Medical product production risk management and control AI robot

Abstract

The invention discloses a medical product production risk management and control AI robot, which comprises a robot body, a processor, a memory and a computer program which is stored on the memory and can run on the processor are arranged in the robot body, and the processor executes the computer program and realizes a medical product production risk management and control method; the medical product production risk management and control method comprises the following steps: step 1, risk factor perception: collecting temperature, humidity, wind speed, dust particle concentration and biological aerosol concentration data of different positions in a production space; collecting pressure difference data between an air inlet and an air outlet of the production workshop; 2, constructing a concentration prediction model, wherein the architecture comprises an ARIMA main model, an LSTM compensation model and an output layer; 3, training a risk prediction network, and training and establishing a risk factor prediction model group; 4, risk factor prediction; 5, production risk cognition and decision making are carried out; and step 6, carrying out production risk dynamic evaluation, alarm and intelligent management and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent management and control of production risks in the pharmaceutical industry, and specifically provides an AI robot for management and control of production risks of pharmaceutical products. Background Art

[0002] Safety risk management in pharmaceutical factories is a key component in ensuring the safety and health of personnel, product quality, and a clean and safe factory environment. Specifically, pharmaceutical factories have strict production environment requirements, including temperature control, humidity control, wind speed control, and monitoring of dust particle concentration and bioaerosol concentration.

[0003] Specifically, in pharmaceutical factories, environmental parameters such as temperature, humidity, wind speed, dust particle concentration, and bioaerosol concentration are not independent of each other, and there is interaction between them. If the temperature is too high (>24°C), it will promote the growth of microorganisms, and it will also cause workers to sweat more, resulting in an increase in the release of microorganisms and an increase in the risk of contamination of pharmaceutical products. If the humidity is too high (>60% RH), it will accelerate the growth of microorganisms (increased activity of bacteria and fungal spores), and may also cause problems such as moisture absorption and agglomeration of materials (such as powder preparations). If the wind speed is too low, the "airflow piston" cannot be effectively formed, and the risk of dust particles and microorganisms being retained increases, which can easily cause pollutants to accumulate in key operating areas (such as filling points) and directly contaminate the product.

[0004] Therefore, in order to achieve production control of pharmaceutical products, it is necessary to monitor the temperature, humidity, wind speed, dust particle concentration and bioaerosol concentration in the production space in real time to ensure that the monitoring data is within the appropriate range. However, in the existing technology, most of these environmental parameters (including temperature, humidity, wind speed, dust particle concentration and bioaerosol concentration) are monitored in real time by sensor measurement. However, due to the limited sensor measurement data, effective monitoring cannot be achieved for areas where sensors are not arranged, resulting in monitoring loopholes in the existing monitoring methods. In particular, since the dust particle concentration and bioaerosol concentration have the characteristics of uneven spatial distribution, the existing monitoring methods using a limited number of sensors may lead to the problem that the local dust particle concentration and bioaerosol concentration are too high but not effectively controlled. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an AI robot for pharmaceutical product production risk management, which adopts a data-driven approach to predict the dust particle concentration and bioaerosol concentration at any location in the production space, and through risk index evaluation of each monitoring data, it can realize dynamic management of pharmaceutical product production risks.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A pharmaceutical product production risk management AI robot includes a robot body, wherein the robot body is provided with a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements a pharmaceutical product production risk management method;

[0008] The pharmaceutical product production risk management method comprises the following steps:

[0009] Step 1: Risk Factor Perception

[0010] Collect temperature data, humidity data, wind speed data, dust particle concentration data, and bioaerosol concentration data at different locations in the production space; collect pressure difference data between the air inlet and outlet of the production workshop;

[0011] A first data set is constructed using temperature data, humidity data, pressure difference data, wind speed data, and dust particle concentration data; a second data set is constructed using temperature data, humidity data, pressure difference data, wind speed data, and bioaerosol concentration data;

