Safety Risk Prediction Method for Basket-Type Steel Box Arch Lifting Process Near Operating Line

Abstract

This invention relates to the field of risk supervision and management, specifically to a method for predicting safety risks during the lifting of basket-type steel box arches near operating lines. The method includes: obtaining an initial estimated risk; using a sensor fusion matrix to weaken the changing trends of sensor-collected data; and updating the risk prediction model and sensor fusion matrix to minimize the difference between the risk prediction model output and the initial estimated risk; re-fusing the estimated risks from all sensors using the updated sensor fusion matrix to obtain a first estimated risk; updating the risk prediction model and sensor fusion matrix again to maximize the difference between the risk prediction model output and the first initial estimated risk; and conducting a risk warning based on the updated sensor fusion matrix. This invention provides a more reliable risk warning and is more suitable for the risk prediction needs when lifting basket-type steel box arches near operating lines.

Description

Technical Field

[0001] This invention relates to the field of risk supervision and management, specifically to a method for predicting safety risks during the lifting process of basket-type steel box arches near operating lines. Background Technology

[0002] Lifting a basket-type steel box arch in an environment adjacent to an operating railway line presents extremely high safety risks and technical challenges. For example, during the lifting process, the steel box arch structure may experience stress concentration, localized deformation, or even instability (such as linear or nonlinear buckling) due to uneven stress distribution or deviations in synchronization control, leading to structural damage or a fall. In the event of such an accident, not only will there be casualties and property damage, but it could also cause far-reaching consequences such as railway operation disruptions, posing significant liability risks to both the construction and operation units.

[0003] To address the aforementioned risks, conventional safety risk prediction methods typically rely on data collected from multiple sensors (such as strain sensors, vibration sensors, and anemometers), and fuse the predicted risks from each sensor using weighted averaging or other methods. However, this approach fails to consider the interrelationships between sensors. These interrelationships manifest in two ways: firstly, different sensors monitor changes in the state of different structures or parts during the lifting process, and changes in the state of some sensor locations may be influenced by the state of other sensor locations (e.g., stress changes in a critical area may trigger a chain reaction in other areas); secondly, changes in the state of some sensor locations may induce more severe risks in other locations (e.g., local instability leads to exponential deterioration of the overall structure). Therefore, these conventional methods have significant limitations, especially in meeting the stringent risk prediction or safety management requirements in environments adjacent to operational lines. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for predicting safety risks during the lifting process of a basket-type steel box arch near an operational railway line.

[0005] The safety risk prediction method for the adjacent operating line basket-type steel box arch lifting process of the present invention adopts the following technical solution:

[0006] One embodiment of the present invention provides a method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating railway line. The method includes the following steps:

[0007] The estimated risk of each sensor is assessed based on the changing trend of the data collected by multiple sensors during the improvement process. The estimated risks of all sensors are then fused by preset weights to obtain the initial estimated risk.

[0008] Initialize a risk prediction model and a sensor fusion matrix. Each element in the sensor fusion matrix represents the fusion coefficient of any two sensors. Randomly select two sensors and first weaken the trend of change in the data collected by the two sensors. The magnitude of the weakening is positively correlated with the fusion coefficients of the two sensors in the sensor fusion matrix. Then, input all the data collected by the sensors, including the data with weakened trend, into the risk prediction model to update the risk prediction model and the sensor fusion matrix, so that the difference between the output of the risk prediction model and the initial risk prediction is minimized.

[0009] The estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the first estimated risk;

[0010] The risk prediction model and sensor fusion matrix are updated again to maximize the difference between the risk prediction model output and the first initial risk prediction. The predicted risks of all sensors are re-fused through the updated sensor fusion matrix to obtain the final predicted risk. The final predicted risk is then used for risk warning.

[0011] Preferably, the step of weakening the trend of change in the data collected by the two sensors, wherein the degree of weakening is positively correlated with the fusion coefficients corresponding to the two sensors in the sensor fusion matrix, includes the following specific steps:

[0012] For either of the two sensors, and the time series formed by the data collected by that sensor; obtain the mean of all elements in the time series, denoted as . The i-th element in this time series is denoted as Normalize all fusion coefficients in the sensor fusion matrix, and use the normalized fusion coefficients corresponding to the two sensors as the attenuation magnitude. The time series obtained after weakening the trend of the data collected by the sensor is used as the corrected time series; the value of the i-th element in the corrected time series is represented as... , .

[0013] Preferably, the specific steps involved in inputting all sensor-collected data, including data after weakening the trend of change, into the risk prediction model, updating the risk prediction model and the sensor fusion matrix, and minimizing the difference between the output of the risk prediction model and the initial risk prediction are as follows:

[0014] The data collected by multi-source sensors at each time point are used as each sample, and the initial estimated risk is used as the label for each sample; the samples and labels obtained at all times are used as the dataset; each sample in the dataset is input into the risk prediction model in sequence; the two randomly selected sensors are used to input data with weakened trend changes.

[0015] The parameters of the risk prediction model are trained using all samples in the dataset and the loss function; the loss function is set as: the loss function P1 is composed of the mean square error of the risk prediction model output and the initial predicted risk; the parameters of the risk prediction model include all fusion coefficients in the sensor fusion matrix.

