Machine Learning-Based Abnormality Identification System and Method for Pharmaceutical Cold Chain Transportation

Abstract

The present invention provides a system and method for identifying anomalies in pharmaceutical cold chain transportation based on machine learning, belonging to the technical field of machine learning. The system includes: an information acquisition module for parsing historical collected data according to predefined metrics during the cold chain transportation process to obtain key information; a model construction module for dividing all the key information into training samples and validation samples, training the initialized neural network model according to the training samples, and validating and correcting the trained model according to the validation samples to obtain an anomaly identification model; an anomaly identification module for real-time collecting current information during the cold chain transportation process, inputting it into the anomaly identification model to obtain all anomaly factors, and giving early warning reminders for each anomaly factor to achieve anomaly control. Ensure the effectiveness of comprehensive analysis of current information and the timeliness of control and early warning, and achieve timely elimination of anomalies generated during the transportation process.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and particularly to a system and method for identifying anomalies in pharmaceutical cold chain transportation based on machine learning. Background Art

[0002] In the field of biological cells, cold chain transportation plays a decisive role in ensuring cell viability and the quality of related biological products. Biological cell products, such as CAR-T cell therapy drugs, stem cell preparations, etc., have extremely strict requirements for temperature, time limit, and status during transportation.

[0003] Currently, cold chain transportation data is complex and lacks effective integration and analysis. Each historical cold chain transportation generates a large amount of historical collected data, covering multi-dimensional information such as temperature, humidity, transportation duration, geographical location, etc. However, during the process of anomaly detection, it is based on the comparison of single-dimensional information with the set range, and there is no comprehensive regulation based on all factors, resulting in the inability to comprehensively analyze and regulate the currently collected information. That is, the existing technology cannot analyze and give early warnings to the current information in the cold chain transportation process in a timely and accurate manner.

[0004] Therefore, the present invention proposes a system and method for identifying anomalies in pharmaceutical cold chain transportation based on machine learning. Summary of the Invention

[0005] The present invention provides a system and method for identifying anomalies in pharmaceutical cold chain transportation based on machine learning. By analyzing the collected historical data to obtain samples for training and correcting the model, the accuracy of the model is guaranteed to ensure the effectiveness of comprehensive analysis of current information and the timeliness of control and early warning, and timely elimination of anomalies generated during transportation is achieved.

[0006] The present invention provides a system for identifying anomalies in pharmaceutical cold chain transportation based on machine learning, including:

[0007] An information acquisition module, configured to acquire the historical collected data of each historical cold chain transportation, and analyze the historical collected data according to predefined indicators of the cold chain target during transportation to obtain key information corresponding to the transportation;

[0008] A model construction module, configured to divide all key information into training samples and verification samples, train the initialized neural network model according to the training samples, and verify and correct the trained model according to the verification samples to obtain an anomaly identification model;

[0009] An anomaly identification module, configured to collect the current information during cold chain transportation in real time, input it into the anomaly identification model to obtain all anomaly factors, and give early warning reminders for each anomaly factor to achieve anomaly control.

[0010] Preferably, the historical collected data is multi-dimensional data, including: the transportation temperature, transportation humidity, driving speed of the transportation vehicle, geographical location, and vibration of the cold chain target during transportation at each transportation time point;

[0011] The number of the predefined indicators is consistent with the dimension, each dimension corresponds to a predefined indicator, and the predefined indicators include: temperature deviation range indicator, humidity qualified interval indicator, transportation duration indicator, and vibration allowable threshold indicator.

