Abnormal state monitoring method and system for medical sterilizer

Abstract

The present invention discloses a method and system for monitoring abnormal conditions of medical sterilizers, comprising: S1: collecting and preprocessing temperature and pressure data from the medical sterilizer to obtain preprocessed temperature and pressure data; S2: dividing the preprocessed temperature and pressure data into time windows, arranging them into two-dimensional temperature and pressure matrices, and calculating temperature and pressure eigenvectors; S3: calculating the difference variation of the temperature and pressure data, identifying inflection points, and dividing the preprocessed temperature and pressure data into three time stages: heating, constant temperature, and cooling, and then extracting background vectors for each stage; S4: eliminating background features from the temperature and pressure eigenvectors to obtain alarm eigenvectors; and S5: calculating the probability of medical sterilizer abnormality based on the alarm eigenvectors, classifying the alarm causes, and outputting inspection advice. This invention can address the problem of poor monitoring effectiveness caused by traditional medical sterilizer monitoring that fails to consider temperature change stages.

Description

Technical Field

[0001] The present invention relates to the field of medical sterilizers, and in particular to a method and system for monitoring abnormal conditions of medical sterilizers. Background Art

[0002] Medical sterilizers are an indispensable part of medical equipment. Their main function is to kill or remove microorganisms on medical devices, dressings, surgical supplies, etc. through high temperature, high pressure or chemical methods to ensure the safety and sterility of the medical process. However, during operation, medical sterilizers may experience abnormal conditions due to various reasons, such as abnormal temperature and pressure control. These abnormal conditions may not only lead to poor sterilization effects, but may also cause damage to the sterilizer itself and even cause safety accidents.

[0003] Traditional methods for detecting abnormal conditions in medical sterilizers rely primarily on manual inspections and threshold alarm systems. Manual inspections typically identify potential problems by regularly checking the sterilizer's operating records and equipment status, but this method suffers from issues like poor timeliness and high labor costs. Threshold alarm systems, on the other hand, set fixed temperature and pressure thresholds, triggering alarms when actual values ​​exceed or fall below these thresholds. However, this method has the following drawbacks:

[0004] First, traditional methods have difficulty in generating alarms based on temperature change phases, and are unable to dynamically adjust according to the sterilizer's different operating phases (such as heating, constant temperature, and cooling). For example, a rapid temperature increase during the heating phase is normal, but if the temperature continues to rise rapidly during the constant temperature phase, it may indicate a sterilizer malfunction, requiring an immediate alarm.

[0005] Secondly, traditional methods cannot achieve joint monitoring of temperature and pressure. The normal operation of medical sterilizers requires the synergy between temperature and pressure. In some cases, the temperature rises but the pressure remains unchanged, or the pressure rises but the temperature does not change accordingly. These may indicate that the sterilizer is in an abnormal state. However, traditional detection methods usually only focus on the changes in a single parameter and ignore the correlation between parameters. Therefore, they cannot accurately identify this joint abnormality. Summary of the Invention

[0006] In view of this, the present invention aims to provide a method and system for monitoring the abnormal state of a medical sterilizer, aiming to solve the problems that the traditional medical sterilizer abnormality monitoring method is difficult to combine with the temperature change stage to issue an alarm, and the abnormality monitoring effect is poor due to the inability to achieve joint monitoring of temperature and pressure.

[0007] A method for monitoring abnormal conditions of a medical sterilizer comprises the following steps:

[0008] S1: collecting and preprocessing the temperature and pressure data of the medical sterilizer to obtain the preprocessed temperature and pressure data;

[0009] S2: Divide the pre-processed temperature and pressure data into time windows and arrange them into a two-dimensional temperature and pressure matrix to calculate the temperature and pressure eigenvectors;

[0010] S3: Calculate the difference change of temperature and pressure data, identify the inflection point, and divide the pre-processed temperature and pressure data into three time stages: heating, constant temperature, and cooling, and then extract the background vector of each stage;

[0011] S4: Eliminate background features of the temperature and pressure feature vectors to obtain an alarm feature vector;

[0012] S5: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector, classify the alarm causes, and output inspection opinions.

[0013] Furthermore, the step S1 collects and pre-processes the temperature and pressure data of the medical sterilizer to obtain the pre-processed temperature and pressure data, including:

[0014] S11: Acquire temperature and pressure data of the medical sterilizer, wherein the temperature and pressure data are time series data in seconds;

[0015] S12: The temperature and pressure data of the medical sterilizer are subjected to outlier elimination using the 3σ principle, and then the missing values ​​are filled using the mean value method to obtain preliminary data of the medical sterilizer;

[0016] S13: The preliminary data of the medical sterilizer were standardized using the Z-score method to obtain the preprocessed temperature and pressure data.