[0016] After training, the risk prediction model and sensor fusion matrix are updated to minimize the difference between the risk prediction model output and the initial risk prediction.

[0017] Preferably, the estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the first estimated risk, including the following specific steps:

[0018] When the difference between the output of the risk prediction model and the initial risk prediction is minimized, the difference between the output of the risk prediction model and the initial risk prediction is recorded as f1. For any sensor, the fusion coefficient of a row in the updated sensor fusion matrix is ​​recorded as f2. The average of all fusion coefficients of the corresponding row in the updated sensor fusion matrix is ​​recorded as f2. The preset weight of the sensor is updated using f2 and f1 to obtain the first weight of the sensor.

[0019] The first weight of all sensors is normalized, and the predicted risk of all sensors is weighted and summed using the normalized first weight of each sensor. The result is denoted as the first predicted risk.

[0020] Preferably, the specific steps involved in updating the risk prediction model and the sensor fusion matrix again to maximize the difference between the risk prediction model output and the first initial risk prediction are as follows:

[0021] The parameters of the risk prediction model are trained using all samples and loss functions in the dataset. The loss function is reset to P2, which is the difference between the output of the risk prediction model and the first initial risk prediction. After training, the risk prediction model and the sensor fusion matrix are updated again to maximize the difference between the output of the risk prediction model and the first initial risk prediction.

[0022] Preferably, the estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the final estimated risk, including the following specific steps:

[0023] When the difference between the risk prediction model output and the first initial risk prediction is the largest, the difference between the risk prediction model output and the first initial risk prediction is recorded as f3; the mean of all fusion coefficients in the corresponding row of any sensor is obtained from the updated sensor fusion matrix and recorded as f4; the first weight of each sensor is updated using f3 and f4 to obtain the second weight of each sensor.

[0024] The second weights of all sensors are normalized, and the predicted risks of all sensors are weighted and summed using the normalized second weights of each sensor. The result is recorded as the final predicted risk.

[0025] Preferably, the step of updating the preset weights of the sensor using f2 and f1 to obtain the first weight of the sensor includes the following specific formula:

[0026] Let W1 = W × (1 - f0), where W1 represents the first weight of the sensor; f0 represents the suppression amplitude of the sensor, which is equal to the ratio of f2 to f1; and W represents the preset weight of the sensor.

[0027] Preferably, the step of updating the first weight of each sensor using f3 and f4 to obtain the second weight of each sensor includes the following specific formula:

[0028] Let W2 = W1 × (1 + g0), where W2 represents the second weight of the sensor; W1 represents the first weight of the sensor; and g0 represents the gain amplitude of the sensor, which is equal to the product of f3 and f4.

[0029] Preferably, the specific steps for assessing the estimated risk of each sensor based on the changing trends of data collected by multiple sensors during the lifting process are as follows:

[0030] For each sensor, the data collected forms a time series, which is a curve that changes over time. The curve is fitted to a straight line using the least squares method, and the slope of the straight line is used as the estimated risk.

[0031] Preferably, the specific steps for risk warning using the final estimated risk are as follows:

[0032] When the steel box arch is lifted to a height less than the preset height, a risk warning is issued using the initial estimated risk; when the steel box arch is lifted to a height greater than or equal to the preset height, a risk warning is issued using the final estimated risk. The timing for issuing a risk warning refers to when the initial estimated risk or the final estimated risk is greater than the preset risk threshold.

[0033] The beneficial effects of the technical solution of the present invention are:

[0034] In this invention, data collected by all sensors, including data after weakening the trend of change, are input into the risk prediction model to update the risk prediction model and the sensor fusion matrix, so that the output of the risk prediction model is as close as possible to the initial predicted risk; the predicted risks of all sensors are re-fused through the updated sensor fusion matrix to obtain the first predicted risk.

[0035] This process involves analyzing the changing trends of data collected by each sensor individually and then directly fusing them to obtain the initial risk estimate. Furthermore, the fusion coefficients in the sensor fusion matrix are used to weaken the changing trends of some sensor data, thereby reducing the role of some sensors in risk assessment. The output of the risk prediction model describes the risk assessment result after deep fusion of data from all sensors. Even after reducing the role of some sensors in risk assessment, the risk assessment result output by the risk prediction model is still as close as possible to the initial risk estimate obtained by directly fusing data after analyzing and fusing the changing trends individually; this ensures that the updated sensor fusion matrix can describe whether it is necessary to analyze and fuse different sensors together when assessing risk. Finally, the initial risk estimate is re-fused based on this updated sensor fusion matrix, avoiding the following problem: because the state changes of some sensor locations are affected by the state of other sensor locations, the risk of data collected by some sensors may include or represent the risk characteristics of data collected by other sensors when describing risk, which is equivalent to double-counting the risk estimates of some sensors.

[0036] Furthermore, the present invention updates the risk prediction model and the sensor fusion matrix again, so that the difference between the output of the risk prediction model and the first initial predicted risk is maximized; the predicted risks of all sensors are re-fused through the updated sensor fusion matrix to obtain the final predicted risk; and the final predicted risk is used for risk warning.

