Construction hoist intelligent safety interlocking control method and device based on multi-source sensing and AI fusion

Abstract

The application discloses a kind of construction hoist intelligent safety interlock control method and device based on multi-source sensing and AI fusion, real-time acquisition vertical distance between cage and floor platform, vibration acceleration, door alignment degree characteristics, monitor the behavior of personnel in cage, environmental harmful gas concentration and door gap clamping resistance;Using improved D-S evidence theory fusion mechanism, real-time confidence evaluation is carried out to each sensor data, and fusion weight is dynamically generated;Based on safety knowledge graph, root cause reasoning is carried out on abnormal events, an AI wear prediction model is constructed, key components are evaluated and predicted, and the root cause reasoning result and the health status of AI wear prediction are quantified as basic probability distribution function respectively, as auxiliary evidence source is embedded into D-S fusion framework, and layer door opening authorization criterion is generated together.Compared with prior art, the application significantly improves the intrinsic safety, intelligent level and predictive operation and maintenance efficiency of construction hoist.

Description

Technical Field

[0001] This invention relates to the field of safety control technology for construction equipment, specifically to an intelligent safety interlocking control method and device for construction hoists based on the fusion of multi-source sensing and AI. Background Technology

[0002] Construction hoists are commonly used vertical transportation equipment in high-rise building construction, widely used for the vertical transport of personnel and materials. However, due to the lack of a reliable linkage control mechanism between the landing doors and the hoist cage, unsafe operating behaviors by machine operators (such as inaccurate stopping, misjudging door opening, and misalignment of the door) and the hoist's reliance on mechanical copper wire brushes contacting the guide rails to determine the cage position result in problems such as slow response, easy wear, high misjudgment rate, and inability to assess dynamic stability. These issues seriously affect the safe operation and management of current construction hoists and have become a significant source of risk in current engineering projects.

[0003] Although some systems have incorporated infrared or ultrasonic ranging technologies, they remain limited to single-sensor perception, making it difficult to comprehensively reflect the actual docking status of the hoist cage. Furthermore, the health status of critical components such as copper wire brushes typically relies on manual inspections, lacking quantitative assessment methods, leading to untimely maintenance and the risk of "operating with defects." Traditional systems generally lack predictive maintenance capabilities and data-driven intelligent decision-making mechanisms, leaving safety protection in a reactive response phase.

[0004] Therefore, there is an urgent need for a new type of intelligent interlocking system that integrates high-precision sensing, multi-source information fusion, intelligent decision-making and predictive maintenance to improve the inherent safety level and intelligent operation and maintenance capabilities of construction hoists. Summary of the Invention

[0005] Purpose of the invention: In view of the problems pointed out in the background technology, the present invention discloses an intelligent safety interlocking control method and device for construction hoists based on multi-source sensing and AI fusion, which integrates high-precision sensing, multi-source information fusion, intelligent decision-making and predictive maintenance, thereby improving the inherent safety level and intelligent operation and maintenance capabilities of construction hoists.

[0006] Technical Solution: This invention discloses an intelligent safety interlocking control method for construction hoists based on the fusion of multi-source sensing and AI, comprising the following steps:

[0007] The system acquires the vertical distance between the bottom plate of the hoist cage and the current floor platform, the hoist cage vibration acceleration, the spatial alignment deviation between the hoist cage door and the floor door, the concentration of combustible or toxic gases, the clamping resistance during the closing process of the floor door and the hoist cage door, and the video stream of personnel behavior.

[0008] Based on the video stream of human behavior, it is determined whether there are dangerous behaviors such as leaning against the door, exceeding the number of people, or carrying long poles and extending them, and the human risk level is output.

[0009] The collected data and human risk levels are fused using a multi-source data fusion algorithm that combines an improved DS evidence theory with dynamic weight allocation to comprehensively determine whether the conditions for opening the floor door are met.

[0010] The physical closure status of the cage door and the landing door is continuously monitored. The cage start restriction is lifted only when both doors are fully closed and locked.

[0011] Furthermore, the multi-source data fusion algorithm also introduces a security knowledge graph module, which infers the most likely root cause of the failure based on the current anomaly pattern, inputs the current anomaly pattern into the pre-constructed security knowledge graph module, outputs the probability distribution of the most likely root cause of the failure, and uses this probability distribution as auxiliary evidence in the multi-source data fusion, specifically:

[0012] The security knowledge graph includes the following entities and relationships:

[0013] Physical components: copper wire brush, guide rail, brake, wind speed, humidity, overload operation;

[0014] Relationship: High humidity leads to guide rail corrosion; frequent start-stop cycles exacerbate wear on the copper wire brush.

