Medical equipment maintenance data monitoring method based on big data

Abstract

The invention provides a medical equipment maintenance data monitoring method based on big data, and the method comprises the steps: collecting the operation data of medical equipment, carrying out the feature extraction, and merging the extracted features into a comprehensive data matrix; constructing an equipment relation graph and carrying out feature extraction; extracting high-order features through neighborhood feature aggregation according to the constructed equipment relation graph and the extracted features; dimensionality reduction and reconstruction are carried out on the high-order feature matrix by using a variational auto-encoder, and potential feature representation is extracted; optimizing a maintenance plan and resource scheduling of the medical equipment according to a prediction result; and constructing a causal relationship model according to a prediction result and an optimization result, carrying out fault tracing analysis based on a causal inference result, optimizing a maintenance strategy by using an optimization algorithm, implementing a maintenance plan according to the optimized maintenance strategy, and monitoring a maintenance effect in real time. According to the invention, the reliability and the maintenance efficiency of equipment operation are improved, and the innovativeness and the practicability are remarkable.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical equipment maintenance data monitoring, and in particular relates to a medical equipment maintenance data monitoring method based on big data. Background Art

[0002] With the rapid development of modern medical technology, advanced medical equipment plays a vital role in clinical diagnosis and treatment, especially large equipment such as MRI, CT and ultrasound, whose efficient and stable operation directly affects the quality of medical care and patient safety. However, existing medical equipment maintenance and repair methods mainly rely on regular inspections and passive repairs, which have significant limitations and are difficult to meet the needs of complex medical environments.

[0003] First, although the traditional regular inspection method can prevent equipment failure to a certain extent, due to the fixed inspection cycle and the random occurrence of equipment failure, potential problems are often not discovered in time, resulting in sudden equipment failure during operation. Such sudden failures not only increase equipment downtime and affect the continuity of diagnosis and treatment, but may also bring risks and inconvenience to patients. Regular inspections can only be performed at scheduled time nodes, and the actual operating status of the equipment may change during the inspection cycle, resulting in potential failures being ignored.

[0004] Secondly, passive maintenance means that the equipment is repaired only after a failure occurs, which not only increases maintenance costs, but may also further damage equipment performance and life. Passive maintenance does not provide early intervention opportunities, and equipment failures require emergency treatment, which not only increases the work pressure of maintenance personnel, but may also cause the equipment to be unusable at critical moments, affecting the quality and efficiency of medical services.

[0005] In addition, existing equipment maintenance methods usually lack data-driven accurate predictions and cannot fully utilize equipment operation data for fault prediction and preventive maintenance. With the development of Internet of Things technology, medical devices are increasingly equipped with sensors that can monitor their operating status in real time and generate a large amount of data. However, the effective use of this data remains a huge challenge. Traditional data analysis methods cannot fully mine the potential information in equipment operation data, making it difficult to achieve high-precision fault prediction and preventive maintenance.

[0006] Finally, as the types and number of medical devices increase, how to efficiently dispatch maintenance resources and optimize maintenance strategies have become urgent issues to be solved. Existing maintenance systems usually operate in isolation, lack cross-device and cross-institutional data sharing and collaboration, and it is difficult to form a global maintenance optimization strategy. This leads to irrational allocation of maintenance resources, further increasing maintenance costs and management difficulties.

[0007] In summary, the current maintenance and repair methods of medical equipment have the following main problems: 1) Regular inspections cannot detect potential faults in time, resulting in sudden equipment shutdown and delayed diagnosis; 2) Passive maintenance increases maintenance costs and equipment operation risks; 3) Traditional data analysis methods fail to fully utilize equipment operation data and cannot achieve high-precision fault prediction; 4) The increase in the number of equipment has led to difficulties in scheduling maintenance resources and insufficient optimization of maintenance strategies; 5) Systems operate in isolation, lacking collaboration and optimization across devices and institutions. Summary of the invention

[0008] The purpose of the present invention is to design a medical equipment maintenance data monitoring method based on big data. Through innovative adaptive graph neural network technology, complex relationship modeling and high-precision fault pattern recognition of medical equipment are realized, effectively solving the deficiencies in the prior art.

[0009] In order to achieve the above object, the present invention provides a medical equipment maintenance data monitoring method based on big data, the method comprising the following steps:

[0010] S1. Collect the operation data of the medical equipment, then pre-process the operation data, extract features from the pre-processed data, and merge the extracted features into a comprehensive data matrix;

[0011] S2. Use an adaptive graph neural network to construct a device relationship graph and extract features of the relationship between medical devices and corresponding components; wherein the adaptive graph neural network aggregates and updates the node features in the device relationship graph through a convolution operation, which is expressed as follows:

[0012] Let H (l) ∈R N×d is the node feature matrix of the lth layer, where d represents the feature dimension. Initially, H (0) =X,X∈R N ×d , which is expressed as follows:

[0013]

[0014] Among them, x i Indicates device v i The feature vector of , i = 1, 2, ..., N, N represents the total number of devices and components, and d represents the feature dimension;

[0015] The update formula of the adaptive graph neural layer is expressed as follows:

[0016] H (l+1) =σ(D -1 AH (l) W (l) )

[0017] Among them, H(l) represents the node feature matrix of the lth layer, σ represents the nonlinear activation function, D represents the degree matrix, and its diagonal element d ii =∑ j a ij ; A represents the adjacency matrix, whose element a ij Indicates device v i and v j Is it connected? (l) Represents the trainable weight matrix of layer l

