A method for predicting the safety performance of rolling shutter doors

Abstract

The present invention proposes a method for predicting the safety performance of rolling shutter doors, which belongs to the field of data prediction and includes dynamic slicing, time-weighted aggregation, adaptive projection and prediction decoder modules. The dynamic slicing module divides the time series into key periods, captures the multi-scale non-uniform changes of rolling shutter door variables, and constructs time series slicing features; the time-weighted aggregation module fuses information from different periods through dynamic weight distribution, and enhances the response capability of key time points; the adaptive projection module maintains the structural similarity of time segments, extracts low-dimensional discriminant features, and improves sensitivity to local changes; the prediction decoder module uses a structural induction modulation mechanism to adjust the influence weight of historical moments on future moments, and combines nonlinear dynamic response functions to achieve accurate prediction. This method fully considers the spatiotemporal dynamic characteristics and structural dependencies of rolling shutter door operation data, and significantly improves the prediction accuracy and robustness compared with traditional methods.

Description

Technical Field

[0001] The present invention belongs to the field of data prediction, and in particular relates to a method for predicting the safety performance of a rolling door. Background Art

[0002] In recent years, with the development of Internet of Things technology, big data analysis and machine learning, researchers have begun to try to collect real-time data of rolling shutter doors, combine it with advanced intelligent algorithms, conduct data analysis and prediction, so as to achieve dynamic evaluation and prediction of the safety performance of rolling shutter doors. These technologies can help realize intelligent monitoring and maintenance of rolling shutter doors, identify potential failure risks in advance, and improve the safety and reliability of rolling shutter doors.

[0003] However, the existing rolling shutter door safety performance prediction technology still faces many challenges, mainly reflected in how to effectively process and integrate these data to improve the accuracy and real-time performance of safety performance prediction. In addition, how to design an efficient safety performance prediction mechanism to adapt to complex and changeable environmental factors and improve the robustness and adaptability of the prediction model is still a key problem that needs to be solved in current technology. Therefore, the present invention proposes a rolling shutter door safety performance prediction method, which can perform dynamic prediction based on real-time data and multi-dimensional information, and has important practical application value. Summary of the Invention

[0004] The present invention provides a method for predicting the safety performance of rolling shutter doors. A prediction model is proposed for complex and multivariate rolling shutter door related data. The prediction model consists of a dynamic slicing module, a time-weighted aggregation module, an adaptive projection module and a prediction decoder module.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps:

[0006] S1. Collect multidimensional data related to rolling shutter doors and construct an initial data set. The features include operation features, environmental features, and structural features.

[0007] S2. A dynamic slicing mechanism is proposed to optimize the initial data set. The first data set is obtained by dynamically calculating the local density of the time series. The dynamic slicing mechanism is: by calculating the density between adjacent time points , smooth the local density field, dynamically divide the slice boundaries according to the density change, and obtain the aggregation characteristics of the entire time series ;

[0008] S3. Construct a time-weighted aggregation module to build an internal time structure potential field, and perform weighted aggregation on the structure of the observation vector to obtain a high-dimensional structural feature set for each time segment of the slice. As the second data set;

[0009] S4. Build an adaptive projection module to construct a low-dimensional structural embedding representation matrix by building a structural similarity graph , based on the significance weight of the time segment Perform weighted fusion on the fragments and output structure-aware fusion features As a third data set, the third data set is divided into a training set and a prediction set;

[0010] S5. A prediction decoder module driven by structure-induced modulation is proposed. The structure-induced modulation function is used to dynamically adjust the time influence weight. Combined with the dynamic response function, the multi-time series features are nonlinearly mapped to achieve a fine expression and modeling of the complex rolling shutter data time series.

[0011] S6. The training set is input into a prediction module, and the basis function weights and bias parameters in the structural induction modulation function parameter response function are jointly optimized through an error feedback mechanism to train a rolling door safety performance prediction model;

[0012] S7: The prediction set is input into the trained safety performance prediction model, and the predicted value of the rolling door safety performance is finally output. .

[0013] Preferably, multidimensional data related to the operation of the rolling shutter door is collected in S1 to construct an initial data set, and the features include operation features, environmental features and structural features, wherein the operation features include switching frequency, operating time, motor current, and voltage; the environmental features include temperature, humidity, and wind speed; and the operation features include door weight and service life.

[0014] Preferably, the prediction of rolling door safety performance has significant spatiotemporal heterogeneity and dynamic complexity. Various variables change frequently on different time scales, and are often accompanied by sudden and local violent fluctuations. Traditional fixed-length time slices are difficult to effectively capture these multi-scale and multi-frequency change characteristics, resulting in the loss of key information and a decrease in prediction accuracy. In view of this feature, the present invention proposes the use of a dynamic slicing module to dynamically calculate the local sampling density and structural changes of the time series, flexibly adjust the boundaries and lengths of the time segments, and accurately divide the time periods with dense information or significant changes.

