CN-121823397-B - Remote operation safety behavior evaluation method of quayside container crane and electronic equipment

CN121823397BCN 121823397 BCN121823397 BCN 121823397BCN-121823397-B

Abstract

The invention relates to a remote operation safety behavior evaluation method and electronic equipment of a quayside container crane, which comprise the steps of obtaining an operation action sequence and an operation environment state track of a driver to be evaluated, constructing a multi-mode time sequence data set, calling a safety behavior reference model obtained through training in advance, calculating to obtain a behavior normalization index of the operation track under an expert strategy, encoding the operation action sequence, quantifying space risk sensitivity and time criticality of different operation stages, constructing a dynamic safety potential field model, calculating repulsive force potential energy between a lifting appliance and a lifting weight and an environment barrier in real time, generating a track safety margin index, and fusing the behavior normalization index, the space risk sensitivity and the time criticality and the track safety margin index to generate a safety behavior comprehensive evaluation result of the driver to be evaluated. The invention can automatically learn the safety intention of an expert, has the capability of differentially evaluating the time-space key characteristics of the operation behaviors, and can effectively fuse the safety constraint of the physical world.

Inventors

HE MENGJIE
ZHU SHIJIE
ZHANG YUJIE
FU CHAO
SHEN YANG
HU JIA
LAI JINTAO
HAO YANGYANG
YANG MAN

Assignees

上海海事大学
宁波北仑第三集装箱码头有限公司
同济大学

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (8)

1. A method for evaluating remote operation safety behavior of a quayside container crane, the method comprising: Acquiring an operation action sequence of a driver to be evaluated and a corresponding operation environment state track to construct a multi-mode time sequence data set; based on a maximum entropy inverse reinforcement learning framework, invoking a safety behavior reference model which is obtained through training of a history operation track of a skilled driver in advance by utilizing the multi-mode time sequence data set so as to calculate and obtain a behavior normalization index of the operation track of the driver to be evaluated under an expert strategy; Encoding the operation action sequence by utilizing a space-time attention mechanism so as to extract and quantify the spatial risk sensitivity and the time criticality corresponding to different operation stages; Constructing a dynamic safety potential field model, and calculating repulsive potential energy between a lifting appliance and a lifting weight and an environmental obstacle in real time based on the operation environment state track to generate a track safety margin index; fusing the behavior normalization index, the space risk sensitivity and time criticality and the track safety margin index to generate a comprehensive safety behavior evaluation result of a driver to be evaluated; the training process of the safety behavior benchmark model comprises the following steps: abstracting a skilled driver history operation track into a Markov decision process, and defining a bottom layer rewarding function of the safety behavior benchmark model as a linear combination of state characteristics, wherein the method is expressed as follows: Wherein, the As a feature weight vector of the model, Feature mapping functions for state-action pairs; Through maximum likelihood estimation, the behavior characteristic expectation generated by the model is enabled to be consistent with the behavior characteristic expectation of the history operation track of the skilled driver, and the characteristic weight vector for representing expert operation logic is obtained through inversion ; The calculation is carried out to obtain the behavior normalization index of the operation track of the driver to be evaluated under the expert strategy, specifically: according to the characteristic weight vector Calculating the operation track of the driver to be evaluated Log likelihood probability under expert policy distribution As an index of the behavior normalization.
2. The method for evaluating the remote operation safety behavior of the shore container crane according to claim 1, wherein the step of obtaining the operation action sequence of the driver to be evaluated and the corresponding operation environment state track comprises the following steps: and synchronously acquiring an operation action sequence comprising the opening degree of a handle, the state of a button and the action duration time, and an operation environment state track comprising the space coordinates of a lifting appliance, the lifting speed, the displacement of a trolley, the swing angle of the lifting appliance and the real-time distance between the lifting appliance and a nearest barrier.
3. The method for evaluating the remote operation safety behavior of a shore container crane according to claim 1, wherein said encoding said sequence of operation actions by means of a spatiotemporal attention mechanism comprises: Encoding the operation action sequence by adopting a two-way long-short-term memory network, and extracting time sequence characteristics; And coupling a space-time attention mechanism at an output layer of the two-way long-short-term memory network, and dynamically calculating attention weights at different sampling moments.
4. The method for evaluating the remote operation safety behavior of a quay container crane according to claim 3, wherein the spatiotemporal attention mechanism is used for identifying key operation transients with high contribution to safety risk in the time dimension, calculating the attention weight The method of (1) comprises: Wherein, the In order to be the hidden state at the instant t, 、 And In order for the parameters to be able to be learned, To correspond to the first Unnormalized attention scores for individual moments.
5. The method for evaluating the remote operation safety behavior of the shore container crane according to claim 3 or 4, wherein the steps of extracting and quantifying the spatial risk sensitivity and the time criticality corresponding to different working phases comprise: based on the attention weight Generating a comprehensive vector embodying key risk stage characteristics , A hidden state at time t; and calculating the time criticality according to the comprehensive vector and the distribution of the attention weight in a high sensitivity stage, and calculating the risk entropy value distribution in different coordinate areas by carrying out dynamic meshing division on a working area so as to quantify the space risk sensitivity.
6. The method for evaluating the remote operation safety behavior of the shore container crane according to claim 1, wherein the constructing the dynamic safety potential field model, calculating repulsive potential energy between the lifting appliance and the lifting weight and the environmental obstacle in real time based on the operation environment state track, comprises: Establishing an exponential repulsive force field decaying along with the distance by taking the center of an obstacle as a potential source, wherein the repulsive force potential energy Expressed as: Wherein, the For the real-time distance between the sling and the obstacle, For the normal velocity component of the spreader near the obstacle, 、、 In order to scale the coefficient of the power consumption, The maximum influence distance of the potential field; integrating along the operation track of the driver to be evaluated based on the repulsive force potential energy to obtain a track safety margin index Or a functional transformation thereof.
7. The method for evaluating the safety behavior of the remote operation of the shore container crane according to claim 1, wherein the step of generating the safety behavior comprehensive evaluation result by fusing the behavior normalization index, the spatial risk sensitivity and the time criticality, and the trajectory safety margin index comprises the steps of: Normalizing the behavior index Critical phase stability index calculated based on the spatial risk sensitivity and time criticality Physical safety redundancy index calculated based on the track safety margin index Carrying out weighted fusion; the comprehensive evaluation result of the safety behavior Is calculated by the following formula: Wherein, the 、、 And for a preset fusion weight coefficient, norm () is a normalization function.
8. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the method for evaluating the remote operation safety behavior of a quay container crane according to any one of claims 1 to 7.

