Forklift self-adaptive safety early warning decision-making method for multiple scenes

Abstract

The invention relates to the technical field of forklift risk management, in particular to a multi-scene-oriented forklift self-adaptive safety early warning decision-making method, which comprises the following steps of: through multi-scene dynamic modeling, constructing multi-dimensional scene feature vectors according to forklift operation environment spatial features, cargo physical attributes and driver behavior parameters; establishing a scene classification model by means of transfer learning; a multi-modal sensor data fusion technology is utilized, a heterogeneous sensor network is deployed to collect data, and data reliability is optimized through space-time alignment and confidence evaluation; a dynamic safety threshold value is generated through fuzzy logic and reinforcement learning, and a driver behavior correction factor is introduced for real-time adjustment; calculating a comprehensive risk index by using a dynamic weight distribution algorithm, and triggering graded early warning and response; a federated learning closed-loop optimization risk assessment model is adopted, robustness is verified in combination with digital twinning, and cooperative obstacle avoidance of multiple forklifts is achieved through inter-vehicle communication. According to the method, the operation safety and decision-making accuracy of the forklift in a complex and changeable scene are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of forklift risk management, and specifically provides a multi-scenario adaptive safety warning and decision-making method for forklifts. Background Technique

[0002] In the fields of modern logistics warehousing and industrial production, forklifts, as key material handling equipment, their operation safety is of crucial importance. The traditional forklift operation environment is complex and changeable, involving various scenarios such as indoor warehouses, factory workshops, and outdoor freight yards. There are significant differences in spatial layout, cargo characteristics, and personnel flow among different scenarios.

[0003] Early forklift safety warnings mainly relied on simple hardware protection devices such as bumpers and reverse alarms. Such devices could only make passive responses to limited risks and could not comprehensively perceive potential dangers in complex scenarios. With the development of technology, some systems have introduced single sensors, such as lidar for detecting obstacles. However, due to the diversity of forklift operation scenarios, the information obtained by a single sensor is limited, making it difficult to accurately assess the overall risk.

[0004] In terms of multi-scenario adaptation, previous methods mostly adopted fixed-parameter models and could not adjust safety strategies according to the dynamic changes of the environment, cargo, and driver behavior. For example, under different warehouse layouts, fixed safety distance settings may be too conservative and affect efficiency when operating in narrow aisles, or may not be sufficient to handle emergencies in open areas.

[0005] The application of data fusion technology in the field of forklift safety also faced challenges in its initial stage. The multi-source sensor data had inconsistent time and space references, making it difficult to fuse and resulting in low data reliability, which was difficult to support accurate safety decisions. In addition, traditional systems lacked effective cooperative control mechanisms. When multiple forklifts operated in the same area, collisions and other accidents were likely to occur due to poor information communication.

[0006] Therefore, a multi-scenario adaptive safety warning and decision-making method for forklifts is proposed to address the above problems. Summary of the Invention

[0007] The purpose of the present invention is to provide a multi-scenario adaptive safety warning and decision-making method for forklifts to solve the problems raised in the above background technique.

[0008] To achieve the above purpose, the present invention provides the following technical solutions:

[0009] A multi-scenario adaptive safety warning and decision-making method for forklifts, comprising the following steps:

[0010] Step S1: Multi-scenario dynamic modeling:

[0011] Based on the spatial characteristics of the forklift operating environment, the physical properties of the goods, and the driver's behavior parameters, construct a multi-dimensional scene feature vector;

[0012] Through transfer learning, map the scene features in the historical accident database to the real-time operating environment and establish a scene classification model;

[0013] Step S2: Multi-modal sensor data fusion:

[0014] Deploy a heterogeneous sensor network, including 3D lidar, vision sensors, pressure sensor arrays, and inertial measurement units (IMUs), to collect forklift operation data in real time;

[0015] Use a spatio-temporal alignment algorithm to fuse multi-source data, and optimize data reliability through a confidence evaluation module;

[0016] Step S3: Generation of dynamic safety thresholds:

[0017] Based on the scene feature vector, combine fuzzy logic and reinforcement learning algorithms to generate dynamic safety thresholds;

[0018] Introduce a driver behavior correction factor to adjust the safety threshold matrix in real time;

[0019] Step S4: Comprehensive risk decision-making and hierarchical response:

[0020] Calculate the comprehensive risk index through a dynamic weight allocation algorithm and trigger a hierarchical early warning strategy;

[0021] Execute multi-level response measures from prompts to emergency braking according to the risk level;

[0022] Step S5: Closed-loop optimization and cooperative control:

[0023] Feed the early warning results back to the model and update the parameters using an online learning algorithm;

[0024] Achieve cooperative obstacle avoidance for multiple forklifts through an inter-vehicle communication module, and deploy edge computing nodes to ensure real-time decision-making.

[0025] As an optimal solution, the multi-dimensional scene feature vector in step S1 includes:

[0026] Environmental complexity coefficient: Calculate the sum of multiplying the ratio of obstacle density to operating area by the first weight coefficient, adding the absolute value of the change rate of ground friction coefficient over time multiplied by the second weight coefficient, and multiplying the light intensity gradient by the third weight coefficient;

[0027] Load stability index: According to the linear combination of the variance of the fork pressure distribution and the height of the center of gravity of the goods, convert it into a stability score through an inverse proportional function, where the larger the variance and the height of the center of gravity, the lower the stability score;

[0028] Operating behavior risk value: A hyperbolic tangent function weighted value based on the steering angle change rate, superimposed with a linear contribution value of the rapid acceleration event frequency, comprehensively reflecting the driver's operation risk.