[0015] Furthermore, in step S2, the implementation process of the dynamic slicing module based on the time density field includes:

[0016] S21. First, integrate the various data of the rolling shutter door into a set of time series data ,in, For the Timestamp of the time point, strictly increasing, For the The rolling shutter door performance observation vector corresponding to each time point, The total number of time points. Data usually have large time intervals or irregular time sampling, which leads to large differences in the time intervals between adjacent sampling points. Therefore, in order to capture the local changes in the data, the present invention defines the local density is the inverse of the time interval, and the mathematical model is:

[0017] ;

[0018] Where, is the time interval between adjacent sampling points, Reflects the The local sparsity of the sampling points. When the density value is large, it means that the time interval is small and the data sampling is dense. When the density value is small, it means that the data in the area is sparse.

[0019] S22. In order to reduce the influence of noise, the density is smoothed and a sliding window is used to smooth the density to obtain a smoothed density field. The mathematical model is:

[0020] ;

[0021] Where, is the density value after smoothing, is the radius of the smoothing window, which indicates the range of the smoothing operation. By smoothing, the influence of short-term abnormal changes on the overall density field can be avoided, making the density change more stable. Then, an adaptive boundary is generated for each slice. Specifically, at each time point, The smoothed density value Define the boundary of the slice, the mathematical model is:

[0022] ;

[0023] Where, and Respectively The starting and ending time points of each slice, Indicates the time interval of the slice, and are the minimum density and maximum density of the density interval respectively;

[0024] S23. To ensure the uniformity of each slice, the present invention adjusts the boundary of each slice by optimizing the target to ensure that the density difference within the slice is minimized. The mathematical model is:

[0025] ;

[0026] Where, is the total number of slices, For the The average density within each slice is minimized to ensure that the density within each slice is as close as possible, thereby enhancing the representativeness of each slice. Then, weighted aggregation is performed according to the time density. For each time point Calculate weighting coefficients , the mathematical model is:

[0027] ;

[0028] Next, use the weighting coefficients Performance observation vector within the slice Perform weighted aggregation to obtain the feature representation of each slice. The mathematical model is:

[0029] ;

[0030] Where, It is the weighted sum of all performance observations in the slice, which can fully reflect the performance characteristics of the slice. Finally, the weighted average of the aggregated features of all slices is used to obtain the aggregated features of the entire time series. , as the first data set, the mathematical model is:

[0031] ;

[0032] Where, The weight of each slice reflects the contribution of the slice to the overall prediction.

[0033] Preferably, the dynamic slicing module can dynamically and adaptively adjust the time segment boundaries according to the local sampling density and structural changes of the time series data, and realize the fine division of the key performance change intervals. It not only effectively focuses on the time periods with drastic changes or information-intensive information, improves the local accuracy of data expression, but also avoids the redundant calculations and information ambiguity problems caused by fixed slicing, and enhances the model's ability to capture sudden safety events. The dynamic slicing module is combined with the real-time requirements of the rolling shutter door safety performance prediction, and can flexibly adjust the slice length and frequency to adapt to the changing environmental conditions and warning time requirements during the use of the flying rolling shutter door, thereby providing more accurate and dynamic structural feature-rich input features for subsequent modules such as time-weighted aggregation and structural significance fusion, significantly improving the stability and robustness of the prediction.

[0034] Preferably, the data during the operation of the rolling shutter door usually exhibits temporal non-uniformity and multi-scale changes, resulting in large differences in the contribution of data information at different time points to the prediction results. Based on this feature, the present invention proposes a time-weighted aggregation module, which automatically allocates weights to each time segment according to the significance of time intervals and structural features by introducing a dynamic weight mechanism, effectively highlighting the influence of key time periods.

[0035] Furthermore, in step S3, the irregular time series in each slice of the first data set is sorted according to the time center. The weighted potential energy function is constructed at the position to obtain the average time deviation scale of each slice. The mathematical model is:

[0036] ;

[0037] Where, is the number of time points, is the slice time center, and the mathematical model is:

[0038] ;

[0039] At the same time, define the time potential energy function calculation time point The aggregation potential energy in the slice is mathematically modeled as:

[0040] ;

[0041] Where, is the average absolute time offset scale of the slice, is the decay rate index, For time point The potential energy value is used to calculate the time point The aggregate potential energy at the slice, and then the potential energy response of each observation point on the time axis As the structural traction force of the observation point on the slice center, for all observation vectors Perform tensor superposition to form a time-aware feature representation , the data model is:

[0042] ;

[0043] Finally, a multi-mode potential energy mechanism is proposed to simulate the aggregate response field under multiple time peaks, and a high-dimensional structural feature set of each time segment of the slice is obtained. , the mathematical model is:

[0044] ;

[0045] Where, is the normalized weight coefficient used to fuse multi-core responses, is the reference time center, It is the time bandwidth parameter, which can be adjusted dynamically. is the exponential parameter, as the second data set.