[0015] When the vibration RMS suddenly increases and the humidity > 80% RH, the graph inference engine outputs the probability of "rail corrosion," and converts the result into a BPA function. It also serves as an auxiliary source of evidence in multi-source data fusion and is used in Dempster synthesis with sensor-based comprehensive evidence.

[0016] Furthermore, the multi-source data fusion algorithm also introduces an AI wear prediction model to predict the wear state of the copper wire brush. When all opening conditions are met and the AI ​​wear prediction model used for predicting the wear state of the copper wire brush does not issue an abnormal warning, an unlocking command is sent to the landing door's electric lock, allowing the landing door to be opened. The AI ​​wear prediction model adopts a dual-channel hybrid model that combines an LSTM neural network and an XGBoost classifier, and its output is the probability distribution of the remaining service life, specifically:

[0017] P1: The probability that the remaining lifespan is greater than 30 days, corresponding to wear level 0;

[0018] P2: The probability that the remaining lifespan is between 10 and 30 days, corresponding to wear level 1;

[0019] P3: The probability that the remaining lifespan is less than 10 days, corresponding to wear level 2;

[0020] The probability distribution is transformed into a basic probability allocation function m4, which serves as the fourth source of evidence in multi-source data fusion, specifically defined as:

[0021] ;

[0022] ;

[0023] ;

[0024] Where A represents "the cage has been stably docked and aligned, meeting the door opening condition", ¬A represents "the door opening condition is not met", and Θ represents an uncertain state; simultaneously, the Shannon entropy H of the basic probability assignment is calculated:

[0025] ;

[0026] when At that time, the door opening authorization judgment is forcibly prohibited from being opened;

[0027] when When this occurs, the cautious operation mode is activated, which includes: extending the stop waiting time to more than 5 seconds, limiting the cage load to no more than 80% of the rated value, and triggering a voice alarm.

[0028] Furthermore, the conditions for opening the door include: vertical distance ≤ 150 mm; root mean square (RMS) values ​​of all three axes of acceleration less than a preset threshold; door alignment deviation less than 10 mm; and wear probability output by the AI ​​wear prediction model lower than a set threshold.

[0029] Furthermore, the multi-source data fusion algorithm employs an improved DS evidence theory, and the specific steps are as follows:

[0030] S1: Define the propositional space ,in: This indicates that "the floor door can be opened safely". This indicates "there is a risk of opening the door," the complete series. The power set is ;

[0031] S2: Based on the laser displacement, RMS acceleration, and visual recognition results, construct basic probability assignment functions for each sensor. ;

[0032] (1) Laser rangefinder sensor, input distance Output BPA:

[0033] ;

[0034] (2) Triaxial accelerometer, input RMS acceleration Output BPA:

[0035] ;

[0036] (3) Visual recognition camera, input door alignment deviation Output BPA:

[0037] ;

[0038] S3: Invoke a lightweight attention network to generate dynamic confidence weights. The revised comprehensive evidence is obtained by weighting and averaging the BPAs.

[0039] ;

[0040] S4: If the security knowledge graph is triggered, quantify its root cause reasoning results into BPA. and with Perform Dempster orthogonal sum operation to obtain intermediate fusion results. Otherwise, ;

[0041] S5: If the AI ​​wear prediction model is enabled, obtain its output health status (BPA). and will and Perform Dempster combinatorial analysis to obtain the final basic probability distribution. Otherwise ;

[0042] S6: Computational Propositions Overall trust level is:

[0043] ;

[0044] in, To meet the trust level required for opening the door, For the degree of trust in an uncertain state, An uncertainty allocation coefficient, set to a fixed value of 0.5, is used to proportionally allocate uncertain evidence to supporting and opposing propositions.

[0045] S7: Set threshold ,like and If the door is opened, the electromagnetic lock on the landing door will be authorized; otherwise, opening the door will be prohibited and an audible and visual alarm will be triggered.

[0046] Furthermore, the dynamic weights are generated by a lightweight attention network deployed on the edge controller; the attention network takes time window segments of the raw sensor data as input and outputs confidence weights corresponding to the laser displacement sensor, triaxial accelerometer, and visual recognition module. ,satisfy and .

[0047] Furthermore, the AI ​​wear prediction model also supports online incremental learning: continuously monitoring the residual between the actual component replacement time and the model's predicted remaining service life; when the absolute value of the residual exceeds the preset tolerance threshold three times in a row, a local fine-tuning mechanism is triggered, updating only the weights of the fully connected layer of the LSTM channel and the leaf node of the XGBoost channel, thereby achieving lightweight online correction of the model parameters.

