Intelligent medical equipment fault diagnosis and prediction method based on deep learning

Abstract

The invention discloses an intelligent medical equipment fault diagnosis and prediction method based on deep learning, and the method comprises the following steps: S1, obtaining the data of medical equipment in an operation process, and constructing an original data set of the medical equipment; s2, preprocessing the original data set of the medical equipment, and extracting an input feature vector of the medical equipment; s3, constructing a multi-task deep neural network model, and combining and constructing an initial result set; s4, on the basis of the initial result set, constructing an overall performance index of the initial result set; s5, based on the overall performance index of the initial result set, a multi-target grey wolf optimization algorithm is adopted to carry out joint optimization on the result, and a final optimization parameter set is obtained; and S6, correcting the result by using the optimization parameter set, and outputting an optimization diagnosis result and an optimization prediction result. According to the method, a joint mechanism of the multi-task deep neural network and the multi-target grey wolf optimization algorithm is adopted, and cooperative output and precision enhancement of medical equipment fault diagnosis and life prediction results are realized.

Description

Technical Field

[0001] The present invention relates to the technical fields of intelligent medical systems and artificial intelligence optimization technologies, and particularly relates to an intelligent medical device fault diagnosis and prediction method based on deep learning. Background Art

[0002] With the increasing popularity and key role of medical devices in various hospitals and intelligent medical systems, ensuring the stable operation and timely maintenance of the devices has become an indispensable part of the modern medical service system. Especially in large comprehensive medical institutions, various high-precision and intelligent medical devices (such as CT, magnetic resonance imaging, life monitors, artificial intelligence-assisted diagnosis platforms, etc.) operate frequently and have complex systems. Once a failure occurs, it may not only affect the daily diagnosis and treatment process, but also pose risks to the lives and safety of patients. Therefore, how to achieve effective prediction before the device fails, and be able to accurately diagnose and respond quickly when a failure occurs, has become the core issue of intelligent operation and maintenance management of medical devices.

[0003] At present, the fault diagnosis and maintenance management of medical devices mainly rely on periodic manual inspections, device-built warning mechanisms, or system judgments based on empirical rules. These methods can play a certain role in scenarios where the devices are relatively simple or have a low usage frequency. However, in actual medical scenarios, the device operating environment is complex, the load fluctuates frequently, and the device state is affected by multi-dimensional factors such as temperature, voltage, current, and mechanical vibration. Traditional methods often cannot reflect the true operating state of the device in a timely and accurate manner, and there are obvious limitations. First, the rule-based method based on experience has a low dependence on data and is difficult to handle complex state combinations and hidden fault modes, resulting in insufficient diagnostic accuracy. Second, the manual intervention mechanism has a lagging response and often only deals with the situation after the device shows obvious failures or even stops running, lacking the ability of active prediction. In addition, for a multi-device collaborative operation system, traditional methods cannot effectively support centralized or unified health status assessment and life prediction.

[0004] In recent years, with the rapid development of artificial intelligence technology, some studies have attempted to introduce deep learning methods into medical equipment status recognition and anomaly detection scenarios, relying on large amounts of historical monitoring data to build neural network models to achieve intelligent judgment of equipment status. Such methods generally use models such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM) to extract features from the device's sensor data and complete classification tasks, achieving certain results. However, existing technologies mostly focus on the "diagnosis" task, that is, judging whether the current device status is abnormal, while paying less attention to the key indicator of "remaining life prediction" of the device. In addition, most methods use a single-task model and only classify whether the device is faulty. They fail to simultaneously consider the regression prediction problem of the future operating trend of the device, resulting in limited predictive and adaptable capabilities of the overall system.

[0005] More importantly, existing research often adopts a single loss function to train models, such as using cross entropy for classification training or using mean square error for regression training, ignoring the trade-off between diagnostic accuracy and prediction accuracy, and lacking a unified multi-task modeling and collaborative optimization mechanism. As a result, when the model performs well on one task, the performance of another task may degrade, making it difficult to meet the actual needs of the medical equipment intelligent operation and maintenance system for "diagnosis + prediction" collaborative optimization.

[0006] In terms of model output processing, most current research adopts a direct output strategy. The diagnostic probabilities and predicted values provided by the model are often used directly without post-processing and without secondary optimization based on task error feedback. Because neural network models are susceptible to input perturbations, their output results may be biased or unstable. This is especially true in scenarios with unbalanced sample distribution, drastic feature fluctuations, or hidden minor faults. The model output may exhibit insufficient diagnostic confidence or large prediction deviations. Traditional strategies such as parameter fine-tuning and posterior adjustment are difficult to adapt to both classification and regression tasks, limiting further improvements in model performance.

[0007] In addition, most existing optimization algorithms for multi-objective tasks use general gradient descent, weighted loss methods, or Pareto frontier search to achieve multi-task balance. However, these methods generally rely on hyperparameter settings and have difficulty automatically searching for nonlinear optimal adjustment paths. In particular, when the nonlinear coupling between task objective functions is strong and the gradient directions are inconsistent, such methods are prone to falling into local optimality, making it difficult to achieve global coordinated optimization of diagnostic and predictive performance. Some studies have also attempted to introduce intelligent algorithms such as genetic algorithms and particle swarm optimization to adjust model parameters or structures, but such methods still have significant deficiencies in optimization efficiency, multi-objective adaptability, and dynamic response capabilities.

[0008] In this context, as an emerging swarm intelligence optimization method, the Grey Wolf Optimization Algorithm has gradually attracted research attention in recent years due to its simple structure, strong search ability, and excellent convergence performance. Although existing literature has applied the Grey Wolf Algorithm to some neural network weight optimization problems, its direct application to multi-objective, structured, and deep output regulation tasks is still relatively scarce. Especially for scenarios such as medical devices with diverse data and complex task coupling, there is still a lack of an optimization method that can simultaneously process classification and regression results and perform joint regulation by combining model feedback errors.

[0009] Generally speaking, the existing intelligent diagnosis and prediction technologies for medical devices have the following prominent problems: First, there is a lack of a unified "diagnosis + prediction" multi-task collaborative modeling mechanism, resulting in fragmented model task division and insufficient information sharing; Second, the model output results lack a regulation mechanism based on feedback performance, making it difficult to ensure the robustness and accuracy of the results in high-risk environments; Third, existing optimization algorithms are difficult to effectively handle multi-objective non-convex tasks, with unstable optimization paths and limited convergence effects, affecting the practicality and deployability of the final system.