[0018] Dynamically update edge weights w based on interaction data between devices ij , which is expressed as follows:

[0019]

[0020] in, represents the edge weight at time t, η represents the learning rate, and h i and h j Respectively represent the device v i and v j The eigenvector of

[0021] Stacking multiple layers of adaptive graph neural layers, aggregating and updating node features layer by layer, is represented as follows:

[0022] H(L)=AGNN(L)(…AGNN(2)(AGNN(1)(X))…)

[0023] Among them, H (L) represents the node feature matrix of the final layer, L represents the number of layers of the adaptive graph neural layer;

[0024] S3. Extract high-order features through neighborhood feature aggregation based on the constructed device relationship graph and the extracted features, construct a high-order feature matrix based on the high-order features, and perform fault identification and fault mode labeling on the high-order features;

[0025] S4. Use variational autoencoders to reduce the dimension and reconstruct high-order feature matrices, extract potential feature representations, design LSTM models to perform time series prediction on potential feature representations, analyze the prediction results, and identify potential faults.

[0026] S5. Based on the prediction results, a reinforcement learning model based on Yebaes optimization is constructed to optimize the maintenance plan and resource scheduling of medical equipment;

[0027] S6. Construct a causal relationship model based on the prediction results and optimization results, use Bayesian network for causal inference, determine the causal influence of each variable on the target variable, and perform fault tracing analysis based on the results of causal inference to determine the main causes and influencing factors of the fault. Use optimization algorithms to optimize maintenance strategies, implement maintenance plans based on the optimized maintenance strategies, and monitor maintenance effects in real time.

[0028] Furthermore, the operating data of the medical equipment is obtained by using sensors, and the sensors include temperature sensors, pressure sensors and vibration sensors; the preprocessing includes data denoising, normalization processing and outlier processing; and the feature extraction includes time domain feature extraction and frequency domain feature extraction.

[0029] Furthermore, the device relationship diagram is represented as where ε = {e ij} indicates device v i and v j The interaction relationship between the edges ij The weight w ij Indicates device v i and v j The intensity of interaction between

[0030] Wherein, the S2 further includes:

[0031] The node features H of the final adaptive graph neural layer (L) As a high-order feature representation of the device, it is used for fault mode identification and prediction, as follows:

[0032] Z=H (L)

[0033] Among them, Z∈R N×d is the final device feature matrix.

[0034] Furthermore, in S3, the neighborhood feature aggregation is performed on each device v i , aggregate the features of its neighboring devices and obtain the neighborhood feature vector n i , which is expressed as follows:

[0035]

[0036] in, Indicates device v i The set of neighboring devices, α ij For device v i and v j The attention weight between devices v j v i The degree of influence; among them, the attention weight α ij Calculated by the following formula, expressed as follows:

[0037]

[0038] Where a represents the trainable attention vector, [z i ||z j ] represents the concatenation operation of feature vectors;

[0039] For each device v i , extract high-order features and form the final feature vector h i , which is expressed as follows:

[0040] h i =σ(W1z i +W2n i )

[0041] Among them, σ represents a nonlinear activation function; W1, W2 represent trainable weight matrices;

[0042] The final feature vector h of all devices i Composition feature matrix H∈R N×d , which is expressed as follows:

[0043]

[0044] Among them, h i Indicates device v i The final eigenvector of , i = 1, 2, ..., N.

[0045] Furthermore, a custom pattern recognition algorithm based on the feature matrix H is used to identify the failure mode of the equipment, including:

[0046] Cluster the feature matrix H to find device groups with similar features;

[0047] By analyzing the device features in each cluster, the Mahalanobis distance is used for anomaly detection to identify abnormal patterns, which are shown as follows:

[0048]

[0049] Among them, D M (h i ) indicates device v i The Mahalanobis distance, μ represents the feature mean vector, ∑ represents the feature covariance matrix;

[0050] Fault mode labeling is to label the identified fault mode on the corresponding device to form a fault mode set Each failure mode corresponds to a specific device group, represented as follows:

[0051]

[0052] Among them, v i Indicates the device, f i Indicates device v i Fault flag, f i =1 means the device is faulty, f i =0 means the device is normal.

[0053] Furthermore, the variational autoencoder comprises an encoder and a decoder; the encoder takes the input feature h i Mapping to the latent space z i , which is expressed as follows:

[0054]

[0055] Among them, μ(h i ) and σ 2 (h i ) represents the feature h i The corresponding mean and variance;

[0056] The decoder transforms the latent space z i Remap back to the feature space to generate reconstructed features It is expressed as follows:

[0057]

[0058] Among them, g(z i ) is the output of the decoder network.

[0059] Furthermore, the LSTM model is designed to perform time series prediction on the potential feature representation, including:

[0060] Assume z i (t+τ) is the device v i The predicted potential features at the future time t+τ, the LSTM model is expressed as follows:

[0061] f t =σ(W f [z i (t), h t-1 ]+b f )

[0062] i t =σ(W i [z i (t), h t-1 ]+b i )

[0063] c t =f t ⊙c t-1 +it ⊙tanh(W c [z i (t), h t-1 ]+b c ))

[0064] o t =σ(W o [z i (t), c t ]+b o )

[0065] h t =o t ⊙tanh(c t )

[0066] Among them, f t ,i t , o t They represent the forget gate, input gate and output gate respectively, and c t represents the state of the memory unit, h t represents the hidden state of LSTM, W f , W i , W c , W o Represents the weight matrix of each gate, b f , b i , b c , b o Represent the bias vector of each gate, σ represents the Sigmoid activation function, tanh represents the hyperbolic tangent activation function, and ⊙ represents element-by-element multiplication.