[0048] Furthermore, the AI ​​wear prediction model adopts a transfer learning framework: a basic model is pre-trained in the cloud based on a large-scale historical operating dataset, and the model is then distributed to the local device for fine-tuning based on recent operating data.

[0049] This invention also discloses an intelligent safety interlocking device for construction hoists based on the fusion of multi-source sensing and AI, comprising:

[0050] Multi-source sensing module: including laser rangefinder, triaxial accelerometer, visual recognition camera, gas concentration sensor and flexible tactile film sensor, which are used to collect the vertical distance between the bottom plate of the hoist cage and the current floor platform, the hoist cage vibration acceleration, the spatial alignment deviation between the hoist cage door and the floor door, the concentration of combustible or toxic gas, the clamping resistance during the closing process of the floor door and the hoist cage door, and the video stream of personnel behavior.

[0051] AI Wear Prediction Module: Used to receive historical operating data and predict the wear status of copper wire brushes;

[0052] Central intelligent control unit: Deployed in the hoist cage or floor control box, it is used to integrate the collected data and human risk level using a multi-source data fusion algorithm that combines improved DS evidence theory with dynamic weight allocation, and then combine the probability distribution of the most likely root cause of the failure with the prediction results of the AI ​​wear prediction module to comprehensively determine whether the conditions for opening the floor door are met.

[0053] Actuators: including floor door electric locks and cage start locks, which are used to open and close based on control commands issued by the central intelligent control unit;

[0054] Electronic door magnetic system: used to detect the opening and closing status of cage doors and landing doors;

[0055] Monitoring and communication module: includes high-definition surveillance cameras and wireless communication units, which connect to the cloud management platform to achieve data synchronization and remote monitoring.

[0056] Beneficial effects:

[0057] 1. This invention uses an improved DS evidence theory and dynamic weight allocation to fuse the collected data and human risk levels. It also introduces a lightweight attention network to evaluate the confidence of each sensor data in real time and dynamically generate fusion weights, which significantly improves the robustness and accuracy of multi-source information fusion under complex working conditions.

[0058] 2. This invention also uses a safety knowledge graph to perform root cause reasoning on abnormal events, and combines human risk levels and equipment health status to participate in safety decision-making. The reasoning based on the knowledge graph is triggered online. When a sensor detects an abnormal pattern, root cause probability calculation is immediately performed. The reasoning result is not used as bypass information, but is rigorously quantified as a BPA in the DS evidence theory, and fused with other sensor evidence within the same mathematical framework to jointly determine the final "overall trust level." This trust level directly serves as one of the hard criteria for authorizing door opening. This means that the output of the knowledge graph can directly prevent a potentially dangerous door opening operation.

[0059] 3. This invention constructs a hybrid AI model that integrates an LSTM neural network and an XGBoost classifier. LSTM+XGBoost, specifically targeting the copper wire brush as a contact and conducting component, combines multi-dimensional heterogeneous features such as circuit current fluctuations, vibration RMS, and start-stop frequency to construct a time-series-static dual-channel model. The output is strictly mapped to the BPA in the DS evidence space, directly participating in safety interlocking decisions. Key components are evaluated and predicted, and an anomaly-driven online incremental learning mechanism is further integrated: when the model's predicted remaining service life deviates significantly from the actual replacement time, local parameter fine-tuning is automatically triggered, achieving lightweight adaptive updates at the edge and ensuring long-term effectiveness of prediction accuracy. Furthermore, the root cause reasoning results from the safety knowledge graph and the health status predicted by AI wear are quantified into basic probability allocation functions, embedded as auxiliary evidence sources within the DS fusion framework, jointly generating the floor door opening authorization criterion. This significantly improves the inherent safety, intelligence level, and predictive maintenance efficiency of the construction hoist operation. Attached Figure Description

[0060] Figure 1 This is a control flowchart of the intelligent safety interlocking system for construction hoists provided in an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the system hardware architecture according to an embodiment of the present invention.

