Breathing machine fault prediction method based on LSTM classifier

Abstract

The invention discloses a breathing machine fault prediction method based on an LSTM classifier, and the method comprises the following steps: collecting historical data of breathing machines of the same type, and building a data set; preprocessing data of the data set, and dividing the processed data set into a training set and a test set; inputting the data of the training set and the test set into an LSTM classifier for training; obtaining an LSTM network model by adjusting the time window; and predicting the fault probability of the breathing machine in a specific time window according to the LSTM network model so as to assist in decision-making of breathing machine maintenance. The invention belongs to the technical field of breathing machine fault medicine. According to the method, the historical data of each sensor of the breathing machine is processed based on the LSTM model, so that the fault probability of the breathing machine is predicted; advanced intervention of major faults of the breathing machine can be achieved, and major medical accidents are further avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of ventilator fault prediction, and in particular to a ventilator fault prediction method based on LSTM classifier. Background Art

[0002] At present, ventilators are widely used in respiratory failure caused by various reasons, anesthesia and respiratory management during major surgery, respiratory support therapy and emergency resuscitation, and they play a very important role in the field of modern medicine. Ventilators are a vital medical device that can prevent and treat respiratory failure, reduce complications, and save and prolong patients' lives.

[0003] During the use of ventilators, it is inevitable that they will encounter malfunctions. At present, ventilator malfunctions are generally discovered during use, after the instrument has a self-test fault alarm or shuts down or other serious faults. Currently, mainstream ventilators all have built-in sensor systems. When the ventilator is used for a certain period of time, the sensor components in its sensor system may have problems. Therefore, it is very necessary to predict the failure of each ventilator sensor in advance. Summary of the invention

[0004] To this end, the present invention provides a ventilator fault prediction method based on LSTM classifier to solve the above problems in the prior art.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] According to a first aspect of the present invention, a ventilator fault prediction method based on an LSTM classifier comprises the following steps:

[0007] S1. Collect historical data of the same type of ventilators and establish a data set;

[0008] S2. preprocess the data of the data set and divide the processed data set into a training set and a test set;

[0009] S3. Input the data of the training set and the test set into the LSTM classifier for training;

[0010] The LSTM layer structure at time t, through the unit state c t , hidden state h t and input gates and output gates to provide richer internal states; assuming that there are N ventilators of the same brand and model during use, a total of m sensors are monitoring the operating status of the ventilators; the sensor data X obtained by each ventilator i (i = 1, 2, ..., N) during its life cycle is represented in the form of a matrix i ;

[0011] Xi =[X 1 ,X 2 ,...X N ], X i ∈R N×m ;

[0012] During training, LSTM uses sensor data X i , and determine whether the true value of RUL belongs to a certain time window;

[0013] During testing, at time t, the constructed LSTM classifier converts the sensor measurement value X t The vector is used as input data and the probability that the RUL belongs to a certain time window is output;

[0014]

[0015] In the above two formulas, X t represents the sensor measurement sequence at time t, X represents the sequence, and N represents the length of the sequence;

[0016] S4. Obtain the LSTM network model by adjusting the time window;

[0017] The input layer brings data into the LSTM network layer. The input data format is three-dimensional, namely: sample, time step, number of features;

[0018] S5. Predict the failure probability of the ventilator in the time window based on the LSTM network model to assist in decision-making on ventilator maintenance.

[0019] Furthermore: the historical data of the ventilator includes relevant data of each sensor of the ventilator (air flow sensor, oxygen flow sensor pressure sensor, oxygen sensor, atmospheric pressure sensor, flow sensor, voltage and current acquisition, etc.) and relevant parameters calculated by the sensors, fault / alarm data and historical maintenance data. The historical maintenance data is the maintenance data of moving parts (such as PEEP valve, flow control valve, etc.) and sensors / circuit boards.