Description

Remote operation safety behavior evaluation method of quayside container crane and electronic equipment Technical Field The invention relates to the technical field of intelligent port and crane safety monitoring, in particular to a remote operation safety behavior evaluation method of a quay container crane and electronic equipment. Background In the construction of intelligent ports, remote control of a shore container crane has become a mainstream operation mode, and scientific and accurate evaluation of safety behaviors of a remote operation driver is a key link for guaranteeing the safety and efficiency of port operation. Currently, a behavior evaluation method based on deep reinforcement learning is considered as an important technical direction in the field. However, the existing methods have at least three technical defects related to each other in practical application, which restrict the accuracy and practicality of the evaluation: First, the construction of the evaluation criterion relies heavily on human experience. The existing method generally requires engineers to preset specific weights of various rewards and punishments rules (such as collision punishment and speed rewards) so as to form a rewarding function for learning by an intelligent agent. Because the crane operating environment is complex and changeable, the safety operation logic relates to global optimization of a long time sequence action strategy, the manually defined and static weight combination is difficult to accurately and completely describe the implicit and dynamic optimal safety decision logic followed by excellent operators, and inherent deviation exists between the inherent standard of the learned evaluation model and the actual optimal safety behavior. Second, the evaluation process lacks a fine sense of temporal and spatial heterogeneity of the operational sequences. One complete working cycle of the crane comprises several phases (e.g. moving, handling boxes, passing on board), the sensitivity of the different phases to safety risks is quite different and the risk distribution of different areas within the working space is also different. The existing methods generally treat continuous operation sequences as homogeneous time steps, and cannot identify and focus on critical timing segments and highly sensitive spatial regions that have decisive influence on the overall safety, so that it is difficult to deeply evaluate the suitability of the operator for behavior in a specific high-risk context. Furthermore, the assessment model is not well-coupled with the safety boundary constraints of the physical world. The pure data driven model mainly relies on historical data for learning, and when the model faces extreme or sporadic working conditions which are not fully covered in training data, the model can be judged against a physical rule due to lack of basic physical common knowledge (such as the relation of the impenetrability of an object and the speed and the braking distance), so that an evaluation suggestion fails under the boundary condition, and the generalization capability and the reliability are insufficient. Disclosure of Invention In view of the above, it is necessary to provide a method, an apparatus and an electronic device for evaluating remote operation safety behavior of a quayside container crane, which can automatically learn expert safety intention, has capability of performing differential evaluation on time-space key characteristics of operation behavior, and can effectively integrate physical world safety constraints. The invention provides a method for evaluating remote operation safety behavior of a quayside container crane, which comprises the following steps: Acquiring an operation action sequence of a driver to be evaluated and a corresponding operation environment state track to construct a multi-mode time sequence data set; based on a maximum entropy inverse reinforcement learning framework, invoking a safety behavior reference model which is obtained through training of a history operation track of a skilled driver in advance by utilizing the multi-mode time sequence data set so as to calculate and obtain a behavior normalization index of the operation track of the driver to be evaluated under an expert strategy; Encoding the operation action sequence by utilizing a space-time attention mechanism so as to extract and quantify the spatial risk sensitivity and the time criticality corresponding to different operation stages; Constructing a dynamic safety potential field model, and calculating repulsive potential energy between a lifting appliance and a lifting weight and an environmental obstacle in real time based on the operation environment state track to generate a track safety margin index; And fusing the behavior normalization index, the space risk sensitivity and the time criticality and the track safety margin index to generate a comprehensive safety behavior evaluation result of the