown.

[0032] Construct an MTS-MSCA prediction model, the structure of which is as Figure 2 shown, and the specific steps are as follows:

[0033] S1. Collect data related to risks in high-altitude construction operations and preprocess the collected data related to risks in high-altitude construction operations.

[0034] Furthermore, collect data on risks in high-altitude construction operations, including environmental parameter data, operator status data, and equipment operation status data. Among them, environmental parameter data includes wind speed data, temperature data, humidity data, air pressure data, and light intensity data in the high-altitude operation area. Operator status data includes the heart rate data, standing stability data, gait data, body surface temperature data, and posture angle data of construction workers. Equipment operation status data includes vibration data, force data, energy consumption data, displacement data, and connection stability data of the operation platform. Use Z-Socre to process the outliers in the data. The specific formula is: In the formula, X i is the value of the i-th data point in the data on risks in high-altitude construction operations, Z i is the standard score of the i-th data point, μ is the mean of this variable, σ is the standard deviation of this variable, and the data point X i where |Z thresh | > Z i is regarded as an outlier. Among them, Z thresh takes the value of 3.

[0035] S2. Standardize the preprocessed data on risks in high-altitude construction operations using the mean normalization method, and divide the data on risks in high-altitude construction operations into a training set and a test set according to a ratio of 7:3.

[0036] Further, the preprocessed data related to the risks of high-altitude construction operations is standardized using the mean normalization method. The specific formula is as follows: In the formula, X is the preprocessed data related to the risks of high-altitude construction operations, μ is the mean of the data related to the risks of high-altitude construction operations, and σ is the standard deviation. Then, the data related to the risks of high-altitude construction operations is divided into a training set and a test set according to a certain proportion.

[0037] S31. Input the data related to the risks of high-altitude construction operations Decompose the data into time blocks of different scales The values of the scales are 4, 8, and 16 respectively.

[0038] Further, input the data related to the risks of high-altitude construction operations Among them, T is the time step, N is the number of variables, and the value is 15. The input data is segmented into time blocks of different scales. The specific formula is as follows: In the formula, T s is the length of the time window of different scales, Patch is the block operation, N is the number of channels of each Patch segment, and the value is 15. P is the number of Patch segments divided by each time series. T p is the length of each Patch segment.

[0039] S32. Use the Mamba network to extract time patterns for different scales and obtain Mamba models of different scales

[0040] Further, for different scales use the Mamba network to extract time patterns and use the state space model SSM(·) to capture time-dependent information and non-linear dynamic features. The specific formula is as follows: In the formula, s is different time scales, and the values are 4, 8, and 16. σ is the non-linear activation function ReLU, and Linear is the linear transformation. The specific formula is as follows: In the formula, W L is the linear projection matrix, b L is the bias term, and Conv is the convolution operation used to extract local time patterns. The specific formula is as follows: In the formula, W c(k) is the convolution kernel, K is the convolution kernel size, with a value of 16, t is the current time, and b c is the bias; SSM(·) uses an explicit state update equation, and the specific formula is: h′(t) = Ah(t) + Bx(t); y(t) = Ch(t); In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, h(t) is the hidden state at the current time step, x(t) is the observed value at the current time step, and y(t) is the predicted value of the state update; The first-order hold method is used to process discrete data, and the specific formula is: h t = Ah t-1 + B 0 x t + B 1 x t-1 ; B 0 = (I - e AΔt )A -1 B; B 1 = ΔtB - B 0 ; Among them, x t-1 is the input value at the previous time step, x t is the input value at the current time step, B 0 is the contribution of the previous time step to the state update, B 1 is the contribution of the current time step to the state update, and Δt is the time step size.

[0041] S33. The Mamba models of different scales are passed through a multi-layer perceptron and a Softmax operation to obtain weight matrices of different scales Then, the weight matrices at different scales are fused to obtain the final weight matrix S c .

