Medical production personnel changing disinfection compliance identification method, device, equipment and medium

Abstract

The application discloses a medical production personnel changing and disinfecting compliance identification method and device, equipment and medium, relates to the technical field of medical intelligent manufacturing, and comprises the following steps: after receiving the edge node encrypted transmission of the monitoring video segment, decrypting and caching to the video stream inference scheduler as an inference request; the monitoring video segment is uniformly sampled and spliced from the ring buffer; each ring buffer is used for buffering each preprocessed video stream of the medical production personnel changing and disinfecting process; the video stream inference scheduler is used for request scheduling processing and input into the preset inference acceleration engine for inferring and optimizing the preset action detection and classification large model; a plurality of candidate action segment proposals are generated for the inference input video sequence by using the large model, and action detection and classification are carried out; based on the preset process compliance determination engine and according to the action detection result, it is judged whether there is a rule-changing disinfection behavior, if there is, the corresponding alarm operation is triggered.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology in the pharmaceutical industry, and in particular to a method, device, equipment, and medium for identifying compliance in the changing and disinfection of pharmaceutical production personnel. Background Technology

[0002] In the pharmaceutical manufacturing industry, personnel are one of the primary sources of contamination in clean areas. All personnel entering clean areas must undergo standardized gowning and disinfection procedures to minimize the contamination of the production environment by particles and microorganisms carried by their bodies. The standardization of the gowning process directly affects the sterility assurance level (SAL) of pharmaceutical products and is a key inspection item in pharmaceutical quality supervision and auditing. Currently, domestic pharmaceutical companies mainly rely on manual inspections and regular training and assessments to supervise gowning and disinfection procedures. However, the frequency of inspections is limited, the judgment standards of inspectors vary subjectively, and supervisory gaps easily occur at night or during overtime work hours. Overall, the supervisory effect is extremely limited, and violations occur frequently.

[0003] In conclusion, optimizing the methods for identifying compliance issues related to gowning and disinfection for pharmaceutical production personnel, thereby improving the accuracy of identifying violations and enhancing the effectiveness of supervision over gowning and disinfection practices, is a pressing issue that needs to be addressed. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for identifying compliance with gown disinfection regulations among pharmaceutical production personnel. This method optimizes the identification of compliance issues related to gown disinfection, thereby improving the accuracy of identifying violations and enhancing the effectiveness of monitoring gown disinfection procedures. The specific solution is as follows: Firstly, this application provides a method for compliant identification of medical production personnel changing clothes and disinfection procedures, applied to a cloud server, including: After receiving the encrypted monitoring video segment transmitted by the edge node, the monitoring video segment is decrypted and cached as an inference request in the video stream inference scheduler; the monitoring video segment is a video segment obtained by the edge node by uniformly sampling from the circular buffer based on a preset sampling period and splicing it together; each of the circular buffers is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process; the preprocessed video stream is a video stream obtained by the edge node by preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process; The video stream inference scheduler schedules the cached inference requests to obtain the inference input video sequence, and then inputs the inference input video sequence into the preset inference acceleration engine. The preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model. Using the preset action detection and classification model in the preset inference acceleration engine, multiple candidate action segment proposals are generated for the inference input video sequence, and action detection and classification are performed on the candidate action segment proposals to obtain the corresponding action detection results. Based on the preset process compliance judgment engine and the action detection results, it is determined whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfecting. If so, the corresponding alarm operation is triggered. The violation of the rules for changing clothes and disinfecting refers to changing clothes and disinfecting that does not meet the preset standard conditions.

[0005] Optionally, the preprocessing of the monitoring video stream of the pharmaceutical production personnel changing clothes and disinfecting includes: The monitoring video stream of the pharmaceutical production personnel changing clothes and disinfecting is preprocessed by decoding, frame extraction, resolution normalization and temporal sampling; the monitoring video stream is any one of the multiple video streams collected in real time by the edge node through each industrial camera deployed in the changing room passage of the pharmaceutical production clean area.

[0006] Optionally, the scheduling process for the cached inference request includes: Based on the location of the changing area monitored by the surveillance video segment in the inference request, determine the corresponding scheduling priority; The inference requests with the same cache time and scheduling priority are merged into a batch request for scheduling processing.

[0007] Optionally, the inference optimization of the preset action detection and classification model includes: The preset action detection classification model is imported using the built-in exchange format parser of the preset inference acceleration engine; the neural network exchange format of the preset action detection classification model is consistent with the exchange format parsed by the exchange format parser. The preset motion detection classification model is subjected to layer fusion and quantization accuracy calibration to obtain the processed preset motion detection classification model. The corresponding kernel tuning scheme is selected according to the graphics processor model of the cloud server, and inference optimization is performed based on the processed preset motion detection classification model, the kernel tuning scheme and the memory pool reuse mechanism.

[0008] Optionally, the training process of the preset action detection and classification large model includes: Using a pre-set human motion video dataset that meets the pre-set large-scale conditions, the initial motion detection and classification network is pre-trained to obtain a pre-trained large-scale motion detection and classification model. Based on a pre-built medical changing action dataset and through a hierarchical learning rate decay mechanism, the pre-trained action detection and classification model is adaptively fine-tuned with a preset number of layers before freezing, so as to obtain the preset action detection and classification model. The medical changing action dataset is a dataset composed of standard changing videos that have been edited, labeled, and augmented; the standard changing videos are videos of different people performing changing and disinfection procedures that meet preset standard conditions.

[0009] Optionally, the step of generating multiple candidate action segment proposals for the inference input video sequence, and performing action detection and classification on the candidate action segment proposals to obtain corresponding action detection results includes: The inference input video sequence is divided into several video slices, and the probability of human motion in each of the several video slices is quantified to obtain the corresponding motion score; the motion score is positively correlated with the probability of human motion. Determine the set of target video slices in the inference input video sequence in which the action score is continuously higher than a preset score threshold, and cluster the set of target video slices into candidate action segment proposals; The spatiotemporal features of the candidate action segment proposals are extracted, and the execution stage of human action is modeled using structured temporal pyramid pooling to obtain the corresponding modeling results. The action category and action integrity are determined by using the dual-branch discriminant head within the preset action detection classification model and based on the modeling results, so as to obtain the corresponding action detection results; the dual-branch discriminant head includes an action classification head and an action integrity quantization head.