[0012] Preferably, the information acquisition module includes:

[0013] A range comparison unit for analyzing whether the corresponding dimension value at each transportation time point is within the standard range under the corresponding predefined indicator;

[0014] If the corresponding predefined indicator is satisfied, assign 1, otherwise, assign 0 to the corresponding predefined indicator, and construct a transportation rough vector at the corresponding transportation time point;

[0015] A normalization unit for performing normalization processing on the dimension values under each predefined indicator, obtaining the normalized values, and constructing a transportation fine vector at the corresponding transportation time point;

[0016]

[0017] Wherein, represents the normalized value at the k-th transportation time point under the corresponding predefined indicator, represents the dimension value at the k-th transportation time point under the corresponding predefined indicator, represents the mean value of all dimension values under the corresponding predefined indicator, represents the standard deviation of all dimension values under the corresponding predefined indicator; R0 corresponds to the reference threshold of the predefined indicator; ln represents the logarithmic function symbol; 、 are set constants;

[0018] A control acquisition unit for performing control analysis and alarm analysis on the transportation rough vector and the transportation fine vector, and obtaining corresponding control information and alarm information;

[0019] A key information acquisition unit for using the transportation rough vector, the transportation fine vector, the control information, and the alarm information as the key information for the corresponding transportation, wherein the control information includes control instructions and abnormal factors for control adjustment.

[0020] Preferably, the model construction module includes:

[0021] A coefficient determination unit for determining the difference coefficient of the trained model based on the verification samples;

[0022]

[0023] Among them, represents the difference coefficient between the actual output and the ideal output of the validation samples based on the trained model, represents the ideal output of the j-th validation sample, represents the actual output of the j-th validation sample, represents the rough transportation vector of the j-th validation sample, represents the fine transportation vector of the j-th validation sample, represents the deviation influence coefficient based on the rough transportation vector, and its value range is (0, 1), represents the deviation influence coefficient based on the fine transportation vector, and its value range is (0, 1); represents the difference function of the j-th validation sample based on , and , , respectively represent the regularization coefficients; N represents the number of validation samples;

[0024] The first judgment unit is used to, when the difference coefficient is greater than or equal to the preset coefficient, adopt the adaptive learning rate optimization algorithm and introduce the regularization technology to update the model parameters and continue to train the model;

[0025] The second judgment unit is used to, when the difference coefficient is less than the preset coefficient, obtain new samples according to the difference coefficient, train the model once and then stop training to obtain the anomaly recognition model.

[0026] Preferably, the second judgment unit includes:

[0027] An update subunit is used to update the standardized values involved in each validation sample based on the difference coefficient;

[0028]

[0029] Among them, represents the o-th standardized value in the corresponding validation sample; , respectively represent the boundary range values of the o-th standardized value in the corresponding validation sample; represents the learning rate of the trained model;

[0030] A new sample acquisition unit is used to construct new samples of the corresponding validation samples based on the updated values, train the model once and then stop training.

[0031] Preferably, the first judgment unit includes:

[0032] An update subunit, configured to introduce an L2 regularization term into a loss function to constrain model parameters, calculate a gradient Ft of the loss function based on all model parameters through a backpropagation algorithm, and update the first moment and the second moment;

[0033]

[0034] Wherein, represents the regularized loss function; L represents the loss function before regularization; represents a regularization parameter; represents a set of model parameters of the trained model; represents a model parameter;

[0035] A quantity determination subunit, configured to perform bias correction on the updated first moment and second moment, and determine an update quantity of the corresponding model parameter to update each model parameter until the model meets a preset stop condition.

[0036] Preferably, the quantity determination subunit includes:

[0037] A calculation block, configured to calculate an update quantity of the corresponding model parameter:

[0038]

[0039]

[0040]

[0041] Wherein, represents the update quantity of the corresponding model parameter after the t-th iteration; represents a learning rate; and respectively represent the first moment and the second moment at the t-th iteration; and represent the attenuation coefficients at the t-th iteration; represents a constant, and the value is ; and respectively represent the bias correction results of the first moment and the second moment at the t-th iteration.

[0042] The present invention provides a method for identifying abnormal cold chain transportation of medicines based on machine learning, including:

[0043] Step 1: Obtain historical collection data of each historical cold chain transportation, and parse the historical collection data according to predefined indicators of a cold chain target during transportation to obtain key information corresponding to the transportation;

[0044] Step 2: Divide all key information to obtain training samples and validation samples, train the initialized neural network model according to the training samples, and verify and correct the trained model according to the validation samples to obtain an anomaly recognition model;

[0045] Step 3: Collect the current information during the cold chain transportation in real time, input it into the anomaly recognition model to obtain all anomaly factors, and give early warning reminders for each anomaly factor to achieve anomaly control.