[0042] Furthermore, the Mamba models of different scales are passed through a multi-layer perceptron and a Softmax operation to obtain weight matrices of different scales The specific formula is: In the formula, the specific formula of the MLP is: MLP(X) = σ(W 2 (σ(W 1 X + b 1 )) + b 2 ); In the formula, W 1, W 2 is a two - layer fully - connected weight matrix, and σ is the non - linear activation function ReLU; Fuse the weight matrices at different scales to obtain the final weight matrix S c , and the specific formula is: In the formula, w s is the learnable weight, and the initial values are 0.3, 0.3, and 0.4 respectively.

[0043] S41. For the historical sequence at each time step, use SSM to extract the historical sequence X to obtain the forward - time feature X f ;

[0044] Furthermore, for the historical sequence at each time step, use the improved dynamic gated convolution to extract local features and use SSM to model time dependence to obtain the forward - time feature X f , and the specific formula is: X f = X gated + X ssm ; In the formula, X gated is the local feature extracted by the gated convolution, and the specific formula is: X gated = G dynamic ⊙ X conv ; G dynamic = σ(Conv g (X forward )⊙ Φ(h t-1 )); X conv = Conv(X forward ); In the formula, X forward is the forward input of the historical input X, Conv g (·) is the gated convolution layer, σ is the non - linear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element - wise multiplication, Φ(h t-1 ) is the non - linear transformation of the fused hidden state h t-1 , and the specific formula is: Φ(h t-1 ) = ReLU(W Φ h t-1 + b Φ ); In the formula, W Φ is the learnable weight parameter, and b Φ is the learnable bias parameter; X ssmThe time-dependent features extracted by SSM, and the specific formula is: X ssm =Ch t ; h t =Ah t-1 +BX forwad ; In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, and h t is the hidden state at the current time step, and h t-1 is the hidden state at the previous time step.

[0045] S42. The historical sequence at each time step. Use SSM to extract the historical sequence X to obtain the reverse time feature X b ;

[0046] Furthermore, for the historical sequence at each time step, use the improved dynamic gated convolution to extract local features, and use SSM to model the time dependence to obtain the reverse time feature X b , and the specific formula is: X b =X gated +X ssm ; In the formula, X gated is the local feature extracted by the gated convolution, and the specific formula is: X gated =G dynamic ⊙X conv ; G dynamic =σ(Conv g (X backward )⊙Φ(h t-1 )); X conv =Conv(X backward ); In the formula, X backward is the reverse input of the historical input X, Conv g (·) is the gated convolution layer, σ is the non-linear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element-wise multiplication, and Φ(h t-1 ) is the non-linear transformation of the fused hidden state h t-1 , and the specific formula is: Φ(h t-1 )=ReLU(W Φ h t-1 +b Φ ); In the formula, W Φ is the learnable weight parameter, bΦ is a learnable paranoia parameter; X ssm is the time-dependent feature extracted by SSM, and the specific formula is: X ssm = Ch t ; h t = Ah t-1 + BX backward ; In the formula, A is the state transition matrix, B is the input weight matrix, C is the input mapping matrix, and h t is the hidden state at the current time step, and h t-1 is the hidden state at the previous time step.

[0047] S43. Combine the forward time feature X f and the reverse time feature X b using cross-variable attention to obtain the combined output X e ;

[0048] Furthermore, use cross-variable attention to combine the forward input X forward and the reverse input X backward to obtain the combined output X e , and the specific formula is: X e = SelfAttention(X f + X b ); Among them, X f is the forward time feature, X b is the reverse time feature, and SelfAttention is the attention mechanism. The specific formula is: X e = A c V; Among them, V is the feature content, and A c is the attention weight. The specific formula is: In the formula, d is the feature dimension, with a value of 15. Q, K, and V are the feature representations after linear transformation. The specific formula is: Q = W Q (X f + X b ); K = W K (X f + X b ); V = W V (X f + X b ); Wherein, W Q , W K , W V is the weight matrix.