[0010] Optionally, the preset process compliance determination engine is built based on a deterministic finite state machine, including a state transition relationship with temporal constraints obtained by modeling the changing and disinfection process that meets preset standard conditions. Accordingly, determining whether current pharmaceutical production personnel have engaged in unauthorized changing and disinfection practices includes: A timeline record of the changing and disinfection process for each pharmaceutical production worker; If the timeline records representing the transition relationship of the current changing and disinfection behavior do not conform to the state transition relationship, it is determined that the current pharmaceutical production personnel have violated regulations regarding changing and disinfection behavior.

[0011] Secondly, this application provides a compliance identification device for pharmaceutical production personnel changing clothes and disinfecting, applied to a cloud server, including: The request caching module is used to decrypt the monitoring video segments transmitted encrypted by the edge node and cache them as inference requests to the video stream inference scheduler after receiving them. The monitoring video segments are video segments that the edge node uniformly samples from a circular buffer and splices based on a preset sampling period. Each circular buffer is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process. The preprocessed video stream is the video stream obtained by the edge node preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process. The sequence input module is used to schedule and process the cached inference requests through the video stream inference scheduler to obtain the inference input video sequence, and input the inference input video sequence into the preset inference acceleration engine; the preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model; The action detection module is used to generate multiple candidate action segment proposals for the inference input video sequence using the preset action detection classification model in the preset inference acceleration engine, and to perform action detection and classification on the candidate action segment proposals to obtain the corresponding action detection results. The operation triggering module is used to determine whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfection based on the preset process compliance judgment engine and the action detection results. If so, the module will trigger the corresponding alarm operation. The violation of the rules for changing clothes and disinfection refers to changing clothes and disinfection that does not meet the preset standard conditions.

[0012] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for compliant identification of pharmaceutical production personnel changing clothes and disinfecting.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for compliant identification of medical production personnel changing clothes and disinfection.

[0014] In this application, after receiving the encrypted monitoring video segment transmitted by the edge node, the monitoring video segment is decrypted and cached as an inference request in the video stream inference scheduler; the monitoring video segment is a video segment obtained by the edge node uniformly sampling and splicing from a circular buffer based on a preset sampling period; each circular buffer is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process; the preprocessed video stream is a video stream obtained by the edge node preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process; the cached inference request is scheduled and processed by the video stream inference scheduler to obtain the inference input video sequence, and the inference input video sequence is input to the preset inference scheduler. In the reasoning acceleration engine, a preset action detection and classification model is deployed to optimize the inference process. Using this model, multiple candidate action segment proposals are generated for the input video sequence, and action detection and classification are performed on these proposals to obtain corresponding action detection results. Based on a preset process compliance judgment engine and the action detection results, it is determined whether the current pharmaceutical production personnel have engaged in unauthorized changing and disinfection procedures. If so, a corresponding alarm is triggered. These unauthorized procedures refer to changing and disinfection behaviors that do not meet preset standard conditions. As can be seen from the above, the edge nodes of this application first preprocess the monitoring video streams of the changing and disinfection process of various medical production personnel. The preprocessed video streams are cached through their respective circular buffers, and then uniformly sampled and spliced ​​into monitoring video segments according to a preset sampling period. After encryption, the video segments are transmitted to the cloud server. The cloud server receives the monitoring video segments, decrypts them, and stores them in the video stream inference scheduler. The video stream inference scheduler uniformly schedules the cached inference requests to obtain the inference input video sequence, which is then sent to the built-in preset action detection classification model. The preset inference acceleration engine is used to optimize the inference process. The preset inference acceleration engine generates multiple candidate action segment proposals, performs action detection classification, and outputs action detection results. The action detection results are sent to the preset process compliance judgment engine to identify illegal changing and disinfection behaviors. If an illegal behavior is detected, an alarm is triggered.In this way, through the process described in this application, the cloud server decrypts the encrypted monitoring video segments and hands them over to the video stream inference scheduler for centralized management of all inference requests. The scheduler can allocate computing power in a coordinated manner, improve hardware resource utilization, and reduce waiting delays caused by concurrent inference of multiple video streams. It configures an inference acceleration engine and builds in an optimized motion detection and classification model for inference optimization, shortening the time for motion recognition and adapting to the real-time analysis needs of multiple video streams in the workshop. It outputs candidate motion segment proposals and then performs detection and classification, performing fine recognition only on valid segments containing human motion, reducing unnecessary computing power consumption. Through a preset process compliance judgment engine combined with motion detection results, it automatically verifies the changing and disinfection procedures, enabling accurate capture of violations of changing and disinfection procedures around the clock. When a violation is detected, an alarm is automatically triggered, realizing unmanned real-time supervision of the medical changing and disinfection process, reducing the risk of oversight by manual inspections, and thus optimizing the method for identifying compliance of medical production personnel changing and disinfection procedures, thereby improving the accuracy of identifying violations and optimizing the supervision effect of changing and disinfection behavior. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This application discloses a flowchart of a method for compliant identification of pharmaceutical production personnel changing clothes and disinfecting. Figure 2 This application discloses a system architecture diagram for a method of identifying compliance in the changing and disinfection of pharmaceutical production personnel. Figure 3 This is a network architecture diagram of an initial action detection and classification network disclosed in this application; Figure 4 This application discloses a finite state machine state transition diagram for determining the compliance of a changing and disinfection process. Figure 5 This is a schematic diagram of a compliance identification device for changing clothes and disinfection of pharmaceutical production personnel disclosed in this application. Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The standardization of the changing-room procedure directly affects the sterility assurance level of pharmaceuticals and is a key inspection item in pharmaceutical quality supervision and auditing. Currently, domestic pharmaceutical companies mainly rely on manual inspections and regular training and assessments to supervise the changing-room disinfection process. However, the frequency of inspections is limited, the judgment standards of inspectors vary subjectively, and supervision gaps easily occur at night or during overtime hours. As a result, the overall supervision effect is extremely limited, and violations occur frequently.

[0019] To overcome the aforementioned technical problems, this application provides a method for identifying compliance with the changing and disinfection procedures of pharmaceutical production personnel. This method optimizes the identification of compliance with the changing and disinfection procedures of pharmaceutical production personnel, thereby improving the accuracy of identifying violations and enhancing the supervisory effect of changing and disinfection practices.