[0046] Compared with the prior art, the beneficial effects of the present application are as follows:

[0047] By analyzing the collected historical data to obtain samples for training and correcting the model, the accuracy of the model is guaranteed to ensure the effectiveness of the comprehensive analysis of the current information and the timeliness of control and early warning, and the timely elimination of anomalies generated during the transportation process is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 is a structural diagram of an abnormal recognition system for pharmaceutical cold chain transportation based on machine learning in an embodiment of the present invention;

[0050] Figure 2 is a flowchart of an abnormal recognition method for pharmaceutical cold chain transportation based on machine learning in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0052] The present invention provides an abnormal recognition system for pharmaceutical cold chain transportation based on machine learning, as Figure 1 shown, including:

[0053] An information acquisition module, configured to acquire historical acquisition data of each historical cold chain transportation, and analyze the historical acquisition data according to predefined indicators of the cold chain target during the transportation process to obtain key information corresponding to the transportation;

[0054] A model construction module, which is used to divide all key information to obtain training samples and validation samples, train the initialized neural network model according to the training samples, and verify and correct the trained model according to the validation samples to obtain an anomaly recognition model;

[0055] An anomaly recognition module, which is used to collect the current information during the cold chain transportation in real time, input it into the anomaly recognition model to obtain all anomaly factors, and give early warning reminders for each anomaly factor to achieve anomaly control.

[0056] Preferably, the historical collected data is multi-dimensional data, including: the transportation temperature, transportation humidity, driving speed of the transportation vehicle, geographical location, and vibration of the cold chain target during transportation at each transportation time point;

[0057] The number of the predefined indicators is consistent with the dimension, each dimension corresponds to a predefined indicator, and the predefined indicators include: temperature deviation range indicator, humidity qualified interval indicator, transportation duration indicator, and vibration allowable threshold indicator.

[0058] In this embodiment, the temperature, humidity, driving speed, geographical location, and vibration are respectively measured by a temperature sensor, a humidity sensor, a speed sensor, a GPS, and a vibration sensor arranged inside the cold chain carriage.

[0059] In this embodiment, for some immune cell products, the predefined indicators may stipulate that the transportation temperature needs to be strictly controlled at 2-8°C, the maximum transportation time cannot exceed 72 hours, and the temperature fluctuation range is within ±1°C, etc.; for stem cell products, it may be required to transport in a cryogenic environment of -150°C to -196°C, and the duration of the temperature deviation from this range shall not exceed 30 minutes.

[0060] In this embodiment, the key information includes: the rough judgment result, the detailed judgment result, the anomaly factor, and the comprehensive control and early warning result of the collected information corresponding to the predetermined indicators.

[0061] In this embodiment, such as temperature anomaly, transportation overtime, route deviation, etc.

[0062] In this embodiment, the process of dividing the training samples and the validation samples generally follows a certain ratio and principle, usually in a ratio of 7:3 or 8:2.

[0063] In this embodiment, during the training process of the neural network model, by learning the key information in the training samples, the model continuously adjusts its own parameters (such as weights and biases) to adapt to the data characteristics under different transportation conditions, and gradually masters the laws of normal transportation and abnormal transportation. For example, the model learns that when the temperature exceeds the predefined range for a certain period of time, it may affect the activity of biological cells, so that similar abnormal situations can be identified in subsequent predictions. After training is completed, the model is verified using verification samples. The verification samples are input into the trained model to obtain the prediction results of the model. Then, the prediction results are compared with the actual abnormal situations in the verification samples, and evaluation indicators such as the prediction accuracy rate and recall rate of the model are calculated. If there is a large deviation between the prediction results of the model and the actual situation, it indicates that there is a problem with the model and corrective measures need to be taken. The corrective measures include adjusting the structure of the model, reselecting training parameters, increasing or decreasing training samples, etc. Through continuous adjustment and optimization, the performance of the model is optimized, and finally an abnormal situation recognition model is obtained.

