Machine learning based smart shelter equipment operation state monitoring method and system

Abstract

The application provides a kind of wisdom shelter equipment operating state monitoring method and system based on machine learning, first, obtain the synchronous operation real-time signal set containing each operating component real-time parameter, interaction between equipment and comprehensive influence of shelter environment, then implement mutual feedback correlation processing to extract mutual feedback correlation feature set, input the set into operating state deduction model to generate short-term operating state evolution sequence, identify potential abnormal evolution trend and determine abnormal correlation influence range based on this, generate collaborative control demand information, then input collaborative control demand information and mutual feedback correlation feature set into collaborative control model to generate multi-device collaborative operation control instruction, collect execution feedback signal after sending multi-device collaborative operation control instruction, update feature extraction rule and adjust model parameters, realize comprehensive, dynamic and accurate monitoring and control of shelter equipment operating state, improve the stability and reliability of equipment operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a smart shelter equipment operation state monitoring method and system based on machine learning. BACKGROUND

[0002] In the field of operation management of shelter equipment, traditional monitoring methods mainly focus on independent monitoring of single equipment components. For example, for power equipment in the shelter, only basic parameters such as voltage and current are usually monitored; for ventilation equipment, only indicators such as wind speed and air volume are concerned. The above isolated monitoring method cannot comprehensively consider the mutual influence between components during equipment operation.

[0003] In actual operation scenarios, various equipment in the shelter does not exist in isolation, but works in coordination with each other. The interaction signals between equipment, such as the influence of power equipment on the operating power of ventilation equipment, and the cross-influence of comprehensive factors of shelter environment (such as temperature, humidity, air pressure, etc.) on equipment operation state, have not been effectively integrated and analyzed. This leads to difficulties in accurately determining the root cause and influence range of the abnormality when the equipment is abnormal, and it is difficult to take effective coordinated control measures in time, thereby increasing the risk of equipment failure and affecting the overall operation stability and reliability of the shelter. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a smart shelter equipment operation state monitoring method based on machine learning, which comprises:

[0005] Obtaining a set of synchronous operation real-time signals, the set of synchronous operation real-time signals comprising real-time parameter signals of each operating component, interaction signals between equipment, and comprehensive influence signals of the shelter environment;

[0006] Performing mutual feedback correlation processing on the set of synchronous operation real-time signals, extracting a set of mutual feedback correlation features in the operating signals, the set of mutual feedback correlation features comprising component parameter coupling features, inter-equipment signal conduction features, and environmental parameter cross features;

[0007] Inputting the set of mutual feedback correlation features into an operation state deduction model to generate a short-term operation state evolution sequence through time sequence feature progressive operation, the short-term operation state evolution sequence comprising predicted values of component parameters at each time period and predicted values of overall operation coordination degree;

[0008] Identifying potential abnormal evolution trends based on the short-term operation state evolution sequence, determining abnormal correlation influence range in combination with the set of mutual feedback correlation features, and generating coordinated control requirement information;

[0009] The cooperative regulation demand information and the mutual feedback correlation feature set are input into a cooperative regulation model to generate a multi-device cooperative operation regulation instruction, the multi-device cooperative operation regulation instruction is sent to a corresponding control module, a synchronous operation feedback signal set after execution is collected, the synchronous operation feedback signal set is input into a mutual feedback correlation processing link, the extraction rule of the mutual feedback correlation feature set is updated, and the timing operation parameter of the operation state deduction model and the instruction generation logic of the cooperative regulation model are adjusted.

[0010] In another aspect, the embodiment of the present application also provides a wisdom shelter equipment operation state monitoring system based on machine learning, characterized by comprising:

[0011] A processor, a machine readable storage medium for storing machine executable instructions of the processor, wherein the processor is configured to execute the machine executable instructions to perform the above-mentioned wisdom shelter equipment operation state monitoring method based on machine learning.

[0012] Based on the above aspects, by acquiring a synchronous operation real-time signal set containing real-time parameter signals of each operating component, interaction signals between devices and comprehensive influence signals of the shelter environment, mutual feedback correlation processing is performed on the synchronous operation real-time signal set, component parameter coupling features, signal conduction features between devices and environmental parameter cross features are extracted, the internal relationship between each element in the device operation process is deeply mined, the mutual feedback correlation feature set is input into an operation state deduction model to generate a short-term operation state evolution sequence, the operation component parameters and the overall operation coordination degree of each period can be predicted in advance, the potential abnormal evolution trend is identified based on the short-term operation state evolution sequence, the abnormal correlation influence range is determined in combination with the mutual feedback correlation feature set, cooperative regulation demand information is generated, the regulation measures are more targeted and effective. Finally, the multi-device cooperative operation regulation instruction is generated through the cooperative regulation model, and the feature extraction rule is updated and the model parameter is adjusted according to the feedback signal after execution, realizing dynamic and accurate monitoring and regulation of the device operation state, greatly improving the stability, reliability and intelligent level of the shelter equipment operation. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is the execution flowchart of the wisdom shelter equipment operation state monitoring method based on machine learning provided by the embodiment of the present application. DETAILED DESCRIPTION

[0014] The present application will be specifically described below in combination with the drawings of the specification, Figure 1 is the flowchart of the wisdom shelter equipment operation state monitoring method based on machine learning provided by an embodiment of the present application, and the wisdom shelter equipment operation state monitoring method based on machine learning will be described in detail below.

[0015] Step S110: Obtain a synchronous running real-time signal set, which contains real-time parameter signals of each running component, inter-device interaction signals, and comprehensive influence signals of the shelter environment.

[0016] In this embodiment, the application scenario is the operation state monitoring of a smart medical shelter device. The smart medical shelter contains multiple key devices, such as a medical diagnostic device group, a life support device group, an environmental control device group, and the like. Each device is composed of multiple running components. For example, a CT scanner in the medical diagnostic device group contains a rotating component, a radiation generating component, a data acquisition component, and the like.

[0017] To obtain the synchronous running real-time signal set, multiple types of sensor networks need to be deployed inside the shelter. For the real-time parameter signals of each running component, a vibration sensor is installed on the rotating component of the CT scanner to collect vibration frequency, amplitude, and other parameters. A temperature sensor is installed on the radiation generating component to collect working temperature parameters. A data transmission rate sensor is installed on the data acquisition component to collect data transmission speed parameters. The above-mentioned sensors continuously collect data at a preset sampling frequency. The sampling frequency is set according to the characteristics of different components. For example, the sampling frequency of the vibration sensor is higher than that of the temperature sensor, so as to ensure that the rapidly changing vibration signals can be captured.

[0018] The inter-device interaction signals are collected through the communication interface between devices. The devices in the shelter exchange data through a dedicated medical device communication protocol. For example, the CT scanner and the image storage server exchange scan image data and control instructions. The life support device and the central monitoring system exchange patient vital sign data and device running state instructions. By deploying a signal collection module on the communication bus, these interaction signals, including the sending time, receiving time, data length, and check code, are captured in real time.

[0019] The comprehensive influence signals of the shelter environment are collected by environmental sensors deployed in different areas of the shelter, including temperature sensors, humidity sensors, air pressure sensors, electromagnetic interference sensors, and the like. The above-mentioned sensors are distributed in key positions such as device-intensive areas, patient areas, and air circulation areas, so as to fully reflect the environmental conditions in the shelter. For example, temperature and humidity sensors are deployed near the CT scanner, an air pressure sensor is deployed at the shelter door, and electromagnetic interference sensors are deployed near the power supply of each device.

[0020] The collected signals of various types are summarized by a data aggregation gateway, which performs time synchronization processing on the signals transmitted by the sensors and communication interfaces. Time synchronization uses the network time protocol to unify the timestamps of all signals to the standard clock system within the shelter, ensuring consistency of the signals in the time dimension, thereby forming a synchronized running real-time signal set. This set is stored in the form of data frames, each of which contains signal type identification, collection timestamp, signal value, and device component identification, etc.

[0021] Step S120: Perform mutual feedback correlation processing on the synchronized running real-time signal set to extract a mutual feedback correlation feature set from the running signals, which includes component parameter coupling features, inter-device signal conduction features, and environmental parameter cross features.

[0022] Step S121: Separate the real-time parameter signals of each running component from the synchronized running real-time signal set and extract the time-domain variation features of each running component parameter, which include parameter fluctuation period features and parameter variation rate features.

[0023] In this embodiment, the synchronized running real-time signal set is first subjected to signal separation. According to the device component identification contained in the signals, the real-time parameter signals belonging to the same running component are extracted. For example, for the rotating component of a CT scanner, all signals with the device component identification "CT-rotating component" are filtered from the synchronized running real-time signal set to obtain the real-time vibration frequency signal and amplitude signal of the rotating component.

[0024] For each extracted running component parameter signal, time-domain variation feature extraction is performed. Taking the vibration frequency signal of the rotating component as an example, the parameter fluctuation period feature is obtained by performing period analysis on the vibration frequency signal. The sliding window method is used to divide the vibration frequency signal into multiple consecutive time windows, and autocorrelation analysis is performed on the signal in each time window to determine the main fluctuation period of the signal. For example, by calculating the autocorrelation coefficients of the signal at different time intervals, the time interval corresponding to the maximum autocorrelation coefficient is the fluctuation period of the signal in that window. Statistical analysis of the fluctuation periods of multiple windows gives the parameter fluctuation period feature of the running component parameter, which is a sequence containing multiple period values, reflecting the fluctuation law of the vibration frequency in different time periods.

[0025] The parameter change rate feature is obtained by calculating the first derivative of the parameter signal. The time series difference of the vibration frequency signal is processed, i.e. the difference value of the vibration frequency of adjacent two sampling points and the sampling time interval are calculated, to obtain the change rate of each sampling point. The obtained change rate sequence is smoothed to remove noise interference, and then the maximum value, minimum value, average value and change trend of the change rate are extracted to form the parameter change rate feature. For example, the change trend is obtained by linear fitting of the smoothed change rate sequence, and the change rate is determined to be increased, decreased or stable according to the slope of the fitting straight line.

[0026] Step S122: separating the inter-device interaction signal in the synchronous running real-time signal set, and extracting the conduction feature of the inter-device interaction signal, the conduction feature including the signal transmission delay feature and the interaction parameter matching deviation feature.

[0027] The inter-device interaction signal is separated from the synchronous running real-time signal set, and is filtered according to the source device identifier and the destination device identifier in the signal. For example, all interaction signals with the source device identifier being "CT scanner" and the destination device identifier being "image storage server" are filtered to obtain the interaction signal sequence between the two devices.

[0028] The signal transmission delay feature is obtained by calculating the difference between the sending time and the receiving time of the interaction signal. In the data frame of the inter-device interaction signal, the sending timestamp and the receiving timestamp are included, and the transmission delay time of each interaction signal is obtained by subtracting the sending timestamp from the receiving timestamp. The transmission delay times of multiple interaction signals are statistically analyzed to extract the average value, variance, maximum value and frequency of delay occurrence of the delay time, to form the signal transmission delay feature. For example, the number of times that the transmission delay exceeds a preset threshold within a period of time is counted to reflect the severity of the signal transmission delay.

[0029] The interaction parameter matching deviation feature is used to measure the deviation degree of the data parameter in the interaction signal from the expected parameter. Taking the transmission of image data from the CT scanner to the image storage server as an example, the expected image data format includes parameters such as image resolution, pixel depth and data compression rate. The actual image data parameters are extracted from the interaction signal and compared with the preset expected parameters. For example, the difference between the actual image resolution and the expected image resolution accounts for a percentage of the expected image resolution, and the difference between the actual pixel depth and the expected pixel depth, etc. After normalization processing of the above deviation values, the interaction parameter matching deviation feature is formed. This feature reflects the consistency and accuracy of the inter-device interaction data.

[0030] Step S123: Separate the shelter environment comprehensive influence signal in the synchronous operation real-time signal set, extract the differentiated action characteristics of the environment signal on different operation component parameters, and the differentiated action characteristics include the influence amplitude characteristics and the influence lag characteristics of the environment parameters on the component parameters.

[0031] The shelter environment comprehensive influence signal is separated from the synchronous operation real-time signal set, and the environment signals such as temperature, humidity, air pressure, and electromagnetic interference are screened out according to the signal type identifier. For example, signals with signal type identifiers such as "temperature" and "humidity" are extracted to obtain the shelter environment comprehensive influence signal.

[0032] The differentiated action characteristics of the environment signal on different operation component parameters are analyzed. Taking the influence of the temperature signal on the vibration frequency of the rotating component of the CT scanner and the working temperature of the ray generating component as an example. The influence amplitude characteristics are obtained by analyzing the correlation between the environment parameter change and the component parameter change. The sliding window method is used to divide the temperature signal and the vibration frequency signal of the rotating component into multiple time windows, and the ratio of the temperature change amount to the vibration frequency change amount is calculated in each window. This ratio reflects the influence amplitude of temperature on vibration frequency. The influence amplitude values of multiple windows are statistically analyzed to obtain the average value, maximum value, and distribution characteristics of the influence amplitude, forming the influence amplitude characteristics of the environment parameters on the component parameters.

