A machine learning-based PIVAS link behavior safety monitoring method and system
By improving the YOLOv8 detection model and multi-frame analysis, and combining it with work process rules, automated behavioral safety monitoring of the PIVAS process was achieved, solving the problems of low efficiency and limited coverage in existing technologies, and improving monitoring coverage and anomaly response capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU FRONTIER MEDICAL PURIFICATION ENG CO LTD
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
Smart Images
Figure CN122337526A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and specifically to a PIVAS (Picture-Based Automation System) method and system for monitoring safety during process behavior based on machine learning. Background Technology
[0002] The Principal Intravenous Medication Preparation Center (PIVAS) is a crucial facility for the centralized preparation and quality control of intravenous medications in hospitals. It handles a large volume of injectable, infusion, and specialty medications. PIVAS operations involve multiple steps, including prescription review, medication dispensing, aseptic preparation, finished product verification, sorting, and packaging. The standardization of its operations directly impacts patient medication safety and medical quality. With the continuous growth in medical demand, the daily preparation volume of PIVAS is constantly increasing. The high workload, repetitive tasks, and high risk of staff fatigue pose significant challenges to behavioral safety monitoring.
[0003] Currently, behavioral safety monitoring in the PIVAS (Personalized Interchange Automation System) process mainly relies on two methods: manual inspection and video spot checks. Manual inspection involves quality control personnel regularly or irregularly entering the dispatching site to observe operations. They can accurately judge operational compliance based on their professional experience, but due to limitations in personnel numbers and inspection frequency, it's difficult to achieve full-time, full-process coverage. Video spot checks, as an important means of daily quality control, involve quality control personnel randomly selecting and reviewing operations at each workstation and in each batch to identify deviations. However, the amount of video footage that quality control personnel can review within their limited working time is far from covering all operational steps, resulting in a low actual spot check rate. Relying on manual review of each monitoring video requires a significant investment of attention and time, and is inefficient, making it difficult to achieve large-scale, routine monitoring of dispatching operations with limited quality control manpower. Summary of the Invention
[0004] The purpose of this invention is to address the problems mentioned in the background art, where existing behavioral safety monitoring methods mainly rely on manual inspection, resulting in low efficiency, limited coverage, and difficulty in adapting to large-scale, routine regulatory needs. Therefore, this invention proposes a machine learning-based PIVAS (Procedure for Monitoring Behavioral Safety) method and system for monitoring behavioral safety.
[0005] A first aspect of this invention provides a machine learning-based method for monitoring behavioral safety in PIVAS processes, the method comprising: Acquire the surveillance video stream of the work area and preprocess it to obtain the target video frame; The target video frame is input into a pre-trained detection model to detect key operational objects that appear during the operation, and a first detection result is obtained. The first detection result includes the category information and location information of the key operational objects. The key operational objects include one or more of the following: the operating subject object, the equipment object, the drug object, and the waste disposal object, which represent the PIVAS dispatching operation process. Based on the first detection result, the corresponding operation status is determined; Based on the operation status corresponding to the local consecutive frames of the target video frame, a comprehensive analysis is performed to determine the basic operation events; Based on the preset work process rules, the detected basic work events are correlated in time and their integrity is determined to obtain a second detection result; the second detection result includes normal and abnormal. When the second detection result is abnormal, an abnormal alarm message is sent to the review terminal.
[0006] Optionally, the detection model is an improvement based on YOLOv8; specific improvements include: The original C2f module is improved by using windmill-shaped convolution PConv to obtain the PC2f module; the C2f module in the backbone network is replaced with the PC2f module to enhance the ability to extract multi-directional features of targets with arbitrary orientation and improve the robustness of the model to targets with variable rotation. A preset feature enhancement module is inserted at the connection between the neck network and the detection head; the feature enhancement module adopts an attention mechanism to enhance the discriminativeness of the target features; A small target detection branch has been added to improve small target detection performance.