[0064] Based on the laws and characteristics learned previously, the model analyzes and judges the current information to identify all possible abnormal factors. For example, when the model detects that the current transportation temperature exceeds the predefined index range, it identifies "temperature anomaly" as an abnormal factor; if it is found that the transportation duration is about to exceed the specified time, it identifies "risk of transportation overtime" as an abnormal factor.

[0065] For each abnormal factor identified by the model, the system will immediately issue a warning reminder. The warning methods can include various forms such as sound alarms, SMS notifications, and APP push messages to ensure that transportation management personnel can obtain abnormal information in a timely manner. At the same time, the system will also provide corresponding abnormal control suggestions according to the type and severity of the abnormal factors. For example, for temperature anomalies, it is recommended to immediately adjust the operating parameters of the refrigeration equipment; for the risk of transportation overtime, it is recommended to plan a better transportation route, etc., so as to effectively control the abnormal situations in the cold chain transportation of biological cells.

[0066] The beneficial effects of the above technical solution are as follows: By analyzing the collected historical data to obtain samples for training and correcting the model, the accuracy of the model is guaranteed to ensure the effectiveness of the comprehensive analysis of current information and the timeliness of control and warning, and the timely elimination of abnormalities generated during the transportation process is realized.

[0067] The present invention provides a cold chain transportation abnormal situation recognition system for pharmaceuticals based on machine learning. The information acquisition module includes:

[0068] A range comparison unit for analyzing whether the corresponding dimension value at each transportation time point is within the standard range under the corresponding predefined index;

[0069] If the corresponding predefined index is assigned 1, otherwise, assign 0 to the corresponding predefined index, and construct a rough transport vector at the corresponding transport time point;

[0070] A normalization unit for normalizing the dimensional values under each predefined index to obtain normalized values, and constructing a fine transport vector at the corresponding transport time point;

[0071]

[0072] Where represents the normalized value at the k-th transport time point under the corresponding predefined index, represents the dimensional value at the k-th transport time point under the corresponding predefined index, represents the mean of all dimensional values under the corresponding predefined index, represents the standard deviation of all dimensional values under the corresponding predefined index; R0 is the reference threshold for the corresponding predefined index; ln represents the logarithmic function symbol; 、 are set constants;

[0073] A control acquisition unit for performing control analysis and alarm analysis on the rough transport vector and the fine transport vector to obtain corresponding control information and alarm information;

[0074] A key information acquisition unit for using the rough transport vector, the fine transport vector, the control information, and the alarm information as key information for the corresponding transport, where the control information includes control instructions and abnormal factors for control adjustment.

[0075] In this embodiment, the transport time point is a specific moment when data is recorded at regular time intervals (such as every minute, every five minutes, etc.) during the cold chain transport of biological cells, and is used to identify the time node of data collection during the transport process. The corresponding dimensional value is the data value related to the transport collected by various sensors and devices at each transport time point. If the temperature value collected at a certain transport time point is 9°C, while the predefined temperature range is 2 - 8°C, it is determined that the temperature value is not within the standard range; if the humidity value collected is 50%, which is within the range of 40% - 60% specified by the predefined index, it is determined that the humidity value is within the standard range. In this way, a comprehensive judgment is made on all dimensional values at each transport time point, providing a basis for subsequent data processing.

[0076] The transport rough vector is a vector composed of 0 or 1 values assigned after judging whether the corresponding dimension values at each transport time point meet the predefined indicators. It reflects the compliance of each key indicator during the transport process and is a preliminary quantitative description of the transport state. Suppose there are three predefined indicators: temperature, humidity, and geographical location at a certain transport time point. After judgment, the temperature value is not within the standard range, while the humidity value and the geographical location value are within the standard range. Then the assigned values corresponding to the three predefined indicators at this transport time point are 0, 1, and 1 respectively. Arranging these assigned values in the order of the predefined indicators successively to construct a vector, which is the transport rough vector at this transport time point.

