Medical equipment operation state monitoring and adjusting system based on big data

Abstract

The application relates to the field of equipment operation state monitoring, in particular to a medical equipment operation state monitoring and adjusting system based on big data, which comprises a multi-source acquisition module, a federal preprocessing module, an equipment multi-feature extraction module, an analysis module, an optimal adjustment strategy module and a database, the multi-source acquisition module acquires operation parameters and generates adversarial samples; the federal preprocessing module embeds a fairness constraint algorithm to output an initial balanced data set; the equipment multi-feature extraction module extracts frequency domain and space-time features and integrates the features into a time feature sequence; the analysis module enables a random greedy algorithm according to a dynamic decision, inputs a processed feature subset into a bidirectional LSTM model to output equipment states; and the optimal adjustment strategy module combines Bayesian network probability reasoning and deterministic rules of a rule engine to generate an optimal adjustment strategy, so that the data robustness and fairness are improved, accurate equipment state determination and optimal strategy generation are realized, and the safe operation of medical equipment is ensured.

Description

TECHNICAL FIELD

[0001] The application relates to the field of equipment operation state monitoring, in particular to a medical equipment operation state monitoring and adjustment system based on big data. BACKGROUND

[0002] In the field of equipment operation state monitoring, the existing medical equipment operation state monitoring system has many technical bottlenecks; in the federated learning scenario, the amount of medical equipment data is significantly different, for example, the amount of data of a monitor in a first-class hospital is much larger than that of a community hospital, and if the data is directly weighted and aggregated, it will lead to the dominance of the global model by large equipment, resulting in low abnormal detection accuracy of small equipment and difficulty in ensuring data fairness.

[0003] The traditional greedy algorithm only takes the current optimal strategy at each step in feature selection, which is easy to fall into local optimization and cannot efficiently obtain the optimal feature subset; in addition, single sensor features cannot fully reflect the equipment state, and the existing system lacks organic integration of frequency domain features and spatiotemporal features in feature extraction, making it difficult to accurately capture the equipment operation state; in terms of adjustment strategy generation, the existing technology lacks probabilistic reasoning of uncertainty or lacks deterministic rule constraints, making it difficult to generate optimal adjustment strategies that take into account reliability and adaptability; therefore, a medical equipment operation state monitoring and adjustment system based on big data is needed to solve the above problems. SUMMARY

[0004] To solve the technical problems raised in the background, the application provides a medical equipment operation state monitoring and adjustment system based on big data.

[0005] The purpose of the application can be achieved by the following technical solutions:

[0006] The application provides a medical equipment operation state monitoring and adjustment system based on big data, which includes a multi-source acquisition module, a federated preprocessing module, a device multi-feature extraction module, an analysis module, an optimal adjustment strategy module and a database.

[0007] The multi-source acquisition module obtains a set of running parameters of each medical equipment and generates an adversarial sample based on the fast gradient sign algorithm, and constructs a data pair set by adding a small perturbation, and the specific process is as follows:

[0008] A multi-source acquisition module is set in each local client to collect a set of running parameters of each medical equipment in real time, and each medical equipment includes a monitor, a ventilator, an infusion pump and a surgical robot, and the set of running parameters includes tidal volume, flow rate, vibration signal, current signal, pressure signal and force feedback; the set of running parameters is cleaned and standardized to obtain each clean sample ;

[0009] The fast gradient sign algorithm is introduced to generate the minimum perturbation to make the model misjudge through the gradient of the loss function of each clean sample, so as to obtain the corresponding adversarial sample of the clean sample The calculation logic is as follows: Wherein is the minimum perturbation strength, is the gradient of the sample to the loss function, is a sign function, is the true label of the device state determination model, is the parameter of the device state determination model, the model parameters include weights and biases, the numerical value of the adversarial sample is limited to the value range of the clean sample, and the clean sample and the corresponding adversarial sample are matched as a data pair, and the data pair is added with device identification and time stamp; the number of data pairs in the hospital scene is counted to obtain a data pair set, and each local client transmits the data pair set to the center server.

[0010] The federal preprocessing module embeds a fairness constraint algorithm in the data pair set, reversely weights according to the data volume, and outputs an initial balanced data set with the global model, and the specific process is as follows:

[0011] The federal preprocessing module is arranged in the center server, collects the data pair set of all local clients, extracts the data volume of each data pair set as , sets its reverse weight in inverse proportion to the data volume, wherein represents the number of data pair sets; the server center uses the reverse weight to update the loss function, and obtains the updated total objective function of each local client , and the specific calculation logic is as follows: Wherein is the loss of the clean sample, and Q is the number of data pair sets, is the loss of the adversarial sample; the total objective function is integrated into the model parameters by the gradient descent method to obtain the processing parameters of each data pair set, and the processing parameters and the reverse weight are weighted and summed by the global aggregation formula to obtain the global model, and the clean sample is extracted and input into the global model to output an initial balanced data set.