[0020] Further: The data preprocessing process is:

[0021] (1) Decompressing the original data: decompressing the compressed file of the original data;

[0022] (2) Normalize the data; the input data is obtained from multiple devices and multiple sensor sources; these data need to be normalized; the Min-Max method is used to normalize the training set and test set data, and the normalized data will be between [0,1];

[0023] (3) Fault type classification: Fault data is classified according to the type of sensor. Existing sensor faults mainly include: flow sensor fault, oxygen sensor fault, pressure sensor control board fault, air flow sensor fault, oxygen flow sensor fault and gas mixture pressure sensor fault;

[0024] (4) Time window processing: A sliding time window is used to obtain a feature vector to define the data label. The sliding time window can obtain more useful time information. The previous time step is used to predict the next time step. The window slides one time unit along the time direction to construct a single training sample. The ventilator sensor RUL corresponding to the last time node is used as the label of the training sample.

[0025] (5) Store in a database; classify the results and store them in a database.

[0026] Further: In the LSTM unit structure,

[0027] Forget Gate t Control which long-term memories of LSTM can be forgotten;

[0028] f t =σ(W f h t-1 +U f X t +b f );

[0029] Input gate g t 、i t is to calculate the information obtained from the input and understand what should be stored in the cell state c t part of;

[0030] g t =tanh(W g h t-1 +U g X t +b g );

[0031] i t =σ(W i h t-1 +U i X t +b i )i t =σ(W i h t-1 +U i X t +b i );

[0032] Update long-term memory in the cell state;

[0033]

[0034] Use output gate o t Update the state h of the current hidden layer according to the input gate, unit state and previous hidden state t ;

[0035] o t =σ(W o h t-1 +U o X t +b o );

[0036]

[0037] Where σ is the Sigmoid function, tanh is the activation function; W i , W f , W g , W o is the hidden layer weight value between the current hidden layer and the previous hidden layer; U i , U f , U g , U o is the weight value between the current input layer and the current hidden layer; b i , b f , b g , b o is the bias vector; is the element-wise multiplication operator.

[0038] Further: The time step refers to the length by which each feature information can be passed to the next feature.

[0039] Going a step further: When training the LSTM classifier, the binary cross entropy function is used as the loss function of the objective function.

[0040] The present invention has the following advantages: based on the LSTM model, the historical data of each sensor of the ventilator is processed to predict the failure probability of the ventilator; it can achieve early intervention of major ventilator failures and further avoid the occurrence of major medical accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flow chart of a ventilator fault prediction method based on LSTM classifier.

[0042] Figure 2 The present invention is a flow chart of data preprocessing in the ventilator fault prediction method.

[0043] Figure 3This is a schematic diagram of data classification in the ventilator fault prediction method of the present invention.

[0044] Figure 4 Schematic diagram of the LST unit structure in the ventilator fault prediction method of the present invention. DETAILED DESCRIPTION

[0045] The following is an explanation of the implementation of the present invention by specific specific examples. People familiar with this technology can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. It should be understood that these embodiments are only to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field can make some non-essential improvements and adjustments to the present invention based on the contents of the above invention; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] See also Figure 1 , a ventilator fault prediction method based on LSTM classifier in the first aspect of the present invention comprises the following steps:

[0047] S1. Collect historical data of the same type of ventilators and establish a data set;

[0048] The historical data of the ventilator includes relevant data of each sensor of the ventilator (air flow sensor, oxygen flow sensor pressure sensor, oxygen sensor, atmospheric pressure sensor, flow sensor, voltage and current acquisition, etc.) and related parameters calculated by the sensors, fault / alarm data and historical maintenance data. The historical maintenance data is the maintenance data of moving parts (such as PEEP valve, flow control valve, etc.) and sensors / circuit boards.

[0049] S2. preprocess the data of the data set and divide the processed data set into a training set and a test set;

[0050] like Figure 2 As shown, the general process of data preprocessing is:

[0051] (1) Decompressing the original data: decompressing the compressed file of the original data;

[0052] (2) Normalize the data; the input data is obtained from multiple devices and multiple sensor sources; these data need to be normalized; the Min-Max method is used to normalize the training set and test set data, and the normalized data will be between [0,1];

[0053] (3) Fault type classification: Fault data is classified according to the type of sensor. Common faults of existing sensors include: flow sensor fault, oxygen sensor fault, pressure sensor control board fault, air flow sensor fault, oxygen flow sensor fault and gas mixture pressure sensor fault. For sensor maintenance data, the following methods can be used: Figure 3 As shown, based on the SVM classifier, a binary tree scheme is used to construct a model, classification is performed based on the above 6 types of sensors, and a total of 5 sub-classifiers SVM1, SVM2, ..., SVM5 are constructed in sequence to classify the above 6 types of sensor fault data; in addition, a sub-classifier SVM6 can also be established for classifying other data.