[0077] The transport fine vector is a vector constructed from the values obtained by standardizing the dimension values under each predefined indicator. It retains the original characteristics of the data and unifies data with different dimensions to the same scale, facilitating more accurate data analysis and comparison. By comparing the dimension values at each transport time point with the predefined indicators, 10,000 transport rough vectors are successfully constructed. After statistics, among them, all the assigned values of the predefined indicators in 8,500 transport rough vectors are 1, indicating that the transport state is normal at these transport time points; there are at least one predefined indicator assigned value of 0 in 1,500 transport rough vectors, that is, there are abnormal situations.

[0078] Control analysis is to comprehensively analyze the transport rough vector and the transport fine vector, judge whether there are abnormal situations during the transport process, and formulate corresponding control strategies for the abnormal situations to ensure that the transport process meets the requirements. During the control analysis process, the system will judge whether there are abnormal situations during the transport process according to the distribution of 0 values in the transport rough vector and the deviation degree of the values in the transport fine vector. For example, if the assigned value corresponding to the temperature indicator in the transport rough vector is 0, and the standardized value of the temperature in the transport fine vector exceeds the normal range by a large margin, the system judges that there is a temperature abnormal situation. At this time, a control instruction will be generated according to the preset strategy, such as "increase the power of the refrigeration equipment by 20%", and the abnormal factor for the control adjustment is determined as "too high temperature". These information together constitute the control information. During the alarm analysis process, the system will set different alarm levels according to the severity of the abnormal situation. For example, when the temperature is abnormal and exceeds the standard range slightly, it may be set as a low-level alarm, sending an alarm message of "temperature is slightly abnormal, please pay attention"; when the temperature is abnormal and exceeds the standard range by a large margin, it may be set as a high-level alarm, sending an alarm message of "temperature is severely abnormal, take measures immediately". Through the comprehensive analysis of the transport rough vector and the transport fine vector, the effective monitoring and timely warning of the transport process are realized.

[0079] In this embodiment, the transport coarse vector, transport fine vector, control information, and alarm information include key states during transportation, detailed data characteristics, strategies for handling anomalies, and early warning prompts. Integrating these information together serves as the key information corresponding to the transportation. These key information can be used for various purposes such as subsequent optimization of the transportation process, tracing of abnormal situations, and model training, providing strong support for ensuring the safety and quality of biological cell cold chain transportation. The system accurately identified 1450 abnormal situations (including 800 temperature anomalies, 400 humidity anomalies, and 250 geographical location anomalies), and the identification accuracy rate reached 96.7%. For these abnormal situations, the system generated corresponding control information and alarm information, and the rationality and effectiveness of the control instructions were verified manually and met the actual transportation requirements.

[0080] The beneficial effects of the above technical solution are as follows: The transport coarse vector can intuitively reflect the compliance of each key indicator at the transport time point, facilitating a quick understanding of the general situation of the transport state. The transport fine vector retains the original characteristics of the data while eliminating the influence of dimensions, making data in different dimensions comparable and providing the possibility for more accurate data analysis. The determination of the control information and alarm information is to ensure the rationality and effectiveness of the control instructions obtained after subsequent analysis of the current information, ensuring compliance with actual transportation requirements.

[0081] The present invention provides an abnormal recognition system for pharmaceutical cold chain transportation based on machine learning. The model construction module includes:

[0082] A coefficient determination unit for determining the difference coefficient of the trained model based on the verification samples;

[0083]

[0084] Wherein, represents the difference coefficient between the actual output and the ideal output of the trained model based on the verification samples, represents the ideal output of the j-th verification sample, represents the actual output of the j-th verification sample, represents the transport coarse vector of the j-th verification sample, represents the transport fine vector of the j-th verification sample, represents the deviation influence coefficient based on the transport coarse vector, and its value range is (0, 1), represents the deviation influence coefficient based on the transport fine vector, and its value range is (0, 1); represents the difference function of the j-th verification sample based on , and , , respectively represent the regularization coefficients; N represents the number of validation samples;

[0085] The first judgment unit is used to, when the difference coefficient is greater than or equal to the preset coefficient, adopt an adaptive learning rate optimization algorithm and introduce a regularization technique to update the model parameters and continue to train the model;

[0086] The second judgment unit is used to, when the difference coefficient is less than the preset coefficient, obtain new samples according to the difference coefficient, train the model once and then stop training to obtain an anomaly recognition model.