[0017] Furthermore, the step S2 divides the pre-processed temperature and pressure data into time windows and arranges them into a two-dimensional matrix of temperature and pressure, and calculates the temperature and pressure feature vectors, including:

[0018] S21: Divide the pre-processed temperature and pressure data into 30-second time windows to obtain data segments within each time window;

[0019] Dividing data into 30-second time windows ensures that the data segments within each time window contain sufficient information to reflect the sterilizer's operating status, while avoiding processing difficulties caused by overly sparse or overly dense data. Furthermore, the 30-second time window is short enough to detect and address anomalies promptly, ensuring the monitoring system's immediate responsiveness. Furthermore, because subsequent feature extraction is based on a neural network, it can simultaneously focus on data trends within 30 seconds, avoiding false alarms caused by anomalies in a single data point.

[0020] S22: For each data segment within each time window, the preprocessed temperature and pressure data are arranged into a two-dimensional temperature and pressure matrix; each row of the two-dimensional temperature and pressure matrix represents the temperature and pressure value at a time point. This processing method unifies the temperature and pressure data in time and space, facilitating subsequent eigenvector calculation and joint monitoring.

[0021] S23: Calculate the temperature and pressure eigenvectors. The calculation method is:

[0022] First, define the temperature sequence T = [t1, t2, ..., t n ], pressure sequence P = [p1, p2, ..., p n ], where [t1, t2, ..., t n ] is the pre-processed temperature data from the first moment to the nth moment, n is the number of moments, [p1, p2, ..., p n ] is the preprocessed pressure data from the first moment to the nth moment;

[0023]

[0024]

[0025] in, are the hidden states of the temperature data from the first moment to the nth moment output by the gated recurrent unit network, GRU() is the gated recurrent unit network operation, are the hidden states of the pressure data from the first moment to the nth moment output by the gated recurrent unit network;

[0026] Secondly, Spliced ​​into temperature hidden state sequence Will Splicing into pressure hidden state sequence

[0027] Finally, the temperature hidden state sequence, temperature sequence, pressure sequence, and pressure hidden state sequence are concatenated in sequence to obtain the temperature and pressure four-dimensional matrix, which is then input into a convolutional neural network with a convolution kernel size of 4 to obtain the temperature and pressure fusion vector. The calculation method is:

[0028]

[0029] Among them, F fusion is the temperature and pressure fusion vector, CNN( ) is the convolutional neural network operation with a convolution kernel size of 4;

[0030] The temperature and pressure fusion vectors, the hidden state of the temperature data at the nth moment, and the hidden state of the pressure data at the nth moment are fused to obtain the temperature and pressure feature vectors. The calculation method is:

[0031]

[0032] Where F is the temperature and pressure feature vector, MLP( ) is the multi-layer perceptron operation;

[0033] Step S23 first uses the GRU network as a feature enhancement method to process the temperature and pressure data, capture the long-term dependencies in the data, and obtain their respective hidden state sequences. Then, the obtained hidden state sequence is spliced ​​with the original data into a four-dimensional matrix and input into the convolutional neural network for feature fusion.

[0034] The convolutional neural network uses a convolution kernel of dimension 4, which perfectly matches the dimensions of the matrix, and can simultaneously focus on the temperature latent state, temperature, pressure, and pressure latent state. Because the latent state sequence contains temporal dependencies, the convolutional neural network actually incorporates previous context information when extracting local features. This makes the extracted features not only local but also temporal. Because the GRU network retains historical information, even if individual outliers in the input data exist, the system can correct or ignore them through contextual information, thereby enhancing the system's robustness.

[0035] Furthermore, the step S3 calculates the difference variation of the temperature and pressure data, identifies the inflection point, and divides the pre-processed temperature and pressure data into three stages: heating, constant temperature, and cooling. The background vectors of each stage are then extracted, including:

[0036] S31: Calculate the difference between temperature and pressure data and identify the inflection point. The calculation method is:

[0037]

[0038]

[0039] Δti =t i+1 -t i

[0040] Δp i =p i+1 -p i

[0041] time=MLP(MLP(Δt i , μ, μ part ), Δp i )

[0042] Where μ is the global standard deviation of temperature, μ part is the local standard deviation of temperature, k is the local length, Δt i is the temperature difference change at the i-th moment, i is the time index, j is the count index, Δp i is the pressure difference change at the i-th moment, and time is the inflection point moment;

[0043] S32: According to the inflection point, the pre-processed temperature and pressure data are divided into three stages: heating, constant temperature and cooling, and the stage labels are added to obtain the temperature and pressure series with stage labels, which are expressed as T′=[t′1, t′2, ..., t′ n ] and P′=[p′1,p′2,…,p′ n ];

[0044] S33: Build a gated recurrent unit network with transit, and extract the background vector of each stage based on the temperature and pressure data with stage labels. The calculation method is:

[0045] B=GRU Q (T′, ΔT, F)

[0046] Among them, B is the background vector of each stage, GRU W () is the gated recurrent unit network operation with transit;

[0047] Q = GRU(T′, ΔT, P′, AP)