[0007] Optionally, the original C2f module consists of the standard convolutional module CBS and the bottleneck structure Bottleneck; The improvement of the original C2f module by using windmill-shaped convolution PConv to obtain the PC2f module includes: A windmill-shaped convolution PConv is inserted into the bottleneck structure, and the structure is transformed into a double residual connection to obtain an improved bottleneck structure, denoted as PBottleneck; the calculation process of PBottleneck includes: ; Where X0 is the input of PBottleneck; X2 is the output of PBottleneck; and X1 is the intermediate feature map generated during the operation. Replace each bottleneck structure Bottleneck in the C2f module with the PBottleneck to obtain the PC2f module.
[0008] Optionally, the calculation process of the feature enhancement module includes: ; Where Y0 is the input of the feature enhancement module; Y4 is the output of the feature enhancement module; GAP represents global average pooling; Conv 1×1 ReLU represents a standard convolution operation with a kernel size of 1×1; ReLU is an activation function. It is the sigmoid function; A C Y1 is the channel attention weight; Y1 is the channel enhanced feature map; DWConv 3×3 This indicates a depthwise separable convolution operation with a kernel size of 3×3; A S Y1 represents spatial attention weights; Y2 represents spatial augmentation feature maps; Concat represents channel concatenation; BN represents batch normalization.
[0009] Optionally, during model training, an improved CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the improved CIoU loss function is as follows: ; Among them, L ECIoU For improved CIoU loss; L CIoU This is the original CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. These are the preset focus parameters.
[0010] Optionally, determining the corresponding work status based on the first detection result includes: Based on the category combination relationship and / or spatial position relationship of the key operation objects in the target video frame, the corresponding operation status is determined by matching them with the preset basic operation conditions.
[0011] Optionally, the step of comprehensively analyzing the job status corresponding to local consecutive frames of the target video frame to determine the basic job events includes: Starting from the target video frame, obtain local continuous frames within a preset time window; Count the cumulative number of times the same job state appears in the local consecutive frames; When the cumulative number of times reaches a preset continuous threshold, a corresponding basic operation event is determined to have occurred.
[0012] Optionally, the step of performing time-series correlation and integrity determination on the detected basic work events according to preset work process rules to obtain the second detection result includes: The detected basic operation events are sorted according to their occurrence time to form the corresponding actual operation event sequence; Match the currently detected basic operation events with the preset operation process rules to determine the necessary preliminary steps; Backtrack through the actual operation event sequence to check whether the necessary preliminary steps have occurred; If the necessary preliminary steps have been taken, the second detection result is determined to be normal; otherwise, the second detection result is determined to be abnormal.
[0013] Optionally, the abnormal alarm information includes the abnormality type, the time of the abnormality, and the corresponding video frame information.
[0014] A second aspect of this invention provides a machine learning-based PIVAS (Procedure for Monitoring Behavioral Safety) monitoring system, the system comprising: The data acquisition module is used to acquire the monitoring video stream of the work area and perform preprocessing to obtain the target video frame; The target recognition module is used to input the target video frame into a pre-trained detection model to detect key operation objects appearing during the operation and obtain a first detection result; the first detection result includes category information and location information of the key operation objects; the key operation objects include one or more of the following: operation subject objects, equipment objects, drug objects, and waste disposal objects that characterize the PIVAS dispensing operation process; The status detection module is used to determine the corresponding operation status based on the first detection result; The event determination module is used to perform comprehensive analysis based on the job status corresponding to local consecutive frames of the target video frame to determine the basic job events; The event review module is used to perform time-series correlation and integrity determination on detected basic operation events according to preset operation process rules, and obtain a second detection result; the second detection result includes normal and abnormal. The anomaly reporting module is used to send an anomaly alarm message to the review terminal when the second detection result is abnormal.
[0015] The beneficial effects of this invention are: 1. By introducing image recognition technology to automatically identify key operational objects in surveillance videos, and combining it with post-processing analysis, automated monitoring and anomaly identification of PIVAS dispatching operations have been achieved, effectively reducing the pressure of manual review and improving monitoring coverage and anomaly response timeliness.