[0033] The influence lag characteristics are used to measure the time delay of the influence of the environment parameter change on the component parameter. It is obtained by calculating the difference between the start time of the environment parameter change and the start time of the component parameter change. For example, when the temperature in the shelter starts to rise from a certain value, the time point at which the temperature starts to rise is recorded, and the time point at which the vibration frequency of the rotating component starts to change is observed. The difference between the two time points is the influence lag time of temperature on vibration frequency. The influence lag times corresponding to multiple environment parameter change events are statistically analyzed to obtain the influence lag characteristics, including the average value, variance, etc.

[0034] Step S124: Perform first-level correlation operation on the time domain change characteristics and the conduction characteristics, calculate the coupling coefficients of different operation component parameters in the same device and the interaction signals between devices, and generate first correlation characteristics.

[0035] In the embodiment, taking a CT scanner as an example, the CT scanner comprises operating components such as a rotating component, a ray generating component, and a data acquisition component, and there is an inter-device interaction signal between the CT scanner and an image storage server. First, time-domain variation characteristics (fluctuation period characteristics and variation rate characteristics of a vibration frequency) of the rotating component, time-domain variation characteristics (fluctuation period characteristics and variation rate characteristics of an operating temperature) of the ray generating component, and time-domain variation characteristics (fluctuation period characteristics and variation rate characteristics of a data transmission rate) of the data acquisition component are associated with conduction characteristics (signal transmission delay characteristics and interaction parameter matching deviation characteristics) of the inter-device interaction signal between the CT scanner and the image storage server.

[0036] For the coupling coefficient calculation of different operating component parameters in the same device and the inter-device interaction signal, a correlation analysis method is adopted. For example, the correlation between the vibration frequency variation rate characteristics of the rotating component and the signal transmission delay characteristics of the inter-device interaction signal is analyzed. The vibration frequency variation rate sequence and the signal transmission delay sequence are time-aligned, and then the correlation coefficient of the two sequences is calculated, which is one of the coupling coefficients of the rotating component parameters and the inter-device interaction signal.

[0037] In the same way, the correlation coefficients of the fluctuation period characteristics of the operating temperature of the ray generating component and the interaction parameter matching deviation characteristics, the correlation coefficients of the variation rate characteristics of the data transmission rate of the data acquisition component and the signal transmission delay characteristics, and the like are calculated. The coupling coefficients are combined to form a first association characteristic, which reflects the association closeness between the operating component parameters in the same device and the inter-device interaction signal.

[0038] Step S125: performing a second-level association operation on the first association characteristic and the differentiation effect characteristic to calculate an indirect influence coefficient of the environmental signal on the operating component parameters through the inter-device interaction, and generating a second association characteristic.

[0039] Taking the temperature environmental signal in a shelter as an example, the differentiation effect characteristic thereof includes an influence amplitude characteristic and an influence lag characteristic on the vibration frequency of the rotating component of the CT scanner, and an influence amplitude characteristic and an influence lag characteristic on the temperature of the hard disk of the image storage server. The first association characteristic comprises the coupling coefficients of the operating component parameters in the CT scanner and the inter-device interaction signal between the CT scanner and the image storage server.

[0040] The indirect influence coefficient of the environmental signal on the operating component parameter through the interaction between devices is calculated by considering the process that the environmental signal first affects the component parameter of a device, and the change of the component parameter affects the component parameter of another device through the interaction signal between devices. For example, the temperature rise affects the temperature of the hard disk of the image storage server, the change of the temperature of the hard disk causes the change of the matching deviation feature of the interaction parameter of the interaction signal between the hard disk and the CT scanner, and then affects the data transmission rate of the data acquisition component of the CT scanner.

[0041] When calculating the indirect influence coefficient, first, the influence amplitude of the environmental signal on the component parameter of the first device is determined (obtained from the differentiated action feature), then the coupling coefficient of the component parameter and the interaction signal between devices is determined (obtained from the first correlation feature), and then the influence amplitude of the interaction signal between devices on the component parameter of the second device is determined (combined with the differentiated action feature and the first correlation feature). The three values are multiplied to obtain the indirect influence coefficient. The indirect influence coefficients under different environmental signals and different interaction paths between devices are calculated and combined to generate the second correlation feature.

[0042] Step S126: After the dimension unification processing of the first correlation feature and the second correlation feature, based on the feature importance evaluation result, the dynamic weight is assigned to the first correlation feature and the second correlation feature after the dimension unification, the first correlation feature and the second correlation feature after the weighting are integrated to generate the mutual feedback correlation feature set containing the component parameter coupling feature, the device interaction signal conduction feature and the environmental parameter cross feature.

[0043] The first correlation feature and the second correlation feature can have different dimensions. For example, the first correlation feature contains multiple coupling coefficients, each coupling coefficient is a numerical value, and the second correlation feature contains multiple indirect influence coefficients, each indirect influence coefficient is also a numerical value, but the number of the two can be different. The dimension unification processing adopts the feature vectorization method to convert the first correlation feature and the second correlation feature into fixed-length feature vectors. For example, if the first correlation feature contains 5 coupling coefficients, it is converted into a 5-dimensional vector; if the second correlation feature contains 4 indirect influence coefficients, it is converted into a 4-dimensional vector, and then the two vectors are adjusted to the same dimension by feature expansion or dimension reduction method, for example, both are adjusted to 10-dimensional vectors. Feature expansion can adopt zero padding method to add zero elements at the end of the vector to reach the target dimension; dimension reduction can adopt principal component analysis method to extract main components to reduce the dimension of the vector.

[0044] The feature importance evaluation adopts a tree model-based feature importance evaluation method. A decision tree model is constructed, with the equipment operation state (normal or abnormal) corresponding to the synchronous running real-time signal set as the label, and the first associated feature and the second associated feature as the input feature, to train the decision tree model. After training, the importance weight of each feature is determined according to the contribution degree of each feature to the node splitting in the decision tree model. For example, a certain coupling coefficient plays a key role in the node splitting of the decision tree, so its importance weight is higher.

[0045] According to the feature importance evaluation result, a dynamic weight is assigned to each feature component in the first associated feature and the second associated feature after unified dimension. The feature component with high importance weight is assigned a larger weight value, and the feature component with low importance weight is assigned a smaller weight value. Then the weighted first associated feature vector and the second associated feature vector are spliced and integrated, i.e. the two vectors are connected in order to form a higher-dimensional feature vector, which is the mutual feedback associated feature set. Among them, the part of the original first associated feature reflecting the association of different component parameters in the same device constitutes the component parameter coupling feature, and the part reflecting the interaction signal association between devices constitutes the inter-device signal conduction feature. The part of the original second associated feature reflecting the association of environmental signals and device component parameters constitutes the environmental parameter cross feature.

[0046] Step S127: associate the mutual feedback associated feature set with the collection time period information of the synchronous running real-time signal set, and extract a core associated feature subset in the mutual feedback associated feature set, the core associated feature subset being a feature combination whose influence weight on the operation state prediction exceeds a preset weight threshold.

[0047] The collection time period information of the synchronous running real-time signal set includes the collection start time and end time of each signal. The generation time of the mutual feedback associated feature set is matched with the collection time period information, so that the mutual feedback associated feature set is associated with the corresponding collection time period. For example, a certain mutual feedback associated feature set is generated based on the synchronous running real-time signal set from 8:00 to 8:30, then the feature set is associated and stored with the collection time period information of "8:00-8:30".

[0048] The extraction of the core correlation feature subset is based on the previously calculated feature importance weights. A weight threshold is preset, for example, the weight threshold is set to the average of all feature importance weights. Each feature component in the mutual feedback correlation feature set is traversed to determine whether its importance weight exceeds the preset weight threshold. The feature components exceeding the threshold are extracted to form the core correlation feature subset. For example, if there are 10 feature components in the mutual feedback correlation feature set, and the importance weights of 6 feature components exceed the preset threshold, then the 6 feature components constitute the core correlation feature subset. The subset contains key features that have a greater impact on the running state prediction and can be used in subsequent running state deduction processes.

[0049] Step S130: input the mutual feedback correlation feature set into the running state deduction model to generate a short-term running state evolution sequence through time series feature progressive operation, the short-term running state evolution sequence including predicted values of component parameters and predicted values of overall running coordination degree in each time period.

[0050] Step S131: input the mutual feedback correlation feature set into the feature preprocessing layer of the running state deduction model, perform time series alignment processing on the mutual feedback correlation feature set, pass the time series aligned mutual feedback correlation feature set to the first time series operation layer of the running state deduction model, extract local time series correlation information of the mutual feedback correlation feature set through convolution operation, and generate a local time series feature vector.

[0051] In this embodiment, each feature component in the mutual feedback correlation feature set is extracted at different time points, and there may be a time misalignment. The feature preprocessing layer first performs timestamp calibration on all feature components in the mutual feedback correlation feature set to ensure that they are consistent in the time dimension. For example, the time series of each feature component is interpolated or resampled to a uniform time interval, so that all feature components have corresponding values at the same time point.

[0052] The time series aligned mutual feedback correlation feature set is represented in the form of a three-dimensional tensor, where the first dimension is the time step, the second dimension is the number of feature channels (i.e. the number of feature components in the mutual feedback correlation feature set), and the third dimension is the number of samples (1 in real-time processing). The three-dimensional tensor is input into the first time series operation layer, which adopts a one-dimensional convolutional neural network structure. The size of the convolution kernel is determined according to the time range of local time series correlation, for example, the time dimension size of the convolution kernel is set to 5, that is, each convolution operation considers 5 consecutive time steps of features.

[0053] The convolution operation is to slide the convolution kernel in the time dimension, weight and sum the features in each sliding window, and perform nonlinear activation processing. For example, for a window at time step t, each element in the convolution kernel is multiplied by the feature value of the corresponding time step and feature channel in the window, then all the product results are added, and finally the convolution output value of the window is obtained through the ReLU activation function. After the convolution operation is performed on all time step windows, a local time sequence feature map is obtained, which is flattened into a vector form, i.e., a local time sequence feature vector is generated. The vector contains the association information of the mutual feedback association feature set in the local time range.

[0054] Step S132: The local time sequence feature vector is transmitted to the second time sequence operation layer of the running state inference model. The long-time association of the local time sequence feature vector is modeled through the loop unit, the dependence relationship of the local time sequence feature vector in the continuous time steps is captured, and a long-time time sequence feature vector is generated.

[0055] The second time sequence operation layer adopts a long short-term memory loop unit structure. The local time sequence feature vector is input into the loop unit in time sequence, and the local time sequence feature vector at each time step is input into the loop unit. The loop unit includes an input gate, a forget gate and an output gate, which selectively remember and forget historical information through these gate mechanisms.

[0056] At each time step, the input gate determines how much information in the current local time sequence feature vector is retained in the cell state; the forget gate determines how much information in the previous cell state is forgotten; and the output gate generates the output of the current time step according to the current cell state and the hidden state. In this way, the loop unit can capture the dependence relationship of the local time sequence feature vector in a long time range, for example, the influence of the change of a feature at multiple time steps before the current time step on the feature of the current time step.

[0057] After processing all time steps, the hidden state sequence of the loop unit constitutes a long-time time sequence feature vector. Each element of the vector corresponds to a hidden state at a time step, and contains long-time time sequence association information from the initial time step to the time step.

[0058] Step S133: The long-time time sequence feature vector is transmitted to the feature abstraction layer of the running state inference model, and high-order feature abstraction is performed on the long-time time sequence feature vector through a multilayer perceptron to generate a high-order running state feature vector.

[0059] The long-time sequence feature vector still contains a large amount of original time sequence information, and the feature abstraction layer is built by using a multi-layer perception mechanism to map the long-time sequence feature vector to a higher-level abstract feature space. The multi-layer perception mechanism includes multiple fully connected layers, the number of input neurons of the first layer is equal to the length of the long-time sequence feature vector, and the number of neurons of each subsequent layer gradually decreases to realize feature dimension reduction and abstraction.

[0060] For example, the first fully connected layer has 256 neurons, which converts the long-time sequence feature vector with a length of 512 into a 256-dimensional feature vector; the second fully connected layer has 128 neurons, which converts the 256-dimensional feature vector into a 128-dimensional feature vector; and the third fully connected layer has 64 neurons, which converts the 128-dimensional feature vector into a 64-dimensional high-order running state feature vector. The output of each fully connected layer is nonlinearly transformed by a ReLU activation function to enhance the expression ability of the model. The high-order running state feature vector contains an abstract representation of the running state and can better reflect the essential features of the running state.

[0061] Step S134: The high-order running state feature vector is transmitted to the prediction output layer of the running state deduction model, and the high-order running state feature vector is converted into running component parameter prediction values of each time period through linear mapping operation, and the coordination coefficients of different running component parameter prediction values of each time period are calculated to generate overall running coordination degree prediction values of each time period.

[0062] The prediction output layer includes two parallel output branches, one branch is used to predict running component parameter prediction values of each time period, and the other branch is used to predict overall running coordination degree prediction values of each time period. For the running component parameter prediction branch, the high-order running state feature vector is linearly mapped through a fully connected layer, and the number of output neurons of the fully connected layer is equal to the number of running component parameters to be predicted. For example, there are 8 parameters to be predicted in the shelter, including the vibration frequency of the CT scanner rotating component, the working temperature of the ray generating component, and the data transmission rate of the data acquisition component, and the fully connected layer has 8 output neurons, and the output value of each neuron is the prediction value of the corresponding running component parameter in a future time period.