[0048] Among them, Q is the transit sequence, including the position identifier; ΔT is the temperature difference sequence, and AP is the pressure difference sequence;

[0049] The gated recurrent unit network GRU with transit W (), taking the gated recurrent unit network GRU as the network basis, according to the position mark in the transit sequence Q, a transit multilayer perceptron is added at the corresponding position of the gated recurrent unit network, and each transit multilayer perceptron divides the gated recurrent unit network into a segment;

[0050] The structure of the relay multilayer perceptron is the same as that of the traditional multilayer perceptron MLP. The difference is that in terms of data flow, the relay multilayer perceptron connects all gated recurrent units that do not cross paragraphs forward and only connects adjacent gated recurrent units backward;

[0051] The GRU network with transit introduces a transit sequence Q, which contains position identifiers that clearly indicate the stage of temperature and pressure data. This allows the network to consciously pay attention to the stage characteristics of the data when processing the data. The transit sequence Q can be regarded as an additional input feature, guiding the GRU network to adopt different processing strategies at different stages. This strategic adjustment helps the network more accurately capture and utilize the stage characteristics of the data.

[0052] The GRU network with transit adds a transit multilayer perceptron to the corresponding position of the GRU network, dividing the GRU network into multiple sections, each of which focuses on processing the data of a certain stage. From a mathematical perspective, this division can be regarded as a segmentation of the GRU network state space. The state space of each section focuses on representing the data features of a specific stage, thus avoiding mutual interference between data features of different stages.

[0053] The relay multilayer perceptron has a special design for data flow. It connects all gated recurrent units that do not cross paragraphs forward and only connects adjacent gated recurrent units backward. This allows the network to maintain smooth information transmission when processing data while avoiding unnecessary cross-stage information transmission.

[0054] Compared with the traditional GRU network that treats data of all time steps equally without considering the stage characteristics of the data, the above design enables the gated recurrent unit network with transit to more accurately extract the background vector of each stage. The background vector not only contains the data features within the stage, but also implies the transition information between stages, laying a solid foundation for subsequent background feature separation and abnormal state alarm.

[0055] Furthermore, the step S4 performs background feature elimination on the temperature and pressure feature vectors to obtain an alarm feature vector, including:

[0056] S41: Perform stage identification on the temperature and pressure feature vectors to obtain temperature and pressure feature vectors with stage labels. The calculation method is:

[0057] F′=MLP(F,t i , p i )

[0058] Where F′ is the temperature and pressure feature vector with stage labels;

[0059] S42: Based on the temperature and pressure feature vectors with stage labels, background features are eliminated to obtain an alarm feature vector. The calculation method is:

[0060] f=W3·MLP(W1·F′, W2·B, (W1·F′-W2·B))

[0061] Among them, f is the alarm feature vector, W3 is the dimension reduction matrix, W1 and W2 are the first dimension increase matrix and the second dimension increase matrix respectively;

[0062] Using W1 and W2 as dimensionality-raising matrices, we first map F′ and the background vector B to a higher-dimensional space to better capture the complex relationship between them. Next, by calculating W1·F′-W2·B, we obtain the difference between the feature vector and the background vector, highlighting features related to the abnormal state. Finally, using a multi-layer perceptron and the dimensionality reduction matrix W3, we map the processed feature vector back to a low-dimensional space to obtain the alarm feature vector, effectively eliminating background information unrelated to the abnormal state and improving the accuracy of anomaly detection.

[0063] Among them, the dimensionality increase operation enables the feature vector to be expanded in a higher-dimensional space, thereby capturing more detailed information. The dimensionality reduction operation retains the most important features and removes redundant information, making the alarm feature vector more compact.

[0064] Due to the existence of the background feature elimination step, the present invention has stronger robustness against noise and interference; in practical applications, the temperature and pressure data of medical sterilizers may be affected by various factors, such as equipment aging, environmental changes, etc., resulting in noise and interference in the data; through background feature elimination, the present invention can effectively suppress these noise and interference and improve the robustness of abnormality monitoring.

[0065] Furthermore, the step S5 calculates the abnormal probability of the medical sterilizer based on the alarm feature vector, classifies the alarm cause, and outputs inspection opinions, including:

[0066] S51: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector. The calculation method is:

[0067] pro=sigmoid(MLP(f,t i , p i ))

[0068] Among them, pro is the abnormal probability of medical sterilizer, sigmoid() sigmoid function;

[0069] S52: Decode the alarm result and classify the alarm cause. The calculation method is:

[0070] pro′=softmax(MLP(f,t i , p i ))

[0071] Among them, pro′ is the probability distribution of the alarm reason, and softmax() is the softmax function operation;

[0072] S53: Output inspection opinions based on the alarm cause classification results.