[0062] In the diagram: 1. Landing door; 2. Landing door electric lock; 3. Control switch; 4. Copper sheet; 5. Cage; 6. Copper wire brush; 7. Wire; 8. Electronic door magnetic sensor; 9. Antenna; 10. Computer (including human safety perception module, self-diagnosis-self-repair edge reasoning engine, safety knowledge graph module, and emergency degradation strategy library); 11. Laser rangefinder sensor; 12. Embedded industrial computer (central intelligent control unit); 13. Visual recognition camera; 14. Three-axis accelerometer; 15. Gas concentration sensor; 16. Flexible tactile film sensor. Detailed Implementation

[0063] To make the technical solution of the present invention clearer and more complete, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] like Figure 1 As shown, this invention provides a method and device for intelligent safety interlocking control of construction hoists based on multi-source sensing and AI fusion. The control method includes the following steps:

[0065] Step 1: Acquire the following data using a multi-source sensing module: vertical distance between the hoist cage bottom plate and the current floor platform; hoist cage vibration acceleration; spatial alignment deviation between the hoist cage door and the floor door; concentration of combustible or toxic gases; clamping resistance during the closing process of the floor door and hoist cage door; and video stream of personnel behavior. The multi-source sensing module includes a laser rangefinder 11, a triaxial accelerometer 14, a visual recognition camera 13, a gas concentration sensor 15, and a flexible tactile film sensor 16. For specific settings, refer to [reference needed]. Figure 2 :

[0066] A laser rangefinder is installed at both the front and rear ends of the hoist cage bottom plate on the side facing the floor platform to measure the vertical distance between the hoist cage bottom plate and the current floor platform in real time. (Unit: mm). A triaxial accelerometer is installed at the center of gravity of the hoisting cage, with a sampling frequency of 100 Hz. This sensor collects vibration acceleration signals of the hoisting cage in the X (lateral), Y (longitudinal), and Z (vertical) directions, and calculates the root mean square (RMS) value as an indicator of docking stability. A high-definition industrial camera with a resolution of at least 1080P and a frame rate of 30 fps is installed on each floor facing the hoist cage docking area. This camera captures images of the relative positions of the hoist cage door and the landing door, and the alignment deviation is extracted using image processing algorithms. Meanwhile, the system reads information from the historical database, such as the cumulative number of start-stop cycles of the copper wire brush, the daily RMS mean value sequence of vibration over the past 30 days, the standard deviation sequence of circuit current fluctuations, the cumulative runtime, and the average ambient temperature, for inference in the AI ​​wear prediction model.

[0067] The original sensor data is filtered, denoised, and normalized; the door contour is extracted using an edge detection algorithm (such as Canny or Sobel), and the spatial alignment deviation is calculated in combination with calibration parameters; a sliding time window (such as N=30 days) is constructed to organize the time-series input feature matrix.

[0068] Step 2: Collect video streams of personnel behavior through high-definition cameras inside the cage, and use a lightweight behavior recognition model to determine whether there are dangerous behaviors such as leaning against the door, overcrowding, or carrying long poles out. Output the human risk level through the human safety perception module.

[0069] Human Factors Safety Perception Module Implementation Details: A wide-angle infrared camera is installed on the top of the cage. A lightweight behavior recognition model (MobileNetV3 + Temporal Shift Module) is deployed in the central intelligent control unit. It takes 5 consecutive frames of images as input and outputs the following behavior probabilities:

[0070] P1: Person leaning against the landing door (threshold > 0.85), P2: Number of people in the cage > 9 (overcrowding), P3: Holding a long pole out of the door. If any of these probabilities exceeds the threshold, a high-risk human factor signal is generated, and a "5-second delay + voice alarm" or "prohibit this door opening" action is forcibly added to the landing door opening authorization judgment. If the human factor risk level is high (e.g., leaning against the door is detected), even if all other sensor conditions are met, the system will still prohibit this door opening operation and play a voice alarm: "Do not lean against the landing door!"

[0071] Step 3: After fusing the collected data and the human risk level using a multi-source data fusion algorithm that combines an improved DS evidence theory with dynamic weight allocation, a comprehensive judgment is made as to whether the door opening conditions are met. Note that the human risk level must be below the high-risk threshold.

[0072] The multi-source data fusion algorithm also incorporates a safety knowledge graph module. Based on the current anomaly pattern, it infers the most likely root cause of the fault and inputs the current anomaly pattern into the pre-built safety knowledge graph module. This module constructs a four-element relationship graph of equipment, fault, environment, and human error based on historical maintenance and accident data. This graph is used for root cause inference of the current anomaly, outputting the probability distribution of the most likely root cause. This probability distribution is then used as auxiliary evidence in the multi-source data fusion process. When all opening conditions are met and the AI ​​wear prediction model used for predicting the wear state of the copper wire brush does not issue an anomaly warning, an unlocking command is sent to the landing door's electronic lock, allowing the landing door to be opened.