[0063] For the overall operation synergy degree prediction branch, first, the synergy coefficients of the prediction values of different operation component parameters in each period are calculated. The synergy coefficients are obtained by calculating the correlation coefficient matrix between the prediction values of different operation component parameters. Each element in the correlation coefficient matrix represents the degree of linear correlation between two operation component parameter prediction values. Then, the eigenvalue decomposition is performed on the correlation coefficient matrix, and the eigenvector corresponding to the maximum eigenvalue is taken as the synergy weight vector. The prediction values of each operation component parameter are weighted and summed with the synergy weight vector to obtain the overall operation synergy degree prediction value. Alternatively, a special fully connected layer is used to process the high-order operation state feature vector to directly output the overall operation synergy degree prediction value. The number of output neurons of the fully connected layer is 1.

[0064] Step S135: The operation component parameter prediction values in each period and the overall operation synergy degree prediction values in the corresponding period are arranged in time sequence to form an initial short-term operation state evolution sequence.

[0065] The operation component parameter prediction values and the overall operation synergy degree prediction values in each period are output in the order of future time. For example, the operation state in each 5-minute period in the next 1 hour is predicted, and there are 12 periods in total. The 8 operation component parameter prediction values and 1 overall operation synergy degree prediction value in each period are combined into a state vector, and then the state vectors of 12 periods are arranged in time sequence to form a two-dimensional matrix, where the rows represent the periods and the columns represent the operation component parameter prediction values and the overall operation synergy degree prediction values. The matrix is the initial short-term operation state evolution sequence.

[0066] Step S136: The initial short-term operation state evolution sequence is subjected to trend smoothing processing, and the key time node features in the smoothed short-term operation state evolution sequence are extracted and stored in association with the smoothed short-term operation state evolution sequence. The key time node features are the features corresponding to the time nodes at which the operation component parameter prediction values or the overall operation synergy degree prediction values change significantly.

[0067] The trend smoothing processing adopts the moving average method to smooth each operation component parameter prediction value sequence and overall operation synergy degree prediction value sequence in the initial short-term operation state evolution sequence. For example, a moving average window with a window size of 3 is used, and for the prediction value at period t, the average value of the prediction values at periods t-1, t, and t+1 is used to replace the original prediction value (for the periods at both ends of the sequence, a one-sided window is used for processing). Through the smoothing processing, the noise and random fluctuations in the prediction values can be removed, and the trend of the sequence becomes more obvious.

[0068] In the smoothed short-term operation state evolution sequence, the identification of the key time node is achieved by detecting the change rate of the predicted value. For each operation component parameter predicted value sequence and the overall operation coordination degree predicted value sequence, the difference between the predicted values of the adjacent two time periods and the time interval are calculated to obtain the change rate. A change rate threshold is preset, and when the change rate of a certain time period exceeds the threshold, the time period is considered as a key time node.

[0069] For the key time node, all operation component parameter predicted values and overall operation coordination degree predicted values corresponding to the node are extracted, as well as the predicted value trend of the time period before and after the node (such as rising trend, falling trend or stable trend). These information together constitute the key time node feature. The key time node feature and the smoothed short-term operation state evolution sequence are stored in the same data structure, and the key time node feature is associated with the corresponding time period in the smoothed sequence through the time stamp, so as to quickly locate the key time point in subsequent identification of potential abnormal evolution trend.

[0070] Step S140: identifying potential abnormal evolution trend based on the short-term operation state evolution sequence, determining abnormal correlation influence range combined with the mutual feedback correlation feature set, and generating coordination control demand information.

[0071] Step S141: extracting the operation component parameter predicted value of each time period in the short-term operation state evolution sequence, comparing it with the preset normal operation parameter range, marking the operation component parameter predicted value and the corresponding time period that exceeds the normal operation parameter range, and counting the number of time periods that exceed the normal parameter range for each operation component in the short-term operation state evolution sequence. Calculate the abnormal time period proportion, which is the ratio of the number of time periods that exceed the normal parameter range to the total number of predicted time periods.

[0072] In this embodiment, the preset normal operation parameter range is determined according to the design specifications and historical operation data of each operation component in the shelter. For example, the normal vibration frequency range of the CT scanner rotating component is [f_min, f_max], the normal working temperature range of the ray generating component is [T_min, T_max], etc. The above normal range is stored in the device parameter database and can be adjusted according to the device model and service life.

[0073] From the short-term operation state evolution sequence, the predicted value of each operation component parameter of each time period is extracted, such as the predicted value f1 of the rotating component vibration frequency of time period 1, the predicted value f2 of the vibration frequency of time period 2, etc. Each predicted value is compared with the corresponding normal operation parameter range. If f1

[0074] For each operating component, the number of time periods in which the predicted value exceeds the normal parameter range is counted. For example, if the vibration frequency of the rotating component exceeds the normal range in 3 of the 12 predicted time periods, then the number of time periods in which the predicted value exceeds the normal parameter range for this operating component is 3. The total number of predicted time periods is 12, so the proportion of abnormal time periods is 3 / 12 = 0.25.

[0075] Step S142: Extract the overall operation synergy degree prediction value of each time period in the short-term operation state evolution sequence, compare it with the preset normal synergy degree range, mark the overall operation synergy degree prediction value and the corresponding time period that is lower than the normal synergy degree range, count the number of time periods in which the overall operation synergy degree prediction value is lower than the normal synergy degree range, and calculate the proportion of synergy abnormal time periods.

[0076] The preset normal synergy degree range is determined according to the requirement for the cooperative work of the equipment in the shelter. For example, the normal synergy degree range is [C_min, C_max], which is obtained by analyzing the distribution of the synergy degree values when the equipment is normally cooperatively operated. The overall operation synergy degree prediction value of each time period is extracted from the short-term operation state evolution sequence, such as the synergy degree prediction value C1 of time period 1, the synergy degree prediction value C2 of time period 2, etc.

[0077] Compare each synergy degree prediction value with the normal synergy degree range. If C1 < C_min, mark C1 as a prediction value lower than the normal range, and record its corresponding time period 1. Count the number of time periods in which the overall operation synergy degree prediction value is lower than the normal synergy degree range. For example, if the synergy degree prediction value of 2 of the 12 predicted time periods is lower than the normal range, then the number of synergy abnormal time periods is 2, and the proportion of synergy abnormal time periods is 2 / 12 ≈ 0.17.

[0078] Step S143: If the proportion of abnormal time periods of any one operating component exceeds the preset component abnormal threshold, or the proportion of synergy abnormal time periods of the overall operation synergy degree exceeds the preset synergy abnormal threshold, it is determined that there is a potential abnormal evolution trend.

[0079] The preset component abnormal threshold and the preset synergy abnormal threshold are set according to the importance and fault tolerance of the equipment in the shelter. For example, the preset component abnormal threshold is set to 0.2, and the preset synergy abnormal threshold is set to 0.15. In this embodiment, the proportion of abnormal time periods of the rotating component is 0.25, which exceeds the preset component abnormal threshold 0.2, so it is determined that there is a potential abnormal evolution trend. Even if the proportion of abnormal time periods of all operating components does not exceed the component abnormal threshold, if the proportion of synergy abnormal time periods of the overall operation synergy degree exceeds the synergy abnormal threshold, it is also determined that there is a potential abnormal evolution trend.

[0080] Step S144: Extract the component parameter prediction value change curve corresponding to the running component with potential abnormal evolution trend, analyze the change slope and change acceleration of the component parameter prediction value change curve, and determine the rate characteristics of abnormal evolution.

[0081] The running component with potential abnormal evolution trend is the rotating component of the CT scanner, and its vibration frequency prediction value change curve in the short-term running state evolution sequence is extracted. The curve takes the time period as the horizontal coordinate and the vibration frequency prediction value as the vertical coordinate. The change slope is obtained by linear fitting of the curve, for example, a straight line is fitted by using the least square method, and the slope of the straight line is the change slope. A positive slope indicates an upward trend in vibration frequency, a negative slope indicates a downward trend, and the absolute value of the slope indicates the steepness of the trend.

[0082] The change acceleration is obtained by calculating the change rate of the change slope. The change curve is divided into multiple continuous subintervals, the linear fitting slope is calculated in each subinterval, and then the ratio of the difference between the slopes of adjacent subintervals to the time length of the subintervals is calculated to obtain the change acceleration. A positive change acceleration indicates that the change slope is increasing, i.e., the change speed of the vibration frequency is accelerating; a negative change acceleration indicates that the change slope is decreasing, i.e., the change speed is slowing down. The change slope and the change acceleration together constitute the rate characteristics of abnormal evolution, reflecting the development speed and change trend of the abnormal trend.

[0083] Step S145: Combine the component parameter coupling characteristics in the mutual feedback association feature set to find other running components that have a coupling relationship with the running component with potential abnormal evolution trend, and form a direct association component list.

[0084] The component parameter coupling characteristics in the mutual feedback association feature set include coupling coefficients between different running component parameters in the same device. In this embodiment, the component parameter coupling characteristics of the rotating component include coupling coefficients with the ray generating component, the data acquisition component, and other running components. The size of the coupling coefficient reflects the closeness of the association between components. A coupling coefficient threshold is preset. When the coupling coefficient between two running components exceeds the threshold, it is considered that they have a coupling relationship.

[0085] For example, the coupling coefficient between the rotating component and the ray generating component is 0.7, which exceeds the preset threshold of 0.5, so the ray generating component is a direct association component that has a coupling relationship with the rotating component. The coupling coefficient between the rotating component and the data acquisition component is 0.6, which also exceeds the threshold, so the data acquisition component is also included in the direct association component list. The identifiers of all running components that have a coupling relationship with the rotating component (such as "CT-ray generating component" and "CT-data acquisition component") are added to the direct association component list.

[0086] Step S146: In combination with the inter-device signal conduction features in the mutual feedback association feature set, find other devices associated with the device with potential abnormal evolution trend through signal conduction, and form an indirect association device list.

[0087] The device with potential abnormal evolution trend is a CT scanner, and the inter-device signal conduction features in the mutual feedback association feature set include signal transmission delay features and interaction parameter matching deviation features between the CT scanner and other devices. By analyzing these features, it can be determined which devices are associated with the CT scanner through signal conduction.

[0088] For example, there is frequent image data interaction between the CT scanner and the image storage server, and the signal transmission delay features and interaction parameter matching deviation features have a high weight in the inter-device signal conduction features, so the image storage server is a device associated with the CT scanner through signal conduction. In addition, the CT scanner and the central monitoring system will transmit patient scan plans and device state information, and there is also signal conduction association, so the central monitoring system is also included in the indirect association device list. Add all device identifiers associated with the CT scanner through signal conduction to the indirect association device list.

[0089] Step S147: Integrate the direct association component list and the indirect association device list to determine the abnormal association impact range, which includes the operating components and devices affected by the potential abnormality.

[0090] The operating components in the direct association component list belong to the CT scanner, and the devices in the indirect association device list are other independent devices associated with the CT scanner through signal conduction. When integrating the two lists, it is necessary to distinguish the hierarchical relationship between the operating components and the devices. Merge the device (i.e. the CT scanner) to which the operating components in the direct association component list belong and the devices in the indirect association device list to obtain a set of devices affected by the potential abnormality, which includes the CT scanner, the image storage server, the central monitoring system, etc.

[0091] In each affected device, in addition to the operating components in the direct association component list, there may be other operating components that have a coupling relationship with these directly associated components, and these operating components may also be affected by the potential abnormality. For example, the ray generating component of the CT scanner is a directly associated component, and the ray generating component has a coupling relationship with the cooling component of the CT scanner, and the cooling component may also be affected. Therefore, it is necessary to expand the operating components in each affected device, and other operating components that have a coupling relationship with the directly associated components are also included in the abnormal association impact range. Finally, the abnormal association impact range is represented in a device-component hierarchical structure, clearly listing all devices affected by the potential abnormality and the relevant operating components in each device.

[0092] Step S148: According to the rate characteristics of abnormal evolution, the size of abnormal correlation influence range and the abnormal degree of overall operation synergy, determine the emergency level of collaborative regulation, and according to the emergency level of collaborative regulation and the abnormal correlation influence range, determine the number of operation components and equipment that need to participate in regulation, and generate the regulation object list.

[0093] The emergency level of collaborative regulation is divided into three levels: high, medium and low. In the rate characteristics of abnormal evolution, the greater the absolute value of the change slope, the positive change acceleration (i.e. the abnormal trend is accelerating), the higher the emergency level; the larger the abnormal correlation influence range, the more equipment and operation components involved, the higher the emergency level; the abnormal degree of overall operation synergy is measured by the proportion of collaborative abnormal period and the deviation degree of abnormal synergy value from the normal range, the higher the proportion, the greater the deviation, the higher the emergency level.

[0094] For example, the absolute value of the change slope of the rotating component abnormal evolution is large and the change acceleration is positive, the abnormal correlation influence range includes 3 devices and 8 operation components, the proportion of collaborative abnormal period of overall operation synergy is 0.17 and the synergy value of part of the period is far below the lower limit of the normal range, and according to these factors, the emergency level of collaborative regulation is determined to be high.

[0095] According to the high level of emergency and the abnormal correlation influence range, the operation components and equipment that need to participate in regulation should include the abnormal source component (rotating component), the directly related component (ray generating component, data acquisition component), the indirectly related equipment (image storage server, central monitoring system) and the related key operation components. For example, the hard disk storage component of the image storage server, the data processing component of the central monitoring system, etc. all need to participate in regulation. The above operation components and equipment are listed to generate the regulation object list.