[0073] The present invention also discloses an abnormal state monitoring system for a medical sterilizer, comprising:

[0074] Sterilizer data processing module: collects temperature and pressure data of medical sterilizers and preprocesses them to obtain preprocessed temperature and pressure data;

[0075] Temperature and pressure feature vector extraction module: divides the pre-processed temperature and pressure data into time windows, arranges them into temperature and pressure two-dimensional matrices, and calculates temperature and pressure feature vectors;

[0076] Background vector extraction module: Calculates the difference change of temperature and pressure data, identifies the inflection point, and divides the pre-processed temperature and pressure data into three stages: heating, constant temperature, and cooling, and then extracts the background vector of each stage;

[0077] Background feature elimination module: eliminates background features of temperature and pressure feature vectors to obtain alarm feature vectors;

[0078] Alarm module: Calculates the abnormal probability of the medical sterilizer based on the alarm feature vector, classifies the alarm causes, and outputs inspection opinions.

[0079] The above-mentioned abnormal state monitoring method of the medical sterilizer is realized through this system.

[0080] Compared with the prior art, the advantages of the present invention are:

[0081] (1) The present invention always takes temperature and pressure data as the objects of joint monitoring. Whether in the calculation of characteristic vectors or in the calculation of abnormality probability and the classification of alarm causes, the correlation between temperature and pressure data is fully considered, thereby improving the comprehensiveness and accuracy of abnormality monitoring.

[0082] (2) The present invention calculates the difference change of temperature and pressure data, captures the change of data between adjacent moments, identifies the inflection point moment, and divides the pre-processed data into three stages: heating, constant temperature, and cooling; secondly, the present invention eliminates background features of temperature and pressure feature vectors, removes background information irrelevant to the abnormal state, and highlights features related to the abnormal state. It can more accurately extract the alarm feature vector related to the abnormal state, thereby improving the accuracy of abnormal monitoring.

[0083] (2) The present invention constructs a gated recurrent unit network with transit. By introducing a transit sequence Q, the sequence contains a position identifier, which can clearly indicate the stage of the temperature and pressure data, so that the network can consciously pay attention to the stage characteristics of the data when processing the data; in addition, the gated recurrent unit network with transit adds a transit multi-layer perceptron at the corresponding position of the gated recurrent unit network, dividing the GRU network into multiple sections, each section focusing on processing the data of one stage, thereby avoiding mutual interference of data features of different stages.

[0084] (3) The present invention proposes an adaptive local standard deviation calculation, which uses a multi-layer perceptron to dynamically generate local lengths, making the calculation of the local standard deviation adaptive. This design can automatically adjust the size of the local window according to the data characteristics at different times, thereby more accurately reflecting the local change characteristics of the data, improving the accuracy of inflection point recognition, and providing a more reliable data basis for stage division and background vector extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A schematic flow chart of the abnormal state monitoring method for a medical sterilizer provided by the present invention;

[0086] Figure 2 Schematic diagram of the process of extracting temperature and pressure feature vectors in the method of the present invention;

[0087] Figure 3 Schematic diagram of the network structure of the gated recurrent unit network with transit in the method of the present invention. DETAILED DESCRIPTION

[0088] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention fall within the scope of protection of the present invention.

[0089] Example 1:

[0090] See also Figure 1 A method for monitoring abnormal conditions of a medical sterilizer comprises the following steps:

[0091] S1: collecting and preprocessing the temperature and pressure data of the medical sterilizer to obtain the preprocessed temperature and pressure data;

[0092] S11: Acquire temperature and pressure data of the medical sterilizer, wherein the temperature and pressure data are time series data in seconds;

[0093] S12: The temperature and pressure data of the medical sterilizer are subjected to outlier elimination using the 3σ principle, and then the missing values ​​are filled using the mean value method to obtain preliminary data of the medical sterilizer;

[0094] S13: The preliminary data of the medical sterilizer were standardized using the Z-score method to obtain the preprocessed temperature and pressure data.

[0095] S2: Divide the pre-processed temperature and pressure data into time windows and arrange them into two-dimensional temperature and pressure matrices to calculate the temperature and pressure eigenvectors. Figure 2 ;

[0096] S21: Divide the pre-processed temperature and pressure data into 30-second time windows to obtain data segments within each time window;

[0097] Dividing data into 30-second time windows ensures that the data segments within each time window contain sufficient information to reflect the sterilizer's operating status, while avoiding processing difficulties caused by overly sparse or overly dense data. Furthermore, the 30-second time window is short enough to detect and address anomalies promptly, ensuring the monitoring system's immediate responsiveness. Furthermore, because subsequent feature extraction is based on a neural network, it can simultaneously focus on data trends within 30 seconds, avoiding false alarms caused by anomalies in a single data point.

[0098] S22: For each data segment within each time window, the preprocessed temperature and pressure data are arranged into a two-dimensional temperature and pressure matrix; each row of the two-dimensional temperature and pressure matrix represents the temperature and pressure value at a time point. This processing method unifies the temperature and pressure data in time and space, facilitating subsequent eigenvector calculation and joint monitoring.

