Medical waste safety disposal risk identification method and system

Abstract

The invention provides a medical waste safety disposal risk identification method and system, and relates to the technical field of risk identification. According to the method, the diffusion model generator is adopted to expand medical waste disposal scene multi-modal samples, and the problems of unbalance and scarcity of different risk type sample sizes are solved; image modal core features are extracted through a Mama model, sensor numerical modal and text description modal data are fused to construct multi-dimensional prompt information, and a comprehensive context is provided for a Lama model; in combination with diffusion model generator data enhancement, multi-modal feature fusion and prompt optimization methods, the credibility and practicability of a risk identification model driven by small sample data in a complex disposal scene are improved, and the application effect of generative artificial intelligence in the field of medical waste safe disposal is enhanced; according to the invention, an efficient, accurate, comprehensive and self-adaptive medical waste disposal risk identification technology is realized, so that higher requirements of public health safety and ecological environment safety on medical waste disposal risk management and control are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk identification, in particular to a medical waste safe disposal risk identification method and system. BACKGROUND

[0002] Medical waste has dangerous characteristics such as infectivity, corrosivity, flammability, and toxicity. Its safe disposal is directly related to public health safety, ecological environment quality, and the health of operating personnel. The medical waste disposal process covers multiple links such as classification, transportation, temporary storage, incineration / burial, etc. Each link is easily affected by factors such as disposal facility status, operation specification execution, and environmental condition changes. The risk hidden danger is highly concealed and the triggering factors are complex.

[0003] Traditional medical waste disposal risk identification mainly relies on manual inspection and single indicator monitoring, which has problems such as low efficiency, high omission rate, and delayed response. With the rapid development of intelligent technologies such as artificial intelligence and the Internet of Things, medical waste disposal risk intelligent identification methods have gradually become a research and application hotspot. Some related patents have proposed risk detection schemes based on single modal data combined with machine learning models, which have improved the automation level of risk identification to a certain extent. However, the existing medical waste safe disposal risk identification technology still has obvious defects: poor sample balance, difficult data acquisition, low multi-modal data fusion, insufficient model generalization ability, incomplete risk identification integrity, and high model update cost.

[0004] From the above description, it can be seen that the existing medical waste safe disposal risk identification technology cannot meet the actual needs of precise and comprehensive control of medical waste disposal risks. SUMMARY

[0005] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides a medical waste safe disposal risk identification method and system, which solves the technical problem that the existing medical waste disposal risk identification technology cannot meet the actual needs of precise and comprehensive control of medical waste disposal risks.

[0006] (II) Technical solutions To achieve the above purpose, the present application is realized by the following technical solutions: In a first aspect, the present application provides a medical waste safe disposal risk identification method, comprising: Obtaining multi-modal data of a medical waste disposal scene to be detected, processing the multi-modal data through a pre-constructed optimal multi-modal feature extraction model to obtain a fusion feature vector Filling the fusion feature vector and its corresponding disposal scene basic information into a pre-set prompt information template to obtain a filled prompt information template; The Llama model interface function is called, the multi-modal data of the medical waste disposal scene to be detected and the filled prompt information template are input into the Llama model, and a medical waste safety disposal risk identification result is generated as an output response. In the training process, the pre-constructed optimal multi-modal feature extraction model expands the medical waste disposal scene multi-modal data sample through a diffusion model generator. The optimal multi-modal feature extraction model is a combined model, which includes a Mamba model and a Qwen-7B-Chat model.

[0007] Preferably, the multi-modal data includes disposal scene image data, sensor monitoring data, and text description data.

[0008] Preferably, the medical waste disposal scene multi-modal data sample is expanded by the diffusion model generator, including: The historical multi-modal data of the medical waste disposal scene is standardized, the scene image data is uniformly resolved, the sensor monitoring data is completed with missing values, the text description data is removed of redundant information, and the risk state is labeled. The standardized historical multi-modal data is randomly divided into an original training set and an original test set. The diffusion model generator hyperparameters are set, the diffusion model generator is trained and optimized based on the original training set and the original test set, and the optimal diffusion model generator parameters are saved. The optimal diffusion model generator parameters are loaded, and the diffusion model generator is used to generate supplementary data for image modalities and sensor numerical modalities, and text modality data is expanded through context-aware rewriting.