[0046] Preferably, the time-weighted aggregation module processes the irregular time series in each time slice in the first data set, constructs a weighted potential function based on the position of the time center, dynamically calculates the average time deviation scale of each slice, and uses the time potential function to calculate the aggregation potential energy of each time point in the slice, and measures the influence of the time point on the slice center. At the same time, in order to further enhance the expression of spatiotemporal features, the present invention introduces a multi-modal potential energy mechanism, which simulates the aggregation response field under multiple time peaks by fusing multi-core responses, thereby generating a high-dimensional structural feature set, providing accurate input data for the subsequent time-weighted aggregation module, and greatly improving the prediction accuracy of the model.

[0047] Preferably, in the prediction of rolling door safety performance, each variable presents complex dynamic fluctuations over time, and the change characteristics of different time periods are significantly different. The traditional static projection method cannot adapt to the diversity and complexity of such time series data. Therefore, the adaptive projection module comes into being. It aims to dynamically adjust the projection process according to the local structure and change law of the time series data, so as to more accurately extract the key time periods and change trends. The module constructs a low-dimensional structure embedding representation matrix by constructing a structural similarity graph, performs weighted fusion on the segments according to the significance weights of the time segments, and outputs structure-aware fusion features.

[0048] Preferably, in step S4, the adaptive projection module implementation process is:

[0049] S41. Each target variable Divided into The time segment feature matrix is ​​constructed by segment feature vectors, and then an adaptive similarity connection mechanism based on feature space structural distance is introduced to calculate the structural similarity between nodes through the relative distribution relationship between structural embedding vectors. , the mathematical model is:

[0050] ;

[0051] Where, For the The feature vector of each time segment, is the L2 norm, In order to introduce the local structure perception scale factor to adaptively perceive the overall discreteness of the feature and prevent overfitting of local abnormal fragments, the mathematical model is:

[0052] ;

[0053] S42, based on the structural similarity between nodes A non-normalized graph Laplacian matrix is ​​constructed by accumulating the structural similarity between each time segment node and all other segments to construct the degree matrix of the graph. , the mathematical model is:

[0054] ;

[0055] Then the degree matrix of the graph is differentially operated with the constructed structural similarity adjacency matrix to obtain the Laplace matrix of the graph , the mathematical model is:

[0056] ;

[0057] Where, is the structural similarity adjacency matrix, which represents the structural connection weights between all time segments. Based on the Laplace structure of the graph, a structure-preserving embedding projection mechanism is introduced. By constructing a structural loss function, the tension imposed by the Laplace matrix on the feature distribution is used as the loss term, and the loss value is minimized in the compression direction to obtain the structure-preserving projection matrix. , the mathematical model is:

[0058] ;

[0059] Where, is a control factor used to adjust tensor operations, is the embedding space dimension, is a unit array, is the Frobenius norm.

[0060] S43, the projection matrix and high-dimensional structural feature matrix Multiply to construct a low-dimensional structure embedding representation matrix , the mathematical model is:

[0061] ;

[0062] Furthermore, the significance weight of each time segment in the embedding space is calculated by calculating the Time segment embedding vector The second norm of The significance amplitude of the time segment , which is used to measure the difference and activation intensity of a fragment from other fragments in the low-dimensional structural space. The larger the response amplitude, the more unique or critical the fragment has. The mathematical model is:

[0063] ;

[0064] Then the weight of each segment is calculated using the proportional distribution mechanism of significance amplitude , the mathematical model is:

[0065] ;

[0066] The proportional weight mechanism avoids simple average fusion, effectively amplifies the dominant contribution of key fragments of local structure, enhances the discrimination ability and expression sparsity of overall fusion, and finally converts the structural feature vectors of all time fragments into According to the weight of each fragment Perform dynamic fusion to obtain structure-aware fusion features , the mathematical model is:

[0067] ;

[0068] Finally, the third data set is obtained, and the third data set is divided into a training set and a prediction set according to a ratio of 7:3.

[0069] Preferably, the adaptive projection module maintains the structural similarity between time segments by constructing local adjacency relationships, and at the same time introduces the Laplace preservation criterion to guide the features to retain the local topological information in the original structure after projection, thereby enhancing the model's sensitivity to local change patterns. The module can flexibly adjust the embedding direction according to the feature distribution of different time periods, effectively improving the ability to express multi-scale operational changes, and providing a more stable and structurally consistent feature basis for subsequent saliency evaluation and fusion.

[0070] Preferably, in order to solve the problem of non-uniformity and structural dependence of the impact of historical observation information on future states in the prediction of rolling shutter door safety performance, the present invention proposes a prediction decoder module driven by structural induction modulation. Traditional prediction methods usually assume that the contribution of historical data to the future is equal weighted or linearly attenuated, which makes it difficult to characterize the dynamic distribution of the influence of complex variables on the time axis, especially in the face of sudden changes and safety accidents. The prediction decoder module adaptively adjusts the weight of the impact of historical moments on future moments according to the overall structural form of the fusion features, thereby realizing time response adjustment that varies with the structure.