[0016] 2. Improvements to multiple modules have enhanced the detection accuracy of small targets such as needles and bottle caps, thereby significantly improving the accuracy and robustness of behavior recognition in complex environments.
[0017] 3. An improved ECIoU loss function is adopted, which introduces a focusing parameter on the basis of CIoU loss, making the model pay more attention to hard-to-distinguish samples, and further improving the accuracy and convergence speed of bounding box regression.
[0018] 4. By mapping single-frame detection results to job status and confirming basic job events based on multi-frame statistics, false event triggering caused by single-frame misdetection is effectively avoided. Furthermore, by combining preset job process rules for time sequence association and pre-step verification, intelligent judgment of operational compliance is achieved, significantly reducing the false alarm rate. Attached Figure Description
[0019] Figure 1 A flowchart illustrating a machine learning-based PIVAS (Procedure for Monitoring Behavioral Safety) method for safety monitoring in an embodiment of the present invention; Figure 2 This invention provides an original network architecture diagram of a YOLOv8 model. Figure 3 A network architecture diagram of a PA-YOLO model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a PC2f module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an FEM module provided in an embodiment of the present invention; Figure 6 This is an architecture diagram of a machine learning-based PIVAS process behavior safety monitoring system provided in an embodiment of the present invention. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.
[0021] This invention provides a machine learning-based method for monitoring behavioral safety in PIVAS (Picture-Based Automation System). See also... Figure 1 , Figure 1 A flowchart illustrating a machine learning-based PIVAS (Procedure for Monitoring Behavioral Safety) process safety according to an embodiment of the present invention. The method includes the following steps: S101: Acquire the monitoring video stream of the work area and perform preprocessing to obtain the target video frame.
[0022] S102, input the target video frame into the pre-trained detection model to detect key work objects that appear during the work process and obtain the first detection result.
[0023] S103, Based on the first detection result, determine the corresponding operation status.
[0024] S104. Based on the operation status corresponding to the local continuous frames of the target video frame, perform comprehensive analysis to determine the basic operation events.
[0025] S105, according to the preset work process rules, perform time sequence correlation and integrity judgment on the detected basic work events to obtain the second detection result.
[0026] S106, When the second detection result is abnormal, send an abnormal alarm message to the review terminal.
[0027] The first detection result includes the category and location information of key operational objects. Key operational objects include the main objects representing the PIVAS dispensing process (hands, face, body, etc.), instruments (syringes, grinding wheels, sterile cotton balls, etc.), medicines (vials, ampoules, etc.), and waste disposal objects (sharps containers, medical waste bins, etc.). The second detection result includes normal and abnormal results; abnormal alarm information includes the abnormality type, the time of occurrence of the abnormality, and the corresponding video frame information.
[0028] The present invention provides a machine learning-based method for monitoring the safety of PIVAS (Portable Intravenous Drug Administration) processes. This method uses image recognition technology to automatically identify key operational objects in monitoring videos, combines multi-frame judgment of basic operational events, and performs integrity verification according to operational process rules. This achieves automated monitoring and anomaly identification of PIVAS dispensing operations, effectively reducing the burden of manual review, improving monitoring coverage and anomaly response timeliness, and ensuring the standardization and safety of intravenous drug dispensing operations.
[0029] In one embodiment, preprocessing includes operations such as video stream decoding, frame extraction, size normalization, and pixel normalization. Specifically, image frames are extracted from the surveillance video stream at a fixed frame rate (e.g., 5 frames per second), the images are scaled to the size required by the detection model input (e.g., 640×640), and the pixel values are normalized to obtain the target video frame.
[0030] In one embodiment, the detection model is an improvement upon YOLOv8, denoted as PA-YOLO. See also Figure 2 , Figure 2This diagram illustrates the original network architecture of a YOLOv8 model provided in this embodiment of the invention. YOLOv8 includes a backbone network, a neck network, and a head. The backbone network extracts features from the input image, the neck network performs feature fusion, and the head performs final object detection based on the features output by the neck network, including category, bounding box location, and confidence score. The modules CBS, C2f, SPPF, Upsample, Concat, and Detect shown in the diagram are inherent to the YOLOv8 model and will not be elaborated upon here.