[0009] Preferably, the process of obtaining the pre-set prompt information template includes: The historical multi-modal data, the fusion feature vector, the disposal scene basic information, and the output format of the risk identification result are constructed into initial prompt text. According to the preset output format requirement, a corresponding Python class is constructed, the attributes in the class correspond one-to-one to the keys of the output JSON object, and the output template construction is completed. The initial prompt text corresponding to each piece of data in the historical multi-modal data is input into the Llama model. The output response of each piece of data is stored as a model response list. The model response list is compared with the real label list, the risk level and risk type are mainly verified, the identification accuracy is calculated, when the identification accuracy is lower than the set threshold, the multi-modal feature extraction model is continuously fine-tuned and the prompt text structure is optimized until the accuracy is higher than the threshold, and the prompt meeting the requirements is stored as a template to obtain the prompt information template.

[0010] Preferably, the treatment scene basis information includes treatment facility number, treatment link, environment temperature and humidity, operation time, and operator number.

[0011] Preferably, the medical waste safe treatment risk identification result meets the following format requirements: determining whether the treatment scene has a safety risk, and identifying the multi-modal data associated with the risk, and the treatment scene basis information; determining and outputting the risk level, the risk type, the risk trigger factor, and the suggestion in the emergency treatment process.

[0012] Preferably, the medical waste safe treatment risk identification method further includes: storing the output response of each treatment scene in a risk response list, and the risk response list supports subsequent query, export, and reference call in the emergency treatment process.

[0013] In a second aspect, the present application provides a medical waste safe treatment risk identification system, comprising: a multi-modal feature extraction module, configured to acquire multi-modal data of a medical waste treatment scene to be detected, process the multi-modal data through a pre-constructed optimal multi-modal feature extraction model, and obtain a fusion feature vector a prompt template filling module, configured to fill the fusion feature vector and its corresponding treatment scene basis information into a pre-set prompt information template, and obtain a filled prompt information template; a Llama model response module, configured to call a Llama model interface function, input the multi-modal data of the medical waste treatment scene to be detected and the filled prompt information template into the Llama model, and generate a medical waste safe treatment risk identification result as an output response; In the training process of the pre-constructed optimal multi-modal feature extraction model, the medical waste treatment scene multi-modal data samples are expanded through a diffusion model generator. The optimal multi-modal feature extraction model is a combined model, and the combined model includes a Mamba model and a Qwen-7B-Chat model.

[0014] In a second aspect, the present application provides a computer readable storage medium storing a computer program for medical waste safe treatment risk identification, wherein the computer program causes a computer to execute the medical waste safe treatment risk identification method as described above.

[0015] In a third aspect, the present application provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for performing the medical waste safe disposal risk identification method as described above.

[0016] (Three) beneficial effects The present application provides a medical waste safe disposal risk identification method and system. Compared with the prior art, the following beneficial effects are achieved: The present application realizes multi-modal data enhancement through the diffusion model generator, the generated samples have higher quality and stronger diversity, and the problems of medical waste disposal risk sample scarcity and uneven distribution are solved in a targeted manner, providing balanced and sufficient data sources for model training; at the same time, the multi-modal feature extractor combined with the Mamba model and the Qwen-7B-Chat model is adopted, the sequence modeling advantage of the Mamba architecture and the semantic understanding ability of the Qwen model are utilized, and the core information of three types of data, i.e., image, sensor and text, is efficiently fused, thereby improving the information utilization rate; in addition, the generative artificial intelligence risk identification framework is constructed based on the Llama model, combined with multi-modal features and prompt optimization, the "hallucination" problem is avoided, and the new risk identification capability is achieved, which can cope with unobserved risk types without retraining. The present application realizes efficient, accurate, comprehensive and adaptive medical waste disposal risk identification technology to meet the higher requirements of public health safety and ecological environment safety on medical waste disposal risk control. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0018] Figure 1 A block diagram of a medical waste safe disposal risk identification method according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] The embodiment of the application provides a medical waste safe disposal risk identification method and system, solves the technical problem that the existing medical waste safe disposal risk identification technology cannot meet the actual demand of precise and comprehensive medical waste disposal risk control, and realizes efficient, precise, comprehensive and self-adaptive medical waste disposal risk identification technology to meet the higher requirement of public health safety and ecological environment safety on medical waste disposal risk control.