[0037] This process ensures that when small state changes occur in some sensor combinations within the updated sensor fusion matrix, they induce larger state changes in other sensor combinations, ultimately leading to a higher risk assessment result from the risk prediction model. This guarantees that the updated sensor fusion matrix can describe whether state changes at the locations of some sensors might induce further risk updates at the locations of other sensors. Based on the updated sensor fusion matrix, the initial risk prediction is re-fused, further improving the reliability of risk warnings.

[0038] In summary, this invention re-integrates the initial estimated risk (i.e., updates the preset weights) for risk warning, taking into account the mutual influence or correlation between data collected from multiple sources of sensors. Compared with directly using the initial estimated risk for warning, this is more reliable and more suitable for the risk prediction needs when lifting basket-type steel box arches near operating lines. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the steps of a safety risk prediction method for the lifting process of a basket-type steel box arch near a service line, as provided in an embodiment of the present invention. Detailed Implementation

[0041] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the safety risk prediction method for the lifting process of a basket-type steel box arch near a business line proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0043] The following description, in conjunction with the accompanying drawings, details the specific scheme of the safety risk prediction method for the lifting process of a basket-type steel box arch near an operating line provided by the present invention.

[0044] Example 1:

[0045] Please see Figure 1 The diagram illustrates a flowchart of a safety risk prediction method for the lifting process of a basket-type steel box arch near a service line, according to an embodiment of the present invention. The method includes the following steps:

[0046] Step S101: Based on the changing trend of the data collected by the multi-source sensors during the lifting process, evaluate the estimated risk of each sensor. The estimated risks of all sensors are fused by preset weights to obtain the initial estimated risk.

[0047] When lifting basket-type steel box arches near operating lines, it is necessary to focus on supervising and managing the safety hazards during the lifting process. In this embodiment, multi-source sensors are used to reasonably predict or assess the risks of the lifting process, so as to achieve early warning of safety hazards.

[0048] Specifically, in this embodiment, strain sensors are attached to key sections of the arch rib, lifting points, and closure joints to collect stress data on important structures or parts. Other embodiments may install other sensors, such as vibration sensors on different parts of the arch rib, lifting equipment, and supports to detect vibrations during the lifting process; temperature and wind sensors on the steel box arch to monitor the environment during the lifting process; and settlement monitoring sensors on the bridge piles and the main bridge structure to detect the impact of the lifting process on the overall bridge structure.

[0049] In this embodiment, each sensor synchronously collects data every 0.2 seconds. All data collected by each sensor within a preset time period prior to the current time constitutes a time series; the estimated risk of each sensor is assessed based on the changing trend of this time series. The higher the estimated risk of each sensor, the more significant the change in the state of the sensor's location. A change in state represents instability and a more serious safety hazard. In this embodiment, the preset time period is set to 10 seconds.

[0050] As an example, the methods for assessing the predicted risk of each sensor based on the changing trend of this time series include:

[0051] For each time series acquired by each sensor at each time point, a Gaussian filter of length 5 is applied to the time series to obtain a filtered sequence. For any two adjacent elements in the filtered sequence, the difference between the element at the later time and the element at the earlier time is calculated. The ratio of this difference to the element at the earlier time is recorded as the difference between the two adjacent elements. The purpose of calculating the ratio with the element at the earlier time is to remove dimensions and orders of magnitude. Specifically, in this example, to avoid the denominator being 0 when calculating the ratio, a constant with a value of 0.1 is added to the denominator.

[0052] Furthermore, the mean of the differences between all adjacent elements in the filtered sequence is taken as the estimated risk for each sensor at each time.

[0053] As another example, assessing the estimated risk of each sensor based on the changing trend of the time series includes the following methods:

[0054] The time series is treated as a curve that changes over time. The curve is fitted to a straight line using the least squares method, and the slope of the straight line is used as the estimated risk.

[0055] Furthermore, each sensor is assigned a weight, and the sum of all sensor weights equals 1. The estimated risks of all sensors are then weighted and summed using the weights of each sensor, and the result is recorded as the initial estimated risk.

[0056] In this embodiment, an expert scoring method is used to assign a weight to each sensor. In other embodiments, the analytic hierarchy process can be used to assign weights. This embodiment does not limit the weight assignment method.

[0057] Thus, after each time interval, an initial risk estimate can be obtained for the current time. The initial risk estimate represents the overall assessment of the risk situation of the current lifting process after the data from all sensors are fused. The higher the initial risk estimate, the more serious the safety risks of the current lifting process.

[0058] In a comparative embodiment of the present invention, a safety warning is issued when the initial estimated risk is greater than a first risk threshold (e.g., 0.4).

[0059] Step S102: Train the risk prediction model and sensor fusion matrix to minimize the difference between the output of the risk prediction model and the initial risk prediction.