[0031] See Figure 3 , Figure 3 This invention provides a network architecture diagram for a PA-YOLO model. Compared to the original YOLOv8 model, the improvements made in this invention include: 1. An improved C2f module, denoted as PC2f module, is proposed. The C2f module in the backbone network is replaced with a PC2f module to enhance the ability to extract multi-directional features from targets with arbitrary orientations and improve the model's robustness to targets with varying rotation. Specifically, the original C2f module is improved using windmill-shaped convolution (PConv) to obtain the PC2f module.
[0032] See Figure 4 , Figure 4 This is a schematic diagram of the structure of a PC2f module provided in an embodiment of the present invention. In the figure, CBS represents Conv (convolution) + BN (batch normalization) + SiLU (activation function); Split represents channel splitting; Bottleneck is a bottleneck structure; Concat represents channel concatenation; Add represents element-wise addition; PBottleneck represents an improved bottleneck structure; and PConv represents a windmill-shaped convolution.
[0033] Figure 4 (a) shows the structure of the original C2f module, which consists of the standard convolutional module CBS and the bottleneck structure Bottleneck; Figure 4 (b) shows the original structure of Bottleneck. Figure 4 As shown in (c), a windmill-shaped convolution PConv is inserted into the bottleneck structure, and the structure is transformed into a double residual connection, thus obtaining the improved bottleneck structure, denoted as PBottleneck. The calculation process of PBottleneck includes: ; Where X0 is the input of PBottleneck; X2 is the output of PBottleneck; and X1 is the intermediate feature map generated during the operation.
[0034] like Figure 4 As shown in (d), replacing each bottleneck structure Bottleneck in the C2f module with PBottleneck yields the PC2f module.
[0035] The windmill-shaped convolution employs a multi-directional convolution branch parallel structure, extracting features along the horizontal, vertical, and diagonal directions to achieve multi-directional feature modeling for targets with arbitrary orientations. This embodiment, by introducing windmill-shaped convolution into the bottleneck structure, enhances the model's ability to extract features from textures, edge contours, and fine-grained local structures in different directions, thereby improving the detection accuracy and localization accuracy of slender and small targets in PIVAS operation scenarios. Simultaneously, the construction of dual residual connections ensures smooth gradient flow, maintaining low computational overhead while improving detection accuracy.
[0036] 2. A Feature Enhancement Module (FEM) is proposed. This feature enhancement module is inserted at the connection between the neck network and the detection head to enhance the features to be input to the detection head, thereby improving the discriminative power of the target features.
[0037] See Figure 5 , Figure 5 This is a schematic diagram of the structure of an FEM module provided in an embodiment of the present invention. In the figure, GAP represents global average pooling; 1×1Conv represents a standard convolution operation with a kernel size of 1×1; ReLU is the activation function; Sigmoid is the normalization function; Mul represents element-wise multiplication; and 3×3DWConv represents a depthwise separable convolution operation with a kernel size of 3×3.
[0038] The calculation process of this feature enhancement module includes: ; Where Y0 is the input of the feature enhancement module; Y4 is the output of the feature enhancement module; It is the Sigmoid function; A C Y1 is the channel attention weight; Y2 is the channel enhanced feature map; A S Y1 is the spatial attention weight; Y2 is the spatial enhancement feature map; Y3 is the attention fusion feature map.
[0039] In this embodiment, the channel attention branch can adaptively enhance the semantic response related to the key task object, suppress background noise and irrelevant feature interference, and the spatial attention branch can highlight the spatial response of the target area and enhance the ability to express the target edge contour, local structure and position information. After the two are fused, the category discrimination ability and localization sensitivity of the input detection head features can be improved at the same time. Meanwhile, the original feature information is preserved by the residual method, which can reduce information loss while strengthening key features and improve the stability and robustness of the feature enhancement process.
[0040] 3. Add a small target detection branch (see...) Figure 3 The part within the dashed box corresponds to a high-resolution feature map, which reduces the loss of small target features during downsampling and improves the detection accuracy of small targets such as needles, cotton balls, and bottle caps.