[0029] As a preferred solution, the method for generating the dynamic safety threshold in step S3 includes:

[0030] Safety distance threshold: According to the fuzzy logic rule base, the typical safety distances in the historical accident data are weighted and averaged, and the weights are the fuzzy membership degrees in each scenario;

[0031] Maximum allowable inclination threshold: The inclination value that maximizes the state-action value function is selected through the reinforcement learning algorithm, and this function is trained by historical operation data.

[0032] As a preferred solution, the dynamic weight allocation algorithm in step S4 is:

[0033] Based on the attention mechanism, the environmental complexity coefficient, load stability index, and operating behavior risk value are input into the multi-layer perceptron, and the attention scores of each dimension are output, and the weights are calculated through the normalization exponential function;

[0034] The comprehensive risk index is composed of the weighted sum of the weights of each dimension and their corresponding eigenvalue. When this index exceeds 0.7, a secondary or higher-level response is triggered.

[0035] As a preferred solution, the closed-loop optimization in step S5 includes:

[0036] Adopt the federated learning framework, aggregate the local model parameters of each edge node proportionally according to the data volume, and update the global model parameters;

[0037] Simulate extreme scenarios through digital twin technology to verify the robustness of the model.

[0038] As a preferred solution, the multi-level response measures are executed according to the hierarchical response strategy, and the hierarchical response strategy includes:

[0039] Primary response: Display the dynamic risk heat map through the HUD interface and give voice prompts to the driver;

[0040] Secondary response: Limit the forklift power output to 50% of the rated value and activate the automatic fork leveling function;

[0041] Tertiary response: Cut off the power transmission, start the electronic parking system, and record the accident prediction data through the blockchain.

[0042] As a preferred solution, the heterogeneous sensor network meets the following performance indicators:

[0043] The 3D lidar scanning frequency is not less than 20 Hz, and the angular resolution is not greater than 0.1 degree;

[0044] The sampling accuracy of the pressure sensor array is not inferior to 0.2% of the full scale;

[0045] The attitude angle error of the 9-axis IMU module does not exceed 0.3 degree, and the data output delay does not exceed 5 milliseconds.

[0046] As a preferred solution, the response delay of the edge computing node does not exceed 30 milliseconds, and it supports the following functions:

[0047] Real-time path planning algorithm;

[0048] Localized model inference;

[0049] Secure data synchronization with the cloud platform.

[0050] As can be seen from the technical solutions provided by the present invention above, a forklift adaptive safety warning and decision-making method for multiple scenarios provided by the present invention has the following beneficial effects:

[0051] Scene adaptability and accurate decision-making:

[0052] Through multi-scenario dynamic modeling, the present invention accurately depicts the operation scenario using multi-dimensional scene feature vectors; it can quickly and accurately identify various complex scenarios based on the spatial layout of the forklift operation environment, the physical characteristics of the goods, and the driver's operating habits, providing a solid foundation for subsequent safety decisions, avoiding decision-making mistakes caused by scene misjudgment, and greatly improving the adaptability and decision-making accuracy of the system under different working conditions;

[0053] The application of transfer learning closely connects the historical accident database with the real-time operation environment; it not only efficiently utilizes the valuable experience in historical data, reduces the cost of data collection and model training in new scenarios, but also enables the scene classification model to quickly adapt to the new environment and give practical scene classification results in a timely manner, facilitating accurate decision-making;

[0054] Data fusion and reliability guarantee:

[0055] The multi-modal sensor data fusion technology comprehensively collects forklift operation data through a heterogeneous sensor network; the 3D lidar, vision sensor, pressure sensor array, and inertial measurement unit (IMU) work together to obtain information from different angles, presenting the operation state of the forklift and the surrounding environment conditions in all directions, providing rich and accurate data support for safety analysis;

[0056] The combination of the spatio-temporal alignment algorithm and the confidence evaluation module effectively eliminates the differences in time and space of multi-source data, removes the interference of noise and incorrect data, and significantly improves the reliability and consistency of the data; it ensures that subsequent safety warnings and decisions based on these data are built on high-quality information, reducing the risks of false alarms and missed alarms;

[0057] Dynamic thresholds and flexible responses:

[0058] The dynamic safety threshold generation mechanism uses fuzzy logic and reinforcement learning algorithms to dynamically adjust the safety threshold according to real-time scene characteristics; in the complex and ever-changing forklift operation scenarios, it can flexibly adapt to environmental changes. For example, in special working conditions such as narrow channels and heavy-load transportation, it can timely and reasonably adjust thresholds such as safety distances and maximum allowable tilting angles, making the safety standards more in line with actual operation requirements and enhancing the system's ability to handle complex situations;

[0059] Introduce a driver behavior correction factor, fully considering the impact of the driver's operation behavior on the safety threshold; if the driver's operation is relatively aggressive, the system automatically increases the safety threshold to strengthen safety precautions; if the operation is stable, the threshold is appropriately optimized to improve operation efficiency, achieving a balance between safety and efficiency;

[0060] High efficiency of risk decision-making and hierarchical response:

[0061] The dynamic weight allocation algorithm combined with the attention mechanism scientifically calculates the comprehensive risk index; through the dynamic weighting of multi-dimensional factors such as environmental complexity, load stability, and operation behavior risk, it accurately evaluates the real-time operation risk of the forklift, providing a quantitative basis for hierarchical early warnings and making the early warnings more accurate and timely;

[0062] The hierarchical response strategy implements multi-level response measures from prompts to emergency braking according to the risk level, with clear pertinence and high efficiency; the first-level response uses the HUD interface and voice prompts to enable the driver to timely understand potential risks; the second-level response restricts power output and levels the forklift forks to reduce the risk level; the third-level response resolutely cuts off power, activates the parking system, and records data to maximize the safety of personnel and equipment and reduce accident losses;

[0063] Continuous improvement of closed-loop optimization and collaborative control:

[0064] Closed-loop optimization uses the federated learning framework to update the risk assessment model, which not only protects the data privacy of each node but also can integrate multi-source data to optimize the global model, improving the accuracy and generalization ability of the model; by continuously feeding back the early warning results and continuously optimizing the model parameters, the system performance is continuously improved over time, better adapting to various complex scenarios and working condition changes;

[0065] The digital twin technology simulates extreme scenarios, effectively verifying the robustness of the model; discovers in advance the problems that may occur in the model under extreme conditions, provides directions for model improvement, and enhances the reliability of the system under special and harsh conditions;

[0066] The vehicle-to-vehicle communication module enables multi-forklift collaborative obstacle avoidance, improving the safety and efficiency of multiple forklifts operating simultaneously in the work area; each forklift shares real-time information such as position and speed, and the collaborative obstacle avoidance algorithm reasonably plans the driving path based on this information to avoid collisions between forklifts and improve the fluency of logistics operations;

[0067] The edge computing node has an extremely short response delay and has multiple key functions to ensure real-time decision-making; the real-time path planning algorithm quickly plans a safe path according to the real-time environment and the state of the forklift; the local model inference quickly analyzes the data to obtain the decision result; the secure data synchronization with the cloud platform ensures data consistency and timeliness of model updates, comprehensively guaranteeing the real-time and stable operation of the system. Brief Description of the Drawings

[0068] Figure 1 It is a schematic diagram of the steps of a forklift adaptive safety warning and decision-making method for multiple scenarios according to the present invention. Detailed Embodiment

[0069] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0070] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings of the specification and the specific embodiments.

[0071] As Figure 1 shown, an embodiment of the present invention provides a forklift adaptive safety warning and decision-making method for multiple scenarios, including the following steps:

[0072] Step S1: Multi-scenario dynamic modeling:

[0073] Based on the spatial characteristics of the forklift operating environment, the physical properties of the goods, and the driver behavior parameters, construct a multi-dimensional scenario feature vector;

[0074] Through transfer learning, map the scenario features in the historical accident database to the real-time operating environment to establish a scenario classification model;

[0075] The multi-dimensional scenario feature vector in step S1 includes:

[0076] Environmental complexity coefficient: It is calculated by multiplying the ratio of obstacle density to the area of the operation area by the first weight coefficient, adding the absolute value of the change rate of the ground friction coefficient over time multiplied by the second weight coefficient, and the light intensity gradient multiplied by the third weight coefficient, and then summing the three items;

[0077] Load stability index: According to the linear combination of the variance of the forklift fork pressure distribution and the height of the center of gravity of the goods, it is converted into a stability score through an inverse proportional function. The larger the variance and the height of the center of gravity, the lower the stability score;

[0078] Operating behavior risk value: Based on the weighted value of the hyperbolic tangent function of the steering angle change rate, adding the linear contribution value of the frequency of sudden acceleration events, comprehensively reflecting the driver's operation risk;

[0079] Step S2: Multi-modal sensor data fusion:

[0080] Deploy a heterogeneous sensor network, including 3D lidar, vision sensors, pressure sensor arrays, and inertial measurement units (IMUs), to collect forklift operation data in real time;

[0081] Use a spatio-temporal alignment algorithm to fuse multi-source data, and optimize data reliability through a confidence evaluation module;

[0082] Step S3: Generation of dynamic safety thresholds:

[0083] Based on the scene feature vector, combine fuzzy logic and reinforcement learning algorithms to generate dynamic safety thresholds;

[0084] Introduce a driver behavior correction factor to adjust the safety threshold matrix in real time;

[0085] The method for generating dynamic safety thresholds in Step S3 includes:

[0086] Safety distance threshold: According to the fuzzy logic rule base, perform weighted averaging on the typical safety distances in historical accident data, with the weights being the fuzzy membership degrees in each scenario;

[0087] Maximum allowable inclination threshold: Select the inclination value that maximizes the state-action value function through a reinforcement learning algorithm, and this function is trained from historical operation data;

[0088] Step S4: Comprehensive risk decision-making and hierarchical response:

[0089] Calculate the comprehensive risk index through a dynamic weight allocation algorithm and trigger a hierarchical early warning strategy;

[0090] Execute multi-level response measures from prompting to emergency braking according to the risk level;

[0091] The dynamic weight allocation algorithm in Step S4 is:

[0092] Based on the attention mechanism, the environmental complexity coefficient, load stability index, and operation behavior risk value are input into a multi-layer perceptron, and the attention scores of each dimension are output, and the weights are calculated through the normalization exponential function;

[0093] The comprehensive risk index is composed of the weighted sum of the weights of each dimension and their corresponding eigenvalue. When this index exceeds 0.7, a secondary or higher-level response is triggered;

[0094] Step S5: Closed-loop optimization and cooperative control:

[0095] Feed the warning result back to the model and update the parameters using the online learning algorithm;

[0096] Implement multi-forklift cooperative obstacle avoidance through the vehicle-to-vehicle communication module, and deploy edge computing nodes to ensure real-time decision-making;

[0097] The closed-loop optimization in step S5 includes:

[0098] Adopt the federated learning framework, weighted aggregate the local model parameters of each edge node according to the data volume ratio, and update the global model parameters;

[0099] Simulate extreme scenarios through digital twin technology to verify the robustness of the model.