[0010] Therefore, how to provide an intelligent fault diagnosis and prediction method for medical devices based on deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0011] An object of the present invention is to propose an intelligent fault diagnosis and prediction method for medical devices based on deep learning. The present invention constructs a multi-task deep neural network model integrating a one-dimensional convolutional network and a bidirectional long short-term memory network to realize the synchronous prediction of the operating state and remaining life of medical devices, and further introduces a multi-objective Grey Wolf Optimization Algorithm to jointly regulate and optimize the model output results, thereby improving the stability and robustness of the prediction results while maintaining a high diagnostic accuracy. The final output can be used to support the preventive maintenance and intelligent operation and maintenance decision-making of key medical devices in hospitals, significantly improving the safety, reliability, and management intelligence level of device use.

[0012] An intelligent fault diagnosis and prediction method for medical devices based on deep learning according to an embodiment of the present invention includes the following steps:

[0013] S1. Obtain the data of the medical device during operation and construct the original dataset of the medical device;

[0014] S2. Preprocess the original dataset of the medical device and extract the input feature vectors of the medical device;

[0015] S3. Construct a multi-task deep neural network model, input the input feature vectors of the medical device, output the initial diagnosis result and the initial prediction result, and combine them to construct the initial result set;

[0016] S4. Based on the initial result set, perform performance evaluation on the initial diagnosis result and the initial prediction result, calculate the first objective function and the second objective function, and construct the overall performance index of the initial result set;

[0017] S5. Based on the overall performance index of the initial result set, use the multi-objective grey wolf optimization algorithm to jointly optimize the initial diagnosis result and the initial prediction result to obtain the final optimized parameter set;

[0018] S6. Use the optimized parameter set to correct the initial diagnosis result and the initial prediction result, and output the optimized diagnosis result and the optimized prediction result.

[0019] Optionally, the original medical device dataset includes temperature data, voltage data, current data, vibration signal data, working duration data, and operation log data of the medical device.

[0020] Optionally, the S2 specifically includes:

[0021] S21. Perform normalization processing on the temperature data, voltage data, current data, vibration signal data, working duration data, and operation log data in the original medical device dataset;

[0022] S22. Perform time series reconstruction processing on the normalized multi-modal data, and use the sliding window mechanism to construct fixed-length sequences for each type of feature to form a sequence matrix;

[0023] S23. Align and fuse the sequence matrices of various features according to time steps to construct the input feature vector of the medical device.

[0024] Optionally, the S3 specifically includes:

[0025] S31. Construct a multi-task deep neural network model including a shared feature extraction layer, a first task output branch, and a second task output branch;

[0026] S32. Input the input feature vector of the medical device into the shared feature extraction layer to obtain the shared feature representation;

[0027] S33. Input the shared feature representation into the first task output branch to output the initial diagnosis result:

[0028]

[0029] where Y cls is the initial diagnosis result, H is the shared feature matrix, W1 is the classification weight matrix, b1 is the classification bias term, c is the number of fault categories, and j is the fault category index;

[0030] S34. Input the shared feature representation into the second task output branch to output the initial prediction result:

[0031] Y reg = max(0, H·W2 + b2)+ ∈;

[0032] Where, Y reg is the initial prediction result, W2 is the regression task weight matrix, b2 is the regression task bias scalar, and ∈ is the prediction error perturbation term;

[0033] S35. Combine the initial diagnosis result Y cls with the initial prediction result Y reg to construct the initial result set.

[0034] Optionally, the S32 specifically includes:

[0035] S321. Input each time series sub-sample in the medical device input feature vector into a one-dimensional convolutional neural network module to perform local pattern extraction on the multi-modal features within each time window, and output the convolutional feature representation;

[0036] S322. Input the convolutional feature representation into a bidirectional long short-term memory network module to perform forward and backward modeling in chronological order, generate a context-related bidirectional hidden state sequence, and output the bidirectional sequence feature;

[0037] S323. Perform a pooling operation on the bidirectional sequence feature, and use the average pooling method to compress the time dimension to obtain the final shared feature vector;

[0038] S324. Process all sequence samples in turn to generate the shared feature representation.

[0039] Optionally, the S4 specifically includes:

[0040] S41. Manually label the device status corresponding to each input sequence in the medical device raw dataset according to the predefined fault type classification standard to generate the original label sequence, vectorize the original label sequence, and use the one-hot encoding method to convert the label of each input sequence into a fault category indicator vector with a length of c to generate the true fault label set;

[0041] S42. Collect the historical operation log data during the operation of the medical device, extract the start timestamp and end timestamp of each medical device during the corresponding operation cycle, and calculate the total life cycle of the device. For each sequence sample generated based on the sliding window mechanism, record the current timestamp, calculate the remaining life label relative to the failure moment of the device, and summarize the life labels of all samples to form the true life label set;

[0042] S43. Compare the initial diagnostic results element by element with the set of true fault labels to construct the first objective function:

[0043]

[0044] where L cls is the first objective function, n is the number of input feature sequences of the medical device, i is the index of the input feature sequence of the medical device, c is the number of fault categories, j is the fault category index, and w j is the weight of the j-th type of fault, is the true label value of the i-th input sequence under the j-th type of fault, is the initial diagnostic result of the i-th input sequence under the j-th type of fault;

[0045] S44. Perform sequence-level matching between the initial prediction result and the actual remaining life label to construct the performance evaluation function of the prediction task:

[0046]

[0047] where L reg is the second objective function, is the initial prediction result of the i-th input sequence, is the actual remaining life value of the i-th input sequence, and δ is the stability factor;

[0048] S45. Weightedly fuse the first objective function and the second objective function to form the overall performance index of the initial result set.