[0060] The method provided in the above comparison example has the following problems:

[0061] The data collected by different sensors describes the state changes of different structures or parts during the lifting process. Since the state changes at some sensor locations are influenced by the states at other sensor locations, the risk described by some sensor data includes or represents the risks present in the data from other sensors. In other words, when obtaining the initial estimated risk through weighted summation using assigned weights in the above steps, the estimated risks from some sensors may be considered repeatedly, leading to an overestimation of the initial risk. Alternatively, the actual risk described by all sensor data may be smaller than the initial estimated risk. Furthermore, since state changes at some sensor locations may induce further risks at other sensor locations (e.g., a state change at one location might cause an exponentially larger state change at other locations), the initial estimated risk obtained in the above steps is not reliable enough.

[0062] In summary, since the initial risk estimate obtained through the above steps does not take into account the mutual influence or correlation between data collected by different sensors, the method of directly using the above comparative examples for risk warning is unreliable and not suitable for the need for strict risk prediction when lifting basket-type steel box arches near operating lines.

[0063] A risk prediction model and a sensor fusion matrix are randomly initialized; in this embodiment, the initial values ​​of the parameters of the risk prediction model and the elements of the sensor fusion matrix are in the range of [0, 1].

[0064] Each element in the sensor fusion matrix represents the fusion coefficient of any two sensors. Two sensors are randomly selected, and the changing trend of the data collected by the two sensors is weakened. The weakening magnitude is positively correlated with the fusion coefficient of the two sensors in the sensor fusion matrix. Then, all the data collected by the sensors, including the data after weakening the changing trend, are input into the risk prediction model to update the risk prediction model and the sensor fusion matrix, so that the difference between the output result of the risk prediction model and the initial predicted risk is minimized.

[0065] In step S101, the changing trends of the data collected by each sensor are analyzed individually and then directly fused to obtain the initial estimated risk. In this step, the fusion coefficients in the sensor fusion matrix are used to weaken the changing trends of some sensor data, thereby reducing the role of some sensors in risk assessment. The output of the risk prediction model describes the risk assessment result after deep fusion of all sensor data. In this embodiment, even after reducing the role of some sensors in risk assessment, the risk assessment result output by the risk prediction model is still ensured to be as close as possible to the initial estimated risk obtained by directly fusing the data based on individual trend analysis in step S101; this allows the updated sensor fusion matrix to describe whether it is necessary to analyze and fuse different sensors together when assessing risk.

[0066] As an example, the changing trends of the data collected by the two sensors are weakened, and the weakening magnitude is positively correlated with the fusion coefficients corresponding to the two sensors in the sensor fusion matrix. The methods include:

[0067] For either of the two sensors, for the time series sequence collected by that sensor within a preset time period prior to the current time, the mean of all elements in the time series sequence is obtained and denoted as . The i-th element in this time series is denoted as .

[0068] It should be noted that the sensor fusion matrix in this embodiment is N×N in size, where N represents the total number of sensors. That is, each row or column of the sensor fusion matrix corresponds to one sensor, and each element of the sensor fusion matrix corresponds to any two sensors, representing the fusion coefficient of any two sensors. This sensor fusion matrix is ​​a diagonal matrix, and in this embodiment, the two elements symmetrical about the main diagonal are always identical.

[0069] All elements (i.e., all fusion coefficients) in the sensor fusion matrix are normalized. In this embodiment, the softmax formula is used for normalization. The normalized elements corresponding to the two sensors are represented as follows: .

[0070] After weakening the trend of the time series acquired by the sensor, the resulting time series is used as the corrected time series; the value of the i-th element in the corrected time series is represented as... , .

[0071] Thus, a time series sequence with weakened change trend of either of the two sensors is obtained, achieving the goal of weakening the change trend of the data collected by the two sensors.

[0072] in, This can also be seen as a reduction in magnitude. The larger the value, the closer the corrected time series sequence is to the mean. At this point, the time series after the trend of change is weakened has a less obvious trend of change compared to the time series before the trend of change is weakened (that is, the trend of change of the time series is greatly weakened), so that when the corrected time series corresponding to the two sensors are deeply fused through the risk prediction model, the role of the two sensors in risk assessment can be reduced.

[0073] As an example, data collected by all sensors, including data after weakening the trend of change, are input into the risk prediction model to update the risk prediction model and the sensor fusion matrix, so that the difference between the output of the risk prediction model and the initial risk prediction is minimized.

[0074] The risk prediction model in this embodiment uses an LSTM neural network structure.

[0075] Each sample is recorded as the time series sequence of all sensors obtained at the current time and every time before the current time, and the initial predicted risk obtained in step S101 is used as the label for each sample. The samples and labels obtained at the current time and all times before the current time during the improvement process are used as the dataset.

[0076] Each sample in the dataset is sequentially input into the risk prediction model. Note that before inputting each sample into the risk prediction model, two sensors are randomly selected from all the sensors included in each sample to weaken the trend of change, following the method described above. The input of the two randomly selected sensors is the corrected time series, while the input of the other sensors is the original time series.

[0077] The risk prediction model outputs a value, denoted as the risk fusion assessment value, which represents the risk assessment result after deep fusion of data collected from all sensors. In this case, the role of some sensors (i.e., two randomly selected sensors) in the risk assessment is weakened.