[0021] The technical solution in the embodiment of the application solves the above technical problem, and the general idea is as follows: The existing medical waste safe disposal risk identification technology still has many defects to be solved: First, the sample balance is poor and the data acquisition is difficult. The low probability of high-risk samples in the medical waste disposal process makes it difficult to construct a balanced sample set. The data acquisition of some key links is limited by the environment and cost, and the data acquisition is difficult. Second, the degree of multi-modal data fusion is low. Multiple only use single type data for risk identification, and cannot effectively fuse multi-dimensional data such as temperature, humidity, image and equipment parameters, and the risk state is not fully described. Third, the model generalization ability is insufficient. Multiple are trained for specific scenes or risk types, and the identification performance decreases greatly when the scene or risk factor changes, which is difficult to adapt to diversified needs. Fourth, the risk identification integrity is lacking. Focus on single link or explicit risk, the identification ability of implicit and complex risks caused by multi-link correlation is insufficient, and the whole process and all types of risks cannot be covered. Fifth, the model updating cost is high. When the process is optimized and new risks appear, large-scale retraining is needed, which consumes a lot of resources and may interrupt the identification continuity, which is not conducive to dynamic optimization and iteration. These defects jointly restrict the application effect of the existing intelligent technology.

[0022] The above problems jointly cause the existing intelligent identification technology to be difficult to meet the actual demand of precise and comprehensive medical waste disposal risk control.

[0023] To solve the above problems, the embodiment of the application provides a medical waste safe disposal risk identification method and system based on multi-modal data fusion and generative artificial intelligence, realizes efficient, precise, comprehensive and self-adaptive medical waste disposal risk identification technology, and meets the higher requirement of public health safety and ecological environment safety on medical waste disposal risk control.

[0024] In order to better understand the above technical solution, the above technical solution will be described in detail in combination with the description of the drawings and the specific embodiment.

[0025] The embodiment of the application provides a medical waste safe disposal risk identification method, which comprises the following steps: Figure 1The method comprises the following steps: S1, acquiring multi-modal data of a medical waste disposal scene to be detected, processing the multi-modal data through a pre-constructed optimal multi-modal feature extraction model to obtain a fusion feature vector S2, filling the fusion feature vector and its corresponding disposal scene basic information into a pre-set prompt information template to obtain a filled prompt information template; S3, calling an Llama model interface function, inputting the multi-modal data of the medical waste disposal scene to be detected and the filled prompt information template into the Llama model to generate a medical waste safety disposal risk identification result as an output response; In the training process of the pre-constructed optimal multi-modal feature extraction model, the medical waste disposal scene multi-modal data samples are expanded through a diffusion model generator. The optimal multi-modal feature extraction model is a combined model, and the combined model comprises a Mamba model and a Qwen-7B-Chat model.

[0026] The embodiment of the present application realizes multi-modal data enhancement through a diffusion model generator, the generated samples have higher quality and stronger diversity, and the problems of medical waste disposal risk sample scarcity and uneven distribution are solved in a targeted manner, balanced and sufficient data sources are provided for model training; at the same time, the multi-modal feature extractor combined with the Mamba model and the Qwen-7B-Chat model utilizes the sequence modeling advantage of the Mamba architecture and the semantic understanding ability of the Qwen model, efficiently fuses the core information of three types of data including images, sensors and texts, and improves the information utilization rate; in addition, the generative artificial intelligence risk identification framework is constructed based on the Llama model, combined with multi-modal features and prompt optimization, the "hallucination" problem is avoided, and the new risk identification capability is possessed, and the unseen risk types can be coped with without retraining. The embodiment of the present application realizes efficient, accurate, comprehensive and adaptive medical waste disposal risk identification technology to meet the higher requirements of public health safety and ecological environment safety on medical waste disposal risk control.