[0020] See Figure 1 As shown, this invention discloses a method for compliant identification of medical production personnel changing clothes and disinfection, applied to a cloud server, including: Step S11: After receiving the encrypted monitoring video segment transmitted by the edge node, the monitoring video segment is decrypted and cached as an inference request in the video stream inference scheduler; the monitoring video segment is a video segment obtained by the edge node by uniformly sampling from the circular buffer and splicing it based on a preset sampling period; each of the circular buffers is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process; the preprocessed video stream is a video stream obtained by the edge node by preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process.

[0021] In this embodiment, the cloud server receives and decrypts the encrypted monitoring video segments transmitted by the edge nodes, and stores the decrypted segments as inference requests in the video stream inference scheduler. The video stream inference scheduler is designed for the concurrent inference needs of multiple cameras (typically 8-32 channels) in a pharmaceutical production workshop, supporting frame buffering and sampling, dynamic batch processing, and priority scheduling. The monitoring video segments are obtained by the edge nodes uniformly sampling a preset number of frames from a circular buffer at a preset sampling period, for example, sampling 16 frames uniformly at a 2-second period, and then stitching them together. Multiple circular buffers are used to cache each pre-processed video stream of the pharmaceutical production personnel's changing and disinfection process. Each video stream maintains a circular buffer (capacity 128 frames), continuously written at a frame rate of 25 fps. The pre-processed video stream is obtained by the edge nodes pre-processing the original monitoring video stream for changing and disinfection.

[0022] It should be noted that in the pharmaceutical manufacturing industry, personnel are one of the main sources of contamination in clean areas. According to the mandatory requirements of Good Manufacturing Practice (GMP) for pharmaceutical production, all personnel entering clean areas must undergo standardized changing and disinfection procedures to minimize the contamination of the production environment by particles and microorganisms carried by the human body. Taking a pharmaceutical A / B level high-risk clean area as an example, the standard changing procedure includes 10 key steps: removing everyday clothing and shoes, putting on shoe covers, six-step hand washing and drying, hand disinfection with ethanol, putting on a clean cap (covering all hair and hairline), putting on a mask (nose strip pressed in), putting on a one-piece cleanroom garment (zipper to the top), putting on sterile gloves (cuffs tucked into gloves), self-inspection in front of a mirror (no exposed skin), and 360° air shower (≥15 seconds). Changing requirements for C / D level clean areas are relatively simplified, but still must cover core steps such as changing shoe covers, hand disinfection, putting on a cap, putting on a mask, putting on a cleanroom garment, and hand disinfection. Table 1 below shows a comparison of the changing and disinfection procedures for different clean area levels provided in this application.

[0023] Table 1 Comparison of Standardized Changing and Disinfection Procedures for Different Cleanroom Levels

[0024] Currently, domestic pharmaceutical companies primarily rely on manual inspections and regular training and assessments to supervise this process. Under traditional manual monitoring, the violation rate of the gowning process is approximately 5%-12%, while violations in the aseptic operation area can lead to the scrapping of an entire batch of products, resulting in economic losses ranging from hundreds of thousands to millions of yuan. To address the shortcomings of current supervision, this application provides a method for identifying compliance in gowning and disinfection procedures for pharmaceutical production personnel. This method is designed for cleanrooms (GMP environments) in pharmaceutical production and encompasses core technologies such as real-time acquisition of multiple industrial monitoring videos, spatiotemporal feature extraction and action detection based on a 3D residual convolutional network (3D ResNet-C3D), state machine compliance determination of process step sequences, and cloud-based GPU (Graphics Processing Unit) inference acceleration and alarm linkage. Figure 2 The diagram shows the system architecture of a compliance identification method for gown disinfection in pharmaceutical production provided in this application. It illustrates a four-layer architecture and its data flow: a perception layer (5 industrial cameras), an edge layer (video acquisition and preprocessing), a cloud inference layer (3D ResNet-C3D + FSM (Finite State Machine) + TensorRT (a high-performance deep learning inference engine)), and an application layer (alarms + auditing). The system employs a four-layer architecture: perception layer, edge layer, cloud inference layer, and application layer. It acquires video streams through industrial cameras deployed in the gown changing corridor of GMP clean areas. A 3D ResNet-C3D model, fine-tuned using a pharmaceutical gown dataset, is used for spatiotemporal feature extraction and action detection. Based on the finite state machine, the system determines the compliance of the gown changing process by considering the step sequence and timing constraints. This allows for accurate identification of violations such as skipped steps, reversed order, timeouts, and incomplete actions. The system utilizes TensorRT inference optimization on an x86 architecture + GPU cloud server to achieve real-time processing of multiple video streams, and integrates real-time alarms, violation recording storage, and audit reporting functions, providing an end-to-end intelligent solution for GMP compliance management in pharmaceutical manufacturing enterprises. This application establishes a standardized 13-step state model for the personnel changing and disinfection process in GMP clean areas. Taking a Class D clean area as the baseline scenario (which can be extended upwards to Classes C / B / A), the state sequence definition of the standard process is shown in Table 2 below: Table 2 Standard Procedure Definitions for Changing and Disinfection in Clean Areas of Pharmaceutical Production

[0025] Furthermore, the perception layer consists of industrial IP (Internet Protocol) cameras deployed at key locations in the cleanroom changing corridors, covering the entrance area, handwashing area, changing area, air shower area, and microscopic inspection area. Industrial-grade network cameras are used. Selection criteria are shown in Table 3 below: Table 3 Industrial Camera Selection Parameter Requirements

[0026] In a typical GMP changing corridor, it is recommended to deploy 5 cameras according to the scheme in Table 4 below: Table 4 Camera Deployment Location Plan

[0027] Each camera pushes its video stream to the edge layer via RTSP (Real Time Streaming Protocol) / ONVIF (Open Network Video Interface Forum). The edge layer is responsible for preprocessing operations such as receiving, decoding, frame extraction, resolution normalization, and temporal sampling of the video stream, and then encrypts and transmits the processed video segments to the cloud inference layer. The cloud inference layer is deployed on an x86 server + GPU hardware platform, configured as shown in Table 5 below. Table 5 Hardware Configuration of Cloud Inference Server

[0028] The cloud-based inference layer is the core computing layer of the system, comprising a 3D ResNet-C3D backbone network, an action detection module (TSN (Temporal Segment Network) / SSN (Structured Segment Network) for time-series segment analysis), a process compliance determination engine (FSM state machine), a TensorRT inference acceleration module, and a Triton (a high-performance inference service framework) model service API (Application Programming Interface). The application layer provides business functions for GMP quality management personnel, such as real-time alerts, violation recording storage, compliance dashboards, audit report generation, and system management.