[0054] Similarly, the maintenance data of moving parts can also be classified using the above-mentioned SVM classifier. For example, sub-classifiers of moving parts such as PEEP valves and flow control valves can be established to classify data.

[0055] (4) Time window processing: A sliding time window is used to obtain a feature vector to define the data label. The sliding time window can obtain more useful time information. The previous time step is used to predict the next time step. The window slides one time unit along the time direction to construct a single training sample. The RUL (remaining useful life) of the ventilator sensor corresponding to the last time node is used as the label of the training sample.

[0056] For example, the probability of a fault occurring in a specified time window T1, the data is marked into two categories: the first category is recorded as D1, indicating that the sensor RUL is greater than T1, that is, RUL>T1; the second category is D2, which represents the device RUL is less than T1, that is, RUL≤T1; if RUL belongs to a given category, its corresponding element will be set to 1, and the remaining elements of the output array will be set to 0;

[0057] (5) Store in a database; classify the results and store them in a database.

[0058] S3. Input the data of the training set and the test set into the LSTM classifier for training;

[0059] like Figure 4 The schematic diagram of the LSTM unit structure shown in Figure 1 shows the LSTM layer structure at time t. The LSTM layer passes through the unit state c t , hidden state h t and various gates to provide richer internal states; Assume that there are N ventilators of the same brand and model during use, and a total of m sensors are monitoring the operating status of the ventilators; The sensor data X obtained by each ventilator i (i = 1, 2, ..., N) during its life cycle is represented in the form of a matrix i ;

[0060] X i =[X 1 ,X 2 ,...X N ], X i ∈R N×m ;

[0061] During training, LSTM uses sensor data X i , and determine whether the true value of RUL belongs to a certain time window;

[0062] During testing, at time t, the constructed LSTM classifier converts the sensor measurement value X t The vector is used as input data and the probability that the RUL belongs to a certain time window is output;

[0063]

[0064] In the formula, X t represents the sensor measurement sequence at time t, and N represents the length of the sequence;

[0065] Among them, the forget gate f t Control which long-term memories of LSTM can be forgotten;

[0066] f t =σ(W f h t-1 +U f X t +b f );

[0067] Input gate g t 、i t It is to calculate the information that can be obtained from the input and understand what should be stored in the cell state c t part of;

[0068] g t =tanh(W g h t-1 +U g X t +b g );

[0069] i t =σ(W i h t-1 +U i X t +b i )i t =σ(W i h t-1 +U i X t +bi );

[0070] Update long-term memory in the cell state;

[0071]

[0072] Finally, use the output gate o t Update the state of the current hidden layer based on the input gate, cell state, and previous hidden state;

[0073] o t =σ(W o h t-1 +U o X t +b o );

[0074]

[0075] Where σ is the Sigmoid function, tanh is the activation function; W i , W f , W g , W o is the hidden layer weight value between the current hidden layer and the previous hidden layer; U i , U f , U g , U o is the weight value between the current input layer and the current hidden layer; b i , b f , b g , b o is the bias vector; Element-wise multiplication operator.

[0076] S4. Obtain the LSTM network model by adjusting the time window;

[0077] The input layer brings the real-time monitoring data into the LSTM network layer. The input real-time monitoring data format is a three-dimensional format, namely: sample, time step, number of features; the time step refers to the length that each feature information can be passed to the next feature.

[0078] S5. Predict the failure probability of the ventilator in a specific time window based on the LSTM model to make maintenance decisions or assist in decision-making on ventilator maintenance;

[0079] Based on the trained LSTM network model, the failure rate of each sensor of the online ventilator in a specific time window can be predicted; then based on this prediction, it can be determined whether the ventilator needs to be maintained or repaired in advance.

[0080] In addition, in this embodiment, when training the LSTM classifier, a binary cross-entropy function is used as the loss of the objective function.