[0071] Furthermore, in step S5, the structure-aware fusion features in the third dataset are As input, the modulation function is induced by the designed structure , to achieve the historical moments in the time series For the future Dynamic weighted adjustment of influence, the mathematical model is:

[0072] ;

[0073] Where, is the core width, which is dynamically calculated using nonlinear compression and coupling enhancement mechanisms, and then the dynamic response function is designed. , the mathematical model is:

[0074] ;

[0075] Where, is the number of basis function groups, that is, the number of multiple bases used to expand and express the nonlinear time-varying relationship of the input features in the dynamic response function. It reflects the complexity and flexibility of the expression of the dynamic response function. The larger the base number, the stronger the model fitting ability and expression details. For the The time-dependent weight function corresponding to the basis function is is the basis matrix, which acts as the projection basis of the nonlinear time-varying mapping and acts on the input features ,Will The basis function expansion is adopted, combined with the time weight of the rolling door related data, to express the dynamic changes of complex variables nonlinearly. Through the weighted summation of each basis function, the dynamic response function realizes the nonlinear transformation and weighted combination of the fusion features at different time points. The final module integrates the structural modulation and dynamic response to obtain the prediction expression suitable for the rolling door safety performance. The mathematical model is:

[0076] ;

[0077] Where, It is a trend item used to capture the dynamic changes in the safety performance of rolling shutter doors, which helps to improve the prediction accuracy of the model, especially in the face of dynamically changing environments and data, and can provide more stable and reliable predictions. For time point Safety performance prediction of rolling shutter doors.

[0078] Preferably, the prediction decoder module driven by structure-induced modulation achieves dynamic adjustment of the influence of historical moments in the prediction by introducing a structure-induced modulation mechanism based on contextual features, overcoming the problem that fixed weights in traditional methods cannot adapt to the non-stationarity of rolling shutter operation-related data. The module can adaptively adjust the modulation intensity according to the overall structural state of the fusion features, highlight the contribution of key time periods, take into account both short-term mutations and long-term trends, and enhance the model's ability to express complex gas data changes. At the same time, the module adopts explicit structure modulation functions and response expansion expressions, which have strong interpretability and computational efficiency, effectively enhancing the accuracy and robustness of rolling shutter safety performance predictions.

[0079] Preferably, in step S6, the mean square error is used to jointly optimize the basis function weights and bias parameters in the parameter response function of the structural induction modulation function to measure the error between the predicted value and the true value. The mathematical model is:

[0080] ;

[0081] Where, For time point The real safety performance value of the rolling shutter door is obtained by using the gradient descent and its variant optimization algorithm. The model gradually converges to effectively capture the complex time dependency relationship. After multiple rounds of training and adjustment, a trained safety performance prediction model is obtained. Finally, the prediction set is input into the trained rolling shutter door safety performance prediction model, and the predicted value of the rolling shutter door safety performance is finally output. .

[0082] In summary, the present invention proposes a method for predicting the safety performance of rolling shutter doors, which includes a dynamic slicing module, a time-weighted aggregation module, an adaptive projection module, and a prediction decoder module driven by structural induction modulation. First, the dynamic slicing module dynamically divides the key time periods of the time series, captures the multi-scale non-uniform changes of the rolling shutter data variables, and constructs high-precision time series slicing features; then, the time-weighted aggregation module finely fuses information from different time periods based on dynamic weight allocation, and enhances the response capability of key time points; the adaptive projection module extracts low-dimensional discriminant features by maintaining the structural similarity between time segments, and improves the model's sensitivity to local changes; finally, the prediction decoder module utilizes the structural induction modulation mechanism to adaptively adjust the influence weight of historical moments on future moments, and combines the nonlinear dynamic response function to achieve accurate modeling and prediction of complex time series. Compared with traditional methods, the present invention fully considers the spatiotemporal dynamic characteristics and structural dependencies of rolling shutter operation-related data, and significantly improves the accuracy and robustness of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is a step-by-step diagram of a rolling shutter door safety performance prediction method.

[0084] Figure 2 This is the structural diagram of the safety performance prediction model.

[0085] Figure 3 This is the structure diagram of the dynamic slicing module.

[0086] Figure 4 This is the structural diagram of the time-weighted aggregation module.

[0087] Figure 5 This is a graph showing how the model effect changes with the number of training times.