[0078] The loss function is obtained based on the mean squared error between the risk fusion assessment value and the initial estimated risk. In this embodiment, the mean squared error is used to represent the difference between the risk fusion assessment value and the initial estimated risk.

[0079] The risk prediction model is trained using all samples and loss functions in the dataset mentioned above. It should be noted that all elements (i.e. all fusion coefficients) in the sensor fusion matrix are also used as parameters in the risk prediction model for training and updating. The parameter update method adopts the stochastic gradient descent algorithm.

[0080] All samples in the dataset are input into the risk prediction model to complete one round of training. In this embodiment, the training process ends after 300 rounds. The parameters of the risk prediction model and the sensor fusion matrix are then updated.

[0081] Throughout the training process, two sensors are randomly selected each time to weaken the trend of change, and the loss function is optimized towards the minimum value throughout the training process (that is, the mean square error between the risk fusion assessment value and the initial predicted risk is minimized). When the training process ends, for a portion of the sensors corresponding to the larger fusion coefficients in the updated sensor fusion matrix, the trend of change of these sensors is greatly weakened, so that these sensors have a smaller role in risk assessment. That is, even without the risk assessment role of these sensors, the risk assessment result after deep fusion can still be close to the initial predicted risk obtained in step S101 when no trend weakening (or no sensor removal) is performed.

[0082] Therefore, the larger the fusion coefficient in the updated sensor fusion matrix, and the smaller the mean square error between the risk fusion assessment value and the initial estimated risk after minimizing the mean square error, the less useful the sensor corresponding to the larger fusion coefficient in the updated sensor fusion matrix is ​​for risk assessment. In other words, for a subset of sensors corresponding to fusion coefficients with larger values, since the state changes of the locations of other sensors are affected by the state of the locations of these sensors, the risk of the data collected by other sensors includes or represents the risk of the data collected by these sensors when describing the risk. Therefore, these sensors should not be overly involved in risk assessment.

[0083] Step S103: The estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the first estimated risk.

[0084] When the root mean squared error between the risk fusion assessment value and the initial predicted risk is minimized (i.e., at the end of the training process), the root mean squared error between the risk fusion assessment value and the initial predicted risk is denoted as f1. The weight assigned to any sensor is denoted as W. The element of a row (i.e., a row of fusion coefficients) in the updated sensor fusion matrix corresponding to that sensor is then calculated. The mean of all fusion coefficients in the corresponding row of the updated sensor fusion matrix is ​​denoted as f2. The ratio of f2 to f1 is denoted as the suppression magnitude of that sensor.

[0085] The greater the suppression magnitude, the less the sensor participates in the risk assessment (i.e., the larger f2 is). This also makes the risk assessment result after deep fusion approach the initial estimated risk obtained in step S101 when no trend weakening (or no sensor removal) is performed (i.e., the smaller f2 is). In particular, in some embodiments, to avoid the denominator being 0, the mean square difference between the risk fusion assessment value and the initial estimated risk is added by 1 and denoted as f1.

[0086] The weights assigned to the sensor are updated, and the updated weights are negatively correlated with the suppression magnitude. A larger suppression magnitude indicates that the weight of the corresponding sensor needs to be reduced to ensure that the sensor is not over-involved in risk assessment and to avoid the problem of some sensors' predicted risks being considered repeatedly.

[0087] As an example, updating the weights assigned to this sensor involves the following formula:

[0088] Let W1 = W × (1 - f0), where W1 represents the first weight of the sensor and f0 represents the suppression amplitude of the sensor.

[0089] The initial weights of all sensors are normalized using the softmax formula. The normalized initial weight of a sensor is used as the updated weight of that sensor and is also denoted as the updated weight. In some embodiments, when f0 is less than 0.3, f0 is set to 0, indicating that the weights of sensors with insufficient suppression are not updated.

[0090] Furthermore, the estimated risks of all sensors are weighted and summed using the updated weights of each sensor, and the result is denoted as the first estimated risk.

[0091] The process of obtaining the first estimated risk avoids to some extent the following problem described in step S101: Since the state changes of some sensor locations are affected by the state of other sensor locations, when describing the risk, the risk of some sensor data includes or represents the risk characteristics of other sensor data. In this case, it is equivalent to the estimated risk of some sensors being considered repeatedly, which leads to the unreliability of the risk warning results of the comparative embodiment.

[0092] Step S104: Continue to update the risk prediction model and sensor fusion matrix to maximize the difference between the risk prediction model output and the first initial risk prediction.

[0093] This embodiment further considers the risk that changes in the state of some sensor locations may induce other sensor locations to update one step (for example, there may be a situation where a change in the state of one location causes an exponential change in the state of other locations). This means that the first estimated risk obtained in the above steps still cannot reliably be used for risk assessment of the improvement process.