[0029] It should be noted that the preprocessing procedure for the monitoring video stream of the pharmaceutical production personnel changing and disinfecting their gowns is as follows: The monitoring video stream undergoes decoding, frame extraction, resolution normalization, and temporal sampling. The monitoring video stream is any one of multiple video streams collected in real-time by the edge node through various industrial cameras deployed in the changing corridor of the pharmaceutical production clean area. That is, the edge node collects multiple monitoring video streams output by various industrial cameras deployed in the changing corridor of the pharmaceutical production clean area, and performs preprocessing operations such as decoding, frame extraction, resolution normalization, and temporal sampling on each video stream. In this way, this embodiment constructs a standard gowning and disinfection procedure library covering clean areas of levels A / B / C / D in GMP, establishes a temporal dependency model between steps, and achieves accurate identification of various violations such as skipped steps, incorrect sequence, and incomplete actions. The edge nodes use a circular buffer to cache the pre-processed video streams of each medical gowning and disinfection process, which can continuously receive uninterrupted real-time video data, avoiding memory overflow and video frame loss problems caused by long-term video caching, and adapting to the 24 / 7 uninterrupted monitoring scenario of medical workshops. The edge nodes uniformly sample the video in the circular buffer based on a preset sampling period and stitch it into video segments, reducing the amount of massive data per frame that needs to be processed in the cloud, balancing the video data transmission and cloud inference computing load, and avoiding network bandwidth congestion caused by real-time full video push. The edge nodes pre-process the original monitoring video. Standardized video streams are generated through preprocessing, including image denoising and size normalization, reducing data processing overhead before cloud inference. Edge nodes encrypt video segments before transmission, and the cloud decrypts them upon receipt, ensuring secure data transmission of monitoring footage from the medical production gowning and disinfection process and mitigating the risk of video leaks from workshop personnel. After decryption, video segments are uniformly encapsulated as inference requests and cached in the video stream inference scheduler, enabling orderly queuing and scheduling of multiple workshop video inference tasks. This prevents concurrent video inference from competing for computing power, ensuring stable and orderly execution of inference tasks for recognizing compliant behavior during gowning and disinfection. Industrial cameras deployed in the gowning corridor of the medical clean area capture multiple gowning and disinfection monitoring video streams in real time, providing complete coverage of the entire gowning and disinfection process, capturing personnel operations without blind spots, and ensuring the integrity of the data source for compliant behavior recognition.

[0030] Step S12: The cached inference request is scheduled and processed by the video stream inference scheduler to obtain the inference input video sequence, and the inference input video sequence is input into the preset inference acceleration engine; the preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model.

[0031] In this embodiment, the video stream inference scheduler schedules and processes the cached inference requests to generate an inference input video sequence, i.e., an input Clip, which is then sent to a preset inference acceleration engine equipped with a preset action detection and classification model for inference optimization.

[0032] It should be noted that the processing flow for scheduling the cached inference requests is as follows: Based on the changing area locations monitored by the surveillance video segments in the inference requests, the corresponding scheduling priority is determined; inference requests with the same cache time and scheduling priority are merged into a batch request for scheduling. That is, priority scheduling is performed, allocating scheduling priorities according to the monitored changing area locations corresponding to the inference requests. For example, for video streams from critical areas such as air showers, a high-priority queue is allocated to ensure that their inference latency is always below 200ms; for video streams from general areas, a normal-priority queue is allocated. Then, dynamic batch processing is performed. Triton InferenceServer's (an open-source inference service software) Dynamic Batcher receives multiple video streams at the same time, merging inference requests with the same scheduling priority into a single batch for scheduling. When GPU memory allows, the batch size can be dynamically expanded to 8-16, significantly reducing the inference overhead per frame.

[0033] It should be noted that the training process of the preset action detection classification model is as follows: Using a preset human action video dataset that meets preset large-scale conditions, the initial action detection classification network is pre-trained to obtain a pre-trained action detection classification model; based on a pre-constructed medical changing action dataset and through a hierarchical learning rate decay mechanism, under the premise of a preset number of layers before freezing, the pre-trained action detection classification model is adaptively fine-tuned to obtain the preset action detection classification model; wherein, the medical changing action dataset is a dataset composed of edited and labeled standard changing videos, and data augmentation is performed; the standard changing videos are videos collected of different personnel performing changing and disinfection procedures that meet preset standard conditions. That is, as... Figure 3 The diagram shown is a network architecture diagram of an initial action detection and classification network provided in this application. It illustrates the complete data flow of the input layer → Conv1 → MaxPool → Conv2_x → Conv3_x → Conv4_x → Conv5_x → GlobalAvgPool → FC → Softmax+NMS, as well as the detailed structure of Conv3D → BN → ReLU → Conv3D → BN → Add → ReLU within the 3D residual block. Table 6 below shows the network layer configuration. Table 6 Network Layer Configuration

[0034] Given the limited size of the medical changing room action dataset (target size: 8000 action segments, ≥500 per category) and the large number of parameters in the 3D ResNet-C3D model (approximately 33 million parameters), direct training on a small dataset can easily lead to severe overfitting. This embodiment employs a two-stage training strategy of pre-training + domain fine-tuning: Stage 1: Large-scale general video pre-training. Pre-training is performed on a pre-defined human action video dataset that meets the pre-defined large-scale conditions, such as the Kinetics-400 dataset (approximately 300,000 videos, 400 action categories). Kinetics-400 contains rich human action samples, and its learned spatiotemporal feature representation has good generalization ability. Training uses the SGD (Stochastic Gradient Descent) optimizer with an initial learning rate of 0.1, a batch size of 128, momentum of 0.9, and weight decay of 0.001, for a total of 200 epochs. Stage 2: Fine-tuning on the medical changing room dataset. Domain-adaptive fine-tuning was performed on a self-built medical changing action dataset (10 action categories, ~5000-10000 samples). During fine-tuning, the `ft_begin_index` parameter (the index of the first layer to be fine-tuned) controlled the number of frozen layers. Fine-tuning began from the 4th residual block of the pre-trained model (starting with Conv5_x), keeping the first 4 layers frozen. A decreasing learning rate was applied to different layers of the network, employing a layered learning rate decay mechanism. The following table (7) illustrates the configuration of the two-stage training strategy: Table 7 Two-stage training strategy configuration