[0012] The device multi-feature extraction module extracts the frequency domain features and the space-time features in the initial balanced data set, and then integrates them into a time feature sequence by using a diffusion model, and the specific process is as follows:

[0013] The frequency domain features of the initial balanced data are extracted, the frequency domain features include envelope spectrum and cepstrum analysis, and the carrier and modulation envelope of each signal are separated by Hilbert transform , the phase shift of the signal is obtained to obtain the quadrature component , and the calculation logic of the envelope signal is as follows: The envelope spectrum is obtained by performing a Fourier transform on the envelope signal. The key information of the envelope spectrum features is extracted to construct the first feature vector set. : ,in The frequency of the main peak in the envelope spectrum. For envelope spectrum energy, The cepstrum is a sideband distribution; it separates the periodic components of the signal by performing an inverse FFT on the logarithmic spectrum of the signal, specifically: Fourier transform The power spectrum is The logarithm is then used to obtain the logarithmic power spectrum. Its calculation logic is as follows: The inverse Fourier transform of the logarithmic power spectrum will be performed. Obtain the cepstral Its calculation logic is as follows: ,in The cepstral frequency is used; key information from the cepstral features is extracted to construct a second feature vector set. : ,in The frequency inverted by the main peak. The amplitude of the main peak Harmonic attenuation rate;

[0014] Extracting the spatiotemporal features of the initial equilibrium data, including local clustering coefficients, degree centrality, and betweenness centrality, based on the node set. Sum of edges Constructing a sensor topology graph The node set represents the sensor set, with each node representing a sensor, and the edge set represents the relationships between the sensors; the weight of each edge is obtained. The edge weights represent the overall correlation strength between sensors, and their calculation logic is as follows: ,in and This can be represented by the numbers of any two sensors. The Pearson correlation coefficient is calculated as follows: ,in For sensors Time series data, for Time series data, Here, n is the ID of the time series data, and n is the total number of time series data. and for and The mean of the corresponding time series data; a physical distance weight, by and Euclidean distance calculation, each edge weight and sensor topology graph matching to get the weighted graph; get the weighted graph node and its neighbor set , for and directly connected nodes, and for any two node numbers; the edge weight between neighbors is , then the local clustering coefficient of node The calculation logic is: , where b is the node degree; the degree centrality measures the connection strength of node , and in the weighted graph, it is the sum of all edge weights of node , and the calculation logic is: ; Set any node pair , and set as the number of shortest paths from m1 to m2, so as to obtain the number of paths passing through node , then the betweenness centrality of node The calculation logic is:; The local clustering coefficient, degree centrality and betweenness centrality are integrated into the third vector feature set ;

[0015] The first vector feature set, the second vector feature set and the third vector feature set are input into the diffusion model, and the diffusion model adds Gaussian noise to each feature set to generate noisy feature points; the noisy feature points are obtained by the reverse denoising process and the gradient descent algorithm to obtain single time point features; in this way, subsequent single time point features are obtained through the diffusion model and integrated into a time feature sequence.

[0016] The analysis module dynamically decides to enable the random greedy algorithm according to the feature complexity, device risk level and historical fluctuations, and the random greedy selects suboptimal features with probability to avoid local optimality; the processed feature subset is input into the device state determination model to capture time dependence, and the medical device running state is output, and the specific process is as follows:

[0017] Determine whether to enable the random greedy algorithm by feature complexity, device risk level and historical fluctuations, extract any two feature vectors in time sequence from the time feature sequence, set them as and , and are feature vector numbers, and the calculation logic of feature complexity is:​​ wherein is the total number of features, is the feature and is the Pearson correlation coefficient of the feature , and is the information entropy of the feature ; the risk level RT is divided into three levels according to the degree of influence of the device on the safety of the patient: the surgical robot and the ventilator, which can directly affect the life safety of the patient or the life support device, are set to the third risk level, the monitor and the infusion pump, which can continuously monitor but are not life support devices, are set to the second risk level, and the ordinary device with auxiliary function is set to the first risk level, the third risk level > the second risk level > the first risk level; the standard deviation of the latest ten determination probabilities is taken as the historical fluctuation value RE; the random greedy algorithm enables the rules as follows: rule 1: if CJ> Cj, RT> RT, and RE> RE, then the strong exploration random greed is enabled, and the exploration probability is ; rule 2: if only one of them is greater than the corresponding threshold value, then the moderate exploration random greed is enabled, and the exploration probability is ; rule 3: if RE> RE, then the weak exploration random greed is enabled, and the exploration probability is ; rule 4: other conditions enable the standard greed, wherein , and are preset threshold values in the database; the feature value is set as