[0012] Step 2: Concentration prediction model construction

[0013] The concentration prediction model architecture includes an ARIMA main model, an LSTM compensation model, and an output layer; the ARIMA main model is used to obtain a preliminary prediction value; the LSTM compensation model uses the standardized residual between the preliminary prediction value of the ARIMA main model and the real data as input, and adopts a multi-input and multi-output iterative multi-step prediction strategy to obtain the compensated prediction value of the LSTM compensation model; the output layer combines and adds the preliminary prediction value of the ARIMA main model and the compensated prediction value of the LSTM compensation model to obtain the final prediction value;

[0014] Step 3: Risk prediction network training

[0015] Using the first data set and the second data set to train concentration prediction models respectively, a dust particle concentration prediction model and a bioaerosol concentration prediction model are obtained to predict corresponding risk factors;

[0016] Step 4: Risk Factor Prediction

[0017] The dust particle concentration prediction model and the bioaerosol concentration prediction model are used to predict the dust particle concentration value and the bioaerosol concentration value respectively;

[0018] Step 5: Production risk awareness and decision-making

[0019] Constructing a production risk index assessment model, wherein the production risk index assessment model obtains a production risk index based on temperature data, humidity data, pressure difference data, wind speed data, and predicted dust particle concentration values ​​and bioaerosol concentration values;

[0020] Step 6: Dynamic production risk assessment, alarm and intelligent management and control

[0021] Calculate the production risk index at different locations in the production space, draw a multi-dimensional risk index radar chart, obtain the risk index radar trend chart that changes over time, and dynamically evaluate and control production risks.

[0022] Furthermore, the ARIMA main model includes an AR model and an MA model, and the prediction formula of the ARIMA main model is:

[0023]

[0024] Among them: s is a constant term; y t is the predicted value of the ARIMA main model at time t; s AR is the AR model coefficient; s MA is the MA model coefficient; ε t is the error term at time t; p and q are model parameters.

[0025] Furthermore, the input of the LSTM compensation model is expressed as:

[0026]

[0027] Where: r is the input of the LSTM compensation model, that is, the predicted value of the ARIMA main model The residual between y and the true data y.

[0028] Furthermore, the principle of the output layer is:

[0029]

[0030] in: is the predicted value of the concentration prediction model; is the predicted value of the ARIMA main model; The predicted value of the LSTM compensation model.

[0031] Furthermore, in the production risk index assessment model, for temperature data, humidity data, wind speed data and pressure difference data, the risk loss function is:

[0032] L1(x)=K(x―m) 2

[0033] Where: L1(x) is the risk loss function; m is the target value of the risk characteristic; x is the independent variable, representing temperature data, humidity data, wind speed data or pressure difference data; K is a constant that does not depend on x;

[0034] In the case of asymmetric two-sided specification limits and unequal losses for exceeding the upper and lower limits, the risk loss function is a piecewise function, expressed as:

[0035]

[0036] Where: C1 and C2 represent the risk loss exceeding the lower limit and the risk loss exceeding the upper limit respectively; U max and U min Indicates that the independent variable exceeds the upper limit and the lower limit respectively; Δ1 and Δ2 are the tolerance of the independent variable exceeding the lower limit and the upper limit respectively;

[0037] The expected value of risk loss for temperature data, humidity data, wind speed data, and pressure difference data is:

[0038]

[0039] Where: E[L1(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics.

[0040] Furthermore, in the production risk index assessment model, the risk loss function for dust particle concentration and bioaerosol concentration is:

[0041] L2(x)=Kx 2

[0042] Where: L2(x) is the risk loss function; x is the independent variable, representing the dust particle concentration or bioaerosol concentration; K is a constant that does not depend on x;

[0043] The risk loss function is a piecewise function, expressed as:

[0044]

[0045] Where: C is the risk loss; Δ is the tolerance of the independent variable exceeding the upper limit; U max The independent variable exceeds the upper limit;

[0046] The expected value of risk loss for dust particle concentration and bioaerosol concentration is:

[0047]

[0048] Where: E[L2(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics.