[0035] The construction process of the medical changing action dataset includes the following steps: Video acquisition: In a laboratory simulating a GMP changing environment, multiple industrial cameras (1080p@25fps) are used to capture videos of different personnel (half male and half female, with varying body types) performing standard changing procedures. Each camera covers a key area (washbasin, changing mirror, air shower, etc.), with a shooting distance of 2-5 meters. Video editing and annotation: Using LabelV (video category annotation software) or VIA (VGG Image Annotator for image data annotation), the videos are time-axis-annotated, labeling the start and end timestamps and category labels for each action segment. At least 500 valid segments are collected for each action category to ensure sample diversity (different personnel, different angles, different speeds). Data augmentation: The following augmentation strategies are applied to the training samples: random horizontal flipping (probability 0.5), multi-scale cropping (scale factor). Random temporal offset (±3 frames) and color jitter (brightness ±20%, contrast ±20%, saturation ±20%). The dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure that samples from the same person do not cross sets.

[0036] It should be further noted that the inference optimization process for the preset action detection classification model is as follows: The preset action detection classification model is imported using the built-in exchange format parser of the preset inference acceleration engine; the neural network exchange format of the preset action detection classification model is consistent with the exchange format parsed by the exchange format parser; layer fusion and quantization accuracy calibration are performed on the preset action detection classification model to obtain the processed preset action detection classification model; and a corresponding kernel tuning scheme is selected according to the graphics processor model of the cloud server to perform inference optimization based on the processed preset action detection classification model, the kernel tuning scheme, and the memory pool reuse mechanism. That is, to maximize GPU inference efficiency, this embodiment uses TensorRT to optimize the trained 3DResNet-C3D model (the preset action detection classification model) for inference. The optimization process includes the following steps: model export, exporting the trained model to ONNX format to ensure cross-framework compatibility. The TensorRT engine is built by importing the model using TensorRT's ONNX parser and performing the following optimizations: Layer Fusion: Convolutional layers, BN layers, and ReLU activation layers are fused into a single CGBN operator, reducing the number of GPU kernel startups; Precision Calibration: Supports three precision modes: FP32, FP16, and INT8. The INT8 mode determines the optimal quantization parameters through a calibration dataset, which can improve inference speed by 3–5 times while maintaining accuracy loss of <1%; Kernel Auto-Tuning: Automatically selects the optimal convolutional algorithm and memory layout for the target GPU model; Dynamic Tensor Memory: Reduces GPU memory usage by 40% through a memory pool reuse mechanism. Inference Engine Deployment: The completed TensorRT engine (.trt file) provides REST / gRPC API services through the Triton Inference Server, supporting Dynamic Batching and Concurrent Model Execution. Table 8 below compares the inference optimization techniques: Table 8 Comparison of TensorRT Inference Optimization Techniques

[0037] This embodiment employs a two-stage training strategy: pre-training on a large-scale general video dataset and domain-adaptive fine-tuning on a self-built medical gowning dataset. This trains a pre-defined action detection and classification model, improving the model's accuracy in recognizing fine-grained medical-specific actions such as wearing gloves, cleanroom gowns, and frontal disinfection under limited labeled data conditions. A multi-stream video inference scheduling scheme is designed to achieve real-time action detection of multiple monitoring video streams on a single GPU, meeting the concurrent performance requirements of large-scale deployment in pharmaceutical production workshops. Inference requests with the same cache time and scheduling priority are merged into batch processing requests, enabling the simultaneous delivery of multiple video sequences to the inference acceleration engine. This fully leverages the advantages of GPU batch parallel computing, increasing the processing throughput of gowning and disinfection action recognition per unit time, reducing frequent memory read / write losses caused by multiple fragmented requests queuing, maximizing computing power utilization while prioritizing high-risk locations, and balancing the timeliness of pharmaceutical production compliance monitoring with system inference efficiency.

[0038] Step S13: Using the preset action detection and classification model in the preset inference acceleration engine, generate multiple candidate action segment proposals for the inference input video sequence, and perform action detection and classification on the candidate action segment proposals to obtain the corresponding action detection results.

[0039] In this embodiment, the preset action detection and classification model is loaded by the preset inference acceleration engine, multiple sets of candidate action segment proposals are generated for the inference input video sequence, and action detection and classification are completed based on the candidate action segment proposals, and the corresponding action detection results are output.

[0040] Specifically, the inference input video sequence is divided into several video slices, and the probability of human action in each video slice is quantified to obtain a corresponding action score; the action score is positively correlated with the probability of human action; a set of target video slices in the inference input video sequence whose action scores are continuously higher than a preset score threshold is determined, and the set of target video slices is clustered into candidate action segment proposals; the spatiotemporal features of the candidate action segment proposals are extracted, and the execution stage of human action is modeled using structured temporal pyramid pooling to obtain the corresponding modeling results; the action category and action integrity are determined using the dual-branch discriminant head in the preset action detection classification model and based on the modeling results to obtain the corresponding action detection results; the dual-branch discriminant head includes an action classification head and an action integrity quantization head. That is, to achieve action localization and classification in unedited long videos, this embodiment adopts a two-stage temporal action detection framework, combining the sparse sampling strategy of temporal segment networks and the action segment proposal mechanism of structured segment networks. First stage: Use a temporal action grouping algorithm to generate action proposals. The Temporal Actionness Grouping (TAG) algorithm is used to scan the video sequence and generate candidate action segment proposals. TAG calculates the actionness score for each time unit, clustering consecutive high-scoring regions into action proposals, effectively eliminating a large number of background interference segments. The second stage involves action classification and boundary refinement using a structured segment network. For each candidate proposal, 3D ResNet-C3D is used to extract its spatiotemporal features, and Structured Temporal Pyramid Pooling (STPP) is used to model the start, course, and end phases of the action. Based on this, a dual-branch discriminant model (activity classifier + integrity classifier) ​​is used to jointly determine the action category and proposal integrity, outputting the final action detection result. The model input format is T×H×W×C=16×112×112×3, where T, H, W, and C correspond to the frame number, height, width, and channel, respectively. The model output is a triple sequence, the specific expression of which is as follows: ; in, The timestamp of the action. For action category tags, Here, M represents the confidence score, and M is the total number of valid action instances in a single video segment. For the medical changing scene, the action category set is defined as: Action detection output parameters: confidence threshold: 0.85 (action detection results below this threshold are not accepted), timestamp accuracy: ±1 frame (approximately ±40ms at 25fps). In this way, this embodiment generates candidate action segment proposals based on a pre-defined action detection classification model optimized for medical changing and disinfection scenarios. It can automatically locate effective temporal intervals within the video containing personnel changing and disinfection behaviors, filtering out inactive or irrelevant blank scenes, and reducing the unnecessary computational cost of subsequent classification calculations.