[0099] S23: Calculate the temperature and pressure eigenvectors. The calculation method is:

[0100] First, define the temperature sequence T = [t1, t2, ..., t n ], pressure sequence P = [p1, p2, ..., p n ], where [t1, t2, ..., t n] is the pre-processed temperature data from the first moment to the nth moment, n is the number of moments, [p1, p2, ..., p n ] is the preprocessed pressure data from the first moment to the nth moment;

[0101]

[0102]

[0103] in, are the hidden states of the temperature data from the first moment to the nth moment output by the gated recurrent unit network, GRU() is the gated recurrent unit network operation, are the hidden states of the pressure data from the first moment to the nth moment output by the gated recurrent unit network;

[0104] Secondly, Spliced ​​into temperature hidden state sequence Will Splicing into pressure hidden state sequence

[0105] Finally, the temperature hidden state sequence, temperature sequence, pressure sequence, and pressure hidden state sequence are concatenated in sequence to obtain the temperature and pressure four-dimensional matrix, which is then input into a convolutional neural network with a convolution kernel size of 4 to obtain the temperature and pressure fusion vector. The calculation method is:

[0106]

[0107] Among them, F fusion is the temperature and pressure fusion vector, CNN( ) is the convolutional neural network operation with a convolution kernel size of 4;

[0108] The temperature and pressure fusion vectors, the hidden state of the temperature data at the nth moment, and the hidden state of the pressure data at the nth moment are fused to obtain the temperature and pressure feature vectors. The calculation method is:

[0109]

[0110] Where F is the temperature and pressure feature vector, MLP() is the multi-layer perceptron operation;

[0111] Step S23 first uses the GRU network as a feature enhancement method to process the temperature and pressure data, capture the long-term dependencies in the data, and obtain their respective hidden state sequences. Then, the obtained hidden state sequence is spliced ​​with the original data into a four-dimensional matrix and input into the convolutional neural network for feature fusion.

[0112] The convolutional neural network uses a convolution kernel of dimension 4, which perfectly matches the dimensions of the matrix, and can simultaneously focus on the temperature latent state, temperature, pressure, and pressure latent state. Because the latent state sequence contains temporal dependencies, the convolutional neural network actually incorporates previous context information when extracting local features. This makes the extracted features not only local but also temporal. Because the GRU network retains historical information, even if individual outliers in the input data exist, the system can correct or ignore them through contextual information, thereby enhancing the system's robustness.

[0113] S3: Calculate the difference change of temperature and pressure data, identify the inflection point, and divide the pre-processed temperature and pressure data into three time stages: heating, constant temperature, and cooling, and then extract the background vector of each stage;

[0114] S31: Calculate the difference between temperature and pressure data and identify the inflection point. The calculation method is:

[0115]

[0116]

[0117] Δt i =t i+1 -t i

[0118] Δp i =p i+1 -p i

[0119] time=MLP(MLP(Δt i , μ, μ part ), Δp i )

[0120] Where μ is the global standard deviation of temperature, μ part is the local standard deviation of temperature, k is the local length, Δt i is the temperature difference change at the i-th moment, i is the time index, j is the count index, Δp i is the pressure difference change at the i-th moment, and time is the inflection point moment;

[0121] In particular, when data fluctuations are large or high accuracy is required, an adaptive local window can be used instead of the fixed local length k. The calculation method is:

[0122] k=MLP(T,i)

[0123] By calculating the difference in temperature and pressure data, step S31 can accurately capture the changes in data between adjacent moments. Unlike traditional methods of calculating local standard deviations with fixed window sizes, the local standard deviation is dynamically generated using a multi-layer perceptron. This makes the calculation of the local standard deviation adaptive, automatically adjusting the size of the local window based on the data characteristics at different moments, thereby more accurately reflecting the local variation characteristics of the data.

[0124] The identification of inflection points combines the global standard deviation, local standard deviation, and difference variation, taking into account not only the global characteristics of the data but also the local variations of the data, thereby improving the accuracy of identification. If only focusing on local features, such as simply calculating local extreme points, while ignoring the guidance of global characteristics, the identified extreme points may not be the true stage demarcation points of temperature changes, which may lead to misjudgment of stage identification in subsequent steps. Therefore, in the present invention, the dual attention to global and local features is indispensable. They complement each other and jointly ensure the accuracy of data processing and stage identification.