[0049] Furthermore, the robot body is a ground inspection robot, and is equipped with a temperature sensor for collecting temperature data, a humidity sensor for collecting humidity data, a wind speed sensor for collecting wind speed data, a dust particle concentration detection sensor for collecting dust particle concentration data, and a bioaerosol concentration detection sensor for collecting bioaerosol concentration data;

[0050] The temperature sensor, humidity sensor, wind speed sensor, dust particle concentration detection sensor, and bioaerosol concentration detection sensor are electrically connected to the processor and store the collected data in the memory;

[0051] Air pressure sensors are respectively provided at the air inlet and the air outlet of the production workshop. The processor obtains pressure difference data based on the air pressure data collected by the two air pressure sensors and stores the data in the memory.

[0052] The beneficial effects of the present invention are:

[0053] The pharmaceutical product production risk management AI robot of the present invention constructs a concentration prediction model by combining the ARIMA main model and the LSTM compensation model. The ARIMA main model is a typical time series prediction model suitable for non-stationary data. As a combination of the autoregressive moving average model and the differential operation, it can effectively fit the linear features in the time series data. However, the ARIMA main model has a strong dependence on parameters and can only construct a linear model, and cannot handle the nonlinearity in the time series; therefore, it is combined with the LSTM compensation model. Compared with the ARIMA main model, the LSTM compensation model can better handle the nonlinear characteristics of the data. By combining the ARIMA main model and the LSTM compensation model to construct a concentration prediction model, it has the linear series prediction of the ARIMA model and the strong nonlinear modeling ability of the LSTM neural network, and can meet the requirements for accurate prediction of dust particle concentrations and bioaerosol concentrations that change dynamically over time;

[0054] By adopting the process evaluation model to construct a production risk index assessment model, the production risk index assessment model is used to obtain the risk indices of temperature data, humidity data, wind speed data, dust particle concentration data and bioaerosol concentration data respectively, and a multi-dimensional risk index radar chart including the temperature data risk index, humidity data risk index, wind speed data risk index, dust particle concentration data risk index and bioaerosol concentration data risk index is drawn to obtain the risk index radar trend that changes over time, and ultimately realize dynamic assessment and control of production risks.

[0055] In summary, the pharmaceutical product production risk management AI robot of the present invention adopts a data-driven approach to predict the dust particle concentration and bioaerosol concentration at any location in the production space, and by performing risk index assessment on each monitoring data, it can realize dynamic management of pharmaceutical product production risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0057] Figure 1 It is a schematic diagram of the structure of the robot body;

[0058] Figure 2 A flowchart of the risk management approach for pharmaceutical product production;

[0059] Figure 3 This is the principle diagram of the concentration prediction model;

[0060] Figure 4 To draw a multi-dimensional risk index radar chart at a certain empty point at a certain moment. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0062] The pharmaceutical product production risk management AI robot of this embodiment includes a robot body, which is equipped with a processor, a memory, and a computer program stored in the memory and runnable on the processor. The processor executes the computer program and implements the pharmaceutical product production risk management method.

[0063] like Figure 1 As shown, in this embodiment, the robot body 1 is a ground inspection robot. It is equipped with a temperature sensor 2 for collecting temperature data, a humidity sensor 3 for collecting humidity data, a wind speed sensor 4 for collecting wind speed data, a dust particle concentration sensor 5 for collecting dust particle concentration data, and a bioaerosol concentration sensor 6 for collecting bioaerosol concentration data. Specifically, the temperature sensor 2, humidity sensor 3, wind speed sensor 4, dust particle concentration sensor 5, and bioaerosol concentration sensor 6 are each electrically connected to the processor and store the collected data in the memory. In this embodiment, air pressure sensors are installed at the air inlet and outlet of the production workshop. The air pressure data detected by the air pressure sensors is transmitted to the processor. The processor then obtains pressure differential data based on the air pressure data collected by the two air pressure sensors and stores it in the memory.

[0064] Of course, in some other embodiments, the robot body 1 can also be a fixed server, while the temperature sensor 2, humidity sensor 3, wind speed sensor 4, dust particle concentration detection sensor 5 and bioaerosol concentration detection sensor 6 are fixedly set in the production room to realize data collection.

[0065] Specifically, such as Figure 2 As shown, the pharmaceutical product production risk management method of this embodiment includes the following steps.