[0049] S5. Construct a prediction module for combining the outputs of the time decomposition module and the channel decomposition module to obtain the final prediction result

[0050] Furthermore, regarding the historical and future sequences as a unified sequence, consistency is achieved through the shared weight matrix S c to reconstruct the historical data and predict the future data for the channels. The specific formula is: X′ c = A c S c ; Wherein, X′ c is the reconstructed channel feature, is the predicted channel feature, A c is the learnable basis signal matrix for reconstruction, A p is the learnable basis signal matrix for prediction, S c is the shared weight matrix, representing the mixing ratio of variables in the feature pattern. In the time dimension, the historical and future sequences are uniformly represented to reconstruct the historical data and predict the future data for time. The specific formula is: X′ t = A t S t ; Wherein, X′ t is the time reconstruction of the historical sequence, is the time prediction of the future sequence, A t is the shared time basis signal matrix, S t is the time coefficient matrix of the historical sequence, S p is the time coefficient matrix of the future sequence. Concatenating the features of the time and channel axes together to obtain the final prediction result The specific formula is: Wherein, is the prediction result on the channel, is the prediction result on the time, Concat(·) is the concatenation operation of the time and channel features, f MLP is the MLP mapping process. The specific formula is: f MLP (x)= W 2·σ(W 1 ·x + b 1 ) + b 2 ; In the formula, W 1 is the weight of the first layer, W 2 is the weight of the second layer, b 1 is the bias of the first layer, b 2 is the bias of the second layer, and σ(·) is the non - linear activation function ReLU.

[0051] Furthermore, the MTS - MSCA prediction model is written in Python. The experiment runs on the Windows operating system. Pytorch is selected as the framework in the CUDA11.27 environment and trained on GeForce RTX3090. The optimizer is selected, the initial learning rate is set to 0.001, the training batch is set to 64, the training cycle is set to 100, and the dataset is the data related to the risk of building high - altitude operations for 60 days. After pre - processing, it is input into the MTS - MSCA prediction model.

[0052] Furthermore, the fitting effect diagram of the building high - altitude operation risk prediction realized by the MTS - MSCA prediction model is as Figure 5 shown. The abscissa in the figure is the high - altitude operation cycle (days), and the ordinate is the operation risk index (%). The gray dotted line and dots are the real data, and the black solid line and squares are the predicted data. It can be seen from the figure that the overall trend of the predicted curve and the real curve is highly consistent. Especially in the early stage of the operation cycle, the two curves basically coincide, indicating that the model can accurately capture the change characteristics of short - term operation risks and accurately model the initial risks. As the operation cycle increases, the operation risk index shows a downward trend, which indicates that the adaptability of the operators to the high - altitude environment is gradually enhanced and the equipment operation tends to be stable. Generally speaking, the overall prediction effect of the model is good, which can better capture the short - term risk characteristics of building high - altitude operations and has good long - cycle prediction ability.

Claims

1. A method for predicting risks of high-altitude construction work, characterized in that: The following steps are involved: S1. Collect risk data of high-altitude construction operations and pre-process the collected data related to risk of high-altitude construction operations; S2. Standardize the preprocessed building height operation risk data using the mean normalization method, and divide the operation risk data into a training set and a test set; S3. Construct a time decomposition module, propose a multi-scale time decomposition strategy, introduce a multi-scale Mamba structure, divide the high-altitude operation risk data into time blocks of different window sizes, and extract local and global time features respectively. The specific steps are as follows: S31. Input risk data of construction high-altitude operations Break the data into time chunks of different sizes S32, different scales Use the Mamba network to extract temporal patterns and obtain Mamba models of different scales S33, will Obtained through multi-layer perceptron and Softmax operation Will Fusion is performed to obtain the final weight matrix S c ; S4. Construct a channel decomposition module, design a cross-variable attention mechanism, capture the interaction pattern between different variables in the time series, design an improved dynamic gated convolution, combine the historical hidden state, propose an adaptive feature adjustment factor, and dynamically adjust the importance of local features. The specific steps are as follows: S41, for each time step of the historical sequence, use SSM to extract the historical sequence X and obtain the positive time feature X f ; S42, for each time step of the historical sequence, use SSM to extract the historical sequence X and obtain the reverse time feature X b ; S43, positive time feature X f and the reverse time feature X b , use cross-variable attention to combine and get the combined output X e ; S5. Construct a prediction module to combine the outputs of the time decomposition module and the channel decomposition module to obtain the final prediction result.