[0125] S32: According to the inflection point, the pre-processed temperature and pressure data are divided into three stages: heating, constant temperature and cooling, and the stage labels are added to obtain the temperature and pressure series with stage labels, which are expressed as T′=[t′1, t′2, ..., t′ n ] and P′=[p′1,p′2,…,p′ n ];

[0126] S33: Build a gated recurrent unit network with transit, and extract the background vector of each stage based on the temperature and pressure data with stage labels. The calculation method is:

[0127] B=GRU Q (T′, ΔT, F)

[0128] Among them, B is the background vector of each stage, GRU Q ( ) is the gated recurrent unit network operation with transit;

[0129] Q = GRU(T′,ΔT,P′,ΔP)

[0130] Among them, Q is the transit sequence, including the position identifier; ΔT is the temperature difference sequence, and ΔP is the pressure difference sequence;

[0131] The gated recurrent unit network GRUQ() with transit, see Figure 3, taking the gated recurrent unit network GRU as the network basis, according to the position identifier in the transit sequence Q, a transit multilayer perceptron is added at the corresponding position of the gated recurrent unit network, and each transit multilayer perceptron divides the gated recurrent unit network into a segment;

[0132] The structure of the relay multilayer perceptron is the same as that of the traditional multilayer perceptron MLP. The difference is that in terms of data flow, the relay multilayer perceptron connects all gated recurrent units that do not cross paragraphs forward and only connects adjacent gated recurrent units backward;

[0133] The GRU network with transit introduces a transit sequence Q, which contains position identifiers that clearly indicate the stage of temperature and pressure data. This allows the network to consciously pay attention to the stage characteristics of the data when processing the data. The transit sequence Q can be regarded as an additional input feature, guiding the GRU network to adopt different processing strategies at different stages. This strategic adjustment helps the network more accurately capture and utilize the stage characteristics of the data.

[0134] The GRU network with transit adds a transit multilayer perceptron to the corresponding position of the GRU network, dividing the GRU network into multiple sections, each of which focuses on processing the data of a certain stage. From a mathematical perspective, this division can be regarded as a segmentation of the GRU network state space. The state space of each section focuses on representing the data features of a specific stage, thus avoiding mutual interference between data features of different stages.

[0135] The relay multilayer perceptron has a special design for data flow. It connects all gated recurrent units that do not cross paragraphs forward and only connects adjacent gated recurrent units backward. This allows the network to maintain smooth information transmission when processing data while avoiding unnecessary cross-stage information transmission.

[0136] Compared with the traditional GRU network that treats data of all time steps equally without considering the stage characteristics of the data, the above design enables the gated recurrent unit network with transit to more accurately extract the background vector of each stage. The background vector not only contains the data features within the stage, but also implies the transition information between stages, laying a solid foundation for subsequent background feature separation and abnormal state alarm.

[0137] S4: Eliminate background features of the temperature and pressure feature vectors to obtain an alarm feature vector;

[0138] S41: Perform stage identification on the temperature and pressure feature vectors to obtain temperature and pressure feature vectors with stage labels. The calculation method is:

[0139] F′=MLP(F,t i ,pi )

[0140] Where F' is the temperature and pressure feature vector with stage labels;

[0141] S42: Based on the temperature and pressure feature vectors with stage labels, background features are eliminated to obtain an alarm feature vector. The calculation method is:

[0142] f=W3·MLP(W1·F′,W2·B,(W1·F′-W2·B))

[0143] Among them, f is the alarm feature vector, W3 is the dimension reduction matrix, W1 and W2 are the first dimension increase matrix and the second dimension increase matrix respectively;

[0144] Using W1 and W2 as dimensionality-raising matrices, we first map F' and the background vector B to a higher-dimensional space to better capture the complex relationship between them. Next, by calculating W1·F' - W2·B, we obtain the difference between the feature vector and the background vector, highlighting features related to the abnormal state. Finally, using a multi-layer perceptron and the dimensionality reduction matrix W3, we map the processed feature vector back to a low-dimensional space to obtain the alarm feature vector, effectively eliminating background information unrelated to the abnormal state and improving the accuracy of anomaly detection.

[0145] Among them, the dimensionality increase operation enables the feature vector to be expanded in a higher-dimensional space, thereby capturing more detailed information. The dimensionality reduction operation retains the most important features and removes redundant information, making the alarm feature vector more compact.

[0146] Due to the existence of the background feature elimination step, the present invention has stronger robustness against noise and interference; in practical applications, the temperature and pressure data of medical sterilizers may be affected by various factors, such as equipment aging, environmental changes, etc., resulting in noise and interference in the data; through background feature elimination, the present invention can effectively suppress these noise and interference and improve the robustness of abnormality monitoring.

[0147] S5: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector, classify the alarm causes, and output inspection opinions;

[0148] S51: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector. The calculation method is:

[0149] pro=sigmoid(MLP(f,t i , p i ))

[0150] Among them, pro is the abnormal probability of the medical sterilizer, sigmoid() sigmoid function;

[0151] When pro>0.6, an alarm is issued;

[0152] When pro≤0.6, no alarm is given;

[0153] S52: Decode the alarm result and classify the alarm cause. The calculation method is:

[0154] pro′=softmax(MLP(f,t i , p i ))

[0155] Among them, pro′ is the probability distribution of alarm cause, softmax( ) is the softmax function operation;

[0156] S53: Output inspection opinions based on the alarm cause classification results;

[0157] For example, within a certain time window, the temperature and pressure data of a medical sterilizer are collected. After preprocessing, feature extraction, and background feature elimination, the alarm feature vector f = [0.5, -0.3, 0.2] is obtained. The alarm feature vector f and the current temperature of 120°C and pressure of 2.5 atm are input into the MLP network, and pro = 0.69 is obtained, indicating that the probability of abnormality of the medical sterilizer at the current moment is 0.69>0.6, so an alarm is issued.