[0088] Figure 6 Fitting effect diagram of the prediction model to predict the safety performance of rolling shutter doors. DETAILED DESCRIPTION

[0089] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0090] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting the safety performance of rolling shutter doors, including a dynamic slicing module, a time-weighted aggregation module, an adaptive projection module, and a prediction decoder module driven by structural induction modulation. First, the dynamic slicing module dynamically divides the key time periods of the time series, captures the multi-scale non-uniform changes of the rolling shutter door data variables, and constructs high-precision time series slicing features; then, the time-weighted aggregation module finely fuses information from different time periods based on dynamic weight allocation to enhance the response capability of key time points; the adaptive projection module extracts low-dimensional discriminant features by maintaining the structural similarity between time segments, thereby improving the model's sensitivity to local changes; finally, the prediction decoder module uses the structural induction modulation mechanism to adaptively adjust the influence weight of historical moments on future moments, and combines nonlinear dynamic response functions to achieve accurate modeling and prediction of complex time series. The specific steps are as follows: Figure 1 shown.

[0091] Construct a rolling door safety performance prediction model, the structure of which is as follows Figure 2 As shown, the specific steps are:

[0092] S1. Collect multidimensional data related to rolling shutter doors and construct an initial data set. The features include operation features, environmental features, and structural features.

[0093] Furthermore, the dataset of the present invention includes 1000 rolling door data, and the dataset is divided into a training set and a test set according to a ratio of 7:3.

[0094] S2. A dynamic slicing mechanism is proposed to optimize the initial data set. The first data set is obtained by dynamically calculating the local density of the time series. The dynamic slicing mechanism is: by calculating the density between adjacent time points , smooth the local density field, dynamically divide the slice boundaries according to the density change, and obtain the aggregation characteristics of the entire time series .

[0095] Furthermore, a dynamic slicing module is constructed, whose structure is as follows Figure 3 As shown, first integrate the rolling shutter door related data into a set of time series data ,in, For the Timestamp of the time point, strictly increasing, For the The performance observation vector corresponding to each time point, The total number of time points is set to 60. The rolling door operation related data are usually sampled in a large time interval, which leads to a large difference in the time intervals between adjacent sampling points. Therefore, in order to capture the local changes of the data, the present invention defines the local density is the inverse of the time interval, and the mathematical model is:

[0096] ;

[0097] Where, is the time interval between adjacent sampling points, the value is set to 1 minute, Reflects the The local sparsity of the sampling points.

[0098] Furthermore, in order to reduce the noise impact caused by local abnormal disturbances in the original density estimation, the density is smoothed and a sliding window is used to smooth the density to obtain the smoothed density field. The mathematical model is:

[0099] ;

[0100] Where, is the density value after smoothing, is the radius of the smoothing window, and the value is set to 3, which indicates the range of the smoothing operation. Through smoothing, the influence of short-term abnormal changes on the overall density field can be avoided, making the density change more stable. Then, an adaptive boundary is generated for each slice. Specifically, at each time point, The smoothed density value Define the boundary of the slice, the mathematical model is:

[0101] ;

[0102] Where, and Respectively The starting and ending time points of each slice, Indicates the time interval of the slice, and are the minimum density and maximum density of the density interval, respectively, and the initial value is set to [0.2, 0.8].

[0103] Furthermore, when the consecutive time points satisfy When , it is represented as a slice interval. If the density value jumps out of This interval is considered the end of the slice, and then a new slice is started. The area with a density greater than 0.5 will be subdivided into multiple small slices to capture these changes more finely, while the area with a density less than 0.3 will be merged into a larger slice to avoid over-slicing and redundant calculations.

[0104] Furthermore, in order to improve the balance and representativeness of slice division, the present invention adjusts the boundary of each slice by optimizing the target, ensuring that the density difference within the slice is minimized and enhancing the ability of each slice to express the area in which it is located. The mathematical model is:

[0105] ;

[0106] Where, is the total number of slices, set to 30, For the The average density within each slice is minimized to ensure that the density within each slice is as close as possible, thereby enhancing the representativeness of each slice. Then, weighted aggregation is performed according to the time density. For each time point Calculate weighting coefficients , the mathematical model is:

[0107] ;

[0108] Next, use the weighting coefficients Safety performance observation vector within the slice Perform weighted aggregation to obtain the feature representation of each slice. The mathematical model is:

[0109] ;

[0110] Where, It is the weighted sum of all performance observations within the slice, which can fully reflect the performance characteristics of the region. Finally, the weighted average of the aggregated features of all slices is used to obtain the aggregated features of the entire time series. , then the first data set is obtained, and the mathematical model is:

[0111] ;

[0112] Where, The weight of each slice reflects the contribution of the slice in the final aggregated feature. It takes a random value between (0, 1) and the initial value is set to 1 / len(p).

[0113] S3. Construct a time-weighted aggregation module to build an internal time structure potential field, and perform weighted aggregation on the structure of the observation vector to obtain a high-dimensional structural feature set for each time segment of the slice. as the second data set.