[0096] Step S149: Based on the emergency level, the regulation object list and the rate characteristics of abnormal evolution, generate the regulation target, which includes the parameter value that needs to be adjusted to the normal range and the synergy value that needs to be restored.

[0097] For the case of high emergency level, the regulation target should be set to adjust the abnormal parameters to the normal range in a short time and restore the overall operation synergy. The parameter value that needs to be adjusted to the normal range is determined according to the preset normal operation parameter range, for example, the vibration frequency of the rotating part needs to be adjusted to the range of [f_min, f_max], the working temperature of the radiation generating part needs to be adjusted to the range of [T_min, T_max], etc. For each operating component parameter that needs to be adjusted, according to the rate characteristics of abnormal evolution, the staged adjustment target is set, for example, the vibration frequency is reduced by a certain amplitude from the current predicted value in the first regulation period, and continues to adjust in the subsequent period until it reaches the normal range.

[0098] The synergy value to be restored is determined according to the preset normal synergy range, and the goal is to adjust the overall operation synergy predicted value to the range of [C_min, C_max]. At the same time, considering the synergy relationship between different devices and operating components, the target parameters of the interaction signals between devices also need to be set, for example, the signal transmission delay between the CT scanner and the image storage server needs to be controlled within the preset normal delay range, and the interaction parameter matching deviation needs to be adjusted to zero or close to zero level.

[0099] Step S1410: integrate the emergency level, the regulation object list and the regulation target to generate the synergy regulation demand information, which also contains the time window requirement of regulation implementation.

[0100] The synergy regulation demand information is a structured data object, which contains the following fields: emergency level field (value "high"), regulation object list field (contains all device and operating component identifiers in the regulation object list), regulation target field (contains the staged adjustment target value of each operating component parameter and the recovery target value of the overall operation synergy), and time window requirement field.

[0101] The time window requirement is determined according to the emergency level, and when the emergency level is high, the time window requirement is a short period of time, for example, 1 hour, within which the regulation is completed. Within the time window, multiple regulation periods are also divided, each corresponding to a period in the short-term operation state evolution sequence, and the regulation tasks that need to be completed in each regulation period are clearly defined. The above information is integrated into the synergy regulation demand information to form a complete regulation instruction generation basis.

[0102] Step S150: input the cooperative regulation demand information and the mutual feedback correlation feature set into the cooperative regulation model to generate a multi-device cooperative operation regulation instruction, send the multi-device cooperative operation regulation instruction to the corresponding control module, collect a synchronous operation feedback signal set after execution, input the synchronous operation feedback signal set into the mutual feedback correlation processing link, update the extraction rule of the mutual feedback correlation feature set, and adjust the timing operation parameter of the running state deduction model and the instruction generation logic of the cooperative regulation model.

[0103] Step S151: input the cooperative regulation demand information into the demand analysis layer of the cooperative regulation model to obtain the emergency level, the regulation object list, the regulation target, and the time window requirement.

[0104] The demand analysis layer of the cooperative regulation model uses the structured information extraction technology in natural language processing to analyze the text description in the cooperative regulation demand information. For example, the information "high level" is extracted from the "emergency level field" of the cooperative regulation demand information; the identification strings of each device and operating component are extracted from the "regulation object list field" and converted into standardized device-component codes; the target value range and phased adjustment requirements of each operating component parameter, as well as the target value range of the overall operation coordination degree, are extracted from the "regulation target field"; and the total time window length and the start time and end time of each regulation period are extracted from the "time window requirement field".

[0105] The analyzed information is stored in a structured data structure, such as a dictionary data type, where the key is the information category (such as "emergency level", "regulation object", etc.) and the value is the specific content obtained by analysis. The demand analysis layer also verifies the analysis results to ensure consistency between the information, such as whether the division of the regulation period matches the time window length and whether the regulation object is within the abnormal correlation influence range. If inconsistencies are found, an error prompt will be returned and the generation of the cooperative regulation demand information will be required.

[0106] Step S152: input the mutual feedback correlation feature set into the correlation feature analysis layer of the cooperative regulation model to extract the component parameter coupling features, the inter-device signal conduction features, and the environmental parameter cross features corresponding to the regulation object list, and generate a regulation correlation feature subset.

[0107] The correlation feature analysis layer filters out relevant features from the mutual feedback correlation feature set according to the device and operating component identification in the regulation object list. For component parameter coupling features, all coupling coefficients containing operating components in the regulation object list are filtered out; for inter-device signal conduction features, signal transmission delay features and interaction parameter matching deviation features between devices in the regulation object list are filtered out; for environmental parameter cross features, influence amplitude features and influence lag features of environmental parameters on operating component parameters in the regulation object list are filtered out.

[0108] For example, the rotating part and the ray generating part of the CT scanner, and the hard disk storage part of the image storage server are included in the regulation object list, the coupling coefficient of the rotating part and the ray generating part, the coupling coefficient of the rotating part and the data acquisition part (the data acquisition part is a direct correlation part) are extracted from the component parameter coupling feature; the signal transmission delay feature and the interaction parameter matching deviation feature between the CT scanner and the image storage server are extracted from the inter-device signal conduction feature; the influence amplitude feature and the influence lag feature of temperature on the vibration frequency of the rotating part, the influence amplitude feature of temperature on the working temperature of the ray generating part, etc. are extracted from the environmental parameter cross feature. These selected features are combined to generate a regulation correlation feature subset.

[0109] Step S153: input the regulation correlation feature subset and the parsed regulation target into the parameter adjustment calculation layer of the cooperative regulation model, calculate the initial parameter adjustment amount of each regulation object, and combine the coupling coefficients in the mutual feedback correlation feature set to calculate the indirect influence value of the initial parameter adjustment amount on the associated regulation object, and analyze whether the indirect influence value exceeds the normal parameter range.

[0110] The parameter adjustment calculation layer first calculates the initial parameter adjustment amount according to the target value of each operating component parameter in the regulation target and the predicted value in the current short-term operating state evolution sequence. The initial parameter adjustment amount is the difference between the target value and the predicted value. For example, the current predicted vibration frequency of the rotating part is f_pred, and the target value is f_target (f_target is within the normal range), and the initial parameter adjustment amount is f_target-f_pred.

[0111] Then, for the initial parameter adjustment amount of each regulation object, find the associated regulation object that has a coupling relationship with the regulation object according to the component parameter coupling feature in the regulation correlation feature subset. For example, the initial parameter adjustment amount of the rotating part will affect the working temperature of the ray generating part through the coupling relationship. According to the coupling coefficient k between the rotating part and the ray generating part, the indirect influence value is calculated as the initial parameter adjustment amount x k.

[0112] Compare the calculated indirect influence value with the normal parameter range of the associated regulation object to determine whether it is out of range. For example, the normal working temperature range of the ray generating part is [T_min, T_max], its current predicted value is T_pred, and the indirect influence value is ΔT. The adjusted predicted value is T_pred+ΔT. If T_pred+ΔT

[0113] Step S154: If the indirect influence value is out of the normal parameter range, adjust the initial parameter adjustment amount, recalculate the indirect influence value until the indirect influence value is within the normal parameter range, and obtain the final parameter adjustment amount.

[0114] When the indirect influence value is out of the normal parameter range, the initial parameter adjustment amount needs to be corrected. The direction of adjustment is determined according to the direction of the indirect influence value out of the normal range, for example, if the indirect influence value causes the predicted value of the associated control object to be higher than the upper limit of the normal range, the initial parameter adjustment amount is reduced; if it is lower than the lower limit of the normal range, the initial parameter adjustment amount is increased.

[0115] The adjustment amplitude is determined according to the degree of out-of-range and the size of the coupling coefficient. The greater the out-of-range degree, the greater the adjustment amplitude; the greater the coupling coefficient, the more sensitive the initial parameter adjustment amount to the associated control object, and the greater the adjustment amplitude. The adjusted initial parameter adjustment amount is denoted as a new initial parameter adjustment amount, the indirect influence value of the associated control object is recalculated, and it is checked whether it is still out of the normal range. This process may need to be iterated several times until the indirect influence values of all associated control objects are within the normal parameter range, and the parameter adjustment amount at this time is the final parameter adjustment amount.

[0116] Step S155: According to the emergency level in the collaborative control demand information, assign a control priority to the final parameter adjustment amount of each control object.

[0117] The emergency level in the collaborative control demand information is high, and under the high emergency level, the assignment of control priority mainly considers the following factors: the importance of the control object in the abnormal associated influence range, the influence degree of the rate characteristics of abnormal evolution on the control object, and the contribution degree of the parameter adjustment of the control object to the recovery of the overall operation synergy.

[0118] For example, the rotating part is the source of abnormality, and its parameter adjustment is most critical to controlling the abnormal evolution trend, so it is assigned the highest priority; the ray generating part is closely coupled with the rotating part, and its abnormal working temperature may directly affect the CT scan image quality, so it is assigned the second highest priority; the hard disk storage part of the image storage server, although it is an indirectly associated device, is responsible for storing scan images, and data security is crucial, so it is assigned a higher priority. According to these factors, a priority level (such as 1st, 2nd, 3rd, etc., with 1st being the highest) is assigned to the final parameter adjustment amount of each control object.

[0119] Step S156: According to the time window requirement in the collaborative control demand information, assign a control execution period to the final parameter adjustment amount of each control object.

[0120] The time window requirement is the total regulation time length and the division of each regulation period. According to the regulation priority and the complexity of parameter adjustment, the final parameter adjustment amount of each regulation object is allocated a regulation execution period. The regulation object with high priority should be allocated in the earlier regulation period to control the abnormal trend as soon as possible; the regulation object with complex parameter adjustment and long execution time (such as physical adjustment involving mechanical parts) should be allocated a longer regulation execution period or multiple consecutive regulation periods.

[0121] For example, the total time window is 1 hour, which is divided into 12 regulation periods of 5 minutes. The regulation priority of the rotating part is the highest, which is allocated in the first and second regulation periods; the radiation generating part is allocated in the second and third regulation periods; the hard disk storage part of the image storage server is allocated in the third and fourth regulation periods, etc. The regulation execution period of each regulation object clearly indicates the start and end period number, ensuring that all regulation tasks are completed within the time window.

[0122] Step S157: According to the regulation priority and the regulation execution period, a regulation execution sequence is generated, and the identification of each regulation object, the final parameter adjustment amount, the regulation execution period and the regulation execution sequence are integrated to generate a single-object regulation instruction.

[0123] The generation of the regulation execution sequence is firstly arranged according to the order of the regulation execution period, and the regulation objects in the same period are arranged according to the regulation priority from high to low. For example, the first period only has the regulation task of the rotating part, and the execution sequence is 1; the second period has the regulation tasks of the rotating part (not completed) and the radiation generating part, and the priority of the rotating part is higher than that of the radiation generating part, and the execution sequences are 2 and 3 respectively; the third period has the regulation tasks of the radiation generating part (not completed) and the hard disk storage part of the image storage server, and the execution sequences are 4 and 5 respectively, etc.

[0124] The single-object regulation instruction is a specific operation instruction for each regulation object, which includes the device-part identification of the regulation object, the final parameter adjustment amount (such as vibration frequency adjustment amount, temperature adjustment amount, etc.), the regulation execution period (start period and end period) and the execution sequence in the period. For example, the single-object regulation instruction of the rotating part may be: device-part identification "CT-rotating part", final parameter adjustment amount "-Δf" (indicating reducing vibration frequency Δf), regulation execution period "period 1-period 2", execution sequence "1".

[0125] Step S158: All single-object regulation instructions are arranged according to the regulation execution sequence to generate a multi-device cooperative running regulation instruction, and a regulation batch identification and a time effectiveness identification are added to the multi-device cooperative running regulation instruction, and the time effectiveness identification is used to indicate the effective execution period of the multi-device cooperative running regulation instruction.

[0126] The multi-device cooperative operation regulation instruction is a sequence of instructions, in which all single-object regulation instructions are arranged in sequence according to the regulation execution order. The position of each single-object regulation instruction in the sequence is determined by its execution order. The regulation batch identifier is generated according to the total regulation time window and the emergency level of the regulation, for example, the regulation batch identifier of the high-level emergency is “E-YYYYMMDD-HHMM”, in which “E” represents emergency, followed by the current timestamp. The time effectiveness identifier is the effective execution time period of the regulation instruction, that is, the time window requirement in the cooperative regulation demand information, for example, “effective period: YYYYMMDD-HHMM to YYYYMMDD-HHMM+1 hour”.

[0127] The multi-device cooperative operation regulation instruction also contains a general control information, such as the target overview of the regulation and the emergency level prompt, which is located at the beginning of the instruction sequence. All single-object regulation instructions, regulation batch identifiers, time effectiveness identifiers, and general control information are combined into a complete instruction document, which adopts a standardized format so that each control module can correctly parse and execute.

[0128] Step S159: Extract the key regulation parameters in the multi-device cooperative operation regulation instruction to generate a regulation parameter summary, which contains the core adjustment amount and the key execution period.

[0129] The core adjustment amount refers to the parameter adjustment amount that plays a decisive role in the overall regulation effect, which is usually the final parameter adjustment amount of the highest priority regulation object and the parameter adjustment amount that has the greatest impact on the overall operation coordination. For example, the final parameter adjustment amount of the rotating part and the adjustment amount of the interaction parameter between the CT scanner and the image storage server (calculated based on the interaction parameter matching deviation characteristics) are selected as the core adjustment amount.