[0073] Implementation details of the security knowledge graph module: The system has a built-in security knowledge graph, which includes the following entities and relationships:

[0074] Physical components: copper wire brush, guide rail, brake, wind speed, humidity, overload operation.

[0075] Relationship: (High humidity leads to guide rail corrosion), (Frequent start-stop cycles exacerbate copper wire brush wear).

[0076] When the vibration RMS suddenly increases and the humidity > 80% RH, the graph inference engine outputs a probability of "rail corrosion" of 82%, which is then converted into a BPA function. ,in It is also input into the DS fusion module as the fourth source of evidence, and combined with sensor evidence. Perform Dempster synthesis; see the DS fusion process below for details.

[0077] The safety knowledge graph module is built based on historical accidents and maintenance records. It includes four types of entities: equipment components, environmental factors, human operations, and failure modes, as well as three types of relationships: 'cause', 'exacerbate', and 'mitigate'. When the system detects abnormal vibration and humidity is higher than the threshold, the graph inference engine outputs a probability of 'rail corrosion' of 0.82 and converts this value into a basic probability assignment function to participate in DS fusion.

[0078] Multi-source sensor data, human risk signals, AI health status probabilities, and knowledge graph inference results are collectively transformed into a basic probability allocation function (BPA), which is then fused using an improved Dempster-Shafer evidence theory to generate a comprehensive trust level. Door opening is authorized only when the comprehensive trust level exceeds a threshold and no high-risk human behavior is observed, with speed, floor, or time constraints implemented in emergency mode.

[0079] Step 4: Continuously monitor the physical closure status of the cage door and the landing door. Only when both doors are fully closed and locked should the cage start restriction be lifted.

[0080] All operation logs and video data are uploaded to a cloud management platform for retraining the AI ​​wear prediction model and remote monitoring. The cloud collaboration module supports federated learning to aggregate model gradients and receives OTA upgrades of the knowledge graph and degradation strategy library.

[0081] The conditions for opening the landing door include: vertical distance ≤ 150 mm; all three-axis acceleration RMS values ​​are less than the preset threshold; door alignment deviation < 10 mm; and the wear probability output by the AI ​​wear prediction model is lower than the set threshold.

[0082] An AI wear prediction model is used to evaluate the wear status of copper wire brushes online. This model is a dual-channel hybrid model that combines an LSTM neural network and an XGBoost classifier. The specific steps include:

[0083] S1: Input feature definition;

[0084] S2: The LSTM channel is used to learn the temporal evolution of component degradation. Its structure includes an input layer, an LSTM layer, a Dropout layer, and a fully connected layer, and outputs a hidden health state vector.

[0085] S3: The XGBoost channel is used to analyze environmental and usage intensity factors and output the probability distribution of each wear level;

[0086] S4: A learnable linear fusion layer is used to perform weighted fusion of the dual-channel output;

[0087] S5: The output layer classifies wear status into three levels: normal, warning, and fault.

[0088] The AI ​​wear prediction model performs online assessment of the wear status of copper wire brushes. It is a dual-channel hybrid model that combines an LSTM neural network and an XGBoost classifier. Its output is the probability distribution of the remaining service life. The specific steps are as follows:

[0089] P1: The probability of remaining lifespan being greater than 30 days, corresponding to wear level 0, which is normal;

[0090] P2: The probability that the remaining lifespan is between 10 and 30 days, corresponding to wear level 1, warning;

[0091] P3: The probability of remaining lifespan being less than 10 days, corresponding to wear level 2, or failure;

[0092] The probability distribution is transformed into a basic probability allocation function m4, which serves as the fourth source of evidence in multi-source data fusion, specifically defined as:

[0093] ;

[0094] ;

[0095] ;

[0096] Where A represents "the cage has been stably docked and aligned, meeting the door opening condition", ¬A represents "the door opening condition is not met", and Θ represents an uncertain state; simultaneously, the Shannon entropy H of the basic probability assignment is calculated:

[0097] ;

[0098] when When the probability of failure exceeds 70%, the door will be forcibly prohibited from opening, regardless of the data from other sensors.

[0099] when When the model height is uncertain but not high risk, activate the cautious operation mode, including: extending the stabilization waiting time to more than 5 seconds, limiting the cage load to no more than 80% of the rated value, and triggering a voice alarm.