[0018] the ratio of its contribution to the classification performance to the cost, and the calculation logic is as follows: wherein , and are balance parameters, is the F1 score of the current selected feature subset ST after adding the feature, which is estimated by cross-validation, is the total calculation time of the feature, from obtaining the feature from the sensor to the total processing time; is the average correlation coefficient of the feature with the existing features in the feature subset ST; the current remaining time feature sequence is set as , the value of each feature is calculated and arranged in descending order to obtain the candidate list , wherein is the optimal, is the suboptimal, and the selection probability is ; the selected feature is added to ST, and is removed from to obtain the processed feature subset ; ​​

[0019] The feature subset is processed and input into the device state determination model, which outputs the anomaly probability. The device state determination model architecture is a bidirectional LSTM, and the input is set to... k2 represents the number of selected features, and the forward and backward hidden states of the bidirectional LSTM are respectively... and The total hidden state is: , Let be the hidden layer dimension of the unidirectional LSTM; recursively obtain the hidden state at each time step, and take the hidden state at the last time step. The data is transmitted to a state classifier, which outputs the state category for each medical device. The status categories include normal, sub-healthy, faulty, and emergency fault.

[0020] The optimal adjustment strategy module combines the probabilistic inference of fault causes from the Bayesian network with the deterministic rules of the rule engine. It calculates the expected reward of actions using posterior probability, resolves conflicts according to rule priority, and generates an optimal adjustment strategy that balances uncertainty and constraints. The specific process is as follows:

[0021] Bayesian networks classify states Time feature series Cause of the malfunction and adjustment strategies Nodes in a directed acyclic graph are used to represent probabilistic dependencies between variables; the causes of failures are calculated using Bayes' theorem. posterior probability Its calculation logic is as follows: ,in Let be the likelihood probability. For prior probability, Evidence probability; the probability of the cause of failure and the payoff value of the action. Get each action Expected returns Its calculation logic is as follows: ,in To adjust the action numbers, To adjust the total number of actions, To determine the execution cost, the action that maximizes the expected return is chosen as the optimal probabilistic strategy. The rule engine maintains its strategy through IF-THEN rule constraints, defined as follows: ; This represents the rule priority; the larger the value, the higher the priority. The execution order of the rules based on priority is as follows: =90 100 is the state enforcement rule, and the next is... =70 89 is a safety-related rule, and so on, =60 69 is a cause-triggered rule, and finally 40 59 is a regular maintenance rule; when multiple rules are triggered, the rule with the highest priority is selected to generate a rule strategy optimal adjustment strategy The high expected return of Bayesian inference and the high priority constraint of the rule engine need to be met at the same time, and the strategy generation logic is: wherein is the rule effective threshold, the malfunctioning medical device is located through the device identifier, and the location and corresponding optimal adjustment strategy are sent to the local client at the location for prompting.

[0022] Compared with the prior art, the beneficial effects of the present application are: the analysis module dynamically enables the random greedy algorithm to avoid feature selection from falling into local optimization, and combines bidirectional LSTM to capture time dependence, so as to realize accurate grading determination of the device state and significantly enhance the early warning capability of potential early faults;

[0023] The optimal adjustment strategy module combines Bayesian network probability inference and rule engine deterministic constraint, which not only quantifies the uncertainty of fault risk, but also guarantees the enforcement of key safety rules, and generates a maintenance strategy that takes into account reliability and adaptability; the system realizes rapid fault positioning through the device identifier and timestamp mechanism, and cooperates with the optimal strategy to realize real-time pushing, which greatly improves the efficiency of medical device maintenance, shortens the fault processing time, and reduces the medical risk; the cross-modal feature fusion and dynamic adaptive decision mechanism make the system have stronger environmental adaptability and explainability, and provide comprehensive protection for the safe operation of medical devices. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description, the following drawings are not deliberately drawn according to the actual size, and the emphasis is on showing the main idea of the present application.

[0025] Figure 1 It is the principle diagram of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings, and obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor also belong to the scope of protection of the present application.