[0130] The key execution period refers to the regulation period corresponding to the key time node in the regulation execution process, for example, the first period (the starting regulation period), the period containing multiple high-priority regulation objects, and the last period when the regulation is about to end. The core adjustment amount and the key execution period are extracted to generate a regulation parameter summary in the form of a simple table (but in the form of a list in the text description), which facilitates the operator to quickly understand the core content of the regulation instruction.

[0131] Step S1510: Send the multi-device cooperative operation regulation instruction to the corresponding control module, and collect the synchronous operation feedback signal set after execution.

[0132] The multi-device cooperative operation regulation instruction is sent to the control modules of the devices to which the regulation objects belong through the control network in the shelter. Each control module identifies the operation components that need to be regulated according to the device-component identification in the instruction, and performs the regulation operation according to the final parameter adjustment amount, the regulation execution period and the execution sequence. For example, after the control module of the CT scanner receives the regulation instruction of the rotating component, it adjusts the driving motor current or control signal of the rotating component within the specified regulation execution period to reduce the vibration frequency.

[0133] After the regulation instruction is executed, the parameters of each operation component change, and the interaction signals between devices and the comprehensive influence signals of the shelter environment may also change. Through the previously deployed sensor network and signal acquisition module, these changed signals are re-acquired, and after time synchronization processing, a synchronous operation feedback signal set is formed. The structure of this set is the same as the synchronous operation real-time signal set obtained in step S110, and contains the real-time parameter signals of each operation component, the interaction signals between devices and the comprehensive influence signals of the shelter environment, but the acquisition period is the period after the regulation instruction is executed.

[0134] Step S1511: input the synchronous operation feedback signal set into the mutual feedback correlation processing link, update the extraction rules of the mutual feedback correlation feature set, and adjust the time sequence operation parameters of the operation state deduction model and the instruction generation logic of the cooperative regulation model.

[0135] Step S1512: input the synchronous operation feedback signal set into the mutual feedback correlation processing link, and update the extraction rules of the mutual feedback correlation feature set.

[0136] For example, step S15121: analyze the synchronous operation feedback signal set, extract the deviation data of the short-term operation state evolution sequence prediction value of the operation state deduction model and the actual operation state value in the synchronous operation feedback signal set, generate a prediction deviation feature set, and the prediction deviation feature set contains the prediction deviation values of each operation component parameter and the overall operation cooperation degree prediction deviation value.

[0137] The synchronous operation feedback signal set contains the actual parameter values of each operation component and the actual operation cooperation degree values of the device after the regulation instruction is executed. Compare the above actual values with the prediction values in the short-term operation state evolution sequence generated by the operation state deduction model for the corresponding period, and calculate the deviation data. The prediction deviation value of each operation component parameter is the actual parameter value minus the predicted parameter value; the overall operation cooperation degree prediction deviation value is the actual cooperation degree value minus the predicted cooperation degree value. Arrange the parameter prediction deviation values of all operation components and the overall operation cooperation degree prediction deviation values in time sequence to form a prediction deviation feature set.

[0138] Step S15122: Perform feature decomposition processing on the predicted deviation feature set to extract the fluctuation period feature of the component parameter predicted deviation value, the change rate feature, and the time sequence change trend feature of the overall operation synergy degree predicted deviation value.

[0139] The fluctuation period feature of each operation component parameter predicted deviation value sequence is extracted, and a similar autocorrelation analysis method as in step S121 is used to determine the main fluctuation period of the deviation value. The change rate feature is obtained by calculating the first derivative of the deviation value sequence, reflecting the speed of change of the deviation value over time.

[0140] The time sequence change trend feature of the overall operation synergy degree predicted deviation value sequence is extracted, and a linear fitting method in a sliding window is used to determine the change trend (rising, falling or stable) of the deviation value in each window, and the frequency and duration of different trends are counted.

[0141] Step S15123: Correlate the fluctuation period feature and change rate feature of the component parameter predicted deviation value with the real-time parameter signal in the synchronous operation feedback signal set, and calculate the deviation sensitive factor of each operation component parameter. The deviation sensitive factor is used to quantify the contribution degree of parameter fluctuation to the predicted deviation.

[0142] The fluctuation period feature of the component parameter predicted deviation value is compared with the fluctuation period feature of the real-time parameter signal of the corresponding operation component in the synchronous operation feedback signal set, and the similarity of the two is calculated. The higher the similarity, the stronger the correlation between the fluctuation of the operation component parameter and the fluctuation of the predicted deviation. At the same time, the change rate feature of the component parameter predicted deviation value is correlated with the change rate feature of the real-time parameter signal, and the correlation coefficient is calculated.

[0143] The deviation sensitive factor considers the fluctuation period similarity and the change rate correlation coefficient, and the two are weighted and summed to obtain. For example, the deviation sensitive factor = α × fluctuation period similarity + (1-α) × change rate correlation coefficient, where α is a weight coefficient determined by experience or cross-validation. The larger the value of the deviation sensitive factor, the higher the contribution degree of the fluctuation of the operation component parameter to the predicted deviation.

[0144] Step S15124: Correlate the time sequence change trend feature of the overall operation synergy degree predicted deviation value with the inter-device interaction signal and the comprehensive environmental influence signal of the shelter in the synchronous operation feedback signal set, extract the collaborative deviation conduction path feature, and the collaborative deviation conduction path feature includes the delay feature of the inter-device interaction signal and the indirect influence feature of the environmental parameter on the synergy degree.

[0145] For each trend segment (such as the rising trend segment above) in the time sequence variation trend characteristics of the overall operation synergy prediction deviation value, analyze the variation of the inter-device interaction signal corresponding to the trend segment, and extract the variation characteristics (such as delay increase, delay decrease, or delay fluctuation) of the signal transmission delay; at the same time, analyze the variation of the comprehensive influence signal of the shelter environment in the trend segment, and extract the variation characteristics of the influence amplitude and influence lag of the environmental parameters on the operating component parameters of each device.

[0146] The synergy deviation conduction path characteristics are represented by constructing a directed graph, where the nodes in the graph are devices or environmental parameters, the edges represent signal conduction or influence relationship, and the attributes on the edges are the delay characteristics of the inter-device interaction signal or the indirect influence characteristics (such as influence amplitude change rate, influence lag change rate) of the environmental parameters on the synergy degree. For example, the edge from the CT scanner to the image storage server is labeled with the characteristic of signal transmission delay increase, and the edge from the temperature parameter to the overall operation synergy degree is labeled with the characteristic of influence amplitude increase.

[0147] Step S15125: Based on the deviation sensitive factor and the synergy deviation conduction path characteristics, a feature correlation strength evaluation matrix is constructed, which is used to describe the correlation tightness between different types of signal characteristics.

[0148] The rows and columns of the feature correlation strength evaluation matrix respectively represent different types of signal characteristics, such as the operating component parameter signal characteristics (such as vibration frequency, working temperature, etc.), the inter-device interaction signal characteristics (such as signal transmission delay, interaction parameter matching deviation, etc.), and the environmental parameter signal characteristics (such as temperature, humidity, etc.). The columns represent the prediction deviation characteristics (such as component parameter prediction deviation, overall operation synergy degree prediction deviation, etc.).

[0149] The element values in the matrix represent the correlation strength between the signal characteristics of the corresponding row and the prediction deviation characteristics of the corresponding column. For the correlation strength between the operating component parameter signal characteristics and the component parameter prediction deviation characteristics, the deviation sensitive factor of the operating component is used; for the correlation strength between the inter-device interaction signal characteristics and the overall operation synergy degree prediction deviation characteristics, the variation degree of the delay characteristics of the inter-device interaction signal in the synergy deviation conduction path characteristics is determined, and the greater the variation degree, the higher the correlation strength; for the correlation strength between the environmental parameter signal characteristics and the overall operation synergy degree prediction deviation characteristics, the variation degree of the indirect influence characteristics of the environmental parameters on the synergy degree in the synergy deviation conduction path characteristics is determined.

[0150] Step S15126: According to the feature correlation strength evaluation matrix, adjust the extraction rules of the mutual feedback correlation feature set, including: for the extraction rules of the component parameter coupling feature, increase the coupling coefficient calculation dimension between the operating component parameters whose deviation sensitive factors exceed the preset threshold; for the extraction rules of the inter-device signal conduction feature, strengthen the weight allocation of the inter-device interaction signal delay feature in the collaborative deviation conduction path feature; for the extraction rules of the environmental parameter cross feature, introduce a dynamic evaluation mechanism for the indirect influence of environmental parameters on the collaborative degree.

[0151] A deviation sensitive factor threshold is preset. When the deviation sensitive factor of a certain operating component parameter exceeds the threshold, it indicates that the parameter is closely related to the prediction deviation, and the coupling coefficient calculation dimension between the parameter and other related operating component parameters needs to be increased in the extraction rules of the component parameter coupling feature. For example, if the deviation sensitive factor of a rotating component exceeds the threshold, the coupling coefficient between the rotating component and the cooling component, which was not previously considered, will be calculated in addition to the coupling coefficient between the rotating component and the original related components when extracting the component parameter coupling feature.

[0152] For the extraction rules of the inter-device signal conduction feature, according to the correlation strength between the inter-device interaction signal delay feature and the overall operation collaborative degree prediction deviation feature in the feature correlation strength evaluation matrix, the weight of the delay feature in feature extraction is adjusted. The higher the correlation strength, the greater the weight, and the easier the feature is selected in subsequent feature importance evaluation. For example, if the correlation strength of the signal transmission delay feature between the CT scanner and the image storage server is high, the weight of the delay feature will be increased when extracting the inter-device signal conduction feature.

[0153] In the extraction rules of the environmental parameter cross feature, only the direct influence of the environmental parameter on the operating component parameter may have been considered before. After introducing the dynamic evaluation mechanism for the indirect influence of the environmental parameter on the collaborative degree in the collaborative deviation conduction path feature, the extraction weight of the cross feature between the environmental parameter and the operating component parameter will be dynamically adjusted according to the indirect influence of the environmental parameter on the collaborative degree in the collaborative deviation conduction path feature. For example, when the indirect influence of temperature on the overall operation collaborative degree increases, the weight of the cross feature between temperature and rotating component vibration frequency will be increased when extracting the cross feature.

[0154] Step S15127: Store the adjusted extraction rules of the mutual feedback correlation feature set and the collection time period information of the synchronous operation feedback signal set in association to form a feature extraction rule library that is dynamically updated over time.

[0155] The adjusted mutual feedback correlation feature set extraction rule is stored in a versioned form, and each version of the rule is associated with the collection time period information (such as the collection start time and end time) of the synchronous operation feedback signal set. The feature extraction rule library adopts a database table structure, and the table contains fields such as rule version number, collection time period start time, collection time period end time, component parameter coupling feature extraction rule, inter-device signal conduction feature extraction rule, and environmental parameter cross feature extraction rule. Each time the extraction rule is updated, a new record is added to the rule library, and historical rule versions are retained for backtracking and comparative analysis when needed.

[0156] Step S15128: Verify the updated mutual feedback correlation feature set extraction rule using the historical synchronous operation data set, calculate the prediction deviation change rate before and after the adjustment of the feature extraction rule, and if the prediction deviation change rate does not reach the preset optimization threshold, re-perform the construction of the feature correlation strength evaluation matrix and the adjustment of the extraction rule until the preset optimization threshold is met.

[0157] The historical synchronous operation data set contains synchronous operation data in the past period of time without regulation or using old extraction rules. The historical synchronous operation data set is used to extract the mutual feedback correlation feature set using the updated extraction rule, and then the extracted feature set is input into the operation state deduction model to obtain the predicted value of the historical data. The deviation between these predicted values and the actual values in the historical data is the prediction deviation of the adjusted rule.

[0158] At the same time, the same historical synchronous operation data set is used to extract features and make predictions using the old extraction rule to obtain the prediction deviation of the rule before adjustment. The prediction deviation change rate is (the prediction deviation before adjustment-the prediction deviation after adjustment) / the prediction deviation before adjustment. A preset optimization threshold, for example 0.1, indicates that the prediction deviation needs to be reduced by at least 10%. If the calculated prediction deviation change rate is greater than or equal to the optimization threshold, it indicates that the adjusted extraction rule is effective; otherwise, return to step S15125 to re-construct the feature correlation strength evaluation matrix and re-adjust the extraction rule according to step S15126, and then verify again until the prediction deviation change rate reaches the optimization threshold.

[0159] Step S1513: Adjust the time sequence operation parameters of the operation state deduction model and the instruction generation logic of the collaborative regulation model.

[0160] Step S15131: Analyze the synchronous operation feedback signal set, extract the deviation data between the short-term operation state evolution sequence prediction value of the operation state deduction model and the actual operation state value in the synchronous operation feedback signal set, generate a prediction deviation feature set, and the prediction deviation feature set contains the prediction deviation value of each operation component parameter and the prediction deviation value of the overall operation synergy degree.

[0161] This step is the same as step S15121, both of which are to parse the synchronous operation feedback signal set, generate a predicted deviation feature set, and use it to adjust the time sequence operation parameter of the running state deduction model.

[0162] Step S15132: Based on the predicted deviation feature set, calculate the time sequence operation parameter adjustment demand index of the running state deduction model, which is determined according to the size of the predicted deviation value and the duration of the deviation.