[0066] Step 1: Risk Factor Perception

[0067] Collect temperature data, humidity data, wind speed data, dust particle concentration data, and bioaerosol concentration data at different locations within the production space; collect pressure difference data between the air inlet and outlet of the production workshop. Specifically, arrange temperature sensors at different locations in the production space to collect temperature data in real time; arrange humidity sensors at different locations in the production space to collect humidity data in real time; arrange wind speed sensors at the air inlet, air outlet, and other locations in the production workshop to collect wind speed data in real time; set air pressure sensors at the air inlet and air outlet of the production workshop to obtain pressure difference data between the air inlet and air outlet of the production workshop; arrange dust particle concentration detection sensors at different locations in the production space to collect dust particle concentration data in real time; arrange bioaerosol concentration detection sensors at different locations in the production space to collect bioaerosol concentration data in real time.

[0068] The first dataset is constructed using temperature data, humidity data, pressure difference data, wind speed data, and dust particle concentration data; the second dataset is constructed using temperature data, humidity data, pressure difference data, wind speed data, and bioaerosol concentration data. The first dataset can be divided into a training set, a test set, and a validation set based on the concentration prediction model training requirements.

[0069] Step 2: Concentration prediction model construction

[0070] The concentration prediction model architecture includes the ARIMA main model, LSTM compensation model and output layer, such as Figure 3 As shown. The ARIMA main model is used to obtain a preliminary prediction value; the LSTM compensation model uses the residual between the preliminary prediction value of the ARIMA main model and the real data after normalization as input, and adopts a multi-input and multi-output iterative multi-step prediction strategy to obtain the compensated prediction value of the LSTM compensation model; the output layer combines and adds the preliminary prediction value of the ARIMA main model and the compensated prediction value of the LSTM compensation model to obtain the final prediction value;

[0071] Specifically, the ARIMA main model includes an AR model and an MA model. The principle of the ARIMA main model is:

[0072]

[0073] Among them: s is a constant term; y t is the predicted value of the ARIMA main model at time t; s AR is the AR model coefficient; s MA is the MA model coefficient; ε t is the error term at time t; p and q are model parameters.

[0074] The input of the LSTM compensation model is represented as:

[0075]

[0076] Where: r is the input of the LSTM compensation model, that is, the predicted value of the ARIMA main model The residual between y and the true data y.

[0077] The principle of the output layer is:

[0078]

[0079] in: is the predicted value of the concentration prediction model; is the predicted value of the ARIMA main model; The predicted value of the LSTM compensation model.

[0080] Step 3: Risk prediction network training

[0081] The first and second datasets are used to train concentration prediction models, respectively, to generate dust particle concentration prediction models and bioaerosol concentration prediction models, which are then used to predict corresponding risk factors. The training methods for the dust particle concentration prediction model and the bioaerosol concentration prediction model are similar to those used for existing deep learning models (such as the LSTM prediction model) and are not detailed here.

[0082] Step 4: Risk Factor Prediction

[0083] The dust particle concentration prediction model and the bioaerosol concentration prediction model were used to predict the dust particle concentration values ​​and the bioaerosol concentration values ​​respectively.

[0084] Step 5: Production risk awareness and decision-making

[0085] A production risk index assessment model is constructed, wherein the production risk index is obtained by using temperature data, humidity data, pressure difference data, wind speed data, and predicted dust particle concentration data values ​​and bioaerosol concentration values.

[0086] In the production risk index assessment model, temperature data, humidity data, wind speed data and pressure difference data are target characteristics, that is, these data all have a risk characteristic target value m and upper and lower specification limits U max and U min Even if the data x is within the specification limits (i.e. U min <x<U max ), as long as x≠m, there will be risks; the greater the deviation of x from the target value m, the greater the risk.

[0087] Specifically, temperature data, humidity data, wind speed data, and pressure difference data are target characteristics, and their risk loss function is:

[0088] L1(x)=K(x―m) 2

[0089] Where: L1(x) is the risk loss function; m is the target value of the risk characteristic; x is the independent variable, representing temperature data, humidity data, wind speed data, or pressure difference data; K is a constant that does not depend on x. When the risk of exceeding the upper limit and the lower limit (C) are equal, K = C / Δ 2 , where Δ is the tolerance.