[0094] In this embodiment, the risk prediction model and sensor fusion matrix are trained again following the method in step S102. That is, based on the trained risk prediction model and sensor fusion matrix obtained after the training process in step S102, the parameters of the trained risk prediction model and sensor fusion matrix are fine-tuned and updated. The difference from step S102 is as follows:

[0095] On the one hand, in this step, all elements in the sensor fusion matrix are normalized using softmax, and the elements (i.e., fusion coefficients) with normalized values ​​greater than the first preset threshold th1 are kept fixed (i.e., no parameter fine-tuning is performed). The purpose is that sensor combinations with larger fusion coefficients do not need to participate excessively in risk assessment (or in other words, the risk assessment process of other sensor combinations includes the risk assessment process of sensor combinations with larger fusion coefficients). This embodiment does not need to further consider the further induced risk of sensor combinations with larger fusion coefficients to other sensors. This embodiment uses th1=0.7 as an example for description.

[0096] On the other hand, the loss function used in this step is obtained by the difference between the output of the risk prediction model and the first initial predicted risk. The purpose is that, considering that the changing trends of the sensors are still randomly weakened each time training continues in this step to reduce the role of some sensors in deep risk assessment, this embodiment obtains the loss function by using the difference between the output of the risk prediction model and the first initial predicted risk. This ensures that after the training process continues, the output of the risk prediction model is as large as possible compared to the first initial predicted risk (i.e., the difference between the output of the risk prediction model and the first initial predicted risk is as large as possible). This allows some sensor combinations (sensor combinations with larger fusion coefficients) in the sensor fusion matrix obtained after the training process to have small state changes (i.e., their changing trends are significantly weakened), inducing other sensor combinations (i.e., sensor combinations with smaller fusion coefficients) to have larger state changes (i.e., their changing trends are not weakened or are weakened less), ultimately causing the risk prediction model to output a larger risk assessment result (i.e., the risk prediction model output result is greater than the first initial predicted risk).

[0097] In summary, the above process ensures that the sensor fusion matrix obtained after the training process can describe the risk that changes in the state of some sensor locations may induce updates in the positions of other sensors.

[0098] Step S105: The estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the second estimated risk; the second estimated risk is used for risk assessment.

[0099] For the sensor fusion matrix after the training process continues in step S104 (i.e., the sensor fusion matrix after the update), after the training process continues in step S104 (i.e., when the difference between the risk prediction model output and the first initial predicted risk is the largest), the difference between the risk prediction model output and the first initial predicted risk is denoted as f3.

[0100] For any given sensor, assign a weight W. Then, in the updated sensor fusion matrix, find the element in one row (i.e., one row of fusion coefficients) corresponding to that sensor. Take the mean of all fusion coefficients in the row corresponding to that sensor in the updated sensor fusion matrix, denoted as f4. The product of f3 and f4 is then taken as the gain amplitude of that sensor.

[0101] A larger gain indicates that when the sensor participates in deep risk assessment with a smaller trend of change, it can induce a larger overall risk assessment result (i.e., a larger f3) for all sensors. The smaller trend of change mentioned above means that when f4 is larger, the trend of change for the sensor is significantly weakened, thus making it exhibit a smaller trend of change.

[0102] The weights assigned to the sensor are further updated. The updated weights of the sensor are positively or negatively correlated with the gain amplitude, indicating that the greater the gain amplitude, the more attention needs to be paid to the risk assessment results of the sensor itself.

[0103] As an example, the weights assigned to this sensor are further updated using the following formula:

[0104] Let W2 = W1 × (1 + g0), where W2 represents the second weight of the sensor; W1 represents the first weight of the sensor; and g0 represents the gain amplitude of the sensor.

[0105] The second weights of all sensors are normalized using the softmax formula. The normalized second weight of a sensor is used as a further update of the sensor's weights and is also denoted as the final weight.

[0106] Furthermore, the estimated risks of all sensors are weighted and summed using the final weight of each sensor, and the result is denoted as the final estimated risk.

[0107] In this embodiment, a safety warning is issued when the final estimated risk exceeds the second risk threshold. This embodiment is described using an example where the second risk threshold is equal to 0.4.

[0108] It should be noted that in this embodiment, by adjusting and updating the weights allocated in step S101, the final risk prediction takes into account the relationship between data collected by different sensors. Specifically, this includes the following two situations: one situation is that the state changes of some sensor locations are affected by the state of other sensor locations, so that the risk of data collected by some sensors includes or represents the risk of data collected by other sensors; the other situation is that the state changes of some sensor locations may induce an update of the risk of other sensor locations. Ultimately, this embodiment is more reliable than the comparative embodiment provided in step S101 and is more suitable for the risk prediction needs when lifting basket-type steel box arches near operating lines.

[0109] Example 2:

[0110] As an example, in step S102, the loss function is obtained based on the mean squared error between the risk fusion assessment value and the initial estimated risk. This includes the following methods:

[0111] Let this loss function be denoted as P1; .

[0112] Where D1 represents the risk fusion assessment value, and D2 represents the initial estimated risk. This represents the mean squared error between the risk fusion assessment value and the initial estimated risk.

[0113] When training using this loss function, denoted as P1, the loss function will optimize towards the minimum value. When optimizing towards the minimum value, This ensures that D1 approaches D2 during training. (Introduction) The goal is to minimize D1 while ensuring it approaches D2. In other words, the loss function P1 set in this embodiment ensures that D1 is less than and as close as possible to D2, so that the deep fusion evaluation result (i.e., D1) output by the risk prediction model after weakening the trend of change in the sensor-collected data is as small as possible compared to the initial predicted risk obtained in step S101 without trend weakening, thereby ensuring the accuracy of the training process.