[0041] 4. Based on CIoU loss and incorporating the focusing concept, an ECIoU loss is proposed. Specifically, the formula for calculating ECIoU loss is: ; Among them, L ECIoU For improved CIoU loss; L CIoU This is the original CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. These are preset focus parameters, which can be adjusted between 0.5 and 3.
[0042] This embodiment introduces a focusing term, assigning higher optimization weights to low IoU samples during training and appropriately reducing the focus on high IoU samples. This concentrates training efforts on samples with larger localization errors and higher regression difficulty, improving the targeting and effectiveness of the bounding box regression process. Compared to directly using CIoU loss, this embodiment can accelerate the convergence speed of difficult samples, reduce the dominant role of easy samples in gradient updates, and improve the model's localization accuracy and regression stability for small targets, occluded targets, and targets with blurred boundaries. It is particularly suitable for the fine localization of key objects such as syringes, ampoules, grinding wheels, and cotton balls in PIVAS operation scenarios, thereby further improving the accuracy and robustness of positional information in the detection results.
[0043] To verify the performance of the detection model (PA-YOLO) proposed in this invention, a comparative experiment was conducted between the improved model and the original YOLOv8n model. The experiment was built using the PyTorch deep learning framework, and all training and validation were performed on the same hardware platform. The experimental parameters were configured as follows: learning rate set to 0.001, batch size set to 16, number of iterations set to 100, and the optimizer used was Adam. After training and validation on the same dataset, the performance metrics of the PA-YOLO model proposed in this invention significantly outperformed the original YOLOv8n model. Specifically, precision increased from 85.3% to 92.1%, recall increased from 81.7% to 88.4%, and mAP@0.5 increased from 85.1% to 91.6%. This indicates that PA-YOLO can more accurately identify key work objects such as syringes, ampoules, grinding wheels, and cotton balls in PIVAS work scene detection.
[0044] In one embodiment, step S103, determining the corresponding job status based on the first detection result, includes: matching the category combination relationship and / or spatial position relationship of key job objects in the target video frame with preset basic job conditions to determine the corresponding job status.
[0045] In one implementation, the preset basic operating conditions are stored in the form of a rule base. For example, when a hand, syringe, and vial are detected, and the distance between the hand and the syringe, and the distance between the syringe and the vial are both less than the corresponding contact determination threshold, the operation is determined to be "transferring medication to vial"; when a hand, grinding wheel, and ampoule are detected, and the distance between the hand and the ampoule, and the distance between the ampoule and the grinding wheel are both less than the corresponding contact determination threshold, the operation is determined to be "opening bottle to ampoule".
[0046] This embodiment further transforms the target detection results into operational status information with clear business semantics, enabling the system not only to identify what objects have appeared, but also to identify what operational relationships are occurring between the objects, thereby effectively improving the interpretability and business usability of the video monitoring results.
[0047] In one embodiment, step S104 involves performing a comprehensive analysis based on the job status corresponding to the local consecutive frames of the target video frame to determine the basic job events, including: Step 1: Starting from the target video frame, acquire local continuous frames within a preset time window. The preset time window can be determined by the operation status corresponding to the target video frame; for example, 3 seconds can be used for disinfection, and 5 seconds can be used for transferring medicine.
[0048] Step 2: Count the cumulative number of times the same job state appears in local consecutive frames.
[0049] Step 3: When the cumulative number of occurrences reaches a preset duration threshold, determine that the corresponding basic job event has occurred. The duration threshold can be set to half the number of frames within the time window.
[0050] This embodiment uses multi-frame event determination to suppress false detections caused by single-frame errors, short-term occlusion, and state jitter, thereby improving the stability and accuracy of event recognition and providing more reliable event input for subsequent process integrity determination and abnormal alarms.
[0051] In one embodiment, step S105, based on preset work process rules, performs time-series correlation and integrity determination on the detected basic work events to obtain a second detection result, including: Step 1: Sort the detected basic operation events according to their occurrence time to form the corresponding actual operation event sequence.