[0090] Next, we will generalize this to the case where the specification limits are asymmetric and the losses from exceeding the upper and lower limits are not equal. max ) when the upper limit is exceeded, and when the lower limit (x min ) when the risk of exceeding the lower limit arises.

[0091] For temperature data, if the temperature data exceeds the upper limit (T>T max ), the risks of exceeding the upper limit include: causing microbial reproduction, increased sweating of personnel leading to increased release of particles and microorganisms, etc.; if the temperature data exceeds the lower limit (T <T min ), the risks of exceeding the lower limit include: causing discomfort to personnel and increasing the rate of operational errors; changes in the viscosity of certain materials (such as syrups and ointments), affecting filling accuracy, etc.

[0092] For humidity data, if the humidity data exceeds the upper limit (R>R max ), the risks of exceeding the upper limit include: accelerated microbial growth (enhanced activity of bacteria and fungal spores); moisture absorption and agglomeration of materials (such as powder preparations), or deformation of packaging materials (aluminum foil, paper boxes); condensation on the surface of equipment, causing circuit short circuit or metal corrosion, etc.; if the temperature data exceeds the lower limit (R <R min ), resulting in risks exceeding the lower limit include: increased static electricity accumulation, adsorption of dust particles and contamination of products; dry skin on personnel, increased dandruff, and increased risk of particulate contamination, etc. ​

[0093] For wind speed data, if the wind speed data exceeds the upper limit (u>u max ), the risks of exceeding the upper limit include: forming a local vortex, which will roll particles from the ground and walls to the key operation area (such as filling needles and open containers); destroying the airflow barrier, allowing external contaminated air to invade the core sterile area (such as Class A area); high-speed airflow will flush the surface of the equipment or the ground, causing the settled particles (including microorganisms) to be re-suspended, resulting in a surge in the risk of contamination, etc. If the temperature data exceeds the lower limit (u min ), resulting in risks exceeding the lower limit include: the inability to effectively form an "air flow piston", which increases the risk of particle and microbial retention; pollutants accumulate in key operating areas (such as filling points), directly contaminating the product; local contamination cannot be discharged in time, forming a continuous source of pollution, etc.

[0094] For differential pressure data, if the differential pressure data exceeds the upper limit (P>P max ), the wind speed data will be too large; if the pressure difference data exceeds the lower limit (P <P min ), the wind speed data will be too small.

[0095] Thus, for temperature data, humidity data, wind speed data, and pressure difference data with visual characteristics, the risk loss function is a piecewise function, which can be expressed as:

[0096]

[0097] Where: C1 and C2 represent the risk loss exceeding the lower limit and the risk loss exceeding the upper limit respectively; U max and U min Indicates exceeding the upper limit and exceeding the lower limit respectively; Δ1 and Δ2 are exceeding the lower limit tolerance and exceeding the upper limit tolerance respectively.

[0098] Therefore, the expected risk loss value of temperature data, humidity data, wind speed data and pressure difference data is:

[0099]

[0100] Where: E[L1(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics.

[0101] Using the expected value of risk loss to evaluate process risk can enable managers to better manage process risk and make different process risks comparable.

[0102] When applying this model, it is necessary to infer the distribution of the process risk characteristics (i.e., f(x)) based on the sample statistics, and then substitute it into the expected value of risk loss to obtain the expected value of risk loss.