2. A method for predicting risks of high-altitude construction work according to claim 1, characterized in that: In order to predict the risk of high-altitude construction operations, the collected risk data of high-altitude construction operations include environmental parameter data, operator status data and equipment operation status data. Among them, the environmental parameter data include wind speed data, temperature data, humidity data, air pressure data and light intensity data in the high-altitude operation area. The operator status data include the heart rate data, standing stability data, gait data, body surface temperature data and posture angle data of the construction workers. The equipment operation status data includes the vibration data, force data, energy consumption data, displacement data and connection stability data of the working platform. The collected relevant data are preprocessed to ensure that there are no outliers in the data, and then the processed data are divided into training set and test set for training and evaluating the performance of the risk prediction model for high-altitude construction operations.

3. A method for predicting risks of high-altitude construction work according to claim 2, characterized in that: In step S31, the risk data of high-altitude construction work is input. Where T is the time step, N is the number of variables, and the input data Divide into time blocks of different scales, the specific formula is: Where, T s is the time window length of different scales, Patch is the block operation, N is the number of channels in each Patch fragment, P is the number of Patch fragments divided into each time series, T p The length of each Patch segment.

4. A method for predicting risks of high-altitude construction work according to claim 3, characterized in that: In step S32, different scales of The Mamba network is used to extract the temporal pattern, and the state space model SSM(·) is used to capture the time-dependent information and nonlinear dynamic characteristics. The specific formula is: In the formula, s is a different time scale, σ is a nonlinear activation function ReLU, and Linear is a linear transformation. The specific formula is: Where W L is the linear projection matrix, b L is the bias term, Conv is the convolution operation, which is used to extract the local time pattern. The specific formula is: Where W c (k) is the convolution kernel, K is the convolution kernel size, t is the current time, b c is the bias; SSM(·) adopts the displayed state update equation, the specific formula is: h′(t)=Ah(t)+Bx(t); y(t) = Ch(t); Where A is the state transfer matrix, B is the input weight matrix, C is the input mapping matrix, h(t) is the hidden state of the current time step, x(t) is the observed value of the current time step, and y(t) is the predicted value of the state update; Use the first-order hold method to process discrete data. The specific formula is: h t =Ah t-1 +B0x t +B1x t-1 ; B0=(I-e AΔt )A -1 B; B1 = ΔtB-B0; Among them, x t-1 is the input value of the previous moment, x t is the input value at the current moment, B0 is the contribution of the previous time step to the state update, B1 is the contribution of the current time step to the state update, and Δt is the time step.

5. A method for predicting risks of high-altitude construction work according to claim 4, characterized in that: In step S33, Mamba models of different scales are The weight matrices of different scales are obtained through multi-layer perceptron and Softmax operation The specific formula is: In the formula, Softmax(·) is the activation function, MLP is the multi-layer perceptron, and the specific formula is: MLP(X)=σ(W2(σ(W1X+b1))+b2); Where W1, W2 are the two-layer fully connected weight matrices, b1, b2 are bias terms, and σ is the nonlinear activation function ReLU; The weight matrices at different scales are fused to obtain the final weight matrix S c , the specific formula is: In the formula, w s are learnable weights.

6. A method for predicting risks of high-altitude construction work according to claim 5, characterized in that: In step S41, the improved dynamic gated convolution is used to extract local features, and the SSM is used to model the time dependency to obtain the forward time feature X f , the specific formula is: X f =X gated +X ssm ; Where, X gated It is the local feature extracted by gated convolution. The specific formula is: X gated＝ G dynamic ⊙ X conv ; G dynamic =σ(Conv g (X forward )⊙Φ(h t-1 )); X conv =Conv(X forward ); Where, X forward is the positive input of the historical input X, Conv g (·) is the gated convolution layer, σ is the nonlinear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element-by-element multiplication, Φ(h t-1 ) is the fusion hidden state h t-1 The nonlinear transformation of is: Φ(h t-1 )=ReLU(W Φ h t-1 +b Φ ); Where V Φ is the learnable weight parameter, b Φ is a learnable bias parameter; X ssm It is the time-dependent feature extracted by SSM. The specific formula is: X ssm =Ch t ; h t =Ah t-1 +BX forward 4 In the formula, A is the state transfer matrix, B is the input weight matrix, C is the input mapping matrix, and h t is the hidden state of the current time step, h t-1 is the hidden state at the previous time step.