[0114] Specifically, in step S103 of Embodiment 1, at the end of the training process, if the risk prediction model output D1 is greater than D2, then the mean square error between the risk fusion evaluation value described in step S103 and the initial predicted risk is set to 0. The purpose is that when the risk prediction model output D1 is greater than D2, it indicates that after weakening the sensor's changing trend, the risk fusion evaluation value is greater than the initial predicted risk without weakening in step S101. This indicates that the necessity for some sensors to participate in risk assessment is almost negligible. In this case, f1 is minimized to ensure the maximum reduction in sensor weights.

[0115] As an example, in step S104, the loss function used is obtained from the difference between the output of the risk prediction model and the first initial predicted risk, and includes the following formula:

[0116] Let this loss function be denoted as P2; Where D3 represents the output of the risk prediction model; p0 represents the difference between the output of the risk prediction model and the first initial risk prediction. D4 represents the first initial estimated risk; exp() represents an exponential function with the natural constant as the base.

[0117] When training with this loss function P2, P2 will optimize towards its minimum value. During optimization towards the minimum value, exp(-p0) is also minimized, making... The goal is to maximize the value of the result, thereby maximizing the difference between the risk prediction model output and the initial risk prediction. This involves introducing... The goal is to increase the value of D3 when the loss function P2 is minimized. Overall, when the loss function P2 is minimized, D3 is guaranteed to be greater than D4, and the difference between D3 and D4 is maximized. This ensures that when the risk prediction model is used as input with a smaller trend of change (i.e., after the trend of sensor change is weakened), the output risk assessment result is maximized. This ensures that when some combinations of sensors in the sensor fusion matrix (sensor combinations with larger fusion coefficients) have small state changes (i.e., after their trend of change is weakened significantly), other combinations of sensors (i.e., sensor combinations with smaller fusion coefficients) still have large state changes (i.e., their trend of change is not weakened or is weakened less). This ensures that the sensor fusion matrix obtained after the training process can describe whether the state change of the position of some sensors may induce the position of other sensors to update one step.

[0118] Specifically, in step S105 of Embodiment 1, at the end of the training process, if the output result of the risk prediction model (i.e., D3) is less than D4, f3 is set to 0. The purpose is that in step S105, the gain amplitude is set to 0, indicating that it is impossible to further improve the risk assessment result output by all sensors after fusion of the risk prediction model by weakening the changing trend of the sensors. At this time, the weights of the sensors are not updated further.

[0119] Example 3:

[0120] This embodiment takes into account that, in the process of Embodiment 1, a sufficient number of samples and time are needed to train or update the risk prediction model and the sensor fusion matrix. This embodiment also takes into account that, in the early stage of the lifting process (e.g., before lifting to half of the expected height), the risk to adjacent operating lines is relatively small, and overly strict risk supervision and control are not required. That is, before the steel box arch is lifted to half of the expected height, this embodiment uses the comparative embodiment provided in step S101 to conduct risk warnings at each time point.

[0121] Furthermore, when the steel box arch is raised to half of the expected height, all the steps in Embodiment 1 above are executed to update the weight of each sensor and obtain the final weight.

[0122] After obtaining the final weights, and once the steel box arch has been raised to half its expected height, further raising it could pose a significant safety hazard to adjacent operating lines if it continues. Therefore, strict risk monitoring and control are necessary. Specifically, after the steel box arch has reached half its expected height, the estimated risks from all sensors at each time point are weighted and summed using the final weights of each sensor. The result is recorded as the final estimated risk for each time point. A safety warning is issued when the final estimated risk exceeds a second risk threshold.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating railway line, characterized in that, The method includes the following steps: The estimated risk of each sensor is assessed based on the changing trend of data collected by multiple sensors during the lifting process. The estimated risks of all sensors are fused by preset weights to obtain the initial estimated risk. The multiple sensors include: strain sensor, vibration sensor, temperature sensor and wind force sensor. Initialize a risk prediction model and a sensor fusion matrix. Each element in the sensor fusion matrix represents the fusion coefficient of any two sensors. Randomly select two sensors and first weaken the trend of change in the data collected by the two sensors. The magnitude of the weakening is positively correlated with the fusion coefficients of the two sensors in the sensor fusion matrix. Then, input all the data collected by the sensors, including the data with weakened trend, into the risk prediction model to update the risk prediction model and the sensor fusion matrix, so that the difference between the output of the risk prediction model and the initial risk prediction is minimized. The estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the first estimated risk; The risk prediction model and sensor fusion matrix are updated again to maximize the difference between the risk prediction model output and the first initial risk prediction; the predicted risks of all sensors are re-fused through the updated sensor fusion matrix to obtain the final predicted risk; the final predicted risk is used for risk warning. Specifically, the trend of change in the data collected by the two sensors is weakened, and the degree of weakening is positively correlated with the fusion coefficients corresponding to the two sensors in the sensor fusion matrix. This includes: for any one of the two sensors, and for the time series formed by the data collected by that sensor, obtaining the mean of all elements in the time series, denoted as . The i-th element in this time series is denoted as Normalize all fusion coefficients in the sensor fusion matrix, and use the normalized fusion coefficients corresponding to the two sensors as the attenuation magnitude. The time series obtained after weakening the trend of the data collected by the sensor is used as the corrected time series; the value of the i-th element in the corrected time series is represented as... , ; The updated risk prediction model includes: incorporating all fusion coefficients in the sensor fusion matrix as part of the parameters of the risk prediction model during training and updating; The method for obtaining the first estimated risk is as follows: when the difference between the output of the risk prediction model and the initial estimated risk is minimal, the difference between the output of the risk prediction model and the initial estimated risk is recorded as f1. For any sensor, the fusion coefficient of a row in the updated sensor fusion matrix is ​​recorded as f2. The average of all fusion coefficients in the row corresponding to that sensor in the updated sensor fusion matrix is ​​recorded as f2. The preset weight of the sensor is updated using f2 and f1 to obtain the first weight of the sensor. The first weights of all sensors are normalized. The estimated risks of all sensors are weighted and summed using the normalized first weights of each sensor. The result is recorded as the first estimated risk.