[0067] Furthermore, the loss functions of the variational autoencoder and LSTM model are optimized simultaneously during the training process. The loss function of the variational autoencoder includes reconstruction error and KL divergence, and the loss function of the LSTM is prediction error.

[0068] Among them, the loss function of the variational autoencoder is expressed as follows:

[0069]

[0070] Among them, L VAE Denotes the loss function of the variational autoencoder, D KL represents the Kullback-Leibler divergence.

[0071] The loss function of the LSTM model is expressed as follows:

[0072]

[0073] Among them, L LSTM represents the loss function of the LSTM model, zi (t) represents the real potential characteristics; represents the prediction potential feature;

[0074] Use the trained variational autoencoder and LSTM model to predict future potential features, as shown below:

[0075]

[0076] in, represents the predicted future potential features, and LSTM represents the trained LSTM model

[0077] Analyze the prediction results, identify potential faults, calculate the difference between the potential features at the future time t+τ and the potential features of the normal state, and set the threshold for fault judgment, which is expressed as follows:

[0078]

[0079] Among them, Δ i (t+τ) represents the potential feature difference, Indicates device v i potential characteristics of the normal state.

[0080] If Δ i (t+τ) exceeds the preset threshold ∈, then the device v is determined i At time t+τ there is a risk of failure.

[0081] Furthermore, the reinforcement learning model based on YeBas optimization includes:

[0082] The deep Q network in deep reinforcement learning is used for training, where the state value function Q(s, a) is approximated by a neural network and is expressed as follows:

[0083]

[0084] Where Q(s, a; θ) is the state value function, which represents the expected reward of selecting action a in state s; θ represents the parameters of the neural network; γ represents the discount factor, which represents the attenuation coefficient of future rewards; s′ represents the state transferred to after executing action a; a′ represents the optimal action in state s′;

[0085] The Bayesian optimization method is used to optimize the hyperparameters of DQN, where the Bayesian optimization objective function is expressed as follows:

[0086]

[0087] Among them, θ * represents the optimal hyperparameters, and E[Q(s, a; θ)] represents the expected reward function.

[0088] Furthermore, the causal relationship model uses a structural equation model to represent the causal relationship between the variables of the equipment, including:

[0089] Let Y be the target variable, X i As the independent variable, the causal relationship model is expressed as follows:

[0090] Y=f(X1,X2,…,X n )+ε

[0091] Among them, f(·) represents the causal relationship function, ε represents the error term, which represents the impact of unobserved factors on Y;

[0092] Use Bayesian network to perform causal inference and determine the causal influence of each variable on the target variable; wherein, the Bayesian network is based on the conditional probability distribution P(Y|X1, X2, ..., X n ) to represent the dependency between variables, as follows:

[0093]

[0094] Where: P(Y, X1, X2, ..., X n ) is the joint probability distribution, P(X1, X2, ..., X n ) is the marginal probability distribution.

[0095] The beneficial technical effects of the present invention are at least as follows:

[0096] (1) The present invention uses adaptive graph neural network (AGN) technology to model the relationship between medical devices and their components, and constructs the operating data of the devices and components and their interactive relationships into a graph structure, where nodes represent devices or components and edges represent interactions or dependencies between devices. The AGN can dynamically adjust feature extraction strategies to capture complex relationships between devices and fault propagation paths. This innovation solves the problem of traditional methods that only focus on single device data and ignore relationships between devices, allowing the system to identify and utilize potential relationships between devices and improve the accuracy and real-time performance of fault prediction.

[0097] (2) The present invention uses an adaptive graph neural network to extract features and recognize patterns in the equipment relationship graph. The system can identify potential fault modes and propagation paths in equipment operation, thereby achieving high-precision fault prediction. The adaptive graph neural network model can dynamically learn and adapt to changes in equipment operating status, discover potential fault hazards in advance, and reduce equipment downtime and diagnostic delays. Through high-precision fault identification and prediction, the reliability and continuity of equipment operation are improved.

[0098] (3) The present invention realizes data-driven fault prediction and preventive maintenance through the application of adaptive graph neural network technology. Adaptive graph neural network can make full use of equipment operation data, combined with time series models (such as LSTM), to predict the future operation status of the equipment, identify potential faults in advance and formulate corresponding maintenance plans, thereby effectively improving maintenance efficiency and equipment operation stability. Through preventive maintenance, the operating efficiency and stability of the equipment are improved.

[0099] (4) The present invention effectively solves the key problems in existing medical equipment maintenance methods by innovatively applying adaptive graph neural network technology, realizes dynamic modeling of relationships between equipment, high-precision fault pattern recognition, and data-driven preventive maintenance. The system significantly improves the reliability of equipment operation and maintenance efficiency, and has significant innovation and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0101] Figure 1 This is a flow chart of a medical equipment maintenance data monitoring method based on big data of the present invention. DETAILED DESCRIPTION

[0102] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0103] In one or more embodiments, Figure 1 As shown, a medical equipment maintenance data monitoring method based on big data is disclosed, and the method includes the following steps S1-S6:

[0104] S1. Collect the operation data of medical equipment, then preprocess the operation data, extract features from the preprocessed data, and merge the extracted features into a comprehensive data matrix.