[0103] Specifically, if the eye characteristics follow the normal distribution N(μ,σ 2 ),Right now:​

[0104]

[0105] Then the expected value of risk loss is as follows:

[0106]

[0107] Wherein:

[0108] In the production risk index evaluation model, the dust particle concentration and the bio-aerosol concentration are characteristics of being desirable to be small. That is, in the production of pharmaceutical products, it is hoped that the data of the dust particle concentration and the bio-aerosol concentration are as small as possible. That is, the expected values of the dust particle concentration and the bio-aerosol concentration have only upper limits and no lower limits. The risk loss function of the characteristic of being desirable to be small x within the specification limit (i.e., 0 ≤ x < USL) is:

[0109] L2(x) = Kx 2

[0110] Wherein: L2(x) is the risk loss function; x is the independent variable, representing the dust particle concentration or the bio-aerosol concentration; K is a constant independent of x, and K = C / Δ 2 ,

[0111] The risk loss function is a piecewise function, expressed as:[[ID=,29]]

[0112]

[0113] Wherein: C is the risk loss; U max is the upper limit exceeded. [[ID=3,8]]

[0114] The expected value of risk loss of the dust particle concentration and the bio-aerosol concentration is:

[0115]

[0116] Wherein: E[L2(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics; Δ is the tolerance.

[0117] Step 6: Dynamic evaluation, alarm and intelligent control of production risk

[0118] Calculate the production risk index at different positions in the production space, draw a multi-dimensional risk index radar chart, and obtain a risk index radar trend chart that changes with time, so as to dynamically evaluate and control the production risk. As Figure 4 shown, it is a multi-dimensional risk index radar chart drawn at a certain moment at a certain empty point position. Through the risk index radar chart, the risk indexes of temperature, humidity, differential pressure, wind speed, dust particle concentration and bio-aerosol concentration can be obtained simultaneously.

[0119] The pharmaceutical product production risk management AI robot of this embodiment constructs a concentration prediction model by combining the ARIMA main model and the LSTM compensation model. The ARIMA main model is a typical time series prediction model suitable for non-stationary data. As a combination of the autoregressive moving average model and the difference operation, it can effectively fit the linear features in the time series data. However, the ARIMA main model has a strong dependence on parameters and can only construct a linear model, and cannot handle the nonlinearity in the time series; therefore, it is combined with the LSTM compensation model. Compared with the ARIMA main model, the LSTM compensation model can better handle the nonlinear characteristics of the data. By combining the ARIMA main model and the LSTM compensation model to construct a concentration prediction model, it has the linear series prediction of the ARIMA model and the strong nonlinear modeling capability of the LSTM neural network, which can meet the requirements for accurate prediction of dust particle concentration and bioaerosol concentration that change dynamically over time.

[0120] By adopting the process evaluation model to construct a production risk index assessment model, the production risk index assessment model is used to obtain the risk indices of temperature data, humidity data, wind speed data, dust particle concentration data and bioaerosol concentration data respectively, and a multi-dimensional risk index radar chart including the temperature data risk index, humidity data risk index, wind speed data risk index, dust particle concentration data risk index and bioaerosol concentration data risk index is drawn to obtain the risk index radar trend that changes over time, and ultimately realize dynamic assessment and control of production risks.

[0121] In summary, the pharmaceutical product production risk management AI robot of this embodiment adopts a data-driven approach to predict the dust particle concentration and bioaerosol concentration at any location in the production space, and by performing risk index assessment on each monitoring data separately, it can realize dynamic management of pharmaceutical product production risks.

[0122] The above-described embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. An AI robot for pharmaceutical product production risk management, characterized by: The robot comprises a robot body, wherein the robot body is provided with a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program and implements a method for intelligent management and control of pharmaceutical product production risks; The method for intelligent management and control of pharmaceutical product production risks comprises the following steps: Step 1: Risk Factor Perception Collect temperature data, humidity data, wind speed data, dust particle concentration data, and bioaerosol concentration data at different spatial locations in the production workshop; collect pressure difference data between the air inlet and outlet of the production workshop; A first data set is constructed using temperature data, humidity data, pressure difference data, wind speed data, and dust particle concentration data; a second data set is constructed using temperature data, humidity data, pressure difference data, wind speed data, and bioaerosol concentration data; Step 2: Concentration prediction model construction The concentration prediction model architecture includes an ARIMA main model, an LSTM compensation model, and an output layer; the ARIMA main model is used to obtain a preliminary prediction value; the LSTM compensation model uses the standardized residual between the preliminary prediction value of the ARIMA main model and the real data as input, and adopts a multi-input and multi-output iterative multi-step prediction strategy to obtain the compensated prediction value of the LSTM compensation model; the output layer combines and adds the preliminary prediction value of the ARIMA main model and the compensated prediction value of the LSTM compensation model to obtain the final prediction value; Step 3: Risk prediction network training Using the first data set and the second data set to train concentration prediction models respectively, a dust particle concentration prediction model and a bioaerosol concentration prediction model are obtained to predict corresponding risk factors; Step 4: Risk Factor Prediction The dust particle concentration prediction model and the bioaerosol concentration prediction model are used to predict the dust particle concentration value and the bioaerosol concentration value respectively; Step 5: Production risk awareness and decision-making Constructing a production risk index assessment model, wherein the production risk index assessment model obtains a production risk index based on temperature data, humidity data, pressure difference data, wind speed data, and predicted dust particle concentration values ​​and bioaerosol concentration values; Step 6: Dynamic production risk assessment, alarm and intelligent management and control Calculate the production risk index at different spatial locations in the production workshop, draw a multi-dimensional risk index radar chart, obtain the risk index radar trend chart that changes over time, conduct dynamic assessment of production risks, and issue alarms and conduct management and control based on risk levels.

2. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: The ARIMA main model includes an AR model and an MA model. The prediction formula of the ARIMA main model is: Among them: s is a constant term; y t is the predicted value of the ARIMA main model at time t; s AR is the AR model coefficient; s MA is the MA model coefficient; ε t is the error term at time t; p and q are model parameters.

3. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: The input of the LSTM compensation model is expressed as: Where: r is the input of the LSTM compensation model, that is, the predicted value of the ARIMA main model The residual between y and the true data y.

4. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: The principle of the output layer is: in: is the predicted value of the concentration prediction model; is the predicted value of the ARIMA main model; The predicted value of the LSTM compensation model.

5. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: In the production risk index assessment model, the risk loss function for temperature data, humidity data, wind speed data, and pressure difference data is: L1(x)=K(x―m) 2 Where: L1(x) is the risk loss function; m is the target value of the risk characteristic; x is the independent variable, representing temperature data, humidity data, wind speed data or pressure difference data; K is a constant that does not depend on x; In the case of asymmetric two-sided specification limits and unequal losses for exceeding the upper and lower limits, the risk loss function is a piecewise function, expressed as: Where: C1 and C2 represent the risk loss exceeding the lower limit and the risk loss exceeding the upper limit respectively; U max and U min Indicates that the independent variable exceeds the upper limit and the lower limit respectively; Δ1 and Δ2 are the tolerance of the independent variable exceeding the lower limit and the upper limit respectively; The expected value of risk loss for temperature data, humidity data, wind speed data, and pressure difference data is: E[L1(x)] Where: E[L1(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics.

6. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: In the production risk index assessment model, the risk loss function for dust particle concentration and bioaerosol concentration is: L2(x)=X 2 Where: L2(x) is the risk loss function; x is the independent variable, representing the dust particle concentration or bioaerosol concentration; K is a constant that does not depend on x; The risk loss function is a piecewise function, expressed as: Where: C is the risk loss; Δ is the tolerance of the independent variable exceeding the upper limit; U max The independent variable exceeds the upper limit; The expected value of risk loss for dust particle concentration and bioaerosol concentration is: Where: E[L2(x)] is the expected value of risk loss; f(x) is the distribution of risk characteristics.

7. The pharmaceutical product production risk management AI robot according to claim 1, characterized in that: The robot body is a ground inspection robot, and is equipped with a temperature sensor for collecting temperature data, a humidity sensor for collecting humidity data, a wind speed sensor for collecting wind speed data, a dust particle concentration detection sensor for collecting dust particle concentration data, and a bioaerosol concentration detection sensor for collecting bioaerosol concentration data; The temperature sensor, humidity sensor, wind speed sensor, dust particle concentration detection sensor, and bioaerosol concentration detection sensor are electrically connected to the processor and store the collected data in the memory; Air pressure sensors are respectively provided at the air inlet and the air outlet of the production workshop. The processor obtains pressure difference data based on the air pressure data collected by the two air pressure sensors and stores the data in the memory.

Images