[0100] In this embodiment, the role of step S1 is to construct a multi-scenario dynamic model. By analyzing and processing the forklift operation environment, goods, and driver-related factors, the characteristics of different scenarios are accurately described, providing a basis and foundation for subsequent safety warning decisions; the following are the detailed steps:

[0101] Step S1-1: Collect data related to forklift operations:

[0102] Collection of environmental space feature data: Use various measurement tools and sensors to collect the spatial feature data of the forklift operation environment; specifically covering information such as the area and shape of the operation area, the distribution, quantity, size, and position of obstacles; these data can intuitively reflect the spatial layout of the operation environment and provide an important basis for constructing the scenario feature vector later;

[0103] Collection of physical property data of goods: Conduct detailed measurement and recording of the physical properties of the goods, including the weight, volume, center of gravity position, shape, etc. of the goods; these properties will affect the load stability and operation difficulty of the forklift and are one of the key factors in constructing the scenario feature vector;

[0104] Collection of driver behavior parameter data: With the help of on-vehicle sensors and monitoring devices, collect the driver's behavior parameters, such as the steering angle change rate, hard acceleration event frequency, operation speed, operation force, etc.; these parameters can reflect the driver's operation habits and risk levels and are crucial for accurately evaluating the safety of the operation scenario;

[0105] Step S1-2: Construct a multi-dimensional scenario feature vector:

[0106] Calculate the environmental complexity coefficient : Based on the collected environmental space feature data, calculate the environmental complexity coefficient according to the following formula : , where is the obstacle density, that is, the ratio of the number of obstacles in the operation area to the area of the operation area; is the area of the operation area; is the ground friction coefficient; is the light intensity gradient; , , are normalization weights used to adjust the proportion of each item in the calculation of the environmental complexity coefficient;

[0107] Calculate the load stability index : According to the physical attribute data of the goods, through the variance of the forklift pressure distribution and the height of the center of gravity of the goods , calculate the load stability index according to the following formula : , where , are calibration coefficients used to adjust the influence degree of the variance of the forklift pressure distribution and the height of the center of gravity of the goods on the load stability index; is the variance of the forklift pressure distribution; is the height of the center of gravity of the goods;

[0108] Calculate the operation behavior risk value : Based on the collected driver behavior parameters, according to the steering angle change rate and the frequency of hard acceleration events , calculate the operation behavior risk value according to the following formula : , where , , are training parameters, and their specific values are determined through training data; is the steering angle change rate; is the frequency of hard acceleration events;

[0109] The calculated environmental complexity coefficient , load stability index and operation behavior risk value Combined into a multi-dimensional scenario feature vector, which comprehensively reflects the characteristics of the forklift operation scenario;

[0110] Step S1-3: Establish a historical accident database:

[0111] Data collection and collation: Collect relevant data on historical forklift accidents, including the time, location, environmental conditions, cargo information, driver operation conditions, and accident type and severity, etc.; Clean and collate these data, remove duplicate, incorrect, and invalid data to ensure the accuracy and consistency of the data;

[0112] Scenario feature extraction: Extract scenario features from the sorted historical accident data, such as environmental complexity, load stability, and operation behavior risk, etc.; Associate these features with the corresponding accident information to form a historical accident database; This database provides rich sample data for subsequent transfer learning;

[0113] Step S1-4: Transfer learning and scenario classification model establishment:

[0114] Feature mapping: Use transfer learning technology to map the scenario features in the historical accident database to the real-time operation environment; By comparing the similarity between historical data and real-time data, find the historical scenario that best matches the current operation scenario, so as to apply the classification information of the historical scenario to the real-time scenario;

[0115] Model training and optimization: Based on the mapped feature data, select a suitable machine learning algorithm (such as decision tree, support vector machine, neural network, etc.) to train the scenario classification model; During the training process, continuously adjust the parameters of the model to improve the classification accuracy and generalization ability of the model; Finally, establish a model that can accurately classify the real-time operation scenario.

[0116] In this embodiment, the function of step S2 is to fuse multi-modal sensor data, collect forklift operation data in real time by deploying multiple sensors, and process and optimize these multi-source data to improve the reliability and availability of the data, providing accurate data support for subsequent safety warning decisions; The following are the detailed steps:

[0117] Step S2-1: Deploy a heterogeneous sensor network:

[0118] Three-dimensional lidar deployment: According to the characteristics and requirements of the forklift operation environment, install a three-dimensional lidar at a suitable position on the forklift; Ensure that its scanning range can cover the main areas around the forklift, the scanning frequency ≥ 20Hz, and the angular resolution ≤ 0.1°, so as to be able to obtain the three-dimensional space information of the surrounding environment in real time and accurately, including the position, shape, and distance of obstacles, etc.;

[0119] Visual sensor deployment: Install visual sensors, such as cameras, at positions that can clearly capture key information during the forklift operation, such as the status of goods, signs on the driving path, etc.; visual sensors can assist 3D lidar in environmental perception and provide richer image information;

[0120] Pressure sensor array deployment: Deploy a pressure sensor array at key parts such as the forklift forks, with a sampling accuracy ≤ 0.2%FS; the pressure sensor array can monitor the pressure distribution of the goods on the forklift forks in real time and provide important data for calculating the load stability index, etc.;

[0121] Inertial measurement unit (IMU) deployment: Install a 9-axis IMU module, ensuring that its attitude angle error ≤ 0.3° and data output delay ≤ 5ms; the IMU module can measure information such as the acceleration, angular velocity, and attitude of the forklift and is used to monitor the motion state of the forklift;

[0122] Step S2-2: Real-time collection of forklift operation data:

[0123] Start the sensor network: Activate the deployed heterogeneous sensor network to make each sensor start working and collect relevant data during the forklift operation in real time;

[0124] Data transmission and storage: Transmit the data collected by the sensors to the data processing center through a suitable communication method (such as wired or wireless communication) and store it; ensure the stable and accurate transmission of data and the standardized storage data format for subsequent processing and analysis;

[0125] Step S2-3: Application of spatio-temporal alignment algorithm:

[0126] Time synchronization: Since there may be differences in the data collection frequencies and times of different types of sensors, it is necessary to perform time synchronization on the multi-source data collected; adopt accurate time synchronization protocols and algorithms to ensure the time consistency of the data collected by different sensors at the same time point for accurate subsequent fusion processing;