[0049] Optionally, the S5 specifically includes:

[0050] S51. Initialize the gray wolf population, set the population size to N, and assign a set of parameter vectors to be optimized to each gray wolf individual. Each set of parameter vectors includes a diagnostic adjustment parameter and a prediction adjustment parameter;

[0051] S52. Based on the position vector corresponding to each gray wolf individual, parametrically adjust the initial diagnostic result and the initial prediction result respectively to obtain the adjusted diagnostic result and the adjusted prediction result corresponding to the k-th gray wolf individual;

[0052] S53. Input the adjusted diagnostic result and the adjusted prediction result into the first objective function and the second objective function respectively, and calculate the fitness values respectively;

[0053] S54. Use the non-dominated sorting method to construct the Pareto optimal solution set and iterate the positions of the gray wolf individuals:

[0054]

[0055] where Xk (t + 1) is the parameter position of the k-th grey wolf individual in the (t + 1)-th round, X α (t), X β (t), X δ (t) are the parameter positions of the three individuals with the best performance in the current iteration respectively. γ is the search step factor, ω k is the adaptive adjustment coefficient of the k-th grey wolf individual, is the partial derivative symbol, is the classification error value of the k-th grey wolf individual, is the regression error value of the k-th grey wolf individual, is the diagnostic adjustment parameter of the k-th grey wolf individual, the prediction adjustment parameter of the k-th grey wolf individual;

[0056] S55. Repeat the operations of parametric adjustment, fitness value calculation, and grey wolf individual position iteration until the optimal individual convergence condition is met. Select the individual with the minimum combined loss from the Pareto optimal solution set to determine the final optimized parameter set.

[0057] Optionally, the S6 specifically includes:

[0058] S61. Use the diagnostic adjustment parameter in the final optimized parameter set to perform parametric confidence reconstruction on the initial diagnostic result to obtain an optimized diagnostic result;

[0059] S62. Use the prediction adjustment parameter in the final optimized parameter set to perform multi-dimensional adjustment on the initial prediction result to obtain an optimized prediction result;

[0060] S63. Output the optimized diagnostic result and the optimized prediction result.

[0061] Optionally, the S61 specifically includes:

[0062] S611. Divide the initial diagnostic result of each fault category of each input sequence by the diagnostic adjustment parameter to obtain a scaled logits expression;

[0063] S612. Calculate the exponential mapping value of each category corresponding to the scaled logits expression;

[0064] S613. Sum up all the exponential mapping values of the same input sample and then perform normalization to obtain a confidence reconstruction result, and form an optimized diagnostic result with all the confidence reconstruction results.

[0065] Optionally, the S62 specifically includes:

[0066] S621. Calculate the mean value of the set of true life labels of all samples in the current batch;

[0067] S622. Perform a sign judgment on the deviation of each initial prediction result from the mean of the true life label set to obtain a directional adjustment sign;

[0068] S623. Construct an adjustment term according to the prediction adjustment parameter and the directional adjustment sign, and add each initial prediction result to the corresponding adjustment term to obtain an optimized prediction result.

[0069] The beneficial effects of the present invention are as follows:

[0070] First of all, the present invention can face the multi-dimensional state perception requirements and complex task coupling problems that exist simultaneously during the operation of medical devices, construct a multi-task neural network structure integrating diagnostic tasks and remaining life prediction tasks, and through the way of sharing feature extraction and task branch decoupling, realize the collaborative modeling and parallel output of fault identification and remaining life assessment. The time series input structure constructed based on multi-source data such as temperature, voltage, current, vibration signal, operation duration, and log information enables the model to capture the dynamic evolution characteristics of the device operation state within the full life cycle range, significantly improving the robustness of the model in complex operation environments and its adaptability to multiple fault types. The introduction of a sliding window mechanism and a sequence alignment and reconstruction method in the feature modeling process further enhances the sensitivity of the model to temporal structure changes and its ability to capture hidden abnormal patterns, providing a solid foundation for highly accurate diagnosis and highly stable prediction.

[0071] Secondly, the present invention improves the model's joint modeling ability for local abnormal fluctuations and global trend changes by introducing a shared coding structure based on one-dimensional convolution and bidirectional long short-term memory networks, and respectively adopts output modules optimized for classification and regression tasks in multiple task branches to achieve an equilibrium coordination between diagnostic accuracy and prediction accuracy. In the model performance evaluation stage, a classification loss function integrating class weights and a normalized proportional deviation regression loss function are constructed to effectively alleviate the performance deviation problems caused by sample imbalance and life label scale differences. By weighted combination of task losses, a unified multi-task total objective function is constructed, enabling the two tasks of diagnosis and prediction to dynamically balance and jointly optimize during the training process, significantly reducing the accuracy decline and learning deviation problems caused by task competition in traditional methods.

[0072] In addition, the present invention innovatively introduces a multi-objective grey wolf optimization algorithm to post-process and adjust the model output. By taking the initial diagnosis result and the prediction result as the optimization objects and constructing a joint fitness function to achieve global search for the adjustment parameters, the problem that traditional optimization methods are difficult to take into account both classification confidence and regression trend simultaneously is solved. During the execution of the algorithm, a Pareto optimal solution set is constructed through non-dominated sorting, and the three grey wolf individuals with the best performance are dynamically selected as the core of population guidance. Further, local error gradient information is fused to construct the parameter update direction, realizing a joint adjustment mechanism based on model feedback. This mechanism avoids the risk of single-objective optimization falling into local extrema, improves the credibility of diagnostic output and the trend consistency of prediction results, and provides more stable and reliable data support for subsequent medical operation and maintenance decisions.

[0073] In the output stage, the present invention constructs an output correction mechanism for the model results based on the optimized parameter set, performs confidence temperature scaling on the classification output respectively, and introduces directional trend adjustment to the prediction output, improving the performance ability of the model results in the marginal distribution area. This correction mechanism not only enhances the interpretability of the results, but also improves the fault tolerance and generalization ability of the model during actual deployment, and is particularly suitable for application scenarios with high requirements for the operation status of high-risk equipment to be highly accurate and robust. Compared with the prior art that only relies on the forward output of the model, ontology structure optimization or simple error post-adjustment, the present invention realizes the systematic collaborative optimization of the whole process of medical device fault diagnosis and life prediction through a closed-loop structure of modeling - optimization - adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0075] Figure 1 is a flowchart of an intelligent medical device fault diagnosis and prediction method based on deep learning proposed by the present invention;

[0076] Figure 2 is a schematic diagram of the multi-task deep neural network structure adopted in an intelligent medical device fault diagnosis and prediction method based on deep learning proposed by the present invention;

[0077] Figure 3 is a structural diagram of the multi-objective grey wolf optimization algorithm of an intelligent medical device fault diagnosis and prediction method based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0079] Reference Figures 1-3 , an intelligent medical device fault diagnosis and prediction method based on deep learning, comprising the following steps:

[0080] S1. Obtain the data of the medical device during operation and construct the original medical device dataset;

[0081] S2. Preprocess the original medical device dataset and extract the input feature vectors of the medical device;

[0082] S3. Construct a multi-task deep neural network model, input the input feature vectors of the medical device, output the initial diagnosis result and the initial prediction result, and combine them to construct the initial result set;

[0083] S4. Based on the initial result set, evaluate the performance of the initial diagnosis result and the initial prediction result, calculate the first objective function and the second objective function, and construct the overall performance index of the initial result set;

[0084] S5. Based on the overall performance index of the initial result set, use the multi-objective grey wolf optimization algorithm to jointly optimize the initial diagnosis result and the initial prediction result to obtain the final optimized parameter set;

[0085] S6. Use the optimized parameter set to correct the initial diagnosis result and the initial prediction result, and output the optimized diagnosis result and the optimized prediction result.