[0105] Specifically, through sensors and Internet of Things technology, the operating data of medical equipment can be collected in real time, including multi-dimensional data such as temperature, pressure, and vibration.

[0106] The temperature sensor collects temperature T(t) in degrees Celsius, where t is time. The pressure sensor collects pressure P(t) in Pascals, where t is time. The vibration sensor collects vibration V(t) in meters per second.2 , t is time. The sampling frequency is set to collect data once per second, that is, the data sequence x i (t), where i represents the type of sensor (e.g., temperature sensor, pressure sensor, and vibration sensor) and t represents time.

[0107] Furthermore, denoising algorithms are used to reduce the noise in the sensor data and improve the data quality.

[0108] For each sensor data x i (t) Perform moving average denoising, using a moving average denoising method with a window size of N:

[0109]

[0110] in, For denoised data, N generally ranges from 5 to 10, and the specific value is adjusted according to the actual data noise situation.

[0111] Furthermore, discrete wavelet transform (DWT) is used to decompose the signal and perform denoising:

[0112]

[0113] Among them, W ψ (a, b) are wavelet transform coefficients, x(t) is the original data, ψ a，b (t) is the wavelet basis function, a and b are the scale and translation parameters respectively.

[0114] Furthermore, sensor data of different dimensions are normalized to the same dimension to facilitate subsequent processing.

[0115] Specifically, for each sensor data x i (t), using the Min-Max normalization method:

[0116]

[0117] Among them, x i，norm (t) is the normalized data, X i，min and x i,max are the minimum and maximum values ​​of the sensor i data respectively.

[0118] For each sensor data x i (t), using the Z-score normalization method:

[0119]

[0120] Among them, x i，norm (t) is the standardized data, μ iis the mean value of sensor i data, σ i is the standard deviation of the sensor i data.

[0121] Furthermore, missing values ​​in the data are filled to ensure data integrity.

[0122] Furthermore, for the time point t where the missing value is located, linear interpolation is performed using the adjacent data points before and after:

[0123]

[0124] Among them, x i (t) is the interpolated data, t1 and t2 are the time points before and after the missing value, respectively, x i (t1) and x i (t2) are the known data at time points t1 and t2 respectively.

[0125] The missing values ​​are filled by optimizing the regularization method and the objective function is constructed:

[0126]

[0127] Among them, J(x) is the objective function, x is the data sequence to be optimized, is a set of known data time points, is the observed data, f(t) is the fitting function, and λ is the regularization parameter.

[0128] Furthermore, key features are extracted from the sensor data for subsequent analysis, including:

[0129] Time domain features: Calculate mean, variance, peak, skewness and kurtosis:

[0130] Mean:

[0131]

[0132] Among them, μ i is the mean value of sensor i data.

[0133] variance:

[0134]

[0135] in, is the variance of sensor i data.

[0136] Peak:

[0137] P peak,i =max(x i (t))

[0138] Among them, P peak,iis the peak value of sensor i data.

[0139] Skewness:

[0140]

[0141] Among them, Skewness i is the skewness of the sensor i data.

[0142] Kurtosis:

[0143]

[0144] Among them, Kurtosis i is the kurtosis of the sensor i data.

[0145] Frequency domain characteristics:

[0146] Perform fast Fourier transform (FFT) to extract frequency domain features:

[0147]

[0148] Among them, X i (f) is the frequency domain signal of sensor i data, f is the frequency, T is the total sampling time, and j is the imaginary unit.

[0149] Furthermore, the processed data from each sensor are merged into a comprehensive data matrix as input for subsequent modeling.

[0150] For each time point t, the normalized data of each sensor are combined:

[0151]

[0152] Among them, T norm (t), P norm (t) and V norm (t) are the normalized data of temperature, pressure and vibration sensors, respectively.

[0153] Combine the feature extraction results:

[0154]

[0155] Among them, μ, σ 2 , P peak , Skewness and Kurtosis are mean, variance, peak, skewness and kurtosis respectively.

[0156] Through the above detailed steps, the collection and preprocessing of multi-dimensional data of medical equipment are realized, including denoising, normalization, missing value filling and feature extraction, to ensure the high quality and consistency of data. This data preprocessing solution is specially designed for the characteristics of medical equipment data. Considering the diversity of sensor data and possible missing value problems, a variety of denoising, normalization and interpolation methods are used to ensure data quality and consistency, providing a solid foundation for subsequent equipment relationship modeling and fault prediction.

[0157] S2. Use adaptive graph neural network to construct a device relationship graph and perform feature extraction on the relationship between medical devices and their corresponding components.

[0158] Specifically, define the device relationship diagram:

[0159] Suppose the set of medical equipment and its components is where v i Represents the i-th device or component, and N is the total number of devices and components. Define the device relationship diagram where ε = {e ij} indicates device v i and v j The interaction between them. ij The weight w ij Indicates device v i and v j The interaction strength between them is initially set to 1 or calculated based on historical interaction data.

[0160] Furthermore, each device v i The corresponding eigenvector is Where d is the feature dimension. The feature vector can contain time domain and frequency domain features extracted from data preprocessing, such as mean, variance, peak, etc. The initial feature matrix X∈R N×d :

[0161]

[0162] Where: x i Indicates device v i The feature vector of , i = 1, 2, ..., N. N represents the total number of devices and components. d represents the feature dimension.