[0027] The following describes each step in detail: In step S1, multi-modal data of a medical waste disposal scene to be detected is acquired, and the multi-modal data is processed through a pre-constructed optimal multi-modal feature extraction model to obtain a fusion feature vector. The specific implementation process is as follows: The multi-modal data comprises disposal scene image data (such as medical waste packaging integrity, facility operation state), sensor monitoring data (such as harmful gas concentration, environmental temperature and humidity, equipment operation pressure) and text description data (such as waste category label, operation record abstract).

[0028] In the embodiment of the present application, the scene image data is converted into Base64 format before input, which facilitates model transmission and analysis; the sensor monitoring data is converted into [0, 1] interval values by using the Min-Max standardization method, so as to eliminate the dimension influence; after the text description data is segmented by jieba and stop words are removed, the Qwen-7B-Chat model is used for vector coding preprocessing.

[0029] The acquisition process of the pre-constructed optimal multi-modal feature extraction model and the pre-set prompt information template mentioned later is as follows: S201, a multi-modal data set of a medical waste disposal scene is acquired, and the multi-modal data set is randomly divided into a training set and a test set; wherein the multi-modal data set is obtained by historical multi-modal data enhancement, and the data enhancement includes the following steps: Firstly, the historical multi-modal data of the medical waste disposal scene is standardized, the scene image data is uniformly set to a resolution of 512*512 pixels, the sensor monitoring data is completed for missing values, the text description data is removed for redundant information, and the risk state is labeled; Secondly, the standardized historical multi-modal data is randomly divided into an original training set and an original test set according to a ratio of 8:2; Thirdly, the diffusion model generator hyperparameters are set, the diffusion model generator training and optimization are performed based on the original training set and the original test set, and the optimal diffusion model generator parameters are saved; Fourthly, the optimal diffusion model generator parameters are loaded, the supplementary data of the image modality and the sensor value modality are respectively generated by the generator, and the text modality data is expanded by context perception rewriting; Fifthly, the generated supplementary data and the standardized historical multi-modal data are combined to construct a complete multi-modal data set.

[0030] The complete multi-modal data set is randomly divided into a training set and a test set according to a ratio of 8:2.

[0031] S202, the pre-trained Mamba model and the text encoder combined model are jointly trained and parameter updated through the training set and the test set, the model optimization is completed, and a multi-modal feature extraction model is obtained. It should be noted that in the embodiment of the present application, the pre-trained Mamba model is a Mamba-L model pre-trained on an ImageNet-21K and a Medical-ImageNet data set, and the text encoder is a text encoding layer of a pre-trained Qwen-7B-Chat model.

[0032] S203, each data in the historical multi-modal data is input into the multi-modal feature extraction model, and a corresponding fusion feature vector is obtained; S204, label risk level, risk type and risk impact range information for each disposal scene data to form a real label list; S205, construct the output format of historical multi-modal data, fusion feature vector, disposal scene basic information and risk identification result as an initial prompt text. The disposal scene basic information includes disposal facility number, disposal link (classification, transfer, temporary storage, incineration / filling), environmental temperature and humidity, operation time and operator number.

[0033] S206, according to the preset output format requirement, construct the corresponding Python class, the attributes in the class correspond one-to-one with the keys of the output JSON object, and the output template construction is completed; S208, input the initial prompt text corresponding to each data in the multi-modal data set into the Llama model; S209, store the output response of each data as a model response list; S210, compare the model response list with the real label list, focus on verifying the risk level and risk type, and calculate the identification accuracy; when the identification accuracy is lower than the set threshold (the set threshold is 0.95), continue to fine-tune the multi-modal feature extraction model and optimize the prompt text structure until the accuracy is higher than the threshold; store the prompt that meets the requirements as a template to obtain the prompt information template; save the optimized multi-modal feature extraction model as the optimal multi-modal feature extraction model.

[0034] The following is an example to illustrate the acquisition process of the pre-constructed optimal multi-modal feature extraction model and the pre-set prompt information template mentioned later: Multi-modal data augmentation: The original multi-modal data set contains 1000 data, of which 850 are normal scenes, and 150 are risk scenes (50 are infectious risks, 30 are corrosive risks, and 70 are other risks). Divide it into 800 original training sets and 200 original test sets according to the ratio of 8:2.