[0114] Furthermore, a time-weighted aggregation module is constructed, whose structure is as follows Figure 4 As shown, for the irregular time series in each slice of the first data set, according to the time center The weighted potential energy function is constructed at the position to obtain the average time deviation scale of each slice. The mathematical model is:

[0115] ;

[0116] Where, is the number of time points, is the slice time center, and the mathematical model is:

[0117] ;

[0118] At the same time, define the time potential energy function calculation time point The aggregation potential energy in the slice is mathematically modeled as:

[0119] ;

[0120] Where, is the average absolute time offset scale of the slice, is the decay rate exponent, the value is set to 2, For time point The potential energy value is used to calculate the time point The aggregate potential energy at the slice, and then the potential energy response of each observation point on the time axis As the structural traction force of the observation point on the slice center, for all observation vectors Perform tensor superposition to form a time-aware feature representation , realizing the aggregation transmission of local time information to the slice level. The data model is:

[0121] ;

[0122] Finally, a multi-mode potential energy mechanism is proposed to simulate the aggregate response field under multiple time peaks, and a high-dimensional structural feature set of each time segment of the slice is obtained. , the mathematical model is:

[0123] ;

[0124] Where, is the normalized weight coefficient used to fuse multi-core responses, is the reference time center, is the time bandwidth parameter, the initial value is set to 3, which can be adjusted dynamically. is the exponential parameter, the value is set to 2, and finally we get It represents the slice-level features after fusing multiple temporal structure potentials as the second data set.

[0125] S4. Build an adaptive projection module to construct a low-dimensional structural embedding representation matrix by building a structural similarity graph , based on the significance weight of the time segment Perform weighted fusion on the fragments and output structure-aware fusion features As the third data set, the third data set is divided into a training set and a prediction set.

[0126] Furthermore, each target variable Divided into segment feature vectors, The value is set to 7, the time segment feature matrix is ​​constructed, and then an adaptive similarity connection mechanism based on the feature space structural distance is introduced to calculate the structural similarity between nodes through the relative distribution relationship between the structural embedding vectors. , the mathematical model is:

[0127] ;

[0128] Where, For the The feature vector of each time segment, is the L2 norm, In order to introduce the local structure perception scale factor to adaptively perceive the overall discreteness of the feature and prevent overfitting of local abnormal fragments, the mathematical model is:

[0129] ;

[0130] Where, The initial value is calculated by the square root of the average Euclidean distance between all the segment feature vectors, which reflects the degree of dispersion of the overall feature. , then based on the structural similarity between nodes A non-normalized graph Laplacian matrix is ​​constructed by accumulating the structural similarity between each time segment node and all other segments to construct the degree matrix of the graph. , the mathematical model is:

[0131] ;

[0132] Then the degree matrix of the graph is differentially operated with the constructed structural similarity adjacency matrix to obtain the Laplace matrix of the graph , the mathematical model is:

[0133] ;

[0134] Where, is the structural similarity adjacency matrix, which represents the structural connection weights between all time segments. Based on the Laplace structure of the graph, a structure-preserving embedding projection mechanism is introduced. By constructing a structural loss function, the tension imposed by the Laplace matrix on the feature distribution is used as the loss term, and the loss value is minimized in the compression direction to obtain the structure-preserving projection matrix. , the mathematical model is:

[0135] ;

[0136] Where, It is a control factor, and its value is set to 0.1, which is used to adjust the tensor operation. is the embedding space dimension, the value is set to 8, is a unit array, is the Frobenius norm.

[0137] Furthermore, the projection matrix and high-dimensional structural feature matrix Multiply to construct a low-dimensional structure embedding representation matrix , the mathematical model is:

[0138] ;

[0139] Furthermore, the significance weight of each time segment in the embedding space is calculated by calculating the Time segment embedding vector The second norm of The significance amplitude of the time segment , which is used to measure the difference and activation intensity of a fragment from other fragments in the low-dimensional structural space. The larger the response amplitude, the more unique or critical the fragment has. The mathematical model is:

[0140] ;

[0141] Then the weight of each segment is calculated using the proportional distribution mechanism of significance amplitude , the mathematical model is:

[0142] ;

[0143] The proportional weight mechanism avoids simple average fusion, effectively amplifies the dominant contribution of key fragments of local structure, enhances the discrimination ability and expression sparsity of overall fusion, and finally converts the structural feature vectors of all time fragments into According to the weight of each fragment Perform dynamic fusion to obtain structure-aware fusion features , which has a high degree of structural expression ability, taking into account the global temporal structure information and local key fragment characteristics. The mathematical model is:

[0144] ;

[0145] Finally, the third data set is obtained, and the third data set is divided into a training set and a prediction set according to a ratio of 7:3.

[0146] S5. A prediction decoder module driven by structure-induced modulation is proposed. The time influence weight is dynamically adjusted through the structure-induced modulation function. The dynamic response function is combined to perform nonlinear mapping of multiple time series features, and the complex rolling shutter data time series is precisely expressed and modeled.