[0027] Please refer to Figure 1 The application provides a medical equipment operation state monitoring and adjusting system based on big data, which comprises a multi-source acquisition module, a federal preprocessing module, an equipment multi-feature extraction module, an analysis module, an optimal adjustment strategy module and a database.

[0028] The multi-source acquisition module acquires a set of operation parameters of each medical equipment, and generates an adversarial sample based on a fast gradient sign algorithm, and constructs a "clean-adversarial" data pair set by adding a slight disturbance, thereby providing a robust training sample for federal adversarial training, and the specific process is as follows:

[0029] A multi-source acquisition module is arranged in each local client to acquire a set of operation parameters of each medical equipment in real time, the medical equipment comprises a monitor, a ventilator, an infusion pump and a surgical robot, and the set of operation parameters comprises a tidal volume, a flow rate, a vibration signal, a current signal, a pressure signal and a force feedback; the set of operation parameters is cleaned and standardized to obtain each clean sample ;

[0030] The fast gradient sign algorithm is introduced to generate a minimum disturbance to make the model misjudge through the gradient of the loss function of each clean sample, so as to obtain an adversarial sample corresponding to the clean sample , and the calculation logic is as follows: , wherein is the minimum disturbance intensity, which determines the difference between the adversarial sample and the original data, is the gradient of the sample to the loss function, is a sign function, is a true label of the equipment state determination model, is a parameter of the equipment state determination model, the model parameter comprises a weight and a bias, and the true label and the parameter are both obtained by the local client from the center server in real time; the value of the adversarial sample is limited in the value range of the clean sample to avoid generating unrealistic parameter values, the clean sample and the corresponding adversarial sample are matched into a data pair, and the data pair is added with an equipment identifier and a time stamp; it should be noted that one local client is arranged in one hospital scene, one medical equipment in the scene corresponds to one data pair, and the equipment identifier and the time stamp are added to facilitate the tracing of the medical equipment when data anomaly occurs; the number of data pairs in the hospital scene is counted to obtain a data pair set, and each local client sends the data pair set to the center server.

[0031] The federal preprocessing module embeds a fairness constraint algorithm in the data pair set, reversely weights according to the data volume, and outputs an initial balanced data set according to the global model, and the specific process is as follows:

[0032] In federated learning, the data volume of medical devices varies significantly, such as the data volume of a monitor in a third-level hospital is several times that of a community hospital. If the data volume is directly weighted and aggregated, it will lead to the dominance of the global model by large devices, and the anomaly detection accuracy of small devices is low. To solve this problem, a fairness constraint algorithm needs to be embedded;

[0033] A federated preprocessing module is arranged in the center server, collects all data pair sets of local clients, extracts the data volume of each data pair set as , sets its reverse weight inversely proportional to the data volume, the smaller the data volume, the higher the corresponding weight, wherein represents the number of data pair sets; the server center uses the reverse weight to update the loss function, and obtains the updated total target function of each local client , and the specific calculation logic is: , wherein is the loss of clean samples, Q is the number of data pair sets, is the loss of adversarial samples; the total target function is integrated into the model parameters by the gradient descent method to obtain the processing parameters of each data pair set, and the processing parameters and the reverse weight are weighted and summed by the global aggregation formula to obtain the global model. The clean sample is extracted and input into the global model to output the initial balanced data set. It should be noted that the finally generated global model can learn the rich features of large devices and focus on strengthening the scarce features of small devices, and because of the adversarial data and fairness constraint algorithm, the robustness and fairness of clean samples are significantly improved.

[0034] The device multi-feature extraction module extracts the frequency domain features and the space-time features in the initial balanced data set, and then integrates them into a time feature sequence using a diffusion model. The specific process is as follows:

[0035] The frequency domain features of the initial balanced data are extracted, which include envelope spectrum and cepstrum analysis. The Hilbert transform is used to separate the carrier and modulation envelope in each signal , and the phase shift of the signal is used to obtain the quadrature component , and the envelope signal is extracted. The calculation logic is: , the envelope spectrum is obtained by Fourier transform of the envelope signal , and the key information of the envelope spectrum feature is extracted to construct the first feature vector set : , wherein is the main peak frequency of the envelope spectrum, is the envelope spectrum energy, The cepstrum is a sideband distribution; by performing an inverse FFT on the logarithmic spectrum of the signal, periodic components in the signal are separated, such as harmonics of gear meshing frequencies and periodicity of bearing impacts. Specifically, it involves: Fourier transform The power spectrum is The logarithm is then used to obtain the logarithmic power spectrum. Its calculation logic is as follows: The inverse Fourier transform of the logarithmic power spectrum will be performed. Obtain the cepstral Its calculation logic is as follows: ,in The cepstral frequency is used; key information from the cepstral features is extracted to construct a second feature vector set. : ,in The frequency inverted by the main peak. The amplitude of the main peak This refers to the harmonic attenuation rate; specifically, the amplitude of the main peak of the envelope spectrum energy reflects the severity of the fault; the larger the amplitude, the more severe the fault. The sideband distribution is smooth for healthy equipment, while it appears more pronounced during faults. The sideband centered on the signal has a main peak frequency in the envelope spectrum that corresponds to the fault characteristic frequency; the reciprocal frequency represents the period of the periodic component in the original signal; the reciprocal frequency of the main peak corresponds to the periodic component in the signal, and the amplitude of the main peak reflects the intensity of the periodic component. The larger the amplitude, the more obvious the periodicity; the harmonic attenuation rate is that the cepstrum harmonic amplitude of medical and health equipment attenuates rapidly with increasing order, and the attenuation slows down during faults.

[0036] In health monitoring of medical devices, the characteristics of a single sensor often cannot fully reflect the device's status. By constructing a sensor topology map and extracting its topological features, spatial correlations and coordinated signal changes between sensors can be captured, thereby locating abnormal clustering areas or key sensor nodes. The spatiotemporal features of the initial equilibrium data can be extracted, including local clustering coefficients, degree centrality, and betweenness centrality. Sum of edges Constructing a sensor topology graph The node set represents the sensor set, with each node representing a sensor, and the edge set represents the relationships between the sensors; the weight of each edge is obtained. The edge weights represent the overall correlation strength between sensors, and their calculation logic is as follows: ,in and This can be represented by the numbers of any two sensors. The Pearson correlation coefficient is calculated as follows: ,in For sensors Time series data, is the time series data, is the number of time series data, n is the total number of time series data, is the time series data, and is the time series data, is the time series data, is the mean of corresponding time series data, which measures the linear correlation degree of two sensor time series data, the value range is , the larger the absolute value, the stronger the correlation; is the physical distance weight, the closer the physical distance of the sensor, the more likely it is to produce correlation due to the same physical process, such as device heating and vibration propagation, is the Euclidean distance calculation of and , and the weighted graph is obtained by matching each edge weight and the sensor topology graph; the node and its neighbor set are obtained, is the node directly connected to , and are the numbers of any two nodes, because the node set is the sensor set, so the numbers are the same; the edge weight between neighbors is , then the local clustering coefficient of node is The calculation logic is: , where b is the node degree; the clustering coefficient reflects the average strength of the edge weight between neighbor nodes, the larger the value, the more closely connected the neighbors of the node are; the degree centrality measures the connection strength of node , which is the sum of all edge weights of node in the weighted graph, and the calculation logic is: ; the betweenness centrality measures the importance of the node as a bridge, that is, the proportion of all shortest paths passing through the node, set any node pair , and set as the number of shortest paths from m1 to m2, so as to obtain the number of paths passing through node ; , then the calculation logic of the betweenness centrality of node is: ; the local clustering coefficient, degree centrality and betweenness centrality are integrated into the third vector feature set ;

[0037] The first, second, and third vector feature sets are input into the diffusion model. The diffusion model adds Gaussian noise to each feature set to generate noisy feature points. The noisy feature points are then processed through a reverse denoising process and a gradient descent algorithm to obtain single-time point features. This process is repeated to obtain subsequent single-time point features through the diffusion model and integrate them into a time feature sequence.

[0038] The analysis module dynamically decides whether to use a randomized greedy algorithm based on feature complexity, equipment risk level, and historical fluctuations. The randomized greedy algorithm selects suboptimal features with probability to avoid local optima. The processed feature subset is input into the equipment status determination model to capture time dependencies and outputs the medical equipment operating status. The specific process is as follows:

[0039] The decision to activate the randomized greedy algorithm is dynamically determined based on feature complexity, equipment risk level, and historical fluctuations to balance algorithm efficiency and accuracy. The feature complexity (CJ) measures the redundancy and distribution complexity between features. Any two feature vectors are extracted sequentially from the time feature sequence and set as... and , and The feature vectors are numbered, and the feature complexity is calculated logically as follows: ,in The total characteristic number, Features and The Pearson correlation coefficient, Features Information entropy; risk levels RT are determined by classifying equipment into three levels based on its impact on patient safety: surgical robots and ventilators, which directly affect patient safety or provide life support, are classified as level 3 risk; monitors and infusion pumps, which provide continuous monitoring but are not life support, are classified as level 2 risk; and ordinary assistive devices are classified as level 1 risk. Level 3 risk > Level 2 risk > Level 1 risk. The standard deviation of the probability of the ten most recent judgments is used as the historical fluctuation value RE. The rule for enabling the random greedy algorithm is: Rule 1: If CJ > And RT> And RE> Then, a strong exploration random greedy algorithm is used, with an exploration probability of 1. Rule 2: If any one of the criteria exceeds the corresponding threshold, a moderately exploratory random greedy algorithm is employed, with an exploration probability of 1 / 2. Rule 3: If RE > Then a weak exploration random greedy algorithm is used, with an exploration probability of . Rule 4: In other cases, use the standard greedy algorithm, with an exploration probability of 0. , and preset threshold value in the database;

[0040] It should be noted that the traditional greedy algorithm selects the optimal strategy at each step, expecting to derive the global optimal solution from the local optimal solution. Each decision is unique and based on the current optimal standard. However, the core of the random greedy algorithm is to select a feature with a probability of selecting a suboptimal feature, with a probability of selecting an optimal feature to avoid falling into a local optimum. Set the feature value as the ratio of its contribution to the classification performance to the cost, and the calculation logic is: where and are balance parameters, is the F1 score of the current selected feature subset ST after adding the feature, which is estimated by cross-validation, is the total calculation time of the feature, from obtaining the feature from the sensor to the total processing time; is the feature and the average correlation coefficient of the existing features in the feature subset ST. It should be noted that at each step of the random greedy feature selection, the algorithm selects a feature from the time feature sequence to add to the selected subset. Therefore, the selected subset is a dynamically updated set, and its meaning changes with the iteration step. Set the remaining time feature sequence as Calculate the value of each feature and arrange it in descending order to get the candidate list where is the optimal, is the suboptimal, and the selection probability is . Add the selected feature to ST and remove it from to get the processing feature subset ;

[0041] Input the processing feature subset into the device state determination model to capture the time dependence and output the anomaly probability. The device state determination model architecture is a bidirectional LSTM, with the input set as , k2 is the number of selected features, and the forward and backward hidden states of the bidirectional LSTM are and , respectively. The total hidden state is: , is the hidden layer dimension of the unidirectional LSTM. Recursively obtain the hidden state of each time step, and take the last time step hidden state to the state classifier. The state classifier outputs each medical device state category , including normal, sub-health, failure, and emergency failure.

[0042] The optimal adjustment strategy module combines the probabilistic inference of fault causes from the Bayesian network with the deterministic rules of the rule engine. It calculates the expected reward of actions using posterior probability, resolves conflicts according to rule priority, and generates an optimal adjustment strategy that balances uncertainty and constraints. The specific process is as follows:

[0043] Bayesian networks classify states Time feature series Cause of the malfunction and adjustment strategies Nodes in a directed acyclic graph are used to represent probabilistic dependencies between variables; the causes of failures are calculated using Bayes' theorem. posterior probability Its calculation logic is as follows: ,in Let be the likelihood probability, the probability of a time-series feature given the cause of failure and the state of the equipment. This represents the prior probability, the probability of a given device state corresponding to a specific cause of failure. The evidence probability is normalized by a factor to ensure the sum of probabilities equals 1; this is achieved by combining the probability of the cause of the failure with the payoff value of the action. Get each action Expected returns If the benefit of replacing the bearing from bearing wear is 100, and the benefit from insufficient lubrication is 20, the calculation logic is as follows: ,in To adjust the action numbers, To adjust the total number of actions, To determine the execution cost, the action that maximizes the expected return is chosen as the optimal probabilistic strategy. The rule engine maintains its strategy through IF-THEN rule constraints, defined as follows: ; This represents the rule priority; the larger the value, the higher the priority. The execution order of the rules based on priority is as follows: =90 100 is triggered directly by the device status, with the highest priority being the status-enforced rule. For example, in an emergency fault, the device will immediately shut down. The next highest priority is... =70 89. Triggered by safety-related rules, such as the requirement to disconnect power for high-voltage equipment maintenance, and so on. =60 Rule 69 is the trigger rule; for example, if the bearing is worn, replace the bearing. =40 Rule 59 represents routine maintenance rules, such as periodic lubrication; when multiple rules are triggered, the rule with the highest priority is selected to generate the rule strategy. Optimal adjustment strategy The high expected reward of Bayesian inference and the high priority constraint of the rule engine need to be satisfied simultaneously, and the strategy generation logic is: wherein is a rule effective threshold, the malfunctioning medical device is located through the device identification, and the location and the corresponding optimal adjustment strategy are sent to a local client at the location for prompting.