[0158] Next, the same alarm feature vector and temperature pressure are input into another MLP network to calculate the probability distribution pro′ of the alarm cause

[0159] The output of the MLP network is [1.5, -0.5, 0.3] (the output is a vector, each element corresponds to the probability of an abnormal cause), and the output is converted into a probability distribution through the softmax function:

[0160] pro′=softmax([1.5,-0.5,0.3])=[0.659, 0.119, 0.222]

[0161] At the current moment, the possible abnormal causes and probabilities of the medical sterilizer are:

[0162] Cause 1 (heating system failure): 65.9%

[0163] Cause 2 (pressure sensor failure): 11.9%

[0164] Cause 3 (cooling system failure): 22.2%

[0165] Since the heating system has the highest probability of failure, the output is the inspection opinion for the heating system failure:

[0166] Check whether the heating element is damaged, such as whether the resistance wire is broken or the heating tube is damaged;

[0167] Check whether the temperature controller is malfunctioning, such as whether the temperature sensor is accurate and whether the controller circuit is normal;

[0168] If there is a problem with the heating system, please replace the damaged parts or adjust the controller parameters in time to ensure that the temperature can reach the expected value and remain stable.

[0169] Example 2:

[0170] A medical sterilizer abnormal state monitoring system, comprising:

[0171] Sterilizer data processing module: collects temperature and pressure data of medical sterilizers and preprocesses them to obtain preprocessed temperature and pressure data;

[0172] Temperature and pressure feature vector extraction module: divides the pre-processed temperature and pressure data into time windows, arranges them into temperature and pressure two-dimensional matrices, and calculates temperature and pressure feature vectors;

[0173] Background vector extraction module: Calculates the difference change of temperature and pressure data, identifies the inflection point, and divides the pre-processed temperature and pressure data into three stages: heating, constant temperature, and cooling, and then extracts the background vector of each stage;

[0174] Background feature elimination module: eliminates background features of temperature and pressure feature vectors to obtain alarm feature vectors;

[0175] Alarm module: Calculates the abnormal probability of the medical sterilizer based on the alarm feature vector, classifies the alarm causes, and outputs inspection opinions.

[0176] The method in Example 1 is implemented by the abnormal state monitoring system for the medical sterilizer in this embodiment.

[0177] It should be noted that the serial numbers of the above-mentioned embodiments of the present invention are for descriptive purposes only and do not represent the advantages or disadvantages of the embodiments. In addition, the terms "including", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method comprising the element.

[0178] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in multiple embodiments of the present invention.

[0179] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for monitoring abnormal conditions of a medical sterilizer, characterized in that: The following steps are involved: S1: collecting and preprocessing the temperature and pressure data of the medical sterilizer to obtain the preprocessed temperature and pressure data; S2: Divide the pre-processed temperature and pressure data into time windows and arrange them into two-dimensional temperature and pressure matrices to calculate the temperature and pressure eigenvectors; including: S21: Divide the pre-processed temperature and pressure data into 30-second time windows to obtain data segments within each time window; S22: For each data segment within each time window, the pre-processed temperature and pressure data are arranged into a temperature and pressure two-dimensional matrix; each row of the temperature and pressure two-dimensional matrix represents the temperature and pressure value at a time point; S23: Calculate the temperature and pressure eigenvectors. The calculation method is: First, define the temperature sequence T=[t1,t2,…,t n ], pressure sequence P=[p1,p2,…,p n ], where [t1,t2,…,t n ] is the pre-processed temperature data from the first moment to the nth moment, n is the number of moments, [p1,p2,…,p n ] is the preprocessed pressure data from the first moment to the nth moment; in, are the hidden states of the temperature data from the first moment to the nth moment output by the gated recurrent unit network, GRU() is the gated recurrent unit network operation, are the hidden states of the pressure data from the first moment to the nth moment output by the gated recurrent unit network; Secondly, Spliced ​​into temperature hidden state sequence Will Splicing into pressure hidden state sequence Finally, the temperature hidden state sequence, temperature sequence, pressure sequence, and pressure hidden state sequence are concatenated in sequence to obtain the temperature and pressure four-dimensional matrix, which is then input into a convolutional neural network with a convolution kernel size of 4 to obtain the temperature and pressure fusion vector. The calculation method is: Among them, F fusion is the temperature and pressure fusion vector, CNN() is the convolutional neural network operation with a convolution kernel size of 4; The temperature and pressure fusion vectors, the hidden state of the temperature data at the nth moment, and the hidden state of the pressure data at the nth moment are fused to obtain the temperature and pressure feature vectors. The calculation method is: Where F is the temperature and pressure feature vector, MLP() is the multi-layer perceptron operation; S3: Calculate the difference change of temperature and pressure data, identify the inflection point, and divide the pre-processed temperature and pressure data into three time stages: heating, constant temperature, and cooling, and then extract the background features of each stage; S4: Eliminate background features of the temperature and pressure feature vectors to obtain an alarm feature vector; S5: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector, classify the alarm causes, and output inspection opinions.