[0081] In addition, the historical maintenance data can be summarized for the number of failures of ventilators of the same brand, and then a gamma distribution model can be established based on the summarized data to analyze the failure time distribution of the ventilator; the reliability indicators of the equipment, such as failure time, failure rate, reliability, etc., can be calculated through the parameters of the gamma distribution; when the failure probability of a certain type of sensor is high in a specific time window, the gamma distribution model can provide further reference for subsequent maintenance; the predicted failure probability of each sensor of the present invention can in turn help maintenance personnel quickly locate and inspect the fault.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A ventilator fault prediction method based on LSTM classifier, characterized in that: The following steps are involved: S1. Collect historical data of the same type of ventilators and establish a data set; S2. preprocess the data of the data set and divide the processed data set into a training set and a test set; S3. Input the data of the training set and the test set into the LSTM classifier for training; S4. Obtain the LSTM network model by adjusting the time window; The input layer brings data into the LSTM network layer. The input data format is three-dimensional, namely: sample, time step, number of features; S5. Predict the failure probability of the ventilator in the time window based on the LSTM network model to assist in decision-making on ventilator maintenance.

2. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: The historical data of the ventilator includes relevant data of each sensor of the ventilator and relevant parameters calculated by the sensors, fault / alarm data and historical maintenance data. The historical maintenance data includes maintenance data of moving parts and sensors / circuit boards.

3. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: The LSTM layer structure at time t, through the unit state c t , hidden state h t and input gates and output gates to provide richer internal states; assuming that there are N ventilators of the same brand and model during use, a total of m sensors are monitoring the operating status of the ventilators; the sensor data X obtained by each ventilator i (i = 1, 2, ..., N) during its life cycle is represented in the form of a matrix i ; X i =[X 1 ,X 2 ,...X N ],X i ∈R N×m ; During training, LSTM uses sensor data X i , and determine whether the true value of RUL belongs to a certain time window; During testing, at time t, the constructed LSTM classifier converts the sensor measurement value X t The vector is used as input data and the probability that the RUL belongs to a certain time window is output; In the formula, X t represents the sensor measurement sequence at time t, X represents the sequence, and N represents the length of the sequence.

4. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: The data preprocessing process is: (1) Decompressing the original data: decompressing the compressed file of the original data; (2) Normalize the data; The input data is obtained from multiple devices and multiple sensor sources; these data need to be normalized; The Min-Max method is used to normalize the training set and test set data, and the normalized data will be between [0,1]; (3) Fault type classification: Fault data is classified according to the type of sensor. Common faults of existing sensors include: flow sensor fault, oxygen sensor fault, pressure sensor control board fault, air flow sensor fault, oxygen flow sensor fault and gas mixture pressure sensor fault; (4) Time window processing: A sliding time window is used to obtain a feature vector to define the data label. The sliding time window can obtain more useful time information. The previous time step is used to predict the next time step. The window slides one time unit along the time direction to construct a single training sample. The ventilator sensor RUL corresponding to the last time node is used as the label of the training sample. (5) Store in a database; classify the results and store them in a database.

5. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: In the LSTM unit structure, Forget Gate t Control which long-term memories of LSTM can be forgotten; f t =σ(W f h t-1 +U f X t +b f ); Input gate g t 、i t is to calculate the information obtained from the input and understand what should be stored in the cell state c t part of; g t =tanh(W g h t-1 +U g X t +b g ); i t =σ(W i h t-1 +U i X t +b i )i t =σ(W i h t-1 +U i X t +b i ); Update long-term memory in the cell state; Use output gate o t Update the state of the current hidden layer based on the input gate, cell state, and previous hidden state; the t =σ(W o h t-1 +U o X t +b o ); Where σ is the Sigmoid function, tanh is the activation function; W i , W f , W g , W o is the hidden layer weight value between the current hidden layer and the previous hidden layer; U i , U f , U g , U o is the weight value between the current input layer and the current hidden layer; b i , b f , b g , b o is the bias vector; is the element-wise multiplication operator.

6. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: The time step refers to the length by which each feature information can be passed to the next feature.

7. A ventilator fault prediction method based on LSTM classifier according to claim 1, characterized in that: When training the LSTM classifier, the binary cross entropy function is used as the loss function of the objective function.

Images