[0044] The above is a description of the present application and should not be considered as a limitation. Although several exemplary embodiments of the present application are described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the present application. Accordingly, all such modifications are intended to be included within the scope of the present application as defined in the claims. It should be understood that the above is a description of the present application and should not be considered as a limitation. Modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present application is defined by the claims and their equivalents.

Claims

1. A medical equipment operation state monitoring and adjustment system based on big data, comprising a multi-source acquisition module, a federal preprocessing module, an equipment multi-feature extraction module, an analysis module, an optimal adjustment strategy module and a database, characterized in that: the multi-source acquisition module acquires a set of operation parameters of each medical equipment, and generates an adversarial sample based on a fast gradient sign algorithm, and constructs a data pair set by adding a small perturbation; the federal preprocessing module embeds a fairness constraint algorithm in the data pair set, reversely weights according to the data volume, and outputs an initial balanced data set with a global model; the equipment multi-feature extraction module extracts frequency domain features and spatio-temporal features in the initial balanced data set, and integrates them into a time feature sequence using a diffusion model; the analysis module dynamically decides to enable a random greedy algorithm according to feature complexity, equipment risk level and historical fluctuations, and the random greedy algorithm selects suboptimal features with a probability to avoid local optimality; the processed feature subset is input into a device state judgment model to capture time dependence, and the medical equipment operation state is output; the optimal adjustment strategy module combines the fault reason inferred by the Bayesian network with the deterministic rules of the rule engine, calculates the expected action reward based on the posterior probability, solves the conflict according to the rule priority, and generates the optimal adjustment strategy. The specific process of the analysis module dynamically deciding to enable the random greedy algorithm, and the random greedy algorithm selecting suboptimal features with a probability to avoid local optimality is as follows:

2. The big data based medical device operational status monitoring and conditioning system as claimed in claim 1, wherein, calculate the ratio of the contribution of the classification performance to the cost to obtain the feature value; obtain the current remaining time feature sequence, calculate the feature value of each feature, and arrange them in descending order to obtain a candidate list, select a probability, add the selected feature to the selected feature subset, and remove it from the remaining time feature sequence to obtain the processed feature subset. Obtaining the feature complexity CJ, the device risk level RT and the historical fluctuation RE in sequence, dynamically determining whether to enable the random greedy algorithm, dividing into three levels according to the influence degree of the device on the safety of the patient to obtain the risk level; taking the standard deviation of the latest ten times of the latest determination probability as the historical fluctuation value; the random greedy algorithm enabling rules are as follows: rule one: if CJ> threshold value, then enabling the strong exploration random greedy, and the exploration probability is =0.3; rule two: if only one item is greater than the corresponding threshold value, then enabling the moderate exploration random greedy, and the exploration probability is =0.2; rule three: if CJ< threshold value, RT< threshold value and RE< threshold value, then enabling the weak exploration random greedy, and the exploration probability is =0.1; rule four: in other cases, enabling the standard greedy, and the exploration probability is 0, wherein,, and are preset threshold values in the database. ​​​​​​​​​​ The analysis module inputs the processed feature subset into the device state judgment model and outputs an abnormal probability, specifically: input the processed feature subset as input, obtain the total hidden state through the forward and backward hidden states of the bidirectional LSTM; recursively obtain the hidden state of each time step, take the last time step hidden state to the state classifier, and the state classifier outputs the state category of each medical equipment, including normal, sub-health, fault and emergency fault.

3. The big data based medical device operational status monitoring and conditioning system as claimed in claim 2, wherein, The specific process of the optimal adjustment strategy module combining the fault reason inferred by the Bayesian network with the deterministic rules of the rule engine, calculating the expected action reward based on the posterior probability, and generating the optimal adjustment strategy by solving the conflict according to the rule priority is as follows: The specific process of the equipment multi-feature extraction module extracting the frequency domain features and spatio-temporal features in the initial balanced data set is as follows:

4. The big data based medical equipment operational status monitoring and conditioning system as claimed in claim 1, wherein, build a sensing topology graph from a node set and an edge set; The Bayesian network sets state category, time feature sequence, fault cause and adjustment strategy as nodes of a directed acyclic graph; the posterior probability of the fault cause is calculated through the Bayesian theorem; the expected return of each action is obtained through the probability of the fault cause and the return value of the action; the action with the maximum expected return is taken as the optimal probability strategy The rule engine maintains the strategy through IF-THEN rule constraints; the rule priority is set as , and the execution order of the rule priority is as follows: =90~100 is a state mandatory rule, followed by =70~89 is a safety-related rule, and so on, =60~69 is a cause-triggered rule, and finally =40~59 is a regular maintenance rule; when multiple rules are triggered, the rule with the highest priority is selected to generate a rule strategy The optimal adjustment strategy needs to meet the high expected return of the Bayesian inference and the high priority constraint of the rule engine at the same time, the fault medical device is located through the device identifier, and the location and the corresponding optimal adjustment strategy are sent to the local client at the location for prompting.