[0127] Spatial alignment: Perform spatial alignment on the data of different sensors and unify them into the same coordinate system; according to information such as the installation position and angle of the sensors, eliminate spatial deviations through methods such as coordinate transformation and calibration to make the data of different sensors consistent in space;

[0128] Step S2-4: Multi-source data fusion:

[0129] Data preprocessing: Perform preprocessing on the spatio-temporally aligned data, including operations such as noise removal, filtering, and normalization, to improve the quality and usability of the data;

[0130] Fusion Algorithm Selection and Application: According to the characteristics and requirements of the data, select appropriate multi-source data fusion algorithms, such as Kalman filtering, Bayesian fusion, etc.; fuse the preprocessed data from different sensors to obtain more comprehensive and accurate forklift operation data;

[0131] Step S2-5: Optimization of the Confidence Evaluation Module:

[0132] Confidence Calculation: Establish a confidence evaluation model to calculate the confidence of the fused data; evaluate the confidence of each data point according to factors such as the accuracy and reliability of the sensor and the consistency of the data;

[0133] Data Optimization: Optimize the fused data according to the confidence evaluation results; for data with low confidence, it can be corrected, supplemented or eliminated to improve the reliability and accuracy of the data; at the same time, continuously update the confidence evaluation model to adapt to different operation scenarios and data changes.

[0134] In this embodiment, the function of step S3 is to generate a dynamic safety threshold. According to the characteristics of the forklift operation scenario, using fuzzy logic and reinforcement learning algorithms, and combining with the driver behavior correction factor, flexibly and accurately determine the safety threshold, providing a more practical evaluation criterion for forklift operation safety; the specific steps are as follows:

[0135] Step S3-1: Initialize the Fuzzy Logic Rule Base and the Reinforcement Learning Environment:

[0136] Construction of the Fuzzy Logic Rule Base: Collect a large amount of historical safe operation data of forklifts in different scenarios, analyze the relationship between scenario characteristics such as environmental complexity, load stability, and operation behavior risk and the safety threshold; based on expert experience and data rules, formulate a series of fuzzy logic rules; for example, when the environmental complexity is high and the load stability is low, the safety distance threshold should be increased, etc. Organize these rules into a fuzzy logic rule base;

[0137] Construction of the Reinforcement Learning Environment: Define the state, action, and reward in reinforcement learning; the state includes the real-time scenario feature vector of the forklift, such as the environmental complexity coefficient 、load stability index and operation behavior risk value etc.; the action corresponds to the adjustment strategy of different safety thresholds; the reward is set based on whether the actual operation of the forklift is safe and efficient. If no accident occurs and the efficiency is high according to the current safety threshold, a higher reward is given, otherwise a lower reward is given; build such a reinforcement learning environment to prepare for subsequent optimization of the safety threshold;

[0138] Step S3-2: Calculate the Preliminary Safety Threshold Based on the Scenario Feature Vector:

[0139] Safety distance threshold Initial calculation: Input the scene feature vector of the current forklift operation into the fuzzy logic rule base;

[0140] For the safety distance threshold , through the fuzzy inference mechanism, calculate the fuzzy membership degree according to the activation degree of different rules in the rule base ; These fuzzy membership degrees correspond to the typical safety distances in the historical accident data ; Then use the formula to calculate the preliminary safety distance threshold; This formula comprehensively considers the weights of different typical safety distances inferred by fuzzy logic to obtain the safety distance threshold adapted to the current scene;

[0141] Maximum allowable inclination threshold Initial calculation: In the reinforcement learning environment, take the current scene state as the input, and use the trained reinforcement learning model (such as the Q-learning model) to calculate the state-action value function to obtain the values corresponding to different inclinations ; Among them, is the current forklift operation scene state, is the inclination; Through the formula , find the inclination that maximizes the value function, that is, the preliminary maximum allowable inclination threshold; This process continuously explores and optimizes through reinforcement learning to find the most suitable maximum allowable inclination threshold in the current scene;

[0142] Step S3-3: Introduce the driver behavior correction factor:

[0143] Driver behavior data acquisition: Continuously collect the real-time operation behavior data of the driver through the forklift on-vehicle equipment, including the steering angle change rate, the frequency of sudden acceleration events, the operation duration, etc.; These data reflect the current operation habits and behavior patterns of the driver;

[0144] Correction factor calculation: According to the collected driver behavior data, establish a driver behavior evaluation model; For example, if the driver frequently accelerates suddenly or the steering angle change rate is too large, it indicates that the operation behavior risk is relatively high, and the corresponding correction factor will increase the adjustment range of the safety threshold; If the driver operates smoothly, the adjustment range of the correction factor for the safety threshold is relatively small; Calculate the correction factor for the current driver behavior through this model;

[0145] Safety threshold matrix adjustment: Apply the calculated driver behavior correction factor to the safety threshold matrix obtained from the initial calculation; For various safety thresholds such as the safety distance threshold and the maximum allowable inclination threshold, make corresponding adjustments according to the size of the correction factor, so as to obtain a dynamic safety threshold matrix that more conforms to the current driver operation behavior characteristics;

[0146] Step S3-4: Real-time update and feedback of dynamic safety threshold:

[0147] Real-time monitoring and update: During the forklift operation, continuously monitor the changes in the scene feature vector and driver behavior data; once these data change, immediately repeat the above steps, recalculate the safety threshold under the fuzzy logic rule base and reinforcement learning environment, and combine the driver behavior correction factor to update the dynamic safety threshold matrix in real time; ensure that the safety threshold can timely adapt to the dynamic changes of the operation scene and driver behavior.

[0148] Feedback and optimization: Apply the updated dynamic safety threshold to the forklift safety warning decision-making system and observe its impact on forklift operation safety; if it is found that the warning is inaccurate or unreasonable, feedback the relevant information to the training process of the fuzzy logic rule base and reinforcement learning model, optimize the model parameters and rules, and further improve the accuracy and reliability of the dynamic safety threshold generation.