[0163] The time sequence operation parameter adjustment demand index is used to measure the urgency of adjusting the time sequence operation parameter of the running state deduction model. For each running component parameter predicted deviation value, calculate the average of its absolute value as the deviation size indicator; count the number of continuous periods when the deviation value exceeds the preset allowed deviation range as the deviation duration indicator. Time sequence operation parameter adjustment demand index = β × deviation size indicator + (1-β) × deviation duration indicator, where β is the weight coefficient. The larger the deviation size indicator and the deviation duration indicator, the higher the time sequence operation parameter adjustment demand index, indicating that the higher the urgency of adjusting the parameter.

[0164] Step S15133: According to the time sequence operation parameter adjustment demand index, determine the adjustment priority order of the time sequence operation parameter in the running state deduction model, and the higher the time sequence operation parameter adjustment demand index, the higher the corresponding adjustment priority.

[0165] Each time sequence operation parameter of the running state deduction model (such as the convolution kernel parameter of the first time sequence operation layer, the recurrent unit parameter of the second time sequence operation layer, and the perceptron parameter of the feature abstraction layer) is associated with each predicted deviation value in the predicted deviation feature set. For example, the convolution kernel parameter of the first time sequence operation layer mainly affects the extraction of local time sequence feature vectors, and is associated with running component parameter predicted deviation values with strong local time sequence correlation. According to the time sequence operation parameter adjustment demand index of the predicted deviation feature set associated with each time sequence operation parameter, the time sequence operation parameters are sorted, and the higher the adjustment demand index, the higher the adjustment priority.

[0166] Step S15134: According to the adjustment priority order, adjust the convolution kernel parameter of the first time sequence operation layer, the recurrent unit parameter of the second time sequence operation layer, and the perceptron parameter of the feature abstraction layer of the running state deduction model in turn, and the adjustment amplitude is dynamically calculated based on the deviation value of the corresponding running component parameter in the predicted deviation feature set.

[0167] For the convolution kernel parameters of the first time sequence operation layer, the adjustment amplitude is proportional to the size of the corresponding operation component parameter prediction deviation value. For example, if the prediction deviation value associated with a certain convolution kernel parameter is large, the adjustment amplitude of the convolution kernel parameter is also large. The adjustment direction is determined according to the positive and negative of the deviation. If the prediction value is continuously higher than the actual value, the value of the corresponding weight in the convolution kernel parameter is reduced. If the prediction value is continuously lower than the actual value, the value of the corresponding weight is increased.

[0168] The loop unit parameters of the second time sequence operation layer include the weights and biases of the input gate, the forget gate, and the output gate. When adjusting these parameters, according to the deviation duration index, for the time region where the deviation continuously occurs, the parameters of the forget gate are adjusted to enhance or weaken the memory of the historical information of the region. According to the deviation size index, the parameters of the input gate and the output gate are adjusted to adjust the contribution proportion of the current input information and the hidden state.

[0169] The adjustment of the perceptron parameters of the feature abstraction layer is similar to that of the first time sequence operation layer. The adjustment amplitude is based on the size of the corresponding prediction deviation value, and the adjustment direction is determined according to the positive and negative of the deviation, so that the high-order running state feature vector output by the perceptron is closer to the real running state feature.

[0170] Step S15135: After adjusting the time sequence operation parameters of the running state deduction model, the prediction accuracy of the adjusted running state deduction model is verified using the historical synchronous running data set. If the prediction accuracy does not reach the preset accuracy threshold, the time sequence operation parameter adjustment requirement index is recalculated, and the time sequence operation parameters are iteratively adjusted until the prediction accuracy meets the standard.

[0171] The same historical synchronous running data set as in step S15128 is used to input the running state deduction model after adjusting the time sequence operation parameters, and the prediction value is obtained. The mean absolute error or root mean square error of the prediction value and the historical actual value is calculated as the prediction accuracy index. A preset accuracy threshold is set, for example, the average absolute error is less than a certain value. If the adjusted prediction accuracy reaches the threshold, the adjustment is stopped; otherwise, return to step S15132 to recalculate the time sequence operation parameter adjustment requirement index, determine the new adjustment priority order according to step S15133, and perform parameter adjustment in step S15134. The prediction accuracy is verified again until the prediction accuracy meets the standard.

[0172] Step S15136: Analyze the synchronous running feedback signal set to extract the execution effect data of the multi-device cooperative running control instruction generated by the cooperative control model, and generate a control effect feature set. The control effect feature set includes the matching degree of the parameter change value of the running component after control and the expected control target, and the matching degree of the overall running cooperation degree change value after control and the expected cooperation degree target.

[0173] The running component parameter change value after regulation is the difference between the actual parameter value in the synchronous running feedback signal set and the predicted value in the short-term running state evolution sequence before regulation. The expected regulation target is the difference between the target value in the regulation target and the predicted value before regulation. The matching degree is the ratio of the actual change value to the expected change value. If the ratio is close to 1, the matching degree is high. The matching degree of the overall running coordination change value and the expected coordination target is calculated in a similar manner.

[0174] Step S15137: Based on the regulation effect feature set, the instruction generation logic adjustment requirement index of the coordination regulation model is calculated. The instruction generation logic adjustment requirement index is determined according to the matching degree and the matching stability.

[0175] The matching degree is measured by the average value of the matching degree value. The closer the average value is to 1, the higher the matching degree. The matching stability is measured by the variance of the matching degree value. The smaller the variance, the higher the matching stability. The instruction generation logic adjustment requirement index = γ × (1-matching degree average value) + (1-γ) × matching degree variance, where γ is a weight coefficient. The lower the matching degree average value and the larger the variance, the higher the instruction generation logic adjustment requirement index, indicating that the instruction generation logic needs to be adjusted more urgently.

[0176] Step S15138: According to the instruction generation logic adjustment requirement index, the adjustment priority order of the instruction generation logic in the coordination regulation model is determined. According to the adjustment priority order, the parameter adjustment algorithm of the parameter adjustment calculation layer, the priority allocation rule of the priority allocation layer, and the time period allocation strategy of the time period allocation layer of the coordination regulation model are adjusted in turn.

[0177] Each instruction generation logic module (parameter adjustment calculation layer, priority allocation layer, and time period allocation layer) of the coordination regulation model is associated with each matching degree index in the regulation effect feature set. For example, the parameter adjustment algorithm of the parameter adjustment calculation layer mainly affects the matching degree of the running component parameter change value, and the priority allocation rule of the priority allocation layer mainly affects the overall matching stability of the multi-device coordination regulation. According to the instruction generation logic adjustment requirement index of the regulation effect feature set associated with each module, the instruction generation logic modules are sorted. The higher the adjustment requirement index of a module, the higher the adjustment priority.

[0178] For the parameter adjustment algorithm of the parameter adjustment calculation layer, if the matching degree of the running component parameter change value is low, it may be that the calculation method of the initial parameter adjustment amount or the calculation method of the indirect influence value has a problem. The adjustment algorithm considers the dynamic changes of the coupling coefficient or the normal parameter range boundary of the associated regulation object more accurately.

[0179] When the priority allocation rule of the priority allocation layer is adjusted, if the matching stability is poor, it may be that the priority division is not fine enough or the mutual influence between the regulated objects is not considered, so the adjustment rule is adjusted to increase more priority division dimensions or dynamically adjust the priority according to the correlation strength of the regulated objects.

[0180] When the time period allocation strategy of the time period allocation layer is adjusted, if the regulation execution time period does not match the actual demand, it may be that the time period length or the time period division method is unreasonable, so the adjustment strategy is adjusted to dynamically allocate the time period according to the parameter adjustment difficulty and the required time of the regulated object.

[0181] Step S15139: After adjusting the instruction generation logic of the cooperative regulation model, the regulation effect of the adjusted cooperative regulation model is verified using the historical cooperative regulation data set. If the regulation effect does not reach the preset effect threshold, the instruction generation logic adjustment requirement index is recalculated, and the instruction generation logic is iteratively adjusted until the regulation effect meets the standard.

[0182] The historical cooperative regulation data set includes cooperative regulation demand information, mutual feedback correlation feature set and corresponding regulation effect data when the old instruction generation logic is used in the past. The cooperative regulation demand information and the mutual feedback correlation feature set in the historical cooperative regulation data set are processed using the adjusted instruction generation logic to generate new multi-device cooperative operation regulation instructions, which are then compared with the actual regulation effect data in the historical cooperative regulation data set to calculate the regulation effect index (such as the average value and variance of the matching degree). A preset effect threshold is set, for example, the average value of the matching degree is greater than 0.9 and the variance is less than 0.05. If the adjusted regulation effect index reaches the threshold, the adjustment is stopped; otherwise, return to step S15137 to recalculate the instruction generation logic adjustment requirement index, adjust the instruction generation logic according to step S15138, and verify again until the regulation effect meets the standard.

[0183] Step S151310: The time sequence operation parameter adjustment process of the running state deduction model and the instruction generation logic adjustment process of the cooperative regulation model are integrated to establish an associated mapping relationship between parameter adjustment and logic adjustment.

[0184] The adjustment amount of each time sequence operation parameter of the running state deduction model and the adjustment content of each instruction generation logic module of the cooperative regulation model are recorded, as well as the changes of the corresponding prediction deviation feature set and regulation effect feature set before and after the adjustment. Through correlation analysis, the correlation between the time sequence operation parameter adjustment and the instruction generation logic adjustment is found out, for example, the adjustment of the first time sequence operation layer convolution kernel parameter may affect the adjustment direction of the parameter adjustment calculation layer algorithm.

[0185] Step S151311: The adjusted time sequence operation parameters of the running state deduction model and the instruction generation logic of the cooperative regulation model are associated with the adjustment time stamp for storage.

[0186] The adjusted timing operation parameters and instruction generation logic are stored in a versioned form, each version being associated with a timestamp of the adjustment operation. The storage is in a model parameter library, which contains fields of model version number, adjustment timestamp, running state inference model timing operation parameters, collaborative control model instruction generation logic, corresponding prediction bias feature set summary, and control effect feature set summary, etc., in order to track the evolution process and trace back the historical version of the model.

[0187] Further, the construction process of the running state inference model includes:

[0188] Step S211: Collect a sample synchronous running data set, which contains running component parameter signals, device interaction signals, and environmental comprehensive influence signals under different running scenarios.

[0189] In this embodiment, the collection of the sample synchronous running data set covers multiple running scenarios of the smart medical shelter device, including normal running scenarios, different degree abnormal running scenarios, and simulated fault scenarios, etc. In the normal running scenario, the running data of each device in the shelter under the conventional working load and standard environmental conditions is collected. The different degree abnormal running scenarios include the cases that part of the running component parameters are close to the boundary of the normal range, part of the device interaction appears obvious delay or parameter matching deviation, etc. The simulated fault scenario is generated by artificially setting the fault condition, such as simulating the case that the cooling system efficiency of a certain device is reduced to cause the abnormal rise of the working temperature of the related components. The collection time span of the sample data covers different time periods such as the initial stage of device startup, continuous running peak period, and low load running period, etc. All sample data are obtained through the same sensor network and data aggregation gateway as the real-time data collection, and are strictly time-synchronized and data-verified, and invalid and abnormal data points are removed.

[0190] Step S212: Time sequence division is performed on the sample synchronous running data set, each continuous sample data is divided into a sample input segment and a corresponding sample output segment according to a fixed time length, and the sample output segment is the running data of the subsequent period of the sample input segment.

[0191] The fixed time length is set, and each piece of continuous data in the sample synchronous running data set is divided into a sample input segment and a sample output segment. For example, a piece of continuous sample data can be divided into a plurality of sample pairs, wherein the sample input segment of a sample pair is the running data of a certain period, and the corresponding sample output segment is the running data of the subsequent period of the period. The division is performed in a sliding window manner, and the window sliding step is set according to the sample size. The sample input segment and the sample output segment each contain complete data of the running component parameter signal, the inter-device interaction signal and the environmental comprehensive influence signal, and the time stamps of the two are continuous. After division, each sample pair is labeled to record the corresponding running scene category.

[0192] Step S213: Perform mutual feedback correlation processing on the sample input segment, and extract a sample mutual feedback correlation feature set, which contains sample component parameter coupling features, sample inter-device signal conduction features and sample environmental parameter cross features.

[0193] The same mutual feedback correlation processing procedure as step S120 is performed on each sample input segment. First, separate each running component parameter signal in the sample input segment to extract time domain variation features, separate the inter-device interaction signal to extract conduction features, and separate the environmental comprehensive influence signal to extract differentiated action features. Then, first-level correlation operation is performed to generate first correlation features, and second-level correlation operation is performed to generate second correlation features. The first correlation features and the second correlation features are subjected to dimension unification processing, dynamic weights are assigned and integrated based on feature importance evaluation results, and a sample mutual feedback correlation feature set is generated. In this process, the parameter settings of all feature extraction and correlation operation are consistent with those in real-time processing.

[0194] Step S214: The sample mutual feedback correlation feature set is used as a running state deduction model input sample, and the running component parameter values and the overall running synergy degree values in the sample output segment are used as running state deduction model output samples.