[0086] An intelligent medical device fault diagnosis and prediction method based on deep learning provided by the present invention can realize the efficient fusion of multi-source state data of medical devices and the collaborative modeling of diagnosis and prediction tasks. By constructing a multi-task neural network model that combines one-dimensional convolution and bidirectional long short-term memory structures, it realizes the intelligent diagnosis of the current operating state of the device and the synchronous prediction of the remaining life. Combining with the multi-objective grey wolf optimization algorithm, it jointly adjusts and optimizes the accuracy of the initial diagnosis result and the prediction result, and constructs a complete model-evaluation-optimization closed loop. This method has the advantages of high diagnostic accuracy, strong prediction stability, and high optimization response efficiency, is applicable to the intelligent operation and maintenance systems of various types of medical devices, and can significantly improve the safety, forward-looking and intelligent level of device management.

[0087] In this embodiment, the original medical device dataset includes the temperature data, voltage data, current data, vibration signal data, working duration data, and operation log data of the medical device.

[0088] In the method for fault diagnosis and prediction of medical devices, an original dataset containing multi-source operation data is constructed, significantly improving the comprehensiveness of modeling of device status information and the predictability of health trends. By integrating key dimensions such as temperature, voltage, current, vibration signals, working hours, and operation logs collected during the operation of the device, a multi-modal data structure covering electrical performance, mechanical status, and historical behavior is formed, effectively enhancing the sensitivity to device status variations. All kinds of data together constitute a complete operation portrait of the medical device, which can not only support abnormal warnings before faults occur but also be used for trend identification of device life changes. This method has the advantages of wide information coverage, rich feature dimensions, and a solid modeling foundation, providing high-quality data support for the accurate training and optimization of subsequent diagnosis and prediction models, and is applicable to intelligent monitoring and operation status analysis scenarios of various types of medical devices.

[0089] In this embodiment, S2 specifically includes:

[0090] S21. Normalize the temperature data, voltage data, current data, vibration signal data, working hour data, and operation log data in the original dataset of the medical device;

[0091] S22. Perform time series reconstruction processing on the normalized multi-modal data, and use a sliding window mechanism to construct fixed-length sequences for each type of feature to form a sequence matrix;

[0092] S23. Align and fuse the sequence matrices of various features according to time steps to construct an input feature vector of the medical device.

[0093] In the process of constructing the input features of an intelligent medical device, the present invention introduces multi-modal normalization processing and a time series reconstruction mechanism, significantly improving the fusion ability of multi-source operation data and the accuracy of time series expression. By uniformly normalizing data such as temperature, voltage, current, vibration, working hours, and logs, the dimension difference is eliminated to ensure data comparability; using the sliding window method to construct fixed-length sequences for various features to capture the continuity and evolution law of device status changes; and then through time step alignment and sequence fusion, a unified input feature vector is constructed, providing high-quality input with high consistency and strong dynamics for the multi-task learning of subsequent models. This method has the advantages of reasonable structure, high data utilization rate, and strong time series expression ability, and can effectively support the high-precision modeling requirements of medical device status perception and health assessment tasks.

[0094] In this embodiment, S3 specifically includes:

[0095] S31. Construct a multi-task deep neural network model including a shared feature extraction layer, a first task output branch, and a second task output branch;

[0096] S32. Input the input feature vector of the medical device into the shared feature extraction layer to obtain the shared feature representation;

[0097] S33. Input the shared feature representation into the first task output branch to output the initial diagnosis result:

[0098]

[0099] where, Y cls is the initial diagnosis result, H is the shared feature matrix, W1 is the classification weight matrix, b1 is the classification bias term, c is the number of fault categories, and j is the fault category index;

[0100] S34. Input the shared feature representation into the second task output branch to output the initial prediction result:

[0101] Y reg = max(0, H·W2 + b2) + ∈;

[0102] where, Y reg is the initial prediction result, W2 is the regression task weight matrix, b2 is the regression task bias scalar, and ∈ is the prediction error perturbation term;

[0103] S35. Combine the initial diagnosis result Y cls and the initial prediction result Y reg to construct the initial result set

[0104] In the design process of the medical device fault diagnosis and life prediction method of the present invention, a multi-task deep neural network model with a shared representation structure is introduced, which significantly improves the model's joint modeling ability for classification and regression tasks under multi-source input data. By constructing a neural network structure including a shared feature extraction layer, a diagnosis task output branch, and a prediction task output branch, the model realizes the joint expression and synchronous learning of the state evolution characteristics of medical devices. After the input feature vector is extracted by the shared layer to obtain the key representation, it can be respectively input into the classification branch and the regression branch to generate the initial diagnosis result and the initial prediction result. The diagnosis output is based on the multi-class classification logic, and the prediction output constructs the trend expression by combining the weight bias and the error perturbation term, and finally synthesizes the initial result set. This method has the characteristics of high expression efficiency, strong task collaboration, and excellent parameter generalization ability, and can effectively improve the accuracy and stability of device anomaly recognition and life prediction.

[0105] In this embodiment, the S32 specifically includes:

[0106] S321. Input each time series sub-sample in the input feature vector of the medical device into a one-dimensional convolutional neural network module to perform local pattern extraction on the multi-modal features within each time window, and output the convolutional feature representation;

[0107] S322, input the convolutional feature representation into a bidirectional long short-term memory network module, perform forward and reverse modeling in chronological order, generate a context-related bidirectional hidden state sequence, and output a bidirectional sequence feature;

[0108] S323, performing a pooling operation on the bidirectional sequence features, compressing the time dimension using an average pooling method, and obtaining a final shared feature vector;

[0109] S324: Process all sequence samples in sequence to generate shared feature representations.