[0035] Set the diffusion model generator hyperparameters: learning rate 0.0001, batch size 16, iteration number 15000 times, and loss function adopts perception loss function. Train the diffusion model generator based on the original training set, generate 300 image supplement data, 200 sensor supplement data and 150 text supplement data after training. Merge the generated data with the original data to obtain a complete multi-modal data set containing 1650 data, of which the risk scene sample proportion increases to 30%, and the sample distribution is more balanced.

[0036] Prompt information template optimization: load the text encoding layer of the pre-trained Mamba-L model and Qwen-7B-Chat model to form a combined model. Divide the complete multi-modal dataset into 1320 training sets and 330 test sets in the ratio of 8:2, input the combined model for joint training, and complete model optimization after 50 iterations as a multi-modal feature extraction model. Construct the initial prompt text and input the Llama model to generate a model response list. Compared with the real label list, the initial recognition accuracy is 89%, which is lower than the threshold value of 0.95. By fine-tuning the attention weight parameters of the feature extraction model and optimizing the logical order of the prompt text, the recognition accuracy is improved to 96.5% after 3 repeated training, meeting the requirements, and the prompt information template and the optimal multi-modal feature extraction model are saved.

[0037] In an embodiment, S2, fill the fusion feature vector and its corresponding disposal scene basic information into the pre-set prompt information template to obtain the filled prompt information template. The specific implementation process is as follows: Fill the fusion feature vector output by the optimal multi-modal feature extraction model, and disposal facility number, disposal link, environmental temperature and humidity, operation time and operator number, and other disposal scene basic information into the pre-set prompt information template to obtain the filled prompt information template.

[0038] In an embodiment, S3, call the Llama model interface function, input the multi-modal data of the medical waste disposal scene to be detected and the filled prompt information template into the Llama model, and generate the medical waste safety disposal risk identification result as the output response. The specific implementation process is as follows: Input the multi-modal data of the medical waste disposal scene to be detected and the filled prompt information template into the Llama model, and output the medical waste safety disposal risk identification result as the output response. The risk identification result meets the following format requirements: judge whether the disposal scene has safety risk, explain the multi-modal data identification associated with the risk, and disposal scene basic information; judge and output risk level (low risk, medium risk, high risk), risk type (infectious risk, corrosive risk, flammable and explosive risk, toxic risk, other risk), risk trigger factor and suggestion in emergency disposal process; no risk scene, no output information; format the response result of each risk scene as a JSON object, with "disposal scene number" as the key.

[0039] In the specific implementation process, the medical waste safety disposal risk identification method further includes: S4, store the output response of each disposal scene in the risk response list to support subsequent query, export and reference call in the emergency disposal process.

[0040] The following is a specific case to illustrate the risk identification process: The multi-modal data of the disposal scene to be detected is as follows: the scene image data is a medical waste transport box damage image (Base64 format), the sensor monitoring data is a harmful gas concentration of 0.8 mg / m³ (standardized to 0.72), the text description data is "infectious waste, transport link", and the disposal scene basic information is "facility number: ZW-012, disposal link: transport, environmental temperature: 25℃, operation time: 2024-XX-XX".

[0041] The multi-modal data is input into the optimal multi-modal feature extraction model to extract a fusion feature vector; the fusion feature vector and the basic information are filled into a prompt information template to generate a complete prompt; the Llama model interface is called, and the multi-modal original data and the complete prompt are input to generate an output response: {"ZW-012": {"risk level": "medium risk", "risk type": "infectious risk", "risk associated data identifier": "IMG-20240510-003, SENSOR-20240510-078, TEXT-20240510-042", "disposal scene basic information": "facility number: ZW-012, disposal link: transport, environmental temperature: 25℃, operation time: 2024-XX-XX", "risk trigger factor": "transport box damage leading to exposure of infectious waste", "suggestions in the emergency disposal process": "stop transport immediately, replace the special sealed transport box, disinfect the contaminated area, and protect the operator"}.