[0052] Step two involves matching the currently detected basic operational events with the preset operational process rules to determine the necessary prerequisite steps. The preset operational process rules are defined by the PIVAS Standard Operating Procedures (SOPs). For example, before the "Transfer Medication - Vial" event, there should be a "Disinfect - Vial Stopper" event.
[0053] Step 3: Backtrack through the actual operation event sequence to check whether the necessary prerequisite steps have occurred.
[0054] Step 4: If the necessary preliminary steps have been completed, the second test result is determined to be normal; otherwise, the second test result is determined to be abnormal.
[0055] This embodiment matches and checks detected job events with job process rules to determine whether the operation follows the standard process, thereby effectively identifying missing or abnormal steps, improving the standardization and reliability of job behavior monitoring, and generating abnormal alarms in a timely manner to improve the work efficiency of reviewers.
[0056] This invention provides a PIVAS (Procedure for Monitoring Behavioral Safety) system based on machine learning. See also... Figure 6 , Figure 6 This invention provides an architecture diagram of a machine learning-based PIVAS (Procedure for Monitoring Behavioral Safety) system for safety monitoring. The system includes: The data acquisition module is used to acquire the monitoring video stream of the work area and perform preprocessing to obtain the target video frame.
[0057] The target recognition module is used to input target video frames into a pre-trained detection model to detect key work objects that appear during the operation and obtain the first detection result.
[0058] The status detection module is used to determine the corresponding job status based on the first detection result.
[0059] The event determination module is used to perform comprehensive analysis based on the job status corresponding to local consecutive frames of the target video frame to determine the basic job events.
[0060] The event review module is used to perform time-series correlation and integrity judgment on the detected basic operation events according to the preset operation process rules, and obtain the second detection result.
[0061] The anomaly reporting module is used to send anomaly alarm information to the review terminal when the second detection result is abnormal.
[0062] The PIVAS process behavior safety monitoring system based on machine learning provided in this invention automatically identifies key operation objects in the monitoring video through a target detection model, determines basic operation events by combining multiple frames, and performs integrity verification according to the operation process rules. This realizes automated monitoring and anomaly identification of PIVAS dispensing operations, effectively reducing the pressure of manual review, improving monitoring coverage and anomaly response time, and ensuring the standardization and safety of intravenous medication dispensing operations.
[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.
Claims
1. A method for monitoring PIVAS link behavior safety based on machine learning, characterized in that, The method includes: Acquire the surveillance video stream of the work area and preprocess it to obtain the target video frame; The target video frame is input into a pre-trained detection model to detect key operational objects that appear during the operation, and a first detection result is obtained. The first detection result includes the category information and location information of the key operational objects. The key operational objects include one or more of the following: the operating subject object, the equipment object, the drug object, and the waste disposal object, which represent the PIVAS dispatching operation process. Based on the first detection result, the corresponding operation status is determined; Based on the operation status corresponding to the local consecutive frames of the target video frame, a comprehensive analysis is performed to determine the basic operation events; Based on the preset work process rules, the detected basic work events are correlated in time and their integrity is determined to obtain a second detection result; the second detection result includes normal and abnormal. When the second detection result is abnormal, an abnormal alarm message is sent to the review terminal.
2. The machine learning based PIVAS link behavior security monitoring method according to claim 1, wherein, The detection model is an improvement upon YOLOv8; the specific improvements include: The original C2f module is improved by using windmill-shaped convolution PConv to obtain the PC2f module; the C2f module in the backbone network is replaced with the PC2f module to enhance the ability to extract multi-directional features of targets with arbitrary orientation and improve the robustness of the model to targets with variable rotation. A preset feature enhancement module is inserted at the connection between the neck network and the detection head; the feature enhancement module adopts an attention mechanism to enhance the discriminativeness of the target features; A small target detection branch has been added to improve small target detection performance.