[0110] The present invention introduces a shared representation extraction mechanism based on convolution and bidirectional time series modeling in the input feature processing method of medical equipment, which significantly improves the modeling accuracy and context understanding ability of the model for multimodal time series features. By inputting each input sequence into a one-dimensional convolutional neural network module in turn, the key change pattern of the equipment operation data in the local time window is extracted; then it is input into a bidirectional long short-term memory network to learn the state evolution trend from the forward and backward time dimensions respectively, and construct a context-related time series expression. Subsequently, average pooling is used to compress the time dimension and extract a stable shared feature vector with global expression ability. Finally, all input samples are uniformly processed to generate a shared feature representation, providing high-quality semantic support for multi-task output branches. This method has the advantages of strong expression ability, scalable structure, and insensitivity to input disturbances, and effectively supports the accurate modeling of subsequent fault diagnosis and life prediction tasks.

[0111] In this embodiment, the S4 specifically includes:

[0112] S41. Manually label the device status corresponding to each input sequence in the original medical device dataset according to a predefined fault type classification standard to generate an original label sequence, vector encode the original label sequence, and use one-hot encoding to convert the label of each input sequence into a fault category indicator vector of length c to generate a true fault label set;

[0113] S42. Collect historical operation log data during the operation of the medical device, extract the activation timestamp and expiration timestamp of each medical device within the corresponding operation cycle, and calculate the total life cycle of the device. For each sequence sample generated based on the sliding window mechanism, record the current timestamp, calculate the remaining life label relative to the time of failure of the device, and aggregate the life labels of all samples to form a true life label set;

[0114] S43: Compare the initial diagnosis result with the real fault label set element by element to construct a first objective function:

[0115]

[0116] Among them, L cls is the first objective function, n is the number of input feature sequences of the medical device, i is the index of the input feature sequence of the medical device, c is the number of fault categories, j is the fault category index, w j is the weight of the j-th type of fault, is the true label value of the i-th input sequence under the j-th type of fault, is the initial diagnosis result of the i-th input sequence under the j-th type of fault;

[0117] S44. Perform sequence-level matching between the initial prediction result and the actual remaining life label to construct a performance evaluation function for the prediction task:

[0118]

[0119] Among them, L reg is the second objective function, is the initial prediction result of the i-th input sequence, is the actual remaining life value of the i-th input sequence, and δ is a stability factor;

[0120] S45. Weightedly fuse the first objective function and the second objective function to form an overall performance index for the initial result set.

[0121] In the process of medical device status monitoring and modeling of the present invention, a diagnosis and prediction dual-task label construction and joint performance evaluation mechanism is introduced, which significantly improves the model's ability to depict the real device status and the comprehensiveness of task evaluation. By manually annotating the input sequences based on predefined fault type criteria and generating fault category indicator vectors in one-hot manner, a real fault label set is established to achieve precise supervision of the classification task; a remaining life label set is constructed by combining the device activation and failure timestamps to ensure that the prediction task has a clear time reference. Subsequently, a diagnosis performance function based on weighted cross-entropy and a prediction performance function based on normalized proportional error are constructed respectively, fully considering class imbalance and prediction scale differences; finally, the two types of objective functions are weightedly fused to form a unified performance index for the dual-task. This method has the advantages of accurate label generation, comprehensive task evaluation, flexible index fusion, etc., and can effectively support the joint training and in-depth performance analysis of multi-task neural networks.

[0122] In this embodiment, the specific steps of S5 include:

[0123] S51. Initialize the gray wolf population, set the population size to N, and assign a set of parameter vectors to be optimized to each gray wolf individual. Each set of parameter vectors includes a diagnosis adjustment parameter and a prediction adjustment parameter;

[0124] S52. Based on the position vectors corresponding to each grey wolf individual, parametrically adjust the initial diagnostic result and the initial prediction result respectively to obtain the adjusted diagnostic result and the adjusted prediction result corresponding to the k-th grey wolf individual;

[0125] S53. Input the adjusted diagnostic result and the adjusted prediction result into the first objective function and the second objective function respectively, and calculate the fitness values respectively;

[0126] S54. Construct a Pareto optimal solution set using the non-dominated sorting method, and iterate the positions of the grey wolf individuals:

[0127]

[0128] where, X k (t + 1) is the parameter position of the k-th grey wolf individual in the (t + 1)-th round, X α (t), X β (t), X δ (t) are the parameter positions of the three individuals with the best performance in the current iteration respectively, γ is the search step factor, ω k is the adaptive adjustment coefficient of the k-th grey wolf individual, is the partial derivative symbol, is the classification error value of the k-th grey wolf individual, is the regression error value of the k-th grey wolf individual, is the diagnostic adjustment parameter of the k-th grey wolf individual, the prediction adjustment parameter of the k-th grey wolf individual;

[0129] S55. Repeat the operations of parametric adjustment, calculating fitness values, and iterating the positions of the grey wolf individuals until the optimal individual convergence condition is satisfied, and select the individual with the minimum combined loss from the Pareto optimal solution set to determine the final optimized parameter set.

[0130] In the process of optimizing the medical device fault diagnosis and prediction results, the present invention introduces a multi-objective grey wolf optimization algorithm, which significantly improves the collaborative adjustment ability between the classification accuracy and prediction accuracy of the model output results. By initializing a grey wolf population including diagnostic adjustment parameters and prediction adjustment parameters, constructing a combined fitness function based on the model output error as the evaluation basis, and evaluating the performance of the classification and regression tasks corresponding to each individual respectively. Combining non-dominated sorting to construct a Pareto optimal solution set, dynamically selecting the grey wolf individuals with the best classification and prediction performance to update the parameter positions iteratively, and integrating the search step, gradient direction and individual adaptive coefficient to realize the guidance and convergence of the optimization process. This method has the advantages of high search accuracy, stable optimization process and strong task adaptability, can effectively improve the overall quality of the diagnostic and prediction results, and enhance the deployability and application robustness of the medical device intelligent evaluation system.

[0131] In this embodiment, S6 specifically includes:

[0132] S61. Using the diagnostic adjustment parameters in the final optimized parameter set to perform parametric confidence reconstruction on the initial diagnostic result to obtain an optimized diagnostic result;

[0133] S62. Using the prediction adjustment parameters in the final optimized parameter set to perform multi-dimensional adjustment on the initial prediction result to obtain an optimized prediction result;

[0134] S63. Outputting the optimized diagnostic result and the optimized prediction result.