[0042] The response is stored in a risk response list, and the risk identification of the scene is completed.

[0043] The embodiment of the application also provides a medical waste safe disposal risk identification system, comprising: A multi-modal feature extraction module is configured to acquire multi-modal data of a medical waste disposal scene to be detected, process the multi-modal data through a pre-constructed optimal multi-modal feature extraction model, and obtain a fusion feature vector A prompt template filling module is configured to fill the fusion feature vector and corresponding disposal scene basic information into a pre-set prompt information template to obtain a filled prompt information template; A Llama model response module is configured to call a Llama model interface function, input the multi-modal data of the medical waste disposal scene to be detected and the filled prompt information template into the Llama model, and generate a medical waste safe disposal risk identification result as an output response; In the training process of the pre-constructed optimal multi-modal feature extraction model, the medical waste disposal scene multi-modal data samples are expanded through a diffusion model generator. The optimal multi-modal feature extraction model is a combination model, and the combination model comprises a Mamba model and a Qwen-7B-Chat model.

[0044] In the specific implementation process, the medical waste safe disposal risk identification system further comprises a risk response storage module, which is used for storing the risk identification response of each disposal scene in a risk response list, supporting subsequent query, export and reference calling in the emergency disposal process.

[0045] It can be understood that the medical waste safe disposal risk identification system provided by the embodiments of the present application corresponds to the medical waste safe disposal risk identification method described above, and the explanation, examples, beneficial effects and the like of related contents can refer to the corresponding contents in the medical waste safe disposal risk identification method, which will not be repeated here.

[0046] The embodiments of the present application also provide a computer readable storage medium storing a computer program for medical waste safe disposal risk identification, wherein the computer program causes a computer to execute the medical waste safe disposal risk identification method described above.

[0047] The embodiments of the present application also provide an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program comprises a program for executing the medical waste safe disposal risk identification method described above.

[0048] In summary, compared with the prior art, the present application has the following beneficial effects: 1. The multi-modal data enhancement is realized by the diffusion model generator, the generated samples have higher quality and stronger diversity, the problems of medical waste disposal risk sample scarcity and uneven distribution are solved, and balanced and sufficient data sources are provided for model training.

[0049] 2. The multi-modal feature extractor combined with the Mamba model and the Qwen-7B-Chat model utilizes the sequence modeling advantage of the Mamba architecture and the semantic understanding ability of the Qwen model, efficiently fuses the core information of three types of data of image, sensor and text, and improves the information utilization rate.

[0050] 3. The generative artificial intelligence risk identification framework is constructed based on the Llama model, combined with multi-modal features and prompt optimization, the "hallucination" problem is avoided, and the new risk identification capability is provided, which can cope with the unobserved risk type without retraining.

[0051] 4. The model is low in updating cost. When the disposal scene or risk type changes, only a small amount of data is needed to fine-tune the prompt template or feature extractor, so that the new demand can be quickly adapted, and the generalization ability is strong.

[0052] 5. Multimodal data fusion and generative artificial intelligence can complete complex tasks such as risk detection, type judgment, grade evaluation and emergency suggestion generation, reduce manual intervention, improve risk identification efficiency and accuracy, and reduce medical waste disposal safety hazards.

[0053] In summary, the embodiment of the application realizes efficient, accurate, comprehensive and adaptive medical waste disposal risk identification technology to meet the higher requirements of public health safety and ecological environment safety on medical waste disposal risk control.

[0054] It should be noted that in this paper, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0055] The above embodiments are only used to illustrate the technical solutions of the application, but not to limit it; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for identifying risks in the safe disposal of medical waste, characterized in that, include: Multimodal data of the medical waste disposal scenario to be tested is acquired, and the multimodal data is processed by a pre-constructed optimal multimodal feature extraction model to obtain a fused feature vector. The fused feature vector and its corresponding basic information of the handling scenario are filled into a pre-set prompt information template to obtain the filled prompt information template; Call the Llama model interface function, input the multimodal data of the medical waste disposal scenario to be detected and the filled prompt information template into the Llama model, and generate the medical waste safe disposal risk identification result as the output response; During the training process, the pre-constructed optimal multimodal feature extraction model expands the multimodal data samples of the medical waste disposal scenario through a diffusion model generator. The optimal multimodal feature extraction model is a combined model, which includes the Mamba model and the Qwen-7B-Chat model.