[0149] In this embodiment, step S4 mainly focuses on comprehensive risk decision-making and hierarchical response, and the detailed steps are as follows:

[0150] Step S4-1: Obtain multi-dimensional risk-related data: Precisely extract the environmental complexity coefficient , load stability index and operation behavior risk value from the data results generated in the previous steps; these data are quantitative presentations of the current actual situation of forklift operation, covering key aspects such as the operation environment, cargo status, and driver operation behavior, providing the original basis for subsequent comprehensive risk assessment.

[0151] Step S4-2: Determine dimension weights based on the attention mechanism:

[0152] Train a multi-layer perceptron (MLP) model: Use a large amount of historical forklift operation data, which includes , , data under different scenarios and the corresponding operation results (whether an accident occurs, the severity of the accident, etc.); through the training of the MLP model, enable it to learn the complex internal relationships between data in each dimension and operation risks.

[0153] Calculate dimension weights : Input the obtained in real time into the trained MLP model to get the output value (corresponding to three dimensions respectively); according to the formula , calculate the weights of each dimension; in this formula, the exponential function can significantly amplify different The differences between the values, thus enabling the dimensions that have a greater impact on the current job risk to obtain higher weights, ensuring that the contributions of all dimensions are reasonably reflected in the comprehensive risk assessment;

[0154] Step S4-3: Calculate the comprehensive risk index :

[0155] Based on the weights calculated above , , , and the corresponding environmental complexity coefficient Load stability index and the operation behavior risk value , use the formula to calculate the comprehensive risk index; this formula organically integrates the risk factors of the three dimensions through weighted summation to generate a value that can comprehensively and quantitatively reflect the current job risk degree of the forklift;

[0156] Step S4-4: Trigger the hierarchical warning strategy: Set clear risk level thresholds in the system in advance according to forklift operation safety standards, past accident statistical analysis, and industry general specifications, etc.; when the calculated comprehensive risk index , the system determines that the forklift operation is in a high-risk state, immediately triggers a response at the second level and above, and starts the corresponding hierarchical warning process;

[0157] Step S4-5: Implement multi-level response measures:

[0158] First-level response: When the risk is at a relatively low level but requires the driver's attention, the system uses the HUD (Head-Up Display) interface of the forklift to visually display the risk status of different areas around the forklift in the form of a dynamic risk heat map, with different colors representing different risk levels; at the same time, with the help of the voice prompt function, clearly broadcast the potential risk points to the driver to remind him to operate carefully and prevent risks in advance;

[0159] Second-level response: Once the risk level rises to the condition for triggering the second-level response, the system takes immediate action; on the one hand, limit the power output of the forklift and reduce it to 50% of the rated value to slow down the running speed of the forklift and reduce the possibility of accidents caused by high-speed driving; on the other hand, automatically activate the automatic fork leveling function, use sensors to monitor the status of the goods on the fork in real time, and automatically adjust the fork angle according to the feedback information to ensure the stability of the goods during handling and reduce the risk of goods falling;

[0160] Level 3 response: If the risk level rises to the highest level 3 response standard, the system will immediately take emergency measures. First, it will decisively cut off the power transmission of the forklift, causing the forklift to stop running instantly to avoid possible serious accidents. Then, it will start the electronic parking system to prevent the forklift from moving accidentally due to inertia or other factors. In addition, blockchain technology is used to record accident prediction data, including the comprehensive risk index when the level 3 response is triggered, the risk value of each dimension, the real-time location of the forklift, the operating status and other detailed information. The tamper-proof nature of blockchain technology ensures the authenticity and integrity of the data, providing a reliable basis for subsequent in-depth analysis of the cause of the accident.

[0161] In this embodiment, the function of step S5 is to continuously improve the accuracy and reliability of the forklift adaptive safety warning decision method through closed-loop optimization and collaborative control, realize collaborative operation between multiple forklifts, and ensure real-time and efficient decision-making; the following are the detailed steps:

[0162] Step S5-1: Early warning result feedback and data collection:

[0163] Establishment of result feedback mechanism: Set up a special feedback channel in the forklift safety warning system to accurately feed back the result of each warning (including whether the warning is triggered, the warning level, the actual situation, etc.) to the core model of the system; ensure that the feedback process is timely and stable to avoid data loss or delay;

[0164] Relevant data collection: Collect various data related to the warning results, including the operating parameters of the forklift (such as speed, acceleration, steering angle, etc.), scene feature data (environmental complexity coefficient, etc.), Load stability index and operational behavior risk value ), sensor data (3D lidar, visual sensor, pressure sensor array and inertial measurement unit data) and final operation results (whether an accident occurred, accident type and severity, etc.); these data are classified, organized and stored to provide rich samples for subsequent model updates;

[0165] Step S5-2: Update the risk assessment model using the federated learning framework:

[0166] Edge node data processing: Each edge computing node (such as a local computing device installed on a forklift) pre-processes the locally collected data, including data cleaning (removing noise and erroneous data), feature extraction, and data encryption to protect data privacy;

[0167] Local model training: Each edge node uses online learning algorithms (such as stochastic gradient descent) to train the local risk assessment model based on the locally processed data and update the local model parameters. Indicates edge nodes);

[0168] Global model parameter aggregation: local model parameters of each edge node Upload to the central server; the central server will upload data based on the local data volume of each edge node. And the total amount of local data of all edge nodes , according to the formula Calculate global model parameters ;in, is the number of edge nodes;

[0169] Global model distribution: The central server sends the updated global model parameters Send it to each edge node; each edge node uses the global model parameters to update the local model, thereby optimizing and updating the entire risk assessment model;

[0170] Step S5-3: Simulate extreme scenarios through digital twin technology:

[0171] Digital twin model construction: Based on the forklift's physical entity, operating environment, and operating rules, an accurate digital twin model is constructed; the model can reflect the actual status and behavior of the forklift in real time, including the forklift's dynamic characteristics, simulation of sensor data, and dynamic changes in the operating scene;

[0172] Definition and simulation of extreme scenarios: Based on historical accident data and industry experience, define a series of extreme scenarios, such as severe weather conditions (heavy rain, heavy fog), complex operating environments (narrow passages, dense obstacles), abnormal driver operations, etc.; simulate these extreme scenarios in the digital twin model to observe the performance of the risk assessment model under extreme conditions;

[0173] Model robustness verification: Analyze simulation results and evaluate the robustness of the risk assessment model in extreme scenarios; check whether the model can accurately warn, whether the graded response is reasonable, and whether the safety threshold can be adjusted in time; if problems are found in the model, record relevant information and feed it back to the model update process to further optimize the model;

[0174] Step S5-4: The inter-vehicle communication module realizes multi-forklift collaborative obstacle avoidance:

[0175] Construction of inter-vehicle communication network: Install inter-vehicle communication modules on each forklift to establish a stable wireless communication network to ensure that forklifts can exchange information in real time and reliably; the communication protocol should have low latency, high bandwidth and anti-interference capabilities to adapt to complex operating environments;

[0176] Information sharing and interaction: Each forklift shares its own position, speed, driving direction, operation status and other information in real time through the vehicle-to-vehicle communication network; at the same time, it receives relevant information of other forklifts to construct a global operation scenario map;

[0177] Execution of collaborative obstacle avoidance algorithm: Based on the shared information, the collaborative obstacle avoidance algorithm runs on each forklift; the algorithm predicts the possible collision risks according to the relative positions, speeds and driving trajectories of the forklifts, and adjusts the driving paths and speeds of the forklifts in a timely manner to avoid collision accidents; for example, when two forklifts approach, the algorithm will command one of the forklifts to decelerate or change the driving direction according to factors such as priority and distance;

[0178] Step S5-5: The edge computing node ensures real-time decision-making:

[0179] Real-time data processing: The edge computing node has powerful computing capabilities and can process the data collected by local sensors and the information of other forklifts received through the vehicle-to-vehicle communication network in real time; quickly analyze and process the data, extract key features, and provide a basis for decision-making;

[0180] Localized model inference: Run the localized risk assessment model and collaborative obstacle avoidance algorithm on the edge computing node for real-time model inference; make quick decisions based on the processed data and model inference results, such as whether to trigger an alarm and how to adjust the operating status of the forklift;

[0181] Secure data synchronization with the cloud platform: The edge computing node regularly uploads the locally processed data and model update information to the cloud platform, and at the same time downloads the latest global model parameters and configuration information from the cloud platform; during the data synchronization process, encryption technology is used to ensure the security of the data, prevent data leakage and malicious attacks; at the same time, ensure the timeliness of data synchronization to ensure that the edge computing node always uses the latest model and data for decision-making.

[0182] In this embodiment, the response delay of the edge computing node does not exceed 30 milliseconds and supports the following functions:

[0183] Real-time path planning algorithm;

[0184] Localized model inference;

[0185] Secure data synchronization with the cloud platform;

[0186] Specifically, as a key component of the forklift adaptive safety warning decision-making system, the response delay of the edge computing node needs to be strictly controlled within 30 milliseconds to ensure the real-time performance and efficiency of the entire system. At the same time, it also needs to support the following important functions:

[0187] Real-time path planning algorithm:

[0188] Environmental perception data processing: The edge computing node receives in real time the surrounding environment information fed back by various forklift sensors, such as 3D lidar, vision sensors, etc.; quickly parses and processes this data to construct an accurate local environment map, which covers key elements such as the position, shape, size of obstacles, and passable areas;

[0189] Forklift status analysis: Synchronously obtains the real-time operating status data of the forklift, including the current position, driving speed, direction, and load condition, etc.; based on this information, accurately predicts the dynamic trend of the forklift;

[0190] Path planning calculation: Utilizes efficient path planning algorithms, such as algorithms, optimized versions of the Dijkstra algorithm, etc., takes the current position of the forklift as the starting point and the target position as the ending point, and at the same time fully considers the obstacle distribution in the environment map and the operating limitations of the forklift itself (such as turning radius, maximum driving speed, etc.), and plans a safe and efficient driving path within a very short time (less than 30 milliseconds);

[0191] Path dynamic adjustment: During the driving process of the forklift, if the sensor detects a change in the environment, such as the emergence of new obstacles or the change in the position of the original obstacles, the edge computing node can immediately re-plan the path to ensure that the forklift always drives along the optimal route, guaranteeing the continuity and safety of the operation;

[0192] Localized model inference:

[0193] Model storage and loading: Pre-stores various key models trained locally on the edge computing node, including multi-scene classification models, risk assessment models, and safety threshold generation models, etc.; when the system starts or model inference is required, the edge computing node can quickly load these models and get ready for inference;

[0194] Real-time data input: Timely inputs the real-time collected forklift operation environment data (such as environmental complexity coefficient, light intensity, etc.), cargo physical attribute data (such as load stability index), and driver behavior parameter data (such as operation behavior risk value) into the corresponding models;

[0195] Fast inference operation: Utilizes the powerful computing power of the edge computing node to perform fast inference operations on the input data within 30 milliseconds; for example, calculates the comprehensive risk index in the current operation state through the risk assessment model to determine whether the forklift is in a safe operation state; generates dynamic safety thresholds according to real-time scene characteristics with the help of the safety threshold generation model;

[0196] Result Output and Application: Quickly output and feedback the results obtained from model inference, such as risk levels, suggestions for adjusting safety thresholds, etc., to the safety warning decision-making system of the forklift, providing key basis for subsequent hierarchical warnings and response measures;