[0100] The multi-source data fusion algorithm employs an improved DS evidence theory, mapping laser ranging, triaxial acceleration, and visual recognition data to basic probability allocation functions (BPAs). A lightweight attention network based on the current raw sensor data is used to generate confidence weights for each sensor in real time; these weights reflect the instantaneous reliability of each sensor under the current operating conditions. The specific steps are as follows:

[0101] S1: Define the propositional space ,in: This indicates that "the floor door can be opened safely". This indicates "there is a risk of opening the door," the complete series. The power set is .

[0102] S2: Based on the laser displacement, RMS acceleration, and visual recognition results, construct basic probability assignment functions for each sensor. .

[0103] (1) Laser rangefinder sensor, input distance Output BPA:

[0104] ;

[0105] (2) Triaxial accelerometer, input RMS acceleration Output BPA:

[0106] ;

[0107] (3) Visual recognition camera, input door alignment deviation Output BPA:

[0108] ;

[0109] S3: A lightweight attention network is invoked to concatenate the distance value output by the laser rangefinder, the RMS value calculated by the triaxial accelerometer, and the alignment deviation extracted by the visual recognition camera into a 3D vector. The attention network takes a time window segment of the original sensor data as input, inputs a fully connected layer containing 16 neurons, followed by a Softmax activation function, and outputs three normalized weights. dynamic confidence weight ,satisfy and The revised comprehensive evidence is obtained by weighting and averaging the BPAs.

[0110] ;

[0111] S4: If the security knowledge graph is triggered, quantify its root cause reasoning results into BPA. and with Perform Dempster orthogonal sum operation to obtain intermediate fusion results. Otherwise, ;

[0112] S5: If the AI ​​wear prediction model is enabled, obtain its output health status (BPA). and will and Perform Dempster combinatorial analysis to obtain the final basic probability distribution. Otherwise ;

[0113] S6: Computational Propositions Overall trust level is:

[0114] ;

[0115] in, To meet the trust level required for opening the door, For the degree of trust in an uncertain state, An uncertainty allocation coefficient, set to a fixed value of 0.5, is used to proportionally allocate uncertain evidence to supporting and opposing propositions.

[0116] S7: Set threshold ,like and If the door is opened, the electromagnetic lock on the landing door will be authorized; otherwise, opening the door will be prohibited and an audible and visual alarm will be triggered.

[0117] In addition, the system continuously records the actual timestamp of each copper wire brush replacement and compares it with the "remaining service life < 10 days" time previously predicted by the AI ​​wear prediction model to calculate the prediction residual. If three times in a row If the model experiences performance drift for a certain number of days (based on a preset tolerance threshold), an online incremental learning process is triggered: 1. Freeze the LSTM backbone layer and XGBoost tree structure; 2. Fine-tune the output layer weights and leaf node scores using only the data from the last 7 days; 3. After fine-tuning, update the local model version number and upload it to the cloud management platform for record-keeping. This mechanism significantly reduces the computing burden on edge devices while ensuring the long-term effectiveness of the model.

[0118] In this embodiment, the system only responds to the door opening request and drives the electromagnetic lock to release when the following conditions are met simultaneously: 1. Vertical distance 2. All three-axis acceleration RMS values ​​are less than 0.5 m / s²; 3. Door alignment deviation. 4. Overall Trust Level 5. AI wear prediction level ≤ Level 1.

[0119] In this embodiment, electronic door magnetic sensors are installed on the cage door and the landing door respectively to continuously monitor the opening and closing status of the two doors; the central control unit controls the main power circuit of the cage through a relay: if either door is not fully closed and locked, the relay is disconnected and the cage cannot be started; after both doors are closed, the relay is closed to release the start restriction; in the power-off state, the electromagnetic lock automatically locks to prevent illegal opening.

[0120] The central intelligent control unit incorporates a Kalman filter to fuse distance measurement results from laser ranging and visual recognition, improving positioning accuracy. The landing door's electric lock is a power-off self-locking electromagnetic lock, automatically locking and preventing opening in the event of a power outage. High-definition surveillance cameras installed inside the cage capture personnel entry and exit footage and behavioral characteristics, supporting local storage and real-time transmission to the cloud management platform via wireless communication modules, enabling remote real-time monitoring, event tracing, and human risk analysis. The central intelligent control unit incorporates a SHAP value analysis module or a LIME interpreter to visualize the decision-making basis of the AI ​​wear prediction model, improving interpretability and reliability.

[0121] If any subsystem failure is detected during operation, the self-diagnosis-self-repair edge inference engine process is initiated: First, it attempts to perform the corresponding repair operation, such as lens cleaning, temperature compensation, or model rollback. If the repair fails, the emergency degradation operation strategy in the emergency degradation strategy library that matches the fault type is activated, including limiting the running speed, prohibiting docking at high levels, or extending the stabilization waiting time.