[0135] In the present invention, a dual-branch reconstruction mechanism based on optimized parameters is introduced in the diagnostic and prediction result output stage, significantly improving the interpretability and practical stability of the model output results. By using the diagnostic adjustment parameters in the final optimized parameter set to perform confidence temperature scaling on the initial diagnostic result, reasonable adjustment of the classification output probability distribution is realized, enhancing the confidence reliability of fault identification; at the same time, using the prediction adjustment parameters to perform multi-dimensional adjustment of the trend direction and error amplitude of the initial remaining useful life prediction result, correcting the prediction deviation and improving the smoothness and forward-looking of the output curve. The optimized diagnostic result and prediction result are used as the final output, which can be directly used for the comprehensive evaluation of the operating state of medical devices. This method has the advantages of simple structure, high adjustment efficiency, strong deployment flexibility, etc., and can effectively improve the output quality and available value of deep models in actual intelligent medical scenarios.

[0136] In this embodiment, S61 specifically includes:

[0137] S611. Dividing the initial diagnostic result of each fault category of each input sequence by the diagnostic adjustment parameter to obtain a scaled logits expression;

[0138] S612. Calculating the exponential mapping value corresponding to each scaled logits expression;

[0139] S613. Summing up all the exponential mapping values of the same input sample and then normalizing to obtain a confidence reconstruction result, and forming the optimized diagnostic result with all the confidence reconstruction results.

[0140] In the process of reconstructing the fault diagnosis output result, the present invention introduces a temperature scaling mechanism based on diagnostic adjustment parameters, which significantly improves the controllability of the multi-class fault recognition confidence and the stability of the classification result. By parametrically scaling the initial diagnostic results of each input sequence for each fault class, a more distinguishable logits expression is obtained; further performing an exponential mapping operation on it to highlight the differences between classes; finally, normalizing the exponential values of all classes to form a confidence output that meets the requirements of the probability distribution. This reconstruction process can effectively alleviate the confidence distortion phenomenon of the original model under problems such as uneven inter-class distribution and amplified prediction deviation, and enhance the credibility and consistency of the diagnostic results. This method has the advantages of simple adjustment method, strong adaptability, flexible deployment, etc., and can be widely applied to the post-processing and output optimization of model results in the multi-class fault recognition scenario of medical devices.

[0141] In this embodiment, the S62 specifically includes:

[0142] S621. Calculate the mean value of the set of true life labels of all samples in the current batch;

[0143] S622. Make a sign judgment on the deviation between each initial prediction result and the mean value of the set of true life labels to obtain a directional adjustment sign;

[0144] S623. Construct an adjustment term according to the prediction adjustment parameter and the directional adjustment sign, and add each initial prediction result and the corresponding adjustment term to obtain an optimized prediction result.

[0145] In the process of correcting the equipment life prediction result, the present invention introduces a dynamic adjustment mechanism based on directional error discrimination, which significantly improves the model's response ability to prediction deviation and the stability of the output result. By calculating the mean value of the true life labels of all samples in the current batch, a global reference benchmark is constructed; then, a sign judgment is made on the deviation between each initial prediction result and this mean value to determine the adjustment direction; finally, an adjustment term is constructed in combination with the prediction adjustment parameter and added to the original prediction result to achieve a prediction correction output with controllable trend direction and adaptive amplitude. This method can effectively alleviate the systematic deviation problem caused by uneven distribution of training data or accumulation of small sample errors in the life prediction task of the model, and improve the stability and clinical usability of the overall prediction result. The method has the characteristics of simple implementation, low calculation cost, and strong robustness, and is suitable for the actual needs of high-precision life prediction in intelligent medical device management.

[0146] Example 1:

[0147] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the intelligent equipment management system of a tertiary comprehensive hospital, and intelligent operation status monitoring and health prediction management were performed on its various key medical equipment such as electrocardiogram monitors, microinjection pumps, ventilators, portable defibrillators and bedside monitoring systems.

[0148] This example scenario uses 42 core medical devices in the hospital's clinical ICU and emergency care areas. Data collection runs from January 1, 2024, to March 31, 2024, at a frequency of every five minutes. This data ultimately captures approximately 750,000 records of device operation data. This data includes device operating time, usage frequency, power supply voltage, current fluctuations, surface temperature, internal vibration sensor signals, system operation logs, fault alarm codes, and maintenance records.

[0149] During the application process, first, all equipment operation data are normalized and a time series feature vector is constructed in a sliding window manner through the data acquisition and preprocessing module proposed in the present invention. Subsequently, the designed multi-task deep neural network model is used to process each feature vector sequence, and the current fault diagnosis results of the equipment and the future operation trend prediction values are output at the same time. The model uses a shared coding structure, and the front end extracts a universal representation vector, which is passed into the classification branch and the regression branch respectively to realize the classification judgment of the equipment status and the prediction of the remaining service life. During the model training process, the weighted cross entropy classification loss function and the normalized proportional regression error function constructed by the present invention are used, and the unified optimization of the diagnosis and prediction goals is achieved through the objective function fusion mechanism.

[0150] To improve the accuracy and stability of the output results, this example further introduces a multi-objective gray wolf optimization algorithm. By using the initial model output as the optimization input, a joint error evaluation metric is constructed, and a non-dominated sorting process is performed on the adjustment parameters of each individual gray wolf to obtain a Pareto optimal solution and guide the model output correction process. Ultimately, the output is a temperature-scaled diagnosis probability and a trend-adjusted lifespan prediction. The optimization process is set to a maximum of 80 iterations, a group size of 30, and an optimization convergence time of less than 4 minutes.

[0151] Results show that in this practical scenario, the proposed method achieved an accuracy rate of 96.8% in identifying medical equipment failure states, a significant improvement over the hospital's existing empirical rule-based judgment method, which relied on logs and alarms (with an accuracy rate of approximately 83.2%). Furthermore, the accuracy rate of predicting the remaining life of the equipment 24 hours before a failure reached 91.4%. In particular, for subtle, hard-to-detect faults such as mild aging of the power module and intermittent jitter in the internal circuit, the prediction curve showed significant fluctuations 18 hours in advance, triggering intelligent alarms. This effectively supported preventive maintenance decisions by maintenance personnel and avoided at least six unexpected downtime incidents.