3. The machine learning based PIVAS link behavior security monitoring method according to claim 2, wherein, The original C2f module consists of the standard convolutional module CBS and the bottleneck structure Bottleneck; The improvement of the original C2f module by using windmill-shaped convolution PConv to obtain the PC2f module includes: A windmill-shaped convolution PConv is inserted into the bottleneck structure, and the structure is transformed into a double residual connection to obtain an improved bottleneck structure, denoted as PBottleneck; the calculation process of PBottleneck includes: ； Where X0 is the input of PBottleneck; X2 is the output of PBottleneck; and X1 is the intermediate feature map generated during the operation. Replace each bottleneck structure Bottleneck in the C2f module with the PBottleneck to obtain the PC2f module.
4. The machine learning based PIVAS link behavior security monitoring method according to claim 2, wherein, The calculation process of the feature enhancement module includes: ； wherein Y0 is the input of the feature enhancement module; Y4 is the output of the feature enhancement module; GAP represents global average pooling; Conv 1×1 represents a standard convolution operation with a kernel size of 1x1; ReLU is an activation function; is a sigmoid function; A C is a channel attention weight; Y1 is a channel-enhanced feature map; DWConv 3×3 represents a depthwise separable convolution operation with a kernel size of 3x3; A S is a spatial attention weight; Y2 is a spatial-enhanced feature map; Concat represents channel concatenation; and BN represents batch normalization.
5. The machine learning based PIVAS link behavior security monitoring method according to claim 2, wherein, During model training, an improved CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the improved CIoU loss function is as follows: ； wherein, L ECIoU is the improved CIoU loss; L CIoU is the original CIoU loss; IoU is the intersection over union of the predicted box and the real box; is a preset focus parameter.
6. The PIVAS process behavior safety monitoring method based on machine learning according to claim 1, characterized in that, Determining the corresponding work status based on the first detection result includes: Based on the category combination relationship and / or spatial position relationship of the key operation objects in the target video frame, the corresponding operation status is determined by matching them with the preset basic operation conditions.
7. The PIVAS process behavior safety monitoring method based on machine learning according to claim 6, characterized in that, The step of comprehensively analyzing the job status corresponding to local consecutive frames of the target video frame to determine the basic job events includes: Starting from the target video frame, obtain local continuous frames within a preset time window; Count the cumulative number of times the same job state appears in the local consecutive frames; When the cumulative number of times reaches a preset continuous threshold, a corresponding basic operation event is determined to have occurred.
8. The PIVAS process behavior safety monitoring method based on machine learning according to claim 1, characterized in that, The step of performing time-series correlation and integrity determination on the detected basic work events according to preset work process rules to obtain the second detection result includes: The detected basic operation events are sorted according to their occurrence time to form the corresponding actual operation event sequence; Match the currently detected basic operation events with the preset operation process rules to determine the necessary preliminary steps; Backtrack through the actual operation event sequence to check whether the necessary preliminary steps have occurred; If the necessary preliminary steps have been taken, the second detection result is determined to be normal; otherwise, the second detection result is determined to be abnormal.
9. The PIVAS process behavior safety monitoring method based on machine learning according to claim 1, characterized in that, The abnormal alarm information includes the abnormality type, the time of the abnormality, and the corresponding video frame information.
10. A PIVAS (Procedure for Monitoring Behavioral Safety) system based on machine learning, characterized in that, The system includes: The data acquisition module is used to acquire the monitoring video stream of the work area and perform preprocessing to obtain the target video frame; The target recognition module is used to input the target video frame into a pre-trained detection model to detect key operation objects appearing during the operation and obtain a first detection result; the first detection result includes category information and location information of the key operation objects; the key operation objects include one or more of the following: operation subject objects, equipment objects, drug objects, and waste disposal objects that characterize the PIVAS dispensing operation process; The status detection module is used to determine the corresponding operation status based on the first detection result; The event determination module is used to perform comprehensive analysis based on the job status corresponding to local consecutive frames of the target video frame to determine the basic job events; The event review module is used to perform time-series correlation and integrity determination on detected basic operation events according to preset operation process rules, and obtain a second detection result; the second detection result includes normal and abnormal. The anomaly reporting module is used to send an anomaly alarm message to the review terminal when the second detection result is abnormal.