2. The method for identifying risks in the safe disposal of medical waste as described in claim 1, characterized in that, The multimodal data includes scene image data, sensor monitoring data, and text description data.

3. The method for identifying risks in the safe disposal of medical waste as described in claim 2, characterized in that, The expansion of multimodal data samples for medical waste disposal scenarios using a diffusion model generator includes: The historical multimodal data of medical waste disposal scenarios are standardized, the resolution of scene image data is unified, missing values ​​of sensor monitoring data are filled in, redundant information is removed from text description data, and risk status is marked. The standardized historical multimodal data was randomly divided into the original training set and the original test set; Set the hyperparameters of the diffusion model generator, train and optimize the diffusion model generator based on the original training set and the original test set, and save the optimal diffusion model generator parameters. Load the optimal diffusion model generator parameters, and generate supplementary data for image modality and sensor numerical modality respectively through the diffusion model generator. Text modality data is expanded by context-aware rewriting.

4. The method for identifying risks in the safe disposal of medical waste as described in claim 1, characterized in that, The process of obtaining the pre-set prompt message template includes: The initial prompt text is constructed by combining historical multimodal data, fused feature vectors, basic information about the handling scenario, and the output format of risk identification results. Based on the preset output format requirements, construct the corresponding Python class, with the attributes in the class corresponding one-to-one with the keys of the output JSON object, thus completing the output template construction; Input the initial prompt text corresponding to each data point in the historical multimodal data into the Llama model; Store the output response for each data point as a list of model responses; Compare the model response list with the real label list, focusing on verifying the risk level and risk type, and calculate the recognition accuracy. When the recognition accuracy is lower than the set threshold, continuously fine-tune the multimodal feature extraction model and optimize the prompt text structure until the accuracy is higher than the threshold. Store the prompts that meet the requirements as templates to obtain prompt information templates.

5. The method for identifying risks in the safe disposal of medical waste as described in any one of claims 1 to 4, characterized in that, The basic information of the disposal scenario includes the disposal facility number, disposal process, ambient temperature and humidity, operation time, and operator number.

6. The method for identifying risks in the safe disposal of medical waste as described in any one of claims 1 to 4, characterized in that, The results of the risk identification for the safe disposal of medical waste must meet the following format requirements: Determine whether there are security risks in the disposal scenario, and explain the multimodal data identifiers associated with the risks and the basic information of the disposal scenario; Assess and output the risk level, risk type, risk triggering factors, and recommendations for emergency response.

7. The method for identifying risks in the safe disposal of medical waste as described in any one of claims 1 to 4, characterized in that, The method for identifying risks in the safe disposal of medical waste also includes storing the output response of each disposal scenario in a risk response list, which supports subsequent querying, exporting, and reference during emergency response.

8. A risk identification system for the safe disposal of medical waste, characterized in that, include: The multimodal feature extraction module is used to acquire multimodal data from the medical waste disposal scenario to be inspected. It processes the multimodal data using a pre-built optimal multimodal feature extraction model to obtain a fused feature vector. The prompt template filling module is used to fill the fused feature vector and its corresponding basic information of the handling scenario into a pre-set prompt information template to obtain the filled prompt information template; The Llama model response module calls the Llama model interface function to input the multimodal data of the medical waste disposal scenario to be detected and the filled prompt information template into the Llama model, and generates the medical waste safe disposal risk identification result as the output response; During the training process, the pre-constructed optimal multimodal feature extraction model expands the multimodal data samples of the medical waste disposal scenario through a diffusion model generator. The optimal multimodal feature extraction model is a combined model, which includes the Mamba model and the Qwen-7B-Chat model.

9. A computer-readable storage medium, characterized in that, It stores a computer program for identifying risks in the safe disposal of medical waste, wherein the computer program causes a computer to execute the risk identification method for the safe disposal of medical waste as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the medical waste safe disposal risk identification method as described in any one of claims 1 to 7.

Images