[0087] In this embodiment, the initial learning rate is set to 0.001. At this time, 、 take the value of 0.001, select 200 historical cold chain transportation data, and divide them into training samples and validation samples according to a ratio of 7:3. The difference coefficient of the model after initial training is 0.35, and the preset coefficient is set to 0.25. Since the difference coefficient is greater than the preset coefficient, the operation of the first judgment unit is triggered.

[0088] The beneficial effects of the above technical solution are: The model optimization method based on the difference coefficient can reasonably select an optimization strategy according to the performance of the model on the validation samples, effectively improve the performance of the biological cell cold chain transportation anomaly recognition model, and has high feasibility and practical value in practical applications.

[0089] The present invention provides a cold chain transportation anomaly recognition system for medicine based on machine learning. The second judgment unit includes:

[0090] An update subunit is used to update the standardized values involved in each validation sample based on the difference coefficient;

[0091]

[0092] wherein, represents the o-th standardized value in the corresponding validation sample; 、 respectively represent the boundary range values of the o-th standardized value in the corresponding validation sample; represents the learning rate of the trained model;

[0093] A new sample acquisition unit is used to construct new samples corresponding to the validation samples based on the updated values, train the model once and then stop training.

[0094] In this embodiment, based on experiments, it is shown that after one training of the model, the comprehensive recognition accuracy of the model for temperature anomalies, humidity anomalies, and geographical location anomalies is 80%, the false negative rate is 12%, and the false positive rate is 8%. After optimization of sample update based on the coefficient of variation, the comprehensive recognition accuracy of the model is increased to 93%, the false negative rate is reduced to 4%, and the false positive rate is reduced to 2%.

[0095] The beneficial effects of the above technical solution are: by updating the samples to ensure the accuracy of subsequent anomaly recognition of the model, and the model can better adapt to the changes in data, improving the recognition ability of abnormal situations in the cold chain transportation of biological cells.

[0096] The present invention provides a cold chain transportation anomaly recognition system for pharmaceuticals based on machine learning. The first judgment unit includes:

[0097] An update subunit, configured to introduce an L2 regularization term into the loss function, constrain the model parameters, calculate the gradient Ft of the loss function based on all model parameters through the backpropagation algorithm, and update the first moment and the second moment;

[0098]

[0099] Wherein, represents the regularized loss function; L represents the loss function before regularization; represents the regularization parameter; represents the set of model parameters of the trained model; represents the model parameter;

[0100] A quantity determination subunit, configured to perform bias correction on the updated first moment and second moment, and determine the update quantity of the corresponding model parameter to update each model parameter until the model meets a preset stop condition.

[0101] Preferably, the quantity determination subunit includes:

[0102] A calculation block, configured to calculate the update quantity of the corresponding model parameter:

[0103]

[0104]

[0105]

[0106] Wherein, represents the update quantity of the corresponding model parameter after the t-th iteration; represents the learning rate; 、 respectively represent the first moment and the second moment at the t-th iteration; 、 represents the attenuation coefficient at the t-th iteration; represents a constant with a value of ; and respectively represent the deviation correction results of the first moment and the second moment at the t-th iteration.

[0107] In this embodiment, a common training method without L2 regularization and adaptive update strategy is used. The recognition accuracy of the model on the training set reaches 85%, but the accuracy on the validation set is only 72%, showing an obvious overfitting phenomenon. Compared with the common training method, the method in this paper has a faster decreasing loss function and reaches a stable state earlier under the same number of training iterations. For example, at the 100th iteration, the loss function value of the common method is 0.8, while the loss function value of the method in this paper has dropped to 0.5.

[0108] The beneficial effects of the above technical solution are as follows: The model parameter optimization method based on L2 regularization and adaptive update can effectively constrain model parameters, avoid overfitting, and at the same time accelerate model convergence by adaptively adjusting the learning rate, significantly improving the performance and reliability of the biological cell cold chain transportation anomaly recognition model, and having high practical value in practical applications.