[0197] Security Data Synchronization with the Cloud Platform:

[0198] Data Encryption and Packing: The edge computing node regularly encrypts important data stored locally, such as the historical operation data of the forklift, model update data, and warning result data, etc., using advanced encryption algorithms, such as the AES encryption algorithm, to ensure the security of the data during transmission and prevent the data from being stolen or tampered with; after encryption, the data is packed into a specific format and prepared to be uploaded to the cloud platform;

[0199] Establishment of a Secure Transmission Channel: Establish a reliable transmission channel with the cloud platform through a secure network protocol, such as the SSL / TLS protocol; during the connection process, strict identity verification and permission management are carried out to ensure that only authorized edge computing nodes can interact with the cloud platform for data;

[0200] Data Upload and Synchronization: Upload the packed and encrypted data to the cloud platform at a predetermined time interval or specific trigger condition; after receiving the data, the cloud platform decrypts and verifies it to ensure the integrity and accuracy of the data; at the same time, update the latest data to the cloud database and model repository to achieve data synchronization;

[0201] Receiving and Processing of Downlink Data: Receive downlink data from the cloud platform, such as global model update parameters, system configuration update information, etc.; after passing through security verification and decryption, apply these data to the local system and model to ensure that the edge computing node always maintains data and functional consistency with the cloud platform, improving the coordination and intelligence level of the entire system.

[0202] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-scenario adaptive safety warning decision method for forklifts, characterized by: The following steps are involved: Step S1: Multi-scenario dynamic modeling: Based on the spatial characteristics of the forklift operating environment, the physical properties of the cargo, and the driver's behavior parameters, a multi-dimensional scene feature vector is constructed; Map the scene features in the historical accident database to the real-time operating environment through transfer learning to establish a scene classification model; Step S2: Multimodal sensor data fusion: Deploy a heterogeneous sensor network, including 3D lidar, visual sensors, pressure sensor arrays, and inertial measurement units (IMUs), to collect forklift operation data in real time; The spatiotemporal alignment algorithm is used to fuse multi-source data, and the confidence assessment module is used to optimize data reliability; Step S3: Dynamic security threshold generation: Generate dynamic safety thresholds based on scene feature vectors by combining fuzzy logic and reinforcement learning algorithms; Introduce driver behavior correction factors to adjust the safety threshold matrix in real time; Step S4: Comprehensive risk decision and graded response: Calculate the comprehensive risk index through a dynamic weight allocation algorithm to trigger a graded warning strategy; Implement multi-level response measures from warning to emergency braking according to the risk level; Step S5: Closed-loop optimization and collaborative control: Feed the warning results back to the model and use the online learning algorithm to update the parameters; Collaborative obstacle avoidance of multiple forklifts is achieved through inter-vehicle communication modules, and edge computing nodes are deployed to ensure real-time decision-making.

2. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The multi-dimensional scene feature vector in step S1 includes: Environmental complexity coefficient: calculated by multiplying the ratio of obstacle density to the area of ​​the working area by the first weight coefficient, superimposing the absolute value of the rate of change of the ground friction coefficient over time by the second weight coefficient, and the light intensity gradient by the third weight coefficient. Load stability index: Based on the linear combination of the fork pressure distribution variance and the cargo center of gravity height, it is converted into a stability score through an inverse proportional function. The larger the variance and center of gravity height, the lower the stability score; Operation behavior risk value: Based on the weighted value of the hyperbolic tangent function of the steering angle change rate, superimposed with the linear contribution value of the frequency of sudden acceleration events, it comprehensively reflects the driver's operation risk.

3. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The method for generating the dynamic safety threshold in step S3 includes: Safety distance threshold: Based on the fuzzy logic rule base, the typical safety distances in historical accident data are weighted averaged, with the weight being the fuzzy membership degree in each scenario; Maximum allowed tilt angle threshold: The tilt angle value that maximizes the state-action value function is selected through the reinforcement learning algorithm, which is trained by historical operation data.

4. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized in that: The dynamic weight allocation algorithm in step S4 is: Based on the attention mechanism, the environment complexity coefficient, load stability index and operation behavior risk value are input into the multi-layer perceptron, and the attention score of each dimension is output. The weight is calculated through the normalized exponential function. The comprehensive risk index is composed of the weighted sum of the weights of each dimension and its corresponding eigenvalue. When the index exceeds 0.7, it triggers a level 2 or higher response.

5. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The closed-loop optimization in step S5 includes: Adopting the federated learning framework, the local model parameters of each edge node are weighted and aggregated according to the proportion of data volume, and the global model parameters are updated; Extreme scenarios are simulated through digital twin technology to verify the robustness of the model.

6. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The multi-level response measures are executed according to a hierarchical response strategy, which includes: Level 1 response: Displays a dynamic risk heat map through the HUD interface and provides voice prompts to the driver; Secondary response: Limit the forklift power output to 50% of the rated value and activate the automatic fork leveling function; Level 3 response: cut off power transmission, activate the electronic parking system, and record accident prediction data through blockchain.

7. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The heterogeneous sensor network meets the following performance indicators: The scanning frequency of the 3D laser radar shall not be less than 20 Hz, and the angular resolution shall not be greater than 0.1 degrees; The sampling accuracy of the pressure sensor array is no less than 0.2% of the full scale; The attitude angle error of the 9-axis IMU module does not exceed 0.3 degrees, and the data output delay does not exceed 5 milliseconds.

8. The method for adaptive safety early warning decision-making for forklifts in multiple scenarios according to claim 1 is characterized by: The edge computing node has a response delay of no more than 30 milliseconds and supports the following functions: Real-time path planning algorithm; Localized model reasoning; Secure data synchronization with cloud platforms.

Images