[0041] Step S14: Based on the preset process compliance judgment engine and the action detection results, determine whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfecting. If so, trigger the corresponding alarm operation. The violation of the rules for changing clothes and disinfecting refers to changing clothes and disinfecting behavior that does not meet the preset standard conditions.

[0042] In this embodiment, the medical production personnel's changing and disinfection operations are verified by relying on the preset process compliance judgment engine and the action detection results output by the model. If any violation that does not conform to the standard process is identified, the corresponding alarm operation is immediately executed.

[0043] It should be noted that the preset process compliance determination engine is built on a deterministic finite state machine, including a time-constrained state transition relationship obtained by modeling the gowning and disinfection process that meets preset standard conditions. Correspondingly, the process for determining whether the current pharmaceutical production personnel have violated gowning and disinfection regulations is as follows: a timeline record of the gowning and disinfection process is made for each pharmaceutical production personnel; if the timeline record represents the transition relationship of the current gowning and disinfection behavior state, it is determined that the current pharmaceutical production personnel have violated the gowning and disinfection regulations. That is, the preset process compliance determination engine is designed based on a deterministic finite state machine (DFA), modeling the gowning process as a time-constrained state transition system. The state transition relationship of the standard gowning and disinfection process in the pharmaceutical production clean area is strictly modeled using a finite state machine. The standard process requires personnel to execute the process sequentially in the order S0→S1→S2→...→S12; any behavior deviating from the standard sequence is judged as a violation. Specific violation types include: step skipping (directly proceeding to Si+2 or higher without executing Si), reversed order (executing Sj after Si is completed and j≤i), action timeout, and incomplete action (the start of an action is detected but its completion is not detected). For example... Figure 4The diagram shows the state transition path of a finite state machine for determining the compliance of a changing and disinfection process, as provided in this application. It illustrates the state transition paths for 13 normal states, from S0 (initial) to S12 (complete), and the triggering conditions for five error states: E1 (reversed order), E2 (step skipped), E3 (timeout), E4 (incomplete action), and E5 (compliance violation). The formal definition of the state machine is a quintuple, and its expression is as follows: ; in, It is a finite set of states (containing 13 normal step states and 5 error states). For the input symbol set, This is the state transition function. This is the initial state. Acceptance status (process compliant completion). FSM core configuration parameters: Status scan cycle: 200ms, Action detection confidence threshold: 0.85, Timeout threshold: See Table 2 for thresholds for each step. The state machine engine maintains the timeline records for each worker. ,in To enter the state The timestamp. Based on this timeline, the system can calculate the following key metrics: Single-step time: Total process time: Waiting timeout: .in, This is an indicator function; it is 1 when the condition is true and 0 otherwise. The state machine engine performs state scanning with a period of 200ms, synchronized with the inference cycle of the action detection model, ensuring real-time state transitions.

[0044] Furthermore, the state transition function The core rules are as follows: Rule 1 (sequential execution): If the current state is ( ), and the input symbol corresponding to was detected. The action category is then transferred to : Rule 2 (Step Skip - E2): If the current state is And the input symbol was detected to correspond to ( If the step is skipped, the process is transitioned to an error state. : Rule 3 (Reversed Order - E1): If the current state is ( ), and the input symbol corresponding to was detected. ( If the order is reversed, the process is transitioned to an error state. : Rule 4 (Timeout Alarm - E3): If in the current state Stay time exceeds the threshold (See Table 2), and no transfer to was detected. If a valid input is received, a timeout alarm will be triggered, and the process will be transferred to... Rule 5 (Incomplete Action - E4): If the start characteristic of an action is detected but the completion characteristic is not detected (e.g., the hand sanitizing action begins but does not reach the 20-second duration), it is determined to be an incomplete action and transferred to... Rule 6 (Global Violation—E5): Any entry into an error state – In all such cases, compliance violation events (E5) are reported to the upper application layer, triggering real-time alerts.

[0045] It should be further noted that when the FSM engine detects any violation (entering E1–E5 status), the system triggers the following linkage mechanisms: Real-time alarm: Within 3 seconds, an alarm notification is sent to the on-duty quality management personnel and the person involved through multiple channels such as audible and visual alarms, workshop electronic dashboards, and message push. The alarm information includes: the violator's ID (with optional facial recognition), violation type, violation time, violation steps, and confidence score. Violation video storage: Automatically saves 30-second video clips before and after the violation, storing them in local NAS (Network Attached Storage) or cloud object storage for a retention period of no less than 5 years to meet GMP audit traceability requirements. Compliance reports: The system automatically generates daily, weekly, and monthly reports, statistically analyzing key indicators such as compliance rate, violation type distribution, and personnel violation frequency ranking for each region and shift, and supports exporting to PDF / Excel format for review by the quality management department. In this way, this embodiment designs a process compliance judgment engine based on finite state machines to address the step sequence dependency characteristics of the medical production changing process. The changing process is modeled as a constrained state transition sequence, realizing a unified judgment framework for four types of violations: step skipping, sequence reversal, action timeout, and incomplete action. This solves the shortcomings of existing technologies that can only judge the standardization of single actions and cannot handle process-level compliance issues, and enables real-time step tracking and compliance judgment of each worker's changing process.