[0109] The present invention provides a method for identifying anomalies in pharmaceutical cold chain transportation based on machine learning, including:

[0110] Step 1: Obtain the historical collection data of each historical cold chain transportation, and parse the historical collection data according to the predefined indicators of the cold chain target during the transportation process to obtain the key information corresponding to the transportation;

[0111] Step 2: Divide all the key information to obtain training samples and validation samples, train the initialized neural network model according to the training samples, and verify and correct the trained model according to the validation samples to obtain an anomaly recognition model;

[0112] Step 3: Real-time collect the current information during the cold chain transportation process, input it into the anomaly recognition model to obtain all anomaly factors, and give early warning reminders for each anomaly factor to achieve anomaly control.

[0113] The beneficial effects of the above technical solution are as follows: By parsing the collected historical data to obtain samples for training and correcting the model, the accuracy of the model is guaranteed to ensure the effectiveness of the comprehensive analysis of the current information and the timeliness of control and early warning, and the timely elimination of anomalies generated during the transportation process is realized.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some 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 various embodiments of the present invention.

Claims

1. A machine learning-based abnormal recognition system for pharmaceutical cold chain transportation, characterized in that, Including: An information acquisition module, configured to acquire historical acquisition data of each historical cold chain transportation, and parse the historical acquisition data according to predefined indicators of a cold chain target during the transportation process to obtain key information corresponding to the transportation; A model construction module, configured to divide all key information to obtain training samples and validation samples, train an initialized neural network model according to the training samples, and validate and correct the trained model according to the validation samples to obtain an anomaly recognition model; An anomaly recognition module, configured to collect current information during the cold chain transportation in real time, input it into the anomaly recognition model to obtain all anomaly factors, and give a warning reminder for each anomaly factor to achieve anomaly control; Wherein, the information acquisition module includes: A range comparison unit, configured to analyze whether the corresponding dimension value at each transportation time point is within the standard range of the corresponding predefined indicator; Assign 1 to the corresponding predefined indicator, otherwise, assign 0 to the corresponding predefined indicator, and construct a rough transportation vector at the corresponding transportation time point; A normalization unit, configured to perform normalization processing on the dimension values under each predefined indicator to obtain normalized values, and construct a fine transportation vector at the corresponding transportation time point; Among them, represents the standardized value of the k-th transportation time point under the corresponding predefined index, represents the dimensional value of the k-th transportation time point under the corresponding predefined index, represents the mean value of all dimensional values under the corresponding predefined index, represents the standard deviation of all dimensional values under the corresponding predefined index; R0 corresponds to the reference threshold of the predefined index; ln represents the logarithmic function symbol; and are set constants; A control acquisition unit, configured to perform control analysis and alarm analysis on the rough transportation vector and the fine transportation vector to obtain corresponding control information and alarm information; A key information acquisition unit, configured to use the rough transportation vector, the fine transportation vector, the control information, and the alarm information as the key information corresponding to the transportation, wherein the control information includes control instructions and anomaly factors for control adjustment.

2. The abnormal recognition system for pharmaceutical cold chain transportation based on machine learning according to claim 1, characterized in that, The historical acquisition data is multi-dimensional data, including: the transportation temperature, transportation humidity, driving speed of the transportation vehicle, geographical location, and vibration of the cold chain target during the transportation at each transportation time point; The number of predefined indicators is the same as the dimension, each dimension corresponds to a predefined indicator, and the predefined indicators include: a temperature deviation range indicator, a humidity qualification interval indicator, a transportation duration indicator, and a vibration allowable threshold indicator.