2. The method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating line as described in claim 1, characterized in that, The specific steps involved in inputting all sensor-collected data, including data after mitigating the changing trend, into the risk prediction model, updating the risk prediction model and the sensor fusion matrix, and minimizing the difference between the risk prediction model output and the initial risk prediction, are as follows: The data collected by multi-source sensors at each time point are used as each sample, and the initial estimated risk is used as the label for each sample; the samples and labels obtained at all times are used as the dataset; each sample in the dataset is then input into the risk prediction model in sequence. The two randomly selected sensors input data with weakened trend of change. The parameters of the risk prediction model are trained using all samples in the dataset and the loss function; the loss function is set as: the loss function P1 is composed of the mean square error of the risk prediction model output and the initial predicted risk; the parameters of the risk prediction model include all fusion coefficients in the sensor fusion matrix. After training, the risk prediction model and sensor fusion matrix are updated to minimize the difference between the risk prediction model output and the initial risk prediction.

3. The method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating line as described in claim 1, characterized in that, The specific steps involved in updating the risk prediction model and sensor fusion matrix again to maximize the difference between the risk prediction model output and the first initial risk prediction are as follows: The parameters of the risk prediction model are trained using all samples and loss functions in the dataset. The loss function is reset to P2, which is the difference between the output of the risk prediction model and the first initial risk prediction. After training, the risk prediction model and the sensor fusion matrix are updated again to maximize the difference between the output of the risk prediction model and the first initial risk prediction.

4. The safety risk prediction method for the lifting process of a basket-type steel box arch near an operating line according to claim 2, characterized in that, The estimated risks of all sensors are re-fused using the updated sensor fusion matrix to obtain the final estimated risk. The specific steps include the following: When the difference between the risk prediction model output and the first initial risk prediction is the largest, the difference between the risk prediction model output and the first initial risk prediction is recorded as f3; the mean of all fusion coefficients in the corresponding row of any sensor is obtained from the updated sensor fusion matrix and recorded as f4; the first weight of each sensor is updated using f3 and f4 to obtain the second weight of each sensor. The second weights of all sensors are normalized, and the predicted risks of all sensors are weighted and summed using the normalized second weights of each sensor. The result is recorded as the final predicted risk.

5. The safety risk prediction method for the lifting process of a basket-type steel box arch near an operating line according to claim 2, characterized in that, The process of updating the preset weights of the sensor using f2 and f1 to obtain the first weight of the sensor includes the following specific formulas: Let W1 = W × (1 - f0), where W1 represents the first weight of the sensor; f0 represents the suppression amplitude of the sensor, which is equal to the ratio of f2 to f1; and W represents the preset weight of the sensor.

6. The method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating line according to claim 4, characterized in that, The process of updating the first weight of each sensor using f3 and f4 to obtain the second weight of each sensor includes the following specific formulas: Let W2 = W1 × (1 + g0), where W2 represents the second weight of the sensor; W1 represents the first weight of the sensor; and g0 represents the gain amplitude of the sensor, which is equal to the product of f3 and f4.

7. The method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating line according to claim 1, characterized in that, The specific steps for assessing the estimated risk of each sensor based on the changing trends of data collected by multiple sensors during the lifting process are as follows: For each sensor, the data collected forms a time series, which is a curve that changes over time. The curve is fitted to a straight line using the least squares method, and the slope of the straight line is used as the estimated risk.

8. The method for predicting safety risks during the lifting process of a basket-type steel box arch near an operating line according to claim 1, characterized in that, The specific steps involved in using the final estimated risk for risk warning are as follows: When the steel box arch is lifted to a height less than the preset height, risk warning is given using the initial risk estimate. When the steel box arch is lifted to a height greater than or equal to the preset height, risk warning is given using the final estimated risk. The timing for issuing a risk warning refers to when the initial or final estimated risk exceeds a preset risk threshold.

Images