[0147] Furthermore, the structure-aware fusion features in the third dataset are As input, the modulation function is induced by the designed structure , to achieve the historical moments in the time series For the future Dynamic weighted adjustment of influence, the mathematical model is:

[0148] ;

[0149] Where, The kernel width is dynamically calculated using a nonlinear compression and coupling enhancement mechanism. The initial value is set to 0.5, and then randomly selected between (0.1, 1). When the kernel width becomes larger, it indicates a more relaxed time-dependent effect, and when it becomes smaller, it emphasizes the response of the local moment. Then, the dynamic response function is designed. , the mathematical model is:

[0150] ;

[0151] Where, is the number of basis function groups, that is, the number of multiple bases used to expand and express the nonlinear time-varying relationship of the input features in the dynamic response function. It reflects the complexity and flexibility of the expression of the dynamic response function. The larger the base number, the stronger the model fitting ability and expression details. The initial value is set to 8. For the The time-dependent weight function corresponding to the basis function is is the basis matrix, which acts as the projection basis of the nonlinear time-varying mapping and acts on the input features ,Will The basis function expansion is used to combine the time weight of the data to express the dynamic changes of the complex rolling door data variables nonlinearly. By weighted summation of each basis function, the dynamic response function realizes the nonlinear transformation and weighted combination of the fusion features at different time points. The final module integrates the structural modulation and dynamic response to obtain the prediction expression of the rolling door safety performance. The mathematical model is:

[0152] ;

[0153] Where, For time point performance prediction, is a trend item used to capture the dynamic changes in the safety performance of rolling shutter doors. The mathematical model is:

[0154] ;

[0155] Where, It is a regulating factor used to control the impact of error changes on structural constraints. Its initial value is 0.2, and it subsequently takes random values ​​between (0.1, 1). Through the integration process combined with dynamic weights and structural modulation, nonlinear modeling of complex temporal relationships is achieved.

[0156] S6. The training set is input into a prediction module, and the basis function weights and bias parameters in the structural induction modulation function parameter response function are jointly optimized through an error feedback mechanism to train the rolling door safety performance prediction model.

[0157] Furthermore, the training set was input into the safety performance prediction model. The model used the PyTorch deep learning framework and was run in a Linux operating system environment. It was accelerated using an NVIDIA V100 32GB GPU. During the training process, the batch size was set to 64. The mean square error was used to jointly optimize the basis function weights and bias parameters in the parameter response function of the structural induction modulation function to measure the error between the predicted value and the true value. The mathematical model is:

[0158] ;

[0159] Where, For time point The real safety performance value of the model is obtained by using gradient descent and its variant optimization algorithm. The model gradually converges to effectively capture the complex time dependency. Figure 5 This is a graph showing the effect of the model as the number of training times increases. It can be seen from the figure that as the number of training times increases, the model prediction effect becomes better and better. Through multiple rounds of training and adjustment, a trained safety performance prediction model is obtained. Finally, the prediction set is input into the trained safety performance prediction model, and the predicted value of the rolling shutter door safety performance is finally output. .

[0160] Furthermore, the safety performance prediction model realizes the rolling door safety performance prediction fitting effect diagram as shown in the figure Figure 6 As shown in the figure, the horizontal axis represents the date, the vertical axis represents the corresponding rolling door safety performance value, the dotted line and the cross represent the model prediction value, and the solid line and the dot represent the actual rolling door safety performance value. It can be seen from the figure that the predicted value and the actual value are highly consistent in the change trend, and the change direction of the two remains synchronized, indicating that the model can effectively capture the time series dynamic characteristics of the variables in the rolling door data and their changing laws, and has strong prediction accuracy and robustness.

Claims

1. A method for predicting the safety performance of a rolling door, characterized in that: The following steps are involved: S1. Collect multidimensional data related to rolling shutter doors and construct an initial data set. The features include operation features, environmental features, and structural features. S2. A dynamic slicing mechanism is proposed to optimize the initial data set. The first data set is obtained by dynamically calculating the local density of the time series. The dynamic slicing mechanism is: by calculating the density between adjacent time points , smooth the local density field, dynamically divide the slice boundaries according to the density change, and obtain the aggregation characteristics of the entire time series ; S3. Construct a time-weighted aggregation module to build an internal time structure potential field, and perform weighted aggregation on the structure of the observation vector to obtain a high-dimensional structural feature set for each time segment of the slice. As the second data set; S4. Build an adaptive projection module to construct a low-dimensional structural embedding representation matrix by building a structural similarity graph , based on the significance weight of the time segment Perform weighted fusion on the fragments and output structure-aware fusion features As a third data set, the third data set is divided into a training set and a prediction set; S5. A prediction decoder module driven by structure-induced modulation is proposed. The structure-induced modulation function is used to dynamically adjust the time influence weight. Combined with the dynamic response function, the multi-time series features are nonlinearly mapped to achieve a fine expression and modeling of the complex rolling shutter data time series. S6. The training set is input into a prediction module, and the basis function weights and bias parameters in the structural induction modulation function parameter response function are jointly optimized through an error feedback mechanism to train a rolling door safety performance prediction model; S7: The prediction set is input into the trained safety performance prediction model, and the predicted value of the rolling door safety performance is finally output. .