3. The abnormal recognition system for pharmaceutical cold chain transportation based on machine learning according to claim 1, wherein The model construction module includes: A coefficient determination unit, configured to determine a difference coefficient of the trained model based on the validation samples; Among them, represents the difference coefficient between the actual output and the ideal output of the trained model based on the validation samples, represents the ideal output of the j-th validation sample, represents the actual output of the j-th validation sample, represents the transportation coarse vector of the j-th validation sample, represents the transportation fine vector of the j-th validation sample, represents the deviation influence coefficient based on the transportation coarse vector, and its value range is (0, 1), represents the deviation influence coefficient based on the transportation fine vector, and its value range is (0, 1); represents the difference function of the j-th validation sample based on and , , respectively represent the regularization coefficients; N represents the number of validation samples; A first judgment unit, configured to, when the difference coefficient is greater than or equal to a preset coefficient, adopt an adaptive learning rate optimization algorithm and introduce a regularization technique to update the model parameters and continue to train the model; A second judgment unit, configured to, when the difference coefficient is less than the preset coefficient, obtain new samples according to the difference coefficient to train the model once and then stop training to obtain an anomaly recognition model.

4. The machine learning-based abnormal recognition system for pharmaceutical cold chain transportation according to claim 3, wherein The second judgment unit includes: An update subunit, configured to update the normalized values involved in each validation sample based on the difference coefficient; Among them, represents the o-th normalized value in the corresponding verification sample; , respectively represent the boundary range values of the o-th normalized value in the corresponding verification sample; represents the learning rate of the trained model; A new sample acquisition unit, configured to construct new samples corresponding to the validation samples based on the updated values, train the model once and then stop training.

5. The machine learning-based abnormal recognition system for pharmaceutical cold chain transportation according to claim 4, wherein The first judgment unit includes: An update subunit, configured to introduce an L2 regularization term into the loss function to constrain model parameters, calculate the gradient Ft of the loss function with respect to all model parameters through the backpropagation algorithm, and update the first moment and the second moment; Among them, represents the regularized loss function; L represents the loss function before regularization; represents the regularization parameter; represents the set of model parameters of the trained model; represents the model parameter; An amount determination subunit, configured to perform bias correction on the updated first moment and second moment, and determine the update amount of the corresponding model parameter to update each model parameter until the model meets a preset stop condition.

6. The machine learning-based abnormal recognition system for pharmaceutical cold chain transportation according to claim 5, wherein The amount determination subunit includes: A calculation block, configured to calculate the update amount of the corresponding model parameter: Among them, represents the update amount after the t-th iteration of the corresponding model parameter; represents the learning rate; , respectively represent the first moment and the second moment at the t-th iteration; , represent the decay coefficient at the t-th iteration; represents a constant, and the value is ; , respectively represent the bias correction results of the first moment and the second moment at the t-th iteration.

7. A method for identifying anomalies in the cold chain transportation of pharmaceuticals based on machine learning, characterized in that, including: Step 1: Obtain the historical acquisition data of each historical cold chain transportation, and parse the historical acquisition data according to predefined metrics during the transportation process according to the cold chain target to obtain key information corresponding to the transportation; Step 2: Divide all the key information into training samples and validation samples, train the initialized neural network model according to the training samples, and verify and correct the trained model according to the validation samples to obtain an anomaly recognition model; Step 3: Real-time collect the current information during the cold chain transportation, input it into the anomaly recognition model to obtain all anomaly factors, and give an early warning reminder for each anomaly factor to achieve anomaly control; Among them, Step 1 includes: Analyze whether the corresponding dimensional value at each transportation time point is within the standard range of the corresponding predefined metric; Assign 1 to the corresponding predefined metric, otherwise, assign 0 to the corresponding predefined metric, and construct a rough transportation vector corresponding to the transportation time point; Perform standardization processing on the dimensional values under each predefined metric to obtain standardized values, and construct a fine transportation vector corresponding to the transportation time point; Among them, represents the standardized value of the k-th transportation time point under the corresponding predefined index, represents the dimensional value of the k-th transportation time point under the predefined index, represents the mean value of all dimensional values under the corresponding predefined index, represents the standard deviation of all dimensional values under the corresponding predefined index; R0 corresponds to the reference threshold of the predefined index; ln represents the logarithmic function symbol; 、 are set constants; Perform control analysis and alarm analysis on the rough transportation vector and the fine transportation vector to obtain corresponding control information and alarm information; Use the rough transportation vector, the fine transportation vector, the control information, and the alarm information as the key information corresponding to the transportation, where the control information includes control instructions and abnormal factors for control adjustment.

Images