[0046] As can be seen from the above, the edge nodes in this embodiment first preprocess the monitoring video streams of the changing and disinfection process of various medical production personnel. The preprocessed video streams are cached through their respective circular buffers, and are uniformly sampled and spliced ​​into monitoring video segments according to a preset sampling period. After encryption, they are transmitted to the cloud server. The cloud server receives the monitoring video segments, decrypts them, and stores them in the video stream inference scheduler. The video stream inference scheduler uniformly schedules the cached inference requests to obtain the inference input video sequence, which is then sent to the built-in preset action detection classification model. The preset inference acceleration engine is used to optimize the inference. The preset inference acceleration engine generates multiple candidate action segment proposals, performs action detection classification, and outputs action detection results. The action detection results are sent to the preset process compliance judgment engine to identify illegal changing and disinfection behaviors. If an illegal behavior is detected, an alarm is triggered. In this way, through the above-described process of this application embodiment, the cloud server decrypts the encrypted transmitted monitoring video segments and hands them over to the video stream inference scheduler for centralized management of all inference requests. The scheduler can allocate computing power in a coordinated manner, improve hardware resource utilization, and reduce waiting delays caused by concurrent inference of multiple video streams. It configures an inference acceleration engine and builds in an optimized action detection classification model for inference optimization, shortens the time for action recognition of the screen, and adapts to the real-time analysis needs of multiple video streams in the workshop. It outputs candidate action segment proposals and then performs detection and classification, performing fine recognition only on valid segments containing human actions, reducing the consumption of ineffective computing power. Through a preset process compliance judgment engine combined with action detection results, it automatically verifies the changing and disinfection specifications, which can accurately capture violations of changing and disinfection operations around the clock. When a violation is detected, an alarm is automatically triggered, realizing unmanned real-time supervision of the medical changing and disinfection process, reducing the risk of omissions in manual inspections, and thus optimizing the method for identifying compliance of medical production personnel changing and disinfection, thereby improving the accuracy of identifying violations and optimizing the supervision effect of changing and disinfection behavior.

[0047] Accordingly, see Figure 5 As shown in the illustration, this application also provides a compliance identification device for pharmaceutical production personnel changing clothes and disinfecting, applied to a cloud server, comprising: The request caching module 11 is used to decrypt the monitoring video segment transmitted encrypted by the edge node and cache it as an inference request to the video stream inference scheduler after receiving the monitoring video segment. The monitoring video segment is a video segment obtained by the edge node by uniformly sampling and splicing it from the circular buffer based on a preset sampling period. Each circular buffer is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process. The preprocessed video stream is a video stream obtained by the edge node by preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process. The sequence input module 12 is used to schedule and process the cached inference requests through the video stream inference scheduler to obtain the inference input video sequence, and input the inference input video sequence into the preset inference acceleration engine; the preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model; The motion detection module 13 is used to generate multiple candidate motion segment proposals for the inference input video sequence using the preset motion detection classification model in the preset inference acceleration engine, and to perform motion detection and classification on the candidate motion segment proposals to obtain the corresponding motion detection results. The operation triggering module 14 is used to determine whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfection based on the preset process compliance judgment engine and the action detection results. If so, the corresponding alarm operation is triggered. The violation of the rules for changing clothes and disinfection refers to changing clothes and disinfection that does not meet the preset standard conditions.

[0048] In some specific embodiments, the request caching module 11 may specifically include: The video stream preprocessing unit is used to perform preprocessing operations such as decoding, frame extraction, resolution normalization, and temporal sampling on the monitoring video stream of the medical production personnel changing clothes and disinfecting process; the monitoring video stream is any one of the multiple video streams collected in real time by the edge node through each industrial camera deployed in the changing room passage of the medical production clean area.

[0049] In some specific embodiments, the sequence input module 12 may specifically include: The priority determination unit is used to determine the corresponding scheduling priority based on the location of the changing area monitored by the monitoring video segment of the inference request. The request merging unit is used to merge inference requests with the same cache time and scheduling priority into batch processing requests for scheduling processing.

[0050] In some specific embodiments, the sequence input module 12 may specifically include: The large model import unit is used to import the preset action detection classification large model using the built-in exchange format parser of the preset inference acceleration engine; the neural network exchange format of the preset action detection classification large model is consistent with the exchange format parsed by the exchange format parser. The scheme selection unit is used to perform layer fusion and quantization accuracy calibration on the preset motion detection classification model to obtain the processed preset motion detection classification model, and select the corresponding kernel tuning scheme according to the graphics processor model of the cloud server, so as to perform inference optimization based on the processed preset motion detection classification model, the kernel tuning scheme and the memory pool reuse mechanism.

[0051] In some specific embodiments, the medical production personnel changing clothes and disinfection compliance identification device may specifically include: The network pre-training unit is used to pre-train the initial action detection and classification network using a pre-set human action video dataset that meets the pre-set large-scale conditions, so as to obtain a pre-trained large-scale action detection and classification model. The large model fine-tuning unit is used to adaptively fine-tune the pre-trained action detection and classification large model based on a pre-built medical changing action dataset and through a hierarchical learning rate decay mechanism, under the premise of a preset number of layers before freezing, so as to obtain the preset action detection and classification large model. The medical changing action dataset is a dataset composed of standard changing videos that have been edited, labeled, and augmented; the standard changing videos are videos of different people performing changing and disinfection procedures that meet preset standard conditions.

[0052] In some specific embodiments, the motion detection module 13 may specifically include: The probability quantization unit is used to divide the inference input video sequence into several video slices and quantify the probability of human motion in each of the several video slices to obtain the corresponding motion score; the motion score is positively correlated with the probability of human motion. A clustering unit is used to determine the set of target video slices in the inference input video sequence in which the action score is continuously higher than a preset score threshold, and to cluster the set of target video slices into candidate action segment proposals. The feature modeling unit is used to extract the spatiotemporal features of the candidate action segment proposals, and to model the execution stage of human action by the spatiotemporal features through structured temporal pyramid pooling to obtain the corresponding modeling results. The action detection unit is used to determine the action category and action integrity by using the dual-branch discriminant head in the preset action detection classification model and based on the modeling results, so as to obtain the corresponding action detection results; the dual-branch discriminant head includes an action classification head and an action integrity quantization head.

[0053] In some specific implementations, the preset process compliance determination engine is built on a deterministic finite state machine, including a state transition relationship with temporal constraints obtained by modeling the changing and disinfection process that meets preset standard conditions. Accordingly, the operation triggering module 14 may specifically include: The timeline recording unit is used to record the timeline of the changing and disinfection process for each medical production worker. The behavior determination unit is used to determine that the current pharmaceutical production personnel have violated the rules regarding changing clothes and disinfection if the transition relationship of the timeline record representing the current state of changing clothes and disinfection does not conform to the state transition relationship.

[0054] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the compliance identification method for medical production personnel changing clothes and disinfection disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be a computer.