[0122] Implementation details of self-diagnosis, self-repair, and emergency degradation mechanisms: The system performs a self-check on the data from each sensor every 10 seconds.

[0123] Laser ranging: fluctuation should be < ±2 mm when stationary; otherwise, it is marked as drift.

[0124] Vision module: Image signal-to-noise ratio < 20 dB is considered lens contamination;

[0125] AI wear prediction model: When the output probability entropy > 1.5, it is judged as having too high uncertainty.

[0126] Once an anomaly is detected, the system queries the preset repair strategy library, performs repair actions according to the fault type, and implements a downgrade strategy if the repair fails.

[0127] The intelligent safety interlocking device for construction hoists disclosed in this invention, based on multi-source sensing and AI fusion, further includes: an actuator comprising an electric lock for the landing door and a cage start lock, which opens and closes based on control commands issued by a central intelligent control unit; an electronic door magnetic system for detecting the opening and closing status of the cage door and landing door; and a monitoring and communication module comprising a high-definition monitoring camera and a wireless communication unit, connected to a cloud management platform to achieve data synchronization and remote monitoring.

[0128] In summary, this invention forms a comprehensive intelligent safety method encompassing the entire process of 'perception → human understanding → self-diagnosis → causal reasoning → decision-making → execution → feedback → optimization,' supporting continuous training and version upgrades of the AI ​​wear prediction model and enabling the system's self-evolution capability.

Claims

1. A method for intelligent safety interlocking control of construction hoists based on multi-source sensing and AI fusion, characterized in that, Includes the following steps: The system acquires the vertical distance between the bottom plate of the hoist cage and the current floor platform, the hoist cage vibration acceleration, the spatial alignment deviation between the hoist cage door and the floor door, the concentration of combustible or toxic gases, the clamping resistance during the closing process of the floor door and the hoist cage door, and the video stream of personnel behavior. Based on the video stream of human behavior, it is determined whether there are dangerous behaviors such as leaning against the door, exceeding the number of people, or carrying long poles and extending them, and the human risk level is output. The collected data and human risk levels are fused using a multi-source data fusion algorithm that combines an improved DS evidence theory with dynamic weight allocation to comprehensively determine whether the conditions for opening the floor door are met. The physical closure status of the cage door and the landing door is continuously monitored. The cage start restriction is lifted only when both doors are fully closed and locked.

2. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 1, characterized in that, The multi-source data fusion algorithm also incorporates a security knowledge graph module. Based on the current anomaly pattern, it infers the most likely root cause of the failure. The current anomaly pattern is then input into the pre-constructed security knowledge graph module, which outputs the probability distribution of the most likely root cause. This probability distribution is used as auxiliary evidence in the multi-source data fusion process. Specifically: The security knowledge graph includes the following entities and relationships: Physical components: copper wire brush, guide rail, brake, wind speed, humidity, overload operation; Relationship: High humidity leads to guide rail corrosion; frequent start-stop cycles exacerbate wear on the copper wire brush. When the vibration RMS suddenly increases and the humidity > 80% RH, the graph inference engine outputs the probability of "rail corrosion", and converts the result into a BPA function. It also serves as an auxiliary source of evidence in multi-source data fusion and is used in Dempster synthesis with sensor-based comprehensive evidence.

3. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 2, characterized in that, The multi-source data fusion algorithm also introduces an AI wear prediction model to predict the wear state of the copper wire brush. When all opening conditions are met and the AI ​​wear prediction model used for predicting the wear state of the copper wire brush does not issue an abnormal warning, an unlocking command is sent to the landing door's electric lock, allowing the landing door to be opened. The AI ​​wear prediction model adopts a dual-channel hybrid model that combines an LSTM neural network and an XGBoost classifier. Its output is the probability distribution of the remaining service life, specifically: P1: The probability that the remaining lifespan is greater than 30 days, corresponding to wear level 0; P2: The probability that the remaining lifespan is between 10 and 30 days, corresponding to wear level 1; P3: The probability that the remaining lifespan is less than 10 days, corresponding to wear level 2; The probability distribution is transformed into a basic probability allocation function m4, which serves as the fourth source of evidence in multi-source data fusion, specifically defined as: ； ； ； Where A represents "the cage has been stably docked and aligned, meeting the door opening condition", ¬A represents "the door opening condition is not met", and Θ represents an uncertain state; simultaneously, the Shannon entropy H of the basic probability assignment is calculated: ； when At that time, the door opening authorization judgment is forcibly prohibited from being opened; when When this occurs, the cautious operation mode is activated, which includes: extending the stop waiting time to more than 5 seconds, limiting the cage load to no more than 80% of the rated value, and triggering a voice alarm.

4. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 3, characterized in that, The conditions for opening the door include: vertical distance ≤ 150 mm; root mean square (RMS) values ​​of all three axes of acceleration less than a preset threshold; door alignment deviation less than 10 mm; and wear probability output by the AI ​​wear prediction model lower than a set threshold.

5. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 3, characterized in that, The multi-source data fusion algorithm employs an improved DS evidence theory, and the specific steps are as follows: S1: Define the propositional space ,in: This indicates that "the floor door can be opened safely". This indicates "there is a risk of opening the door," the complete series. The power set is ; S2: Based on the laser displacement, RMS acceleration, and visual recognition results, construct basic probability assignment functions for each sensor. ; (1) Laser rangefinder sensor, input distance Output BPA: ； (2) Triaxial accelerometer, input RMS acceleration Output BPA: ； (3) Visual recognition camera, input door alignment deviation Output BPA: ； S3: Invoke a lightweight attention network to generate dynamic confidence weights. The revised comprehensive evidence is obtained by weighting and averaging the BPAs. ； S4: If the security knowledge graph is triggered, quantify its root cause reasoning results into BPA. and with Perform Dempster orthogonal sum operation to obtain intermediate fusion results. Otherwise, ; S5: If the AI ​​wear prediction model is enabled, obtain its output health status (BPA). and will and Perform Dempster combinatorial analysis to obtain the final basic probability distribution. Otherwise ; S6: Computational Propositions Overall trust level is: ； in, To meet the trust level required for opening the door, For the degree of trust in an uncertain state, An uncertainty allocation coefficient, set to a fixed value of 0.5, is used to proportionally allocate uncertain evidence to supporting and opposing propositions. S7: Set threshold ,like and If the door is opened, the electromagnetic lock on the landing door will be authorized; otherwise, opening the door will be prohibited and an audible and visual alarm will be triggered.

6. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 5, characterized in that, The dynamic weights are generated by a lightweight attention network deployed on the edge controller; the attention network takes time window segments of the raw sensor data as input and outputs confidence weights corresponding to the laser displacement sensor, triaxial accelerometer, and visual recognition module. ,satisfy and .

7. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 3, characterized in that, The AI ​​wear prediction model also supports online incremental learning: continuously monitoring the residual between the actual component replacement time and the model's predicted remaining service life; when the absolute value of the residual exceeds the preset tolerance threshold three times in a row, a local fine-tuning mechanism is triggered, updating only the weights of the fully connected layer of the LSTM channel and the leaf node of the XGBoost channel, thereby achieving lightweight online correction of the model parameters.

8. The intelligent safety interlocking control method for construction hoists based on multi-source sensing and AI fusion as described in claim 3, characterized in that, The AI ​​wear prediction model adopts a transfer learning framework: the basic model is pre-trained in the cloud based on a large-scale historical running dataset, and the model is then distributed to the local device for fine-tuning based on recent running data.

9. A construction hoist intelligent safety interlocking device based on multi-source sensing and AI fusion, characterized in that, include: Multi-source sensing module: including laser rangefinder, triaxial accelerometer, visual recognition camera, gas concentration sensor and flexible tactile film sensor, which are used to collect the vertical distance between the bottom plate of the hoist cage and the current floor platform, the hoist cage vibration acceleration, the spatial alignment deviation between the hoist cage door and the floor door, the concentration of combustible or toxic gas, the clamping resistance during the closing process of the floor door and the hoist cage door, and the video stream of personnel behavior. AI Wear Prediction Module: Used to receive historical operating data and predict the wear status of copper wire brushes; Central intelligent control unit: Deployed in the hoist cage or floor control box, it is used to integrate the collected data and human risk level using a multi-source data fusion algorithm that combines improved DS evidence theory with dynamic weight allocation, and then combine the probability distribution of the most likely root cause of the failure with the prediction results of the AI ​​wear prediction module to comprehensively determine whether the conditions for opening the floor door are met. Actuators: including floor door electric locks and cage start locks, which are used to open and close based on control commands issued by the central intelligent control unit; Electronic door magnetic system: used to detect the opening and closing status of cage doors and landing doors; Monitoring and communication module: includes high-definition surveillance cameras and wireless communication units, which connect to the cloud management platform to achieve data synchronization and remote monitoring.

Images