[0152] In addition, by comparing the equipment maintenance efficiency in the ICU area before and after the deployment of the present invention, it is found that after the introduction of the system of the present invention, the equipment failure response time is shortened from an average of 3.1 hours to 28 minutes, and the early warning accuracy rate is increased by nearly 19 percentage points, greatly alleviating the pressure on ensuring the stability of the equipment in the intensive care unit. In the performance comparison experiment, the method of the present invention was also compared with the traditional single-task CNN model, LSTM model and random forest method, and the results are shown in Table 1.

[0153] Table 1 Comparison table of experimental effects of intelligent fault diagnosis and life prediction of medical equipment

[0154]

[0155]

[0156] The experimental results show that with the support of the intelligent medical equipment fault diagnosis and prediction method based on deep learning proposed by the present invention, the operation safety, maintenance response efficiency and management intelligence level of medical equipment have been significantly improved. Under the synergistic effect of the multi-task deep neural network structure and the multi-objective grey wolf optimization mechanism in the system, it can effectively fuse multi-source perception data, accurately identify potential equipment failures, and predict the trend of life decay in advance, thus overcoming the problems of limited diagnostic ability, lack of prediction mechanism and unstable model output in traditional methods.

[0157] In the actual deployment process, as an equipment with a very high application frequency in the front line of medical treatment, the accuracy of the state recognition of the electrocardiogram monitor is of great significance for clinical continuous monitoring. The comparison data shows that with the support of the method of the present invention, the fault recognition accuracy rate of the electrocardiogram monitor is increased from 90.5% of the traditional LSTM model to 96.2%, and the improvement amplitude reaches 6.3%, effectively reducing the situation of misjudgment and missed judgment. At the same time, the remaining life prediction error is significantly reduced from 3.1 hours to 1.5 hours, and the early warning time in advance is also increased from less than 8 hours to more than 17.3 hours, enhancing the ability to capture potential faults with slight fluctuations. This effect benefits from the joint expression ability of the shared coding structure in the present invention for sensitive modeling of time series features and trend analysis.

[0158] In the application of ventilator equipment, the method of the present invention exhibits better performance. Ventilators operate with a high load in the ICU scenario, and the environmental temperature and voltage fluctuate frequently. Traditional algorithms often suffer from prediction instability problems. However, the system of the present invention can extract the multi-period operation characteristics of the equipment through a sliding window mechanism, and combine grey wolf optimization to adjust the classification confidence and trend deviation, effectively improving the reliability of diagnosis. Specifically, the diagnostic accuracy rate of the ventilator is increased from 89.7% of the traditional CNN model to 97.5%, and the error is reduced by more than 50%. The life prediction error is controlled within 1.2 hours, far better than other methods, and the early warning ability is improved from 5.6 hours to 19.4 hours, showing great value in preventing sudden shutdowns.

[0159] For precision equipment such as syringe pumps, the present invention also achieves good compatibility and adaptability. The equipment itself is extremely sensitive to parameters such as temperature and current, and the prediction curve is prone to violent fluctuations due to noise interference. The multi-objective grey wolf regulation mechanism adopted by the present invention can achieve a balance between diagnosis and prediction loss, avoiding the degradation of the overall performance of the model caused by the dominance of a certain task. The comparison results show that the prediction error of the syringe pump is reduced from 5.2 hours of the traditional random forest model to 1.7 hours, and the early warning time is increased from 2.3 hours to 15.1 hours, indicating that the method of the present invention can effectively solve the problems of model stability and expressiveness in the case of small sample size and sparse fault distribution.

[0160] In the application test of the bedside monitoring system, the method of the present invention demonstrates the optimal comprehensive performance. This device belongs to a comprehensive monitoring platform with complex operation data dimensions and frequent state switches. Traditional empirical warning systems can often only trigger responses after a fault occurs. The present invention realizes the early perception of the changes in the microscopic parameters of the system through multi-source data modeling and optimization regulation. Its diagnostic accuracy rate reaches 98.1%, and the remaining life prediction error is only 1.0 hour. It can achieve an average early warning of more than 20 hours, greatly enhancing the foresight of system operation and maintenance.

[0161] In addition to the improvement in diagnostic and prediction capabilities, the present invention also shows significant advantages in system response efficiency and deployment feasibility. The model inference time is all controlled within 40 milliseconds, supporting real-time deployment; the iteration of the multi-objective grey wolf optimization process is controlled within 4 minutes, suitable for daily update or batch update scenarios; and in the practical application in the ICU area, the equipment fault response time is shortened from the original 3.1 hours to 28 minutes, and the equipment shutdown risk is significantly reduced.

[0162] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An intelligent medical device fault diagnosis and prediction method based on deep learning, characterized in that, It includes the following steps: S1. Obtain the data during the operation of the medical device and construct the original dataset of the medical device; S2. Preprocess the original dataset of the medical device and extract the input feature vectors of the medical device; S3. Construct a multi-task deep neural network model, input the input feature vectors of the medical device, output the initial diagnosis result and the initial prediction result, and combine them to construct the initial result set; S4. Based on the initial result set, evaluate the performance of the initial diagnosis result and the initial prediction result, calculate the first objective function and the second objective function, and construct the overall performance index of the initial result set; S5. Based on the overall performance index of the initial result set, use the multi-objective grey wolf optimization algorithm to jointly optimize the initial diagnosis result and the initial prediction result to obtain the final optimized parameter set; S6. Use the optimized parameter set to correct the initial diagnosis result and the initial prediction result, and output the optimized diagnosis result and the optimized prediction result.

2. The intelligent medical device fault diagnosis and prediction method based on deep learning according to claim 1, characterized in that, The original dataset of the medical device includes the temperature data, voltage data, current data, vibration signal data, working duration data, and operation log data of the medical device.

3. A method for fault diagnosis and prediction of intelligent medical devices based on deep learning according to claim 1, characterized in that, The specific content of S2 includes: S21. Normalize the temperature data, voltage data, current data, vibration signal data, working duration data, and operation log data in the original dataset of the medical device; S22. Perform time series reconstruction processing on the normalized multi-modal data, use the sliding window mechanism to construct fixed-length sequences for each type of feature, and form a sequence matrix; S23. Align and fuse the sequence matrices of various features according to time steps to construct the input feature vectors of the medical device.