7. A method for predicting risks of high-altitude construction work according to claim 6, characterized in that: In step S42, the improved gated convolution is used to extract local features, and the SSM is used to model the time dependency to obtain the reverse time feature X b , the specific formula is: X b =X gated +X ssm ; Where, X gated It is the local feature extracted by gated convolution. The specific formula is: X gated＝ G dynamic ⊙X conv ; G dynamic =σ(Conv g (X backward )⊙Φ(h t-1 )); X conv =Conv(X backward ); Where, X backward is the reverse input of the historical input X, Conv g (·) is the gated convolution layer, σ is the nonlinear activation function Sigmoid, G dynamic is the adaptive feature adjustment factor, ⊙ is the element-by-element multiplication, Φ(h t-1 ) is the fusion hidden state h t-1 The nonlinear transformation of is: Φ(h t-1 )=ReLU(W Φ h t-1 +b Φ ); Where V Φ is the learnable weight parameter, b Φ is a learnable bias parameter; X ssm It is the time-dependent feature extracted by SSM. The specific formula is: X ssm =Ch t ; h t =Ah t-1 +BX backward 4 In the formula, A is the state transfer matrix, B is the input weight matrix, C is the input mapping matrix, and h t is the hidden state of the current time step, h t-1 is the hidden state at the previous time step.

8. A method for predicting risks of high-altitude construction work according to claim 7, characterized in that: In step S43, cross-variable attention is used to analyze the positive input X forward and the reverse input X backward Combine and get the combined output X e , the specific formula is: X e =SelfAttention(X f +X b ); Among them, X f is the forward time characteristic, X b is the reverse time feature, SelfAttention is the attention mechanism, and the specific formula is: X e =A c V; Among them, V is the characteristic content, A c is the attention weight, and the specific formula is: In the formula, d is the feature dimension, Q, K, and V are the feature representations after linear transformation. The specific formula is: Q=W Q (X f +X b ); K=W K (X f +X b ); V=W V (X f +X b ); Where W Q , W K , W V is the weight matrix.

9. A method for predicting risks of high-altitude construction work according to claim 8, characterized in that: In step S5, the historical and future sequences are regarded as a unified sequence, and the shared weight matrix S c To achieve consistency, the channel is reconstructed for historical data and predicted for future data. The specific formula is: X′ c =A c S c ; In the formula, X′ c is the reconstructed channel feature, To predict channel features, A c is the learnable basis signal matrix for reconstruction, A p is the learnable basis signal matrix for prediction, S c is a shared weight matrix, which indicates the mixing ratio of variables in the feature pattern. In the time dimension, the historical and future sequences are uniformly represented, and the historical data are reconstructed and the future data are predicted. The specific formula is: X′ t =A t S t ; In the formula, X′ t For the temporal reconstruction of historical sequences, For the time prediction of future sequences, A t is the shared time base signal matrix, S t is the time coefficient matrix of the historical sequence, S p The time coefficient matrix of the future sequence is concatenated with the features of the two axes of time and channel to obtain the final prediction result. The specific formula is: In the formula, is the prediction result on the channel, is the prediction result in time, Concat(·) is the concatenation operation of time and channel features, and f MLP It is the MLP mapping process, and the specific formula is: f MLP (x)=W2·σ(W1·x+b1)+b2; Where W1 is the weight of the first layer, W2 is the weight of the second layer, b1 is the bias of the first layer, b2 is the bias of the second layer, and σ(·) is the nonlinear activation function ReLU.

Images