[0163] Furthermore, the adaptive graph neural network (AGNN) layer aggregates and updates the node features in the graph through convolution operations. is the node feature matrix of the first layer, initially H (0) =X. The update formula of AGNN layer:

[0164] H (l+1) =σ(D -1 AH(l )W (l )) (11)

[0165] Among them, H (l) represents the node feature matrix of the first layer. σ is a nonlinear activation function, such as ReLU. D is the degree matrix, whose diagonal elements d ii =∑ j a ij . A is an adjacency matrix, whose element a ij Indicates device v i and v j Whether connected. (l) is the trainable weight matrix of the lth layer.

[0166] Furthermore, the edge weight w is dynamically updated according to the interaction data between devices. ij :

[0167]

[0168] in, is the edge weight at time t. η is the learning rate. h i and h j Device v i and v j The feature vector of .

[0169] Furthermore, multiple layers of AGNN are stacked to aggregate and update node features layer by layer:

[0170] H (L) =AGNN (L) (…AGNN (2) (AGNN (1) (X))…) (13)

[0171] Among them, H (L) is the node feature matrix of the final layer. L is the number of layers of AGNN.

[0172] The node features H of the final layer (L) As a high-level feature representation of the device, it is used for subsequent failure mode identification and prediction:

[0173] Z=H (L) (14)

[0174] Among them, Z∈R N×d is the final device feature matrix.

[0175] Through the above detailed steps, the device relationship modeling of multi-dimensional data of medical devices is realized, including defining the device relationship graph, constructing the initial feature matrix, applying the adaptive graph neural network layer, adaptive weight update, multi-layer AGNN stacking and output feature integration. This method can effectively capture the complex interaction relationship between devices and provide a solid foundation for subsequent fault mode identification and prediction. This device relationship modeling solution is specially designed for the diversity and complexity of medical device data, ensuring efficient data utilization and feature extraction.

[0176] S3. According to the constructed device relationship diagram and the extracted features, high-order features are extracted through neighborhood feature aggregation, and a high-order feature matrix is ​​constructed based on the high-order features, and fault identification and fault mode labeling are performed on the high-order features.

[0177] Specifically, for each device v i , using the feature matrix obtained from device relationship modeling Perform feature extraction. Feature extraction includes aggregation operations and high-order feature extraction. Let z i For device v i The feature vector of , whose dimension is d.

[0178] For each device v i , aggregate the features of its neighboring devices and obtain the neighborhood feature vector n i :

[0179]

[0180] in, Indicates device v i The set of neighboring devices. ij For device v i and v j The attention weight between devices v j v i The attention weight is calculated by the following formula:

[0181]

[0182] Where: a is a trainable attention vector. i ||z j ] represents the concatenation operation of feature vectors.

[0183] For each device v i , extract high-order features and form the final feature vector h i :

[0184] h i =σ(W1z i +W2n i ) (17)

[0185] Where: σ is a nonlinear activation function, such as ReLU. W1 and W2 are trainable weight matrices.

[0186] Furthermore, the final feature vector h of all devices i Composition feature matrix H∈R N×d :

[0187]

[0188] Among them, h i Indicates device v i The final eigenvector of , i = 1, 2, ..., N.

[0189] Use a custom pattern recognition algorithm based on the feature matrix H to identify the failure mode of the equipment. The specific steps of pattern recognition are as follows:

[0190] Feature clustering: Cluster the feature matrix H to find groups of devices with similar features.

[0191] Anomaly detection: By analyzing the characteristics of the devices in each cluster, abnormal patterns are identified. Anomaly detection using Mahalanobis distance:

[0192]

[0193] Among them, D M (h i ) is the device v i The Mahalanobis distance of . μ is the feature mean vector. ∑ is the feature covariance matrix.

[0194] Furthermore, the identified fault modes are annotated on the corresponding devices to form a set of fault modes. Each failure mode corresponds to a specific device group:

[0195]

[0196] Among them, v i Indicates the device. i Indicates device v i Fault flag, f i =1 means the device is faulty, f i =0 means the device is normal.

[0197] It can be understood that through the above detailed steps, the present invention realizes the extraction of features from the device relationship diagram and pattern recognition, including neighborhood feature aggregation, high-order feature extraction, feature matrix construction and pattern recognition. This method can effectively extract the key features of the device, identify potential failure modes, and provide a solid foundation for subsequent failure prediction. This feature extraction and pattern recognition solution is specially designed for the diversity and complexity of medical device data, ensuring efficient use of data and accurate identification of failure modes.

[0198] S4. Use variational autoencoders to reduce the dimension and reconstruct high-order feature matrices, extract potential feature representations, design LSTM models to perform time series predictions on potential feature representations, analyze the prediction results, and identify potential faults.

[0199] For each device v i , extract its eigenvector h from the feature matrix H i , and combined with historical feature data, construct time series data. Let the time window be T, and the time series data be represented by h i (tT),h i (t-T+1),...,h i (t).

[0200] Furthermore, the variational autoencoder (VAE) feature extraction is used. VAE reduces the dimension and reconstructs the time series features to extract the potential feature representation. VAE consists of two parts: encoder and decoder:

[0201] The encoder takes as input the feature h i Mapping to the latent space z i :

[0202]

[0203] Among them, μ(h i ) and σ 2 (h i ) is the characteristic h i The corresponding mean and variance are learned through neural network.