[0195] The sample mutual feedback correlation feature set is directly used as the input sample of the running state deduction model, and its data format is the same as that of the mutual feedback correlation feature set input into the model in real-time processing. The running component parameter values in the sample output segment need to be extracted corresponding to the parameters in the input sample. The overall running synergy degree values are calculated in the same way as in step S134, and the synergy coefficients of different running component parameter values in the sample output segment are obtained, which are used as part of the output sample. The input sample and the output sample are one-to-one corresponding, forming a sample pair required for model training.

[0196] Step S215: The running state deduction model input sample and the running state deduction model output sample are divided into a training sample set, a verification sample set and a test sample set according to a preset ratio.

[0197] The preset proportion is set according to the model training requirement, and the division process adopts a stratified sampling manner to ensure that the sample proportion of different running scene categories in each sample set is consistent with the proportion in the original sample synchronous running data set. After the division is completed, each sample set is independently stored, and data format checking is performed to ensure that the dimensions of the input sample and the output sample are matched.

[0198] Step S216: An initial network structure of the running state deduction model is constructed, and the initial network structure includes a feature preprocessing layer, a first time sequence operation layer, a second time sequence operation layer, a feature abstraction layer, and a prediction output layer.

[0199] The feature preprocessing layer is used for time sequence alignment and standardization processing on the input sample mutual feedback correlation feature set. The first time sequence operation layer adopts a one-dimensional convolutional neural network structure, sets multiple convolution kernels to capture local time sequence correlation information of different time scales, and is connected with a batch normalization layer and an activation function after the convolution layer. The second time sequence operation layer adopts a recurrent unit structure, sets multiple hidden layer neurons to capture long-time time sequence dependence, and adopts a dropout technology to prevent overfitting. The feature abstraction layer is a multilayer perceptron structure, includes multiple fully connected layers, maps a long-time time sequence feature vector to a high-order feature space, and is connected with a batch normalization layer and an activation function after each layer. The prediction output layer includes two parallel fully connected sublayers, one sublayer outputs a running component parameter prediction value of each period, and the other sublayer outputs an overall running coordination degree prediction value, and both adopt a linear activation function.

[0200] Step S217: The initial values of the parameters of each layer of the initial network structure are set, including the initial parameters of the convolution kernel of the first time sequence operation layer, the initial parameters of the recurrent unit of the second time sequence operation layer, and the initial parameters of the perceptron of the feature abstraction layer.

[0201] The initial parameters of the convolution kernel of the first time sequence operation layer adopt an initialization method suitable for a ReLU activation function, and the bias parameters of the convolution kernel are initialized to zero. In the initial parameters of the recurrent unit of the second time sequence operation layer, each weight matrix adopts an initialization method that is helpful to keep the stability of the gradient in the cycle process, the bias parameters are initialized to zero, and the cell state is initialized to a zero vector. In the initial parameters of the perceptron of the feature abstraction layer, the weight matrix adopts an initialization method suitable for an activation function, and the bias parameters are initialized to zero.

[0202] Step S218: The training sample set is input into the initial network structure, the error between the prediction output of the running state deduction model and the output sample of the training sample set is calculated through a back propagation algorithm, and the parameters of each layer are adjusted.

[0203] The mean square error is used as a loss function to calculate the error between the model prediction output and the output sample of the training sample set, and the total loss function is obtained by weighting the parameter error and the synergy error of the running component. Select a suitable optimizer, set the initial learning rate and learning rate decay strategy. During the training process, the training sample set is input into the model in batches. For each batch of data, the prediction output is calculated by forward propagation, and then the gradient of each layer parameter is calculated by the back propagation algorithm, and the parameter is updated using the optimizer.

[0204] Step S219: During the training process, after completing a preset training round, the prediction accuracy of the running state deduction model is evaluated using the validation sample set, the validation accuracy change curve is recorded, and if the validation accuracy has not improved for a preset number of consecutive rounds, the early stopping strategy is used to stop training to avoid overfitting of the running state deduction model.

[0205] The preset training round is set according to the model complexity, and the validation sample set is used for evaluation after completing the preset training round. The validation accuracy is measured by calculating the prediction mean square error of the model on the validation sample set, and the validation accuracy of each evaluation is recorded and the change curve is plotted. The preset consecutive non-improvement rounds, when the validation accuracy of the consecutive preset rounds does not exceed the highest validation accuracy, the training is stopped, and the model parameters with the highest validation accuracy are saved. During the training process, the model checkpoint mechanism is used, and the current model parameters are saved after completing a training round.

[0206] Step S220: The trained running state deduction model is tested for performance using the test sample set, the average prediction error of the running state deduction model on the test sample set is calculated, the prediction accuracy is calculated, and if the average prediction error exceeds the preset test error threshold or the prediction accuracy is lower than the preset test accuracy threshold, the network structure of the running state deduction model is adjusted, including increasing the number of convolution kernels of the first time operation layer, adjusting the number of cycle units of the second time operation layer, and retraining and testing.

[0207] The test sample set is used to evaluate the generalization ability of the model. When calculating the average prediction error, the average absolute error of each running component parameter and the overall running synergy prediction value is calculated, and then the average value is taken. The calculation method of the prediction accuracy is: for each period of prediction value, if it is within the preset error range of the actual value, it is considered to be accurate, and the prediction accuracy is the proportion of the number of periods of accurate prediction to the total number of test periods. The preset test error threshold and the preset test accuracy threshold are set according to the application requirements. If the model performance does not meet the requirements, the network structure is adjusted, such as increasing the number of convolution kernels of the first time operation layer, adjusting the number of cycle units of the second time operation layer, and retraining and testing until the model performance meets the threshold requirements.

[0208] Step S221: operation efficiency optimization is performed on the running state inference model that passes the test, so that the running state inference model completes the generation of the short-term running state evolution sequence within the preset operation time, and the optimized running state inference model and the training parameters, verification results, and test results in the running state inference model construction process are stored in association.

[0209] The operation efficiency optimization includes model pruning and quantization processing. Model pruning removes connections or neurons with small absolute weight values in the network to reduce the number of model parameters and the amount of calculation. Quantization processing converts model parameters from higher-bit floating-point numbers to lower-bit floating-point numbers or integers to reduce the consumption of computing resources. The preset operation time is set according to real-time monitoring requirements. The operation time of the optimized model is measured by running the optimized model in a test environment. If the preset requirement is not met, the pruning threshold or the quantization bit number is further adjusted. The optimized model is stored in the form of a binary file, and the hyperparameters, verification results, and test results in the training process are stored in association with the model file in the model repository to establish a model version management record.

[0210] The construction process of the collaborative regulation model includes:

[0211] Step S311: Collect a sample collaborative regulation data set, which includes sample collaborative regulation demand information, a sample mutual feedback correlation feature set, a sample multi-device collaborative operation regulation instruction, and a corresponding sample synchronous operation feedback signal set.

[0212] The collection of the sample collaborative regulation data set is aimed at the collaborative regulation process of the smart medical shelter device, covers multiple regulation scenarios, including regulation requirements of different emergency levels, regulation objects of different abnormal correlation influence ranges, and scenarios of different regulation targets. The sample collaborative regulation demand information includes different emergency levels, a regulation object list, regulation targets, and time window requirements. The sample mutual feedback correlation feature set is the mutual feedback correlation feature under the corresponding regulation scenario. The sample multi-device collaborative operation regulation instruction is a regulation instruction generated for the sample collaborative regulation demand information. The sample synchronous operation feedback signal set is the feedback data after executing the sample regulation instruction. All sample data are collected through standardized processes to ensure the integrity and accuracy of the data.

[0213] Step S312: Data arrangement is performed on the sample collaborative regulation data set, and the sample collaborative regulation demand information and the corresponding sample mutual feedback correlation feature set are used as the input samples of the collaborative regulation model, and the sample multi-device collaborative operation regulation instruction is used as the output sample of the collaborative regulation model.

[0214] The various types of data in the sample synergistic regulation data set are associated and sorted to ensure that the sample synergistic regulation demand information, the sample mutual feedback correlation feature set, the sample multi-device synergistic operation regulation instruction, and the sample synchronous operation feedback signal set are one-to-one corresponding. The sample synergistic regulation demand information and the corresponding sample mutual feedback correlation feature set are combined into the input sample of the synergistic regulation model, and the sample multi-device synergistic operation regulation instruction is used as the output sample. The input sample and the output sample are data cleaned and format unified for model training.

[0215] Step S313: The sample synergistic regulation demand information in the synergistic regulation model input sample is structured and processed to extract the sample emergency level, the sample regulation object list, the sample regulation target, and the sample time window requirement, and generate structured input features.

[0216] The sample synergistic regulation demand information can be unstructured or semi-structured data, which is converted into features recognizable by the model through structured processing. The sample emergency level is converted into a numerical or categorical feature; the sample regulation object list is converted into an encoded form; the sample regulation target includes the normal range value of each parameter to be adjusted and the synergistic degree value to be restored; the sample time window requirement includes the total time window length and the division of each regulation period. The above extracted information is combined into structured input features.

[0217] Step S314: The sample mutual feedback correlation feature set in the synergistic regulation model input sample is feature screened to extract a sample regulation correlation feature subset corresponding to the sample regulation object list as a correlation input feature, and the structured input feature and the correlation input feature are integrated to form the synergistic regulation model input sample. The parameter adjustment amount, the regulation priority, and the regulation execution period in the sample multi-device synergistic operation regulation instruction are used as the key features of the synergistic regulation model output sample.

[0218] According to the device and component identifiers in the sample regulation object list, the relevant features are screened from the sample mutual feedback correlation feature set to generate a sample regulation correlation feature subset. The structured input feature and the correlation input feature are integrated in the feature dimension to form the input sample of the synergistic regulation model. The parameter adjustment amount, the regulation priority, and the regulation execution period are extracted from the sample multi-device synergistic operation regulation instruction as the key features of the synergistic regulation model output sample.

[0219] Step S315: The complete synergistic regulation model input sample and the synergistic regulation model output sample are divided into a model training set, a model validation set, and a model test set according to a preset ratio.

[0220] The preset proportion is determined according to the training requirement of the collaborative regulation model, and is divided in a stratified sampling manner to ensure that the sample proportions of different regulation scene categories in each sample set are consistent. After the division is completed, each sample set is independently stored and data is verified.

[0221] Step S316: An initial structure of the collaborative regulation model is constructed, and the initial structure includes a requirement analysis layer, an associated feature analysis layer, a parameter adjustment calculation layer, a priority allocation layer, and a time period allocation layer.

[0222] The requirement analysis layer is used to analyze the collaborative regulation requirement information and output structured information such as emergency level and regulation object list. The associated feature analysis layer is used to extract a regulation associated feature subset from a mutual feedback associated feature set. The parameter adjustment calculation layer is used to calculate an initial parameter adjustment amount and a final parameter adjustment amount of the regulation object. The priority allocation layer is used to allocate a regulation priority to the regulation object. The time period allocation layer is used to allocate a regulation execution time period to the regulation object. The layers are connected through a data interface to ensure smooth transmission of information.

[0223] Step S317: Initial parameters of each layer of the initial structure are set, including an initial value of a coupling coefficient of the parameter adjustment calculation layer, an initial value of a weight of the priority allocation layer, and initial parameters of an algorithm of the time period allocation layer.

[0224] The initial value of the coupling coefficient of the parameter adjustment calculation layer is set based on historical data or domain knowledge and can be initialized as an average coupling coefficient obtained by statistical analysis of sample data. The initial value of the weight of the priority allocation layer is set according to the importance of the regulation object and other factors and is initialized as an equal weight or an experience-based allocated weight. The initial parameters of the algorithm of the time period allocation layer include a time interval of time period division and an initial rule of regulation execution order, which are set according to a conventional regulation strategy.

[0225] Step S318: The model training set is input into the initial structure, and the error between the regulation instruction generated by the collaborative regulation model and the output sample of the model training set is calculated by a gradient descent algorithm to adjust the parameters of each layer.

[0226] A suitable loss function, such as a mean square error of a regulation parameter adjustment amount or a ranking loss of a regulation priority, is used to calculate the error between the regulation instruction generated by the model and the output sample of the model training set. A gradient descent algorithm and its variants are selected as the optimizer, and an initial learning rate and a training batch size are set. The model training set is input into the initial structure in batches, and the regulation instruction is generated by forward propagation. After the error is calculated, the parameters of each layer are adjusted by backward propagation.

[0227] Step S319: During the training process, after each preset training batch is completed, the matching degree between the control instructions generated by the collaborative control model and the output samples of the validation set is evaluated using the matching degree evaluation algorithm on the model validation set. The trend of matching degree change is recorded. If the matching degree does not improve for consecutive preset batches, the weight of the loss function of the collaborative control model is adjusted and retraining is performed until the matching degree reaches the preset validation matching threshold.

[0228] Preset training batches are set according to the model training progress. After each preset training batch is completed, the model is evaluated using a validation set. The matching degree evaluation algorithm is used to measure the similarity between the control instructions generated by the model and the output samples of the validation set, and can be evaluated from multiple dimensions such as parameter adjustment amount, priority, and execution time. The matching degree of each evaluation is recorded and the trend of change is observed. A preset batch with no improvement is set. When the matching degree of consecutive preset batches does not improve, the weights of each part of the loss function are adjusted, such as increasing the weight of parameter adjustment error, and training is re-run until the matching degree reaches the preset validation matching threshold.