2. A rolling door safety performance prediction method according to claim 1, characterized in that: The dynamic slicing mechanism construction process in S2 is specifically as follows: calculating the local time density for each sampling point , and then use the weighted sliding window method to smooth the density. The mathematical model is: ； Where, is the density value after smoothing, is the radius of the smoothing window, and then the slice is determined according to the smoothed density value The border , so that the sampling density value inside the slice The average density of the slice The sum of squared deviations between the two is minimized. At the same time, weighted aggregation is performed within each slice according to the time density. Calculate weighting coefficients , and then use the weighted coefficient to slice The observation values ​​within are weighted and integrated to generate a high-order representation vector Finally, the weighted average of the aggregate features of all slices is used to obtain the aggregate features of the entire time series , as the first dataset.

3. A rolling door safety performance prediction method according to claim 2, characterized in that: For the irregular time series in each slice of the first dataset, according to the time center The weight potential function is constructed at the position of Center of slice time Sum the absolute time deviations between the two and divide by the number of observation points , get the average time deviation scale of each slice, and define the time potential energy function to calculate the time point The aggregate potential energy in the slice, the mathematical model is: ； Where, is the average absolute time offset scale of the slice, is the decay rate index, For time The potential energy value is then converted into the potential energy response of each observation point on the time axis. Assume that the structural traction of the observation point to the slice center is applied to all observation vectors. Perform tensor superposition to form a time-aware feature representation Finally, a multi-mode potential energy mechanism is proposed to simulate the aggregate response field under multiple time peaks, and a high-dimensional structural feature set of each time segment of the slice is obtained. as the second data set.

4. A rolling door safety performance prediction method according to claim 3, characterized in that: Each target variable Divided into The time segment feature matrix is ​​constructed by segment feature vectors, and then an adaptive similarity connection mechanism based on feature space structural distance is introduced to calculate the structural similarity between nodes through the relative distribution relationship between structural embedding vectors. , obtain the Euclidean distance of any two time segments in the structural feature space, then normalize the Euclidean distance through the local scale adjustment term, and finally map it to the structural connection weight through the exponential function, in which the local structure perception scale factor is introduced The response width of the connection graph is automatically adjusted according to the overall distribution discreteness of the fragment set. The mathematical model of the local structure perception scale factor is: ； Where, For the The feature vector of each time segment, is the L2 norm, and then based on the structural similarity between nodes A non-normalized graph Laplacian matrix is ​​constructed by accumulating the structural connection strength between each time segment node and all other segments to construct the degree matrix of the graph. , perform a difference operation between the degree matrix of the graph and the constructed structural similarity adjacency matrix to obtain the Laplace matrix of the graph .

5. A rolling door safety performance prediction method according to claim 4, characterized in that: Based on the Laplace structure of the graph, a structure-preserving embedding projection mechanism is introduced. By constructing a structural loss function, the tension exerted by the Laplace matrix on the feature distribution is used as the loss term, and the loss value is minimized in the compression direction to obtain the structure-preserving projection matrix , the learning model is: ； Where, is the control factor, is the embedding space dimension, is a unit array, is the Frobenius norm, and then the projection matrix and high-dimensional structural feature matrix Multiply to construct a low-dimensional structure embedding representation matrix , further, calculate the significance weight of each time segment in the embedding space, by calculating the Time segment embedding vector The second norm of The significance amplitude of the time segment , and then use the proportional distribution mechanism of significance amplitude to calculate the weight of each segment , and finally the structural feature vectors of all time segments According to the weight of each fragment Perform dynamic fusion to obtain structure-aware fusion features As the third data set, the third data set is divided into a training set and a prediction set.

6. A rolling door safety performance prediction method according to claim 5, characterized in that: The structure-aware modulation-driven predictive decoder module is based on the structure-aware fusion features from the third dataset. As input, the modulation function is induced by the designed structure , to achieve the historical moments in the time series For the future Dynamic weighted adjustment of influence, the mathematical model is: ； Where, is the core width, which is dynamically calculated using nonlinear compression and coupling enhancement mechanisms, and then the dynamic response function is designed. ,Will The basis function expansion is used, combined with the time weight of the rolling shutter door operation related data, to express the dynamic changes of complex rolling shutter door data variables nonlinearly. The final module integrates structural modulation and dynamic response to obtain a predictive expression of the rolling shutter door safety performance. The mathematical model is: ； Where, is the trend item, For time point Safety performance prediction.

Images