2. The abnormal state monitoring method of a medical sterilizer according to claim 1, characterized in that: The S1 step includes: S11: Obtain temperature and pressure data of the medical sterilizer; S12: The temperature and pressure data of the medical sterilizer are subjected to outlier elimination using the 3σ principle, and then the missing values ​​are filled using the mean value method to obtain preliminary data of the medical sterilizer; S13: The preliminary data of the medical sterilizer were standardized using the Z-score method to obtain the preprocessed temperature and pressure data.

3. The abnormal state monitoring method of a medical sterilizer according to claim 1, characterized in that: The S3 step includes: S31: Calculate the difference between temperature and pressure data and identify the inflection point. The calculation method is: Δt i =t i+1 -t i Δp i =p i+1 -p i time=MLP(MLP(Δt i ,μ,μ part ),Δp i ) Among them, μ is the global mean temperature, μ part is the local mean temperature, k is the local length, Δt i is the temperature difference change at the i-th moment, i is the time index, j is the count index, Δp i is the pressure difference change at the i-th moment, and time is the inflection point moment; S32: According to the inflection point, the pre-processed temperature and pressure data are divided into three stages: heating, constant temperature and cooling, and the stage labels are added to obtain the temperature and pressure series with stage labels, which are expressed as T'=[t'1,t'2,…,t' n ] and P'=[p'1,p'2,…,p' n ]; S33: Build a gated recurrent unit network with transit, and extract the background features of each stage based on the temperature and pressure data with stage labels. The calculation method is: B=GRU Q (T′,ΔT,F) Among them, B is the background feature of each stage, GRU Q () is the gated recurrent unit network operation with transit; Q = GRU(T′,ΔT,P′,ΔP) Among them, Q is the transit sequence, including the position identifier; ΔT is the temperature difference change sequence, and ΔP is the pressure difference change sequence; The gated recurrent unit network GRU with transit Q (), taking the gated recurrent unit network GRU as the network basis, according to the position identifier in the transit sequence Q, a transit multilayer perceptron is added at the corresponding position of the gated recurrent unit network, and each transit multilayer perceptron divides the gated recurrent unit network into a segment.

4. The abnormal state monitoring method of a medical sterilizer according to claim 3, characterized in that: The S4 step includes: S41: Perform stage identification on the temperature and pressure feature vectors to obtain temperature and pressure feature vectors with stage labels. The calculation method is: F′=MLP(F,t i ,p i ) Where F' is the temperature and pressure feature vector with stage labels; S42: Based on the temperature and pressure feature vectors with stage labels, background features are eliminated to obtain an alarm feature vector. The calculation method is: f=W3·MLP(W1·F′,W2·B,(W1·F′-W2·B)) Among them, f is the alarm feature vector, W3 is the dimension reduction matrix, W1 and W2 are the first dimension increase matrix and the second dimension increase matrix respectively.

5. The abnormal state monitoring method of a medical sterilizer according to claim 4, characterized in that: The S5 step includes: S51: Calculate the abnormal probability of the medical sterilizer based on the alarm feature vector. The calculation method is: pro=sigmoid(MLP(f,t i ,p i )) Among them, pro is the abnormal probability of the medical sterilizer, sigmoid() is the sigmoid function; S52: Decode the alarm result and classify the alarm cause. The calculation method is: pro′=softmax(MLP(f,t i ,p i )) Among them, pro' is the probability distribution of the alarm reason, and softmax() is the softmax function operation; S53: Output inspection opinions based on the alarm cause classification results.

6. A medical sterilizer abnormal state monitoring system, used in the medical sterilizer abnormal state monitoring method according to any one of claims 1 to 5, characterized in that: include: Sterilizer data processing module: collects temperature and pressure data of medical sterilizers and preprocesses them to obtain preprocessed temperature and pressure data; Temperature and pressure feature vector extraction module: divides the pre-processed temperature and pressure data into time windows, arranges them into temperature and pressure two-dimensional matrices, and calculates temperature and pressure feature vectors; Background feature extraction module: Calculates the difference change of temperature and pressure data, identifies the inflection point, and divides the pre-processed temperature and pressure data into three time stages: heating, constant temperature, and cooling, and then extracts the background features of each stage; Background feature elimination module: eliminates background features of temperature and pressure feature vectors to obtain alarm feature vectors; Alarm module: Calculates the abnormal probability of the medical sterilizer based on the alarm feature vector, classifies the alarm causes, and outputs inspection opinions.

Images