[0055] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0056] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0057] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the compliance identification method for pharmaceutical production personnel changing clothes and disinfection as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0058] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for compliant identification of pharmaceutical production personnel changing clothes and disinfection. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0060] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0062] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying compliance with regulations regarding the disinfection and changing of clothes by pharmaceutical production personnel, characterized in that, Applications to cloud servers include: After receiving the encrypted monitoring video segment transmitted by the edge node, the monitoring video segment is decrypted and cached as an inference request in the video stream inference scheduler; the monitoring video segment is a video segment obtained by the edge node by uniformly sampling from the circular buffer based on a preset sampling period and splicing it together; each of the circular buffers is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process; the preprocessed video stream is a video stream obtained by the edge node by preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process; The video stream inference scheduler schedules the cached inference requests to obtain the inference input video sequence, and then inputs the inference input video sequence into the preset inference acceleration engine. The preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model. Using the preset action detection and classification model in the preset inference acceleration engine, multiple candidate action segment proposals are generated for the inference input video sequence, and action detection and classification are performed on the candidate action segment proposals to obtain the corresponding action detection results. Based on the preset process compliance judgment engine and the action detection results, it is determined whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfection. If so, the corresponding alarm operation is triggered. The violation of the rules for changing clothes and disinfection refers to changing clothes and disinfection behavior that does not meet the preset standard conditions.

2. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to claim 1, characterized in that, The preprocessing of the monitoring video stream of the pharmaceutical production personnel changing clothes and disinfecting process includes: The monitoring video stream of the pharmaceutical production personnel changing clothes and disinfecting is preprocessed by decoding, frame extraction, resolution normalization and temporal sampling; the monitoring video stream is any one of the multiple video streams collected in real time by the edge node through each industrial camera deployed in the changing room passage of the pharmaceutical production clean area.

3. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to claim 1, characterized in that, The scheduling process for the cached inference requests includes: Based on the location of the changing area monitored by the surveillance video segment in the inference request, determine the corresponding scheduling priority; The inference requests with the same cache time and scheduling priority are merged into a batch request for scheduling processing.

4. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to claim 1, characterized in that, The inference optimization of the preset action detection and classification model includes: The preset action detection classification model is imported using the built-in exchange format parser of the preset inference acceleration engine; the neural network exchange format of the preset action detection classification model is consistent with the exchange format parsed by the exchange format parser. The preset motion detection classification model is subjected to layer fusion and quantization accuracy calibration to obtain the processed preset motion detection classification model. The corresponding kernel tuning scheme is selected according to the graphics processor model of the cloud server, and inference optimization is performed based on the processed preset motion detection classification model, the kernel tuning scheme and the memory pool reuse mechanism.

5. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to claim 1, characterized in that, The training process of the pre-defined action detection and classification model includes: Using a pre-set human motion video dataset that meets the pre-set large-scale conditions, the initial motion detection and classification network is pre-trained to obtain a pre-trained large-scale motion detection and classification model. Based on a pre-built medical changing action dataset and through a hierarchical learning rate decay mechanism, the pre-trained action detection and classification model is adaptively fine-tuned with a preset number of layers before freezing, so as to obtain the preset action detection and classification model. The medical changing action dataset is a dataset composed of standard changing videos that have been edited, labeled, and augmented; the standard changing videos are videos of different people performing changing and disinfection procedures that meet preset standard conditions.

6. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to claim 1, characterized in that, The process involves generating multiple candidate action segment proposals for the inference input video sequence, performing action detection and classification on the candidate action segment proposals, and obtaining corresponding action detection results, including: The inference input video sequence is divided into several video slices, and the probability of human motion in each of the several video slices is quantified to obtain the corresponding motion score; the motion score is positively correlated with the probability of human motion. Determine the set of target video slices in the inference input video sequence in which the action score is continuously higher than a preset score threshold, and cluster the set of target video slices into candidate action segment proposals; The spatiotemporal features of the candidate action segment proposals are extracted, and the execution stage of human action is modeled using structured temporal pyramid pooling to obtain the corresponding modeling results. The action category and action integrity are determined by using the dual-branch discriminant head within the preset action detection classification model and based on the modeling results, so as to obtain the corresponding action detection results; the dual-branch discriminant head includes an action classification head and an action integrity quantization head.

7. The method for compliant identification of pharmaceutical production personnel changing clothes and disinfection according to any one of claims 1 to 6, characterized in that, The preset process compliance determination engine is built on a deterministic finite state machine, including a state transition relationship with temporal constraints obtained by modeling the changing and disinfection process that meets preset standard conditions. Accordingly, determining whether current pharmaceutical production personnel have engaged in unauthorized changing and disinfection practices includes: A timeline record of the changing and disinfection process for each pharmaceutical production worker; If the timeline records representing the transition relationship of the current changing and disinfection behavior do not conform to the state transition relationship, it is determined that the current pharmaceutical production personnel have violated regulations regarding changing and disinfection behavior.

8. A device for identifying compliance with medical production personnel's changing and disinfection procedures, characterized in that, Applications to cloud servers include: The request caching module is used to decrypt the monitoring video segments transmitted encrypted by the edge node and cache them as inference requests to the video stream inference scheduler after receiving them. The monitoring video segments are video segments that the edge node uniformly samples from a circular buffer and splices based on a preset sampling period. Each circular buffer is used to cache each preprocessed video stream of the medical production personnel changing and disinfecting process. The preprocessed video stream is the video stream obtained by the edge node preprocessing the monitoring video stream of the medical production personnel changing and disinfecting process. The sequence input module is used to schedule and process the cached inference requests through the video stream inference scheduler to obtain the inference input video sequence, and input the inference input video sequence into the preset inference acceleration engine; the preset inference acceleration engine is equipped with a preset action detection classification model, which is used to optimize the inference of the preset action detection classification model; The action detection module is used to generate multiple candidate action segment proposals for the inference input video sequence using the preset action detection classification model in the preset inference acceleration engine, and to perform action detection and classification on the candidate action segment proposals to obtain the corresponding action detection results. The operation triggering module is used to determine whether the current pharmaceutical production personnel have violated the rules for changing clothes and disinfection based on the preset process compliance judgment engine and the action detection results. If so, the module will trigger the corresponding alarm operation. The violation of the rules for changing clothes and disinfection refers to changing clothes and disinfection that does not meet the preset standard conditions.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the pharmaceutical production personnel changing clothes and disinfection compliance identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the compliance identification method for medical production personnel changing clothes and disinfection as described in any one of claims 1 to 7.

Images