[0229] Step S320: Test the trained collaborative regulation model using the model test set, calculate the synergy compliance rate and regulation effect compliance rate of the regulation instructions generated by the collaborative regulation model on the test set. If the synergy compliance rate is lower than the preset synergy threshold or the regulation effect compliance rate is lower than the preset effect threshold, adjust the network structure of the collaborative regulation model, including increasing parameters to adjust the computational dimension of the computation layer, optimizing the computational logic of the priority allocation layer, and retraining and testing.

[0230] The model test set is used to evaluate the generalization performance of the collaborative regulation model. The synergy achievement rate is the proportion of coordinated actions among the regulated objects in the generated regulation instructions, such as reasonable execution order and no conflicting parameter adjustments. The regulation effect achievement rate is the proportion of the regulation objectives achieved after the regulation instructions are executed. Preset synergy thresholds and preset effect thresholds are set according to actual regulation needs. If the model fails to meet the test criteria, the network structure is adjusted, such as increasing the computational dimension of the parameter adjustment layer to consider more factors, optimizing the computational logic of the priority allocation layer to improve the accuracy of priority allocation, and then retraining and testing.

[0231] Step S321: Optimize the logic of the tested collaborative control model, add an exception handling branch, so that the collaborative control model can generate reasonable control instructions when the control demand information is incomplete or the mutual feedback correlation feature set is abnormal. Then, store the optimized collaborative control model together with the training parameters, verification results and test results in the collaborative control model construction process.

[0232] The logic optimization includes increasing abnormal situation processing mechanism, such as when the regulation object list in the regulation demand information is incomplete, the model can automatically supplement the possible regulation object according to the inter-feeding correlation feature set; when the inter-feeding correlation feature set has an abnormal value or is missing, the model can use a fault-tolerant mechanism for feature processing. The abnormal situation processing branch is realized by adding the corresponding judgment logic and processing algorithm in the model. The optimized collaborative regulation model is associated with the parameters, verification results, test results, etc. in the training process, and a model archive is established.

[0233] It should be noted that the foregoing description of the embodiments of the application has been presented for the purposes of simplifying the disclosure of the present application and aiding in the understanding of one or more embodiments of the application, and, as such, various features can be combined into a single embodiment, figure or description of the embodiments of the application.

Claims

1. A method for monitoring the running state of a smart shelter equipment based on machine learning, characterized in that, The method comprises: acquiring a synchronous operation real-time signal set, the synchronous operation real-time signal set comprising real-time parameter signals of each operation component, inter-device interaction signals and shelter environment comprehensive influence signals; performing mutual feedback correlation processing on the synchronous operation real-time signal set, extracting a mutual feedback correlation feature set in the operation signal, the mutual feedback correlation feature set comprising component parameter coupling features, inter-device signal conduction features and environment parameter cross features; inputting the mutual feedback correlation feature set into an operation state deduction model, generating a short-term operation state evolution sequence through time sequence feature progressive operation, the short-term operation state evolution sequence comprising predicted values of each time period operation component parameter and predicted values of overall operation coordination degree; identifying a potential abnormal evolution trend based on the short-term operation state evolution sequence, determining an abnormal correlation influence range in combination with the mutual feedback correlation feature set, and generating a cooperative control demand information; inputting the cooperative control demand information and the mutual feedback correlation feature set into a cooperative control model, generating a multi-device cooperative operation control instruction, sending the multi-device cooperative operation control instruction to a corresponding control module, collecting a synchronous operation feedback signal set after execution, inputting the synchronous operation feedback signal set into the mutual feedback correlation processing link, updating the extraction rule of the mutual feedback correlation feature set, and simultaneously adjusting the time sequence operation parameters of the operation state deduction model and the instruction generation logic of the cooperative control model. 2.The method of claim 1, wherein, The mutual feedback correlation processing on the synchronous operation real-time signal set, extracting a mutual feedback correlation feature set in the operation signal, comprises: separating the real-time parameter signals of each operation component from the synchronous operation real-time signal set, extracting the time domain variation features of each operation component parameter, the time domain variation features comprising parameter fluctuation period features and parameter variation rate features; separating the inter-device interaction signals in the synchronous operation real-time signal set, extracting the conduction features of the inter-device interaction signals, the conduction features comprising signal transmission delay features and interaction parameter matching deviation features; separating the shelter environment comprehensive influence signals in the synchronous operation real-time signal set, extracting the differential action features of the environment signals on different operation component parameters, the differential action features comprising influence amplitude features and influence lag features of the environment parameters on the component parameters; performing first-level correlation operation on the time domain variation features and the conduction features, calculating the coupling coefficients of different operation component parameters in the same device and the inter-device interaction signals, and generating first correlation features; performing second-level correlation operation on the first correlation features and the differential action features, calculating the indirect influence coefficients of the environment signals on the operation component parameters through inter-device interaction, and generating second correlation features; after performing dimension unification processing on the first correlation features and the second correlation features, based on the feature importance evaluation results, assigning dynamic weights to the first correlation features and the second correlation features after dimension unification, integrating the weighted first correlation features and the second correlation features, and generating a mutual feedback correlation feature set comprising component parameter coupling features, inter-device signal conduction features and environment parameter cross features; The collection time period information of the inter-feedback correlation feature set and the real-time signal set in synchronous operation is associated, and a core correlation feature subset in the inter-feedback correlation feature set is extracted, the core correlation feature subset being a feature combination whose influence weight on operation state prediction exceeds a preset weight threshold. 3.The method of claim 1, wherein, The inter-feedback correlation feature set is input into an operation state deduction model to generate a short-term operation state evolution sequence through time sequence feature progressive operation, including: The inter-feedback correlation feature set is input into a feature preprocessing layer of the operation state deduction model, time sequence alignment processing is performed on the inter-feedback correlation feature set, the time sequence aligned inter-feedback correlation feature set is transmitted to a first time sequence operation layer of the operation state deduction model, local time sequence correlation information of the inter-feedback correlation feature set is extracted through convolution operation to generate a local time sequence feature vector; The local time sequence feature vector is transmitted to a second time sequence operation layer of the operation state deduction model, long-time correlation modeling is performed on the local time sequence feature vector through a loop unit to capture the dependency of the local time sequence feature vector within continuous time steps, and a long-time time sequence feature vector is generated; The long-time time sequence feature vector is transmitted to a feature abstraction layer of the operation state deduction model, high-order feature abstraction is performed on the long-time time sequence feature vector through a multilayer perceptron to generate a high-order operation state feature vector; The high-order operation state feature vector is transmitted to a prediction output layer of the operation state deduction model, the high-order operation state feature vector is converted into operation component parameter prediction values of each time period through linear mapping operation, and the coordination coefficients of different operation component parameter prediction values in each time period are calculated to generate overall operation coordination degree prediction values in each time period; The operation component parameter prediction values in each time period and the overall operation coordination degree prediction values in the corresponding time period are arranged in time sequence to form an initial short-term operation state evolution sequence; Trend smoothing processing is performed on the initial short-term operation state evolution sequence, key time node features in the smoothed short-term operation state evolution sequence are extracted, and the key time node features are stored in association with the smoothed short-term operation state evolution sequence, the key time node features being features corresponding to time nodes at which operation component parameter prediction values or overall operation coordination degree prediction values change significantly.

4. The machine learning based SHELTER equipment operation state monitoring method of claim 3, wherein, The construction process of the operation state deduction model includes: A sample synchronous operation data set is collected, the sample synchronous operation data set containing operation component parameter signals, device interaction signals and environmental comprehensive influence signals in different operation scenarios; The sample synchronous operation data set is time sequence divided, each continuous sample data being divided into a sample input segment and a corresponding sample output segment according to a fixed time length, the sample output segment being operation data in a subsequent time period of the sample input segment; Inter-feedback correlation processing is performed on the sample input segment to extract a sample inter-feedback correlation feature set, the sample inter-feedback correlation feature set containing sample component parameter coupling features, sample device signal conduction features and sample environmental parameter cross features; The sample inter-feedback correlation feature set is taken as an operation state deduction model input sample, and operation component parameter values and overall operation coordination degree values in the sample output segment are taken as operation state deduction model output samples; and The running state inference model input samples and the running state inference model output samples are divided into a training sample set, a verification sample set and a test sample set according to a preset ratio; An initial network structure of the running state inference model is constructed, and the initial network structure includes a feature preprocessing layer, a first time sequence operation layer, a second time sequence operation layer, a feature abstraction layer and a prediction output layer; Initial values of parameters of each layer of the initial network structure are set, including initial parameters of a convolution kernel of the first time sequence operation layer, initial parameters of a cycle unit of the second time sequence operation layer and initial parameters of a perception machine of the feature abstraction layer; The training sample set is input into the initial network structure, and an error between a prediction output of the running state inference model and an output sample of the training sample set is calculated by using a back propagation algorithm, and parameters of each layer are adjusted; During the training process, after a preset training round is completed, the prediction accuracy of the running state inference model is evaluated by using the verification sample set, a verification accuracy change curve is recorded, if the verification accuracy does not improve for a continuous preset round, an early stopping strategy is adopted to stop the training, and overfitting of the running state inference model is avoided; Performance of the trained running state inference model is tested by using the test sample set, an average prediction error of the running state inference model on the test sample set is calculated, a prediction accuracy is calculated, if the average prediction error exceeds a preset test error threshold or the prediction accuracy is lower than a preset test accuracy threshold, a network structure of the running state inference model is adjusted, including increasing a number of convolution kernels of the first time sequence operation layer, adjusting a number of cycle units of the second time sequence operation layer, and retraining and testing are performed again; The running state inference model that passes the test is optimized in operation efficiency, the running state inference model generates a short-term running state evolution sequence within a preset operation time, and the optimized running state inference model and training parameters, verification results and test results in the running state inference model construction process are stored in association.

5. The machine learning based SHEL method of claim 1, wherein, The short-term running state evolution sequence is used to identify a potential abnormal evolution trend, a mutual feedback correlation feature set is used to determine an abnormal correlation influence range, cooperative control demand information is generated, including: Running component parameter prediction values of each period in the short-term running state evolution sequence are extracted, compared with a preset normal running parameter range, and running component parameter prediction values and corresponding periods that exceed the normal running parameter range are marked, a number of periods in which each running component exceeds the normal parameter range in the short-term running state evolution sequence is counted, an abnormal period proportion is calculated, and the abnormal period proportion is a ratio of the number of periods exceeding the normal parameter range to a total number of prediction periods; Overall running cooperation degree prediction values of each period in the short-term running state evolution sequence are extracted, compared with a preset normal cooperation degree range, and overall running cooperation degree prediction values and corresponding periods that are lower than the normal cooperation degree range are marked, a number of periods in which the overall running cooperation degree prediction values are lower than the normal cooperation degree range is counted, and a cooperation abnormal period proportion is calculated; If the abnormal period proportion of any one running component exceeds a preset component abnormal threshold, or the cooperation abnormal period proportion of the overall running cooperation degree exceeds a preset cooperation abnormal threshold, it is determined that there is a potential abnormal evolution trend. The change curve of the component parameter prediction value corresponding to the operation component with the potential abnormal evolution trend is extracted, the change slope and change acceleration of the component parameter prediction value change curve are analyzed, and the rate characteristic of abnormal evolution is determined; The other operation components having a coupling relationship with the operation component with the potential abnormal evolution trend are found out by combining the component parameter coupling characteristics in the mutual feedback correlation characteristic set, and a directly related component list is formed; The other devices having a correlation with the device with the potential abnormal evolution trend through signal conduction are found out by combining the inter-device signal conduction characteristics in the mutual feedback correlation characteristic set, and an indirectly related device list is formed; The abnormal correlation influence range is determined by integrating the directly related component list and the indirectly related device list, and the abnormal correlation influence range includes the operation components and devices affected by the potential abnormality; According to the rate characteristic of abnormal evolution, the size of the abnormal correlation influence range and the abnormal degree of the overall operation synergy, the emergency level of the collaborative control is determined, and the number of operation components and devices that need to participate in the control is determined according to the emergency level of the collaborative control and the abnormal correlation influence range, and a control object list is generated; Based on the emergency level, the control object list and the rate characteristic of abnormal evolution, a control target is generated, and the control target includes the parameter value that needs to be adjusted to the normal range and the synergy value that needs to be restored; The emergency level, the control object list and the control target are integrated to generate a collaborative control demand information, and the collaborative control demand information also includes the time window requirement of the control implementation.

6. The machine learning based SHEL method of claim 1, wherein, The collaborative control demand information and the mutual feedback correlation characteristic set are input into the collaborative control model to generate a multi-device collaborative operation control instruction, which includes: The demand analysis layer of the collaborative control model is input with the collaborative control demand information, and the emergency level, the control object list, the control target and the time window requirement are obtained by analysis; The correlation characteristic analysis layer of the collaborative control model is input with the mutual feedback correlation characteristic set, the component parameter coupling characteristics, the inter-device signal conduction characteristics and the environmental parameter cross characteristics corresponding to the control object list are extracted, and a control correlation characteristic subset is generated; The parameter adjustment calculation layer of the collaborative control model is input with the control correlation characteristic subset and the control target obtained by analysis, the initial parameter adjustment amount of each control object is calculated, and the indirect influence value of the initial parameter adjustment amount on the related control objects is calculated by combining the coupling coefficients in the mutual feedback correlation characteristic set, and whether the indirect influence value exceeds the normal parameter range is analyzed; If the indirect influence value exceeds the normal parameter range, the initial parameter adjustment amount is adjusted, the indirect influence value is