5. The big data based medical equipment operational status monitoring and regulating system as claimed in claim 1, wherein, obtain the weight of each edge through the Pearson correlation coefficient and the physical distance weight, match the edge weight and the sensing topology graph to obtain a weighted graph, obtain the nodes and their neighbor sets of the weighted graph, and extract the spatio-temporal features of the initial balanced data, including the local clustering coefficient, the degree centrality and the betweenness centrality; integrate the local clustering coefficient, the degree centrality and the betweenness centrality into a third vector feature set. The frequency domain features of the initial equalization data are extracted, the frequency domain features include envelope spectrum and cepstrum analysis, the carrier and modulation envelope in each signal are separated through Hilbert transform, the phase offset of the signal is obtained to obtain orthogonal components, the envelope signal is extracted, the envelope spectrum is obtained through Fourier transform, the key information of the envelope spectrum features is extracted to construct a first feature vector set, the first feature vector set includes envelope spectrum main peak frequency, envelope spectrum energy and sideband distribution; the cepstrum is obtained by inverse Fourier transform on the logarithmic spectrum of the signal, the periodic component in the signal is separated, specifically: the power spectrum is calculated after Fourier transform of the signal, the logarithmic power spectrum is obtained after taking logarithm, the logarithmic power spectrum is inverse Fourier transformed to obtain the cepstrum; the key information of the cepstrum features is extracted to construct a second feature vector set, the second feature vector set includes main peak inverse frequency, main peak amplitude and harmonic attenuation rate; The specific process of the equipment multi-feature extraction module integrating the time feature sequence using the diffusion model is as follows: ​ 6. The big data based medical device operational status monitoring and conditioning system as claimed in claim 5, wherein, ​ The first vector feature set, the second vector feature set and the third vector feature set are input to a diffusion model, the diffusion model adds Gaussian noise to each feature set to generate noisy feature points; the noisy feature points obtain single time point features through a reverse denoising process and a gradient descent algorithm; in this way, subsequent single time point features are obtained through the diffusion model and integrated into a time feature sequence.

7. The big data based medical equipment operational status monitoring and regulating system as claimed in claim 1, wherein, The multi-source acquisition module obtains a set of running parameters of each medical device, and generates an adversarial sample based on a fast gradient sign algorithm, and constructs a data pair set by adding a small perturbation, and the specific process is as follows: A multi-source acquisition module is arranged in each local client, and a set of running parameters of each medical device is collected in real time, the medical devices include a monitor, a ventilator, an infusion pump and a surgical robot, and the set of running parameters includes a tidal volume, a flow rate, a vibration signal, a current signal, a pressure signal and a force feedback; the set of running parameters is cleaned and standardized to obtain each clean sample; The fast gradient sign algorithm is introduced, and the minimum perturbation is generated by the gradient of the loss function of each clean sample to make the model misjudge, so as to obtain the adversarial sample corresponding to the clean sample, the value of the adversarial sample is limited in the value range of the clean sample, and then the clean sample and the corresponding adversarial sample are matched into a data pair, and the device identifier and the time stamp are added to the data pair. The number of data pairs in the hospital scene is counted to obtain a data pair set, and each local client sends the data pair set to the center server.

8. The big data based medical device operational status monitoring and conditioning system as claimed in claim 7, wherein, The federal preprocessing module embeds a fairness constraint algorithm in the data pair set, reversely weights according to the data volume, and outputs an initial balanced data set by using a global model, and the specific process is as follows: The federal preprocessing module is arranged in the center server, collects the data pair set of all local clients, extracts the data volume of each data pair set, and sets the reverse weight to be inversely proportional to the data volume; the server center uses the reverse weight to update the loss function, and obtains the updated total objective function of each local client; the total objective function is integrated into the model parameter by using the gradient descent method to obtain the processing parameter of each data pair set, the processing parameter and the reverse weight are weighted and summed by using a global aggregation formula to obtain a global model, and the clean sample is input into the global model to output an initial balanced data set.

Images