[0204] The decoder transforms the latent space z i Remap back to the feature space to generate reconstructed features

[0205]

[0206] Among them, g(z i ) is the output of the decoder network.

[0207] Furthermore, the LSTM model is used to train the latent features z extracted from VAE iFor time series forecasting, let z i (t+τ) is the device v i The predicted potential features at the future time t+τ. The LSTM model is defined as follows:

[0208]

[0209] Among them, f t ,i t , o t They are forget gate, input gate and output gate respectively. t is the memory unit state. t is the hidden state of LSTM. f , W i , W c , W o are the weight matrices of each gate. f , b i , b c , b o are the bias vectors of each gate. σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function. ⊙ represents element-by-element multiplication.

[0210] Furthermore, the loss functions of VAE and LSTM are optimized simultaneously during the training process. The loss function of VAE includes reconstruction error and KL divergence, and the loss function of LSTM is prediction error. VAE loss function:

[0211]

[0212] Among them, L VAE is the loss function of VAE. KL is the Kullback-Leibler divergence.

[0213] LSTM loss function:

[0214]

[0215] Among them, L LSTM is the loss function of LSTM. i (t) is the real potential feature. To predict potential features.

[0216] Furthermore, the trained VAE and LSTM joint model is used to predict future potential features:

[0217]

[0218] in, is the predicted future potential feature. LSTM is the trained LSTM model.

[0219] Analyze the prediction results to identify potential faults. Calculate the difference between the potential features at the future time t+τ and the potential features of the normal state, and set the threshold to make fault judgments. Fault judgment formula:

[0220]

[0221] Among them, Δ i (t+τ) is the potential feature difference, For device v i potential characteristics of the normal state.

[0222] If Δ i (t+τ) exceeds the preset threshold ∈, then the device v is determined i At time t+τ there is a risk of failure:

[0223]

[0224] It can be understood that through the above detailed steps, data-driven fault prediction based on the feature matrix H is realized, including constructing time series data, defining and training the VAE and LSTM joint prediction model, and performing fault prediction and result analysis. This method can effectively combine historical data and real-time data to accurately predict the future operating status of the equipment, identify potential faults in advance, and provide an important basis for subsequent maintenance scheduling. This data-driven fault prediction solution is specially designed for the time series characteristics of medical equipment data, ensuring the efficiency and accuracy of the prediction.

[0225] S5. According to the prediction results, a reinforcement learning model based on Yebaes optimization is constructed to optimize the maintenance plan and resource scheduling of medical equipment.

[0226] Specifically, the equipment maintenance scheduling problem is modeled as a Markov decision process (MDP), which is expressed as follows:

[0227] State Space Including the health status of the equipment, current time, maintenance resource availability, etc.

[0228] Action Space Includes the decision to maintain or not maintain the equipment.

[0229] State transition function P(s′|s, a): describes the probability of transitioning to state s′ after executing action a in state s.

[0230] Reward function R(s, a): describes the reward obtained by performing action a in state s. The reward function needs to consider factors such as maintenance costs and equipment failure risks.

[0231] Furthermore, initialize the state space Action Space State transition function P(s′|s,a) and reward function R(s,a).

[0232] Furthermore, the DQN model parameters θ and the Bayesian optimization objective function are initialized. The health status data of the equipment is obtained from the fault prediction model as the state space Input.

[0233] Furthermore, the Deep Q-Network (DQN) in Deep Reinforcement Learning (DRL) is used for training.

[0234] The state value function Q(s, a) is approximated by a neural network:

[0235]

[0236] Where Q(s, a; θ) is the state value function, which represents the expected reward of selecting action a in state s. θ is the parameter of the neural network. γ is the discount factor, which represents the decay coefficient of future rewards. s′ is the state transferred to after executing action a. a′ is the optimal action in state s′.

[0237] Furthermore, the Bayesian optimization method is used to optimize the hyperparameters of DQN to improve the model performance. Bayesian optimization objective function:

[0238]

[0239] Among them, θ * is the optimal hyperparameter. is the expected reward function.

[0240] Furthermore, the DQN model is used for training to update the state value function Q(s, a; θ). During the training process, the hyperparameters of the DQN model are adjusted through the Bayesian optimization method to improve the model performance.

[0241] It can be understood that based on the trained DQN model and optimized hyperparameters, an intelligent scheduling strategy is generated. The intelligent scheduling strategy π(s) is defined as:

[0242]

[0243] Among them, π(s) is the optimal action selected in state s, which represents the specific maintenance decision.

[0244] S6. Construct a causal relationship model based on the prediction results and optimization results, use Bayesian network for causal inference, determine the causal influence of each variable on the target variable, and perform fault tracing analysis based on the results of causal inference to determine the main causes and influencing factors of the fault. Use optimization algorithms to optimize maintenance strategies, implement maintenance plans based on the optimized maintenance strategies, and monitor maintenance effects in real time.

[0245] Specifically, by collecting sensor data, equipment historical operation data and environmental data, a causal relationship model between various variables of the equipment is constructed.

[0246] Structural equation model (SEM) is used to represent the causal relationship between the variables of the equipment. Let Y be the target variable (such as equipment failure status), X i is the independent variable (such as sensor data, environmental factors, etc.), the causal relationship model is expressed as:

[0247] Y=f(X1,X2,…,X n )+ε (36)

[0248] Among them, f(.) is the causal function, and ε is the error term, which represents the impact of unobserved factors on Y.

[0249] Furthermore, the Bayesian Network is used for causal inference to determine the causal influence of each variable on the target variable. The Bayesian Network is based on the conditional probability distribution P(Y|X1,X2,...,X n ) to express the dependency between variables:

[0250]

[0251] Among them, P(Y,X1,X2,…,X n ) is the joint probability distribution, P(X1,X2,…,X n ) is the marginal probability distribution.

[0252] Based on the causal inference results, the fault source analysis is carried out to determine the main causes and influencing factors of the fault. i is the i-th cause variable, Y is the fault state, and the influence degree of each cause variable on the fault state is calculated by i :

[0253]

[0254] Where: Δ i is the cause variable X i The degree of impact on fault state Y.

[0255] Furthermore, the maintenance strategy is optimized using an optimization algorithm, with the goal of minimizing the risk of failure and maintenance cost. Assume that the maintenance strategy S includes variables such as maintenance time t and maintenance resource allocation r, and the optimization objective function is expressed as:

[0256]

[0257] in, is the expected value of the failure cost, which represents the expected cost of a failure under maintenance strategy S. is the expected value of the maintenance cost, which represents the maintenance resource consumption under the maintenance strategy S, and λ is the adjustment parameter used to balance the failure cost and maintenance cost.

[0258] Understandably, according to the optimized maintenance strategy, the maintenance plan is implemented and the maintenance effect is monitored in real time. Feedback is provided based on the maintenance results, the causal relationship model and optimization algorithm are adjusted, and the maintenance strategy is continuously optimized.

[0259] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0260] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A medical equipment maintenance data monitoring method based on big data, characterized in that: The method comprises the following steps: S1. Collect the operation data of the medical equipment, then pre-process the operation data, extract features from the pre-processed data, and merge the extracted features into a comprehensive data matrix; S2. Use an adaptive graph neural network to construct a device relationship graph and perform feature extraction on the relationship between medical devices and corresponding components; wherein the adaptive graph neural network aggregates and updates the node features in the device relationship graph through a convolution operation, which is expressed as follows: Let H (l) ∈R N×d is the node feature matrix of the lth layer, where d represents the feature dimension. Initially, H (0) =X,X∈R N×d , which is expressed as follows: Among them, x i Indicates device v i The feature vector of , i = 1, 2, ..., N, N represents the total number of devices and components, and d represents the feature dimension; The update formula of the adaptive graph neural layer is expressed as follows: H (l+1) =σ(D -1 AH (l) W (l) ) Among them, H (l) represents the node feature matrix of the lth layer, σ represents the nonlinear activation function, D represents the degree matrix, and its diagonal element d ii =∑ j a ij ; A represents the adjacency matrix, whose element a ij Indicates device v i and v j Is it connected? (l) Represents the trainable weight matrix of layer l Dynamically update edge weights w based on interaction data between devices ij , which is expressed as follows: in, represents the edge weight at time t, η represents the learning rate, and h i and h j Respectively represent the device v i and v j The eigenvector of Stacking multiple layers of adaptive graph neural layers, aggregating and updating node features layer by layer, is represented as follows: H (L) =AGNN (L) (…AGNN (2) (AGNN (1) (X))…) Among them, H (L) represents the node feature matrix of the final layer, L represents the number of layers of the adaptive graph neural layer; S3. Extract high-order features through neighborhood feature aggregation based on the constructed device relationship graph and the extracted features, construct a high-order feature matrix based on the high-order features, and perform fault identification and fault mode labeling on the high-order features; S4. Use variational autoencoders to reduce the dimension and reconstruct high-order feature matrices, extract potential feature representations, design LSTM models to perform time series prediction on potential feature representations, analyze the prediction results, and identify potential faults. S5. Based on the prediction results, a reinforcement learning model based on Yebaes optimization is constructed to optimize the maintenance plan and resource scheduling of medical equipment; S6. Construct a causal relationship model based on the prediction results and optimization results, use Bayesian network for causal inference, determine the causal influence of each variable on the target variable, and perform fault tracing analysis based on the results of causal inference to determine the main causes and influencing factors of the fault. Use optimization algorithms to optimize maintenance strategies, implement maintenance plans based on the optimized maintenance strategies, and monitor maintenance effects in real time.

2. A method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: The operating data of the medical equipment is obtained by using sensors, wherein the sensors include temperature sensors, pressure sensors and vibration sensors; the preprocessing includes data denoising, normalization processing and outlier processing; and the feature extraction includes time domain feature extraction and frequency domain feature extraction.

3. A method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: The device relationship diagram is represented as where ε = {e ij } indicates device v i and v j The interaction relationship between the edges ij The weight w ij Indicates device v i and v j The intensity of interaction between Wherein, the S2 further includes: The node features H of the final adaptive graph neural layer (L) As a high-order feature representation of the device, it is used for fault mode identification and prediction, as follows: Z=H (L) Among them, Z∈R N×d is the final device feature matrix.

4. The method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: In S3, the neighborhood feature aggregation is performed on each device v i , aggregate the features of its neighboring devices and obtain the neighborhood feature vector n i , which is expressed as follows: in, Indicates device v i The set of neighboring devices, α ij For device v i and v j The attention weight between devices v j v i The degree of influence; among them, the attention weight α ij Calculated by the following formula, expressed as follows: Where a represents the trainable attention vector, [z i ||z j ] represents the concatenation operation of feature vectors; For each device v i , extract high-order features and form the final feature vector h i , which is expressed as follows: h i =σ(W1z i +W2n i ) Among them, σ represents a nonlinear activation function; W1, W2 represent trainable weight matrices; The final feature vector h of all devices i Composition feature matrix H∈R N×d , which is expressed as follows: Among them, h i Indicates device v i The final eigenvector of , i = 1, 2, …, N.

5. A method for monitoring medical equipment maintenance data based on big data according to claim 4, characterized in that: Use a custom pattern recognition algorithm based on the feature matrix H to identify the failure modes of the equipment, including: Cluster the feature matrix H to find device groups with similar features; By analyzing the device features in each cluster, the Mahalanobis distance is used for anomaly detection to identify abnormal patterns, which are shown as follows: Among them, D M (h i ) indicates device v i The Mahalanobis distance, μ represents the feature mean vector, and Σ represents the feature covariance matrix; Fault mode labeling is to label the identified fault mode on the corresponding device to form a fault mode set Each failure mode corresponds to a specific device group, represented as follows: Among them, v i Indicates the device, f i Indicates device v i Fault flag, f i =1 means the device is faulty, f i =0 means the device is normal.

6. A method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: The variational autoencoder comprises two parts: an encoder and a decoder; the encoder takes the input feature h i Mapping to the latent space z i , which is expressed as follows: Among them, μ(h i ) and σ 2 (h i ) represents the feature h i The corresponding mean and variance; The decoder transforms the latent space z i Remap back to the feature space to generate reconstructed features It is expressed as follows: Among them, g(z i ) is the output of the decoder network.

7. A method for monitoring medical equipment maintenance data based on big data according to claim 6, characterized in that: The designed LSTM model performs time series prediction on the potential feature representation, including: Assume z i (t+τ) is the device v i The predicted potential features at the future time t+τ, the LSTM model is expressed as follows: f t =σ(W f [z i (t),h t-1 ]+b f ) i t =σ(W i [z i (t),h t-1 ]+b i ) c t =f t ⊙c t-1 +i t ⊙tanh(W c [z i (t),h t-1 ]+b c )) o t =σ(W o [z i (t),c t ]+b o ) h t =o t ⊙tanh(c t ) Among them, f t ,i t ,o t They represent the forget gate, input gate and output gate respectively, and c t represents the state of the memory unit, h t represents the hidden state of LSTM, W f ,W i ,W c ,W o Represents the weight matrix of each gate, b f ,b i ,b c ,b o Represent the bias vector of each gate, σ represents the Sigmoid activation function, tanh represents the hyperbolic tangent activation function, and ⊙ represents element-by-element multiplication.

8. A method for monitoring medical equipment maintenance data based on big data according to claim 7, characterized in that: During the training process, the loss functions of the variational autoencoder and LSTM model are optimized simultaneously. The loss function of the variational autoencoder includes reconstruction error and KL divergence, and the loss function of the LSTM is prediction error. Among them, the loss function of the variational autoencoder is expressed as follows: Among them, L VAE Denotes the loss function of the variational autoencoder, D KL represents the Kullback-Leibler divergence. The loss function of the LSTM model is expressed as follows: Among them, L LSTM represents the loss function of the LSTM model, z i (t) represents the real potential characteristics; represents the prediction potential feature; Use the trained variational autoencoder and LSTM model to predict future potential features, as shown below: in, Represents the predicted future potential features, and LSTM represents the trained LSTM model Analyze the prediction results, identify potential faults, calculate the difference between the potential features at the future time t+τ and the potential features of the normal state, and set the threshold for fault judgment, which is expressed as follows: Among them, Δ i (t+τ) represents the potential feature difference, Indicates device v i potential characteristics of the normal state. If Δ i (t+τ) exceeds the preset threshold ∈, then the device v is determined i At time t+τ there is a risk of failure.

9. The method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: The reinforcement learning model based on YeBas optimization includes: The deep Q network in deep reinforcement learning is used for training, where the state value function Q(s,a) is approximated by a neural network and is expressed as follows: Among them, Q(s,a;θ) is the state value function, which represents the expected reward of selecting action a in state s; θ represents the parameters of the neural network; γ represents the discount factor, which represents the attenuation coefficient of future rewards; s′ represents the state transferred to after executing action a; a′ represents the optimal action in state s′; The Bayesian optimization method is used to optimize the hyperparameters of DQN, where the Bayesian optimization objective function is expressed as follows: Among them, θ * represents the optimal hyperparameters, and E[Q(s,a;θ)] represents the expected reward function.

10. The method for monitoring medical equipment maintenance data based on big data according to claim 1, characterized in that: The causal relationship model uses a structural equation model to represent the causal relationship between the variables of the equipment, including: Let Y be the target variable, X i As the independent variable, the causal relationship model is expressed as follows: Y=f(X1,X2,…,X b )+ε Among them, f(·) represents the causal function, ε represents the error term, which represents the impact of unobserved factors on Y; Use Bayesian network to perform causal inference to determine the causal influence of each variable on the target variable; wherein, the Bayesian network is based on the conditional probability distribution P(Y|X1,X2,…,X n ) to represent the dependency between variables, as follows: Where: P(Y,X1,X2,…,X n ) is the joint probability distribution, P(X1,X2,…,X n ) is the marginal probability distribution.

Images