4. A method for intelligent medical device fault diagnosis and prediction based on deep learning according to claim 1, characterized in that, The specific content of S3 includes: S31. Construct a multi-task deep neural network model including a shared feature extraction layer, a first task output branch, and a second task output branch; S32. Input the input feature vectors of the medical device into the shared feature extraction layer to represent the shared features; S33. Input the shared feature representation into the first task output branch to output the initial diagnosis result; Among them, Y cls is the initial diagnosis result, H is the shared feature matrix, W1 is the classification weight matrix, b1 is the classification bias term, c is the number of fault categories, and j is the fault category index; S34. Input the shared feature representation into the second task output branch to output the initial prediction result; Y reg = max(0, H·W2 + b2) + ∈; Among them, Y reg is the initial prediction result, W2 is the regression task weight matrix, b2 is the regression task bias scalar, and ∈ is the prediction error perturbation term; S35. Combine the initial diagnosis result Y cls with the initial prediction result Y reg to construct an initial result set.

5. A method for fault diagnosis and prediction of intelligent medical devices based on deep learning according to claim 4, characterized in that, The specific content of S32 includes: S321. Input each time series subsample in the input feature vectors of the medical device into a one-dimensional convolutional neural network module to extract local patterns of the multi-modal features within each time window, and output the convolutional feature representation; S322. Input the convolutional feature representation into a bidirectional long short-term memory network module to perform forward and backward modeling in chronological order, generate a context-related bidirectional hidden state sequence, and output the bidirectional sequence features; S323. Perform a pooling operation on the bidirectional sequence features, use the average pooling method to compress the time dimension, and obtain the final shared feature vector; S324. Process all sequence samples in turn to generate the shared feature representation.

6. A method for intelligent medical device fault diagnosis and prediction based on deep learning according to claim 1, characterized in that, The specific content of S4 includes: S41. Manually label the device status corresponding to each input sequence in the original medical device dataset according to a predefined fault type classification standard to generate an original label sequence, vector encode the original label sequence, and use one-hot encoding to convert the label of each input sequence into a fault category indicator vector of length c to generate a true fault label set; S42. Collect historical operation log data during the operation of the medical device, extract the activation timestamp and expiration timestamp of each medical device within the corresponding operation cycle, and calculate the total life cycle of the device. For each sequence sample generated based on the sliding window mechanism, record the current timestamp, calculate the remaining life label relative to the time of failure of the device, and aggregate the life labels of all samples to form a true life label set; S43: Compare the initial diagnosis result with the real fault label set element by element to construct a first objective function: Among them, L cls is the first objective function, n is the number of medical device input feature sequences, i is the medical device input feature sequence index, c is the number of fault categories, j is the fault category index, w j is the weight of the j-th type of fault, is the true label value of the i-th input sequence under the j-th type of fault, is the initial diagnosis result of the i-th input sequence under the j-th type of fault; S44. Perform sequence-level matching between the initial prediction results and the actual remaining lifespan labels, and construct a performance evaluation function for the prediction task: Among them, L reg is the second objective function, is the initial prediction result of the i-th input sequence, is the actual remaining life value of the i-th input sequence, and δ is the stability factor; S45. Weighted fusion of the first objective function and the second objective function to form an overall performance indicator of the initial result set.

7. A method for intelligent medical device fault diagnosis and prediction based on deep learning according to claim 1, characterized in that, The S5 specifically includes: S51, initializing a gray wolf population, setting the population size to N, and assigning a set of parameter vectors to be optimized to each gray wolf individual, where each set of parameter vectors includes a diagnostic adjustment parameter and a predictive adjustment parameter; S52, based on the position vector corresponding to each gray wolf individual, respectively perform parameter adjustment on the initial diagnosis result and the initial prediction result to obtain an adjusted diagnosis result and an adjusted prediction result corresponding to the k-th gray wolf individual; S53, inputting the adjusted diagnosis result and the adjusted prediction result into the first objective function and the second objective function respectively, and calculating the fitness values respectively; S54. Use the non-dominated sorting method to construct the Pareto optimal solution set and iterate the individual positions of the gray wolf: Among them, X k (t + 1) is the parameter position of the k-th gray wolf individual in the (t + 1)-th round, X α (t), X β (t), X δ (t) are the parameter positions of the three individuals with the best performance in the current iteration, γ is the search step factor, ω k is the adaptive adjustment coefficient of the k-th gray wolf individual, is the partial derivative symbol, is the classification error value of the k-th gray wolf individual, is the regression error value of the k-th gray wolf individual, is the diagnostic adjustment parameter of the k-th gray wolf individual, The prediction adjustment parameter of the k-th gray wolf individual; S55, repeatedly perform parameter adjustment, calculate fitness value and iterate gray wolf individual position operation until the optimal individual convergence condition is met, select the individual with the smallest joint loss from the Pareto optimal solution set, and determine the final optimization parameter set.

8. A method for fault diagnosis and prediction of intelligent medical devices based on deep learning according to claim 1, characterized in that, The S6 specifically includes: S61, using the diagnostic adjustment parameters in the final optimized parameter set, reconstructing the initial diagnostic result with parameterized confidence to obtain an optimized diagnostic result; S62, using the final optimized parameters to centrally predict the adjustment parameters, perform multi-dimensional adjustment on the initial prediction results, and obtain an optimized prediction result; S63. Output the optimized diagnosis result and the optimized prediction result.

9. A method for fault diagnosis and prediction of intelligent medical devices based on deep learning according to claim 9, characterized in that, The S61 specifically includes: S611, dividing the initial diagnosis result of each fault category of each input sequence by the diagnosis adjustment parameter to obtain a scaled logits expression; S612, calculating the exponential mapping value of each scaled logits expression corresponding to the category; S613: All exponential mapping values of the same input sample are summed up and normalized to obtain a confidence reconstruction result, and all confidence reconstruction results are combined to form an optimized diagnosis result.

10. A method for intelligent medical device fault diagnosis and prediction based on deep learning according to claim 9, characterized in that, The S62 specifically includes: S621. Calculate the mean of the true lifespan label set of all samples in the current batch; S622. Determine the sign of the deviation of each initial prediction result from the mean of the true lifespan label set to obtain a directional adjustment sign; S623. Construct an adjustment term based on the prediction adjustment parameter and the directional adjustment sign, and add each initial prediction result to the corresponding adjustment term to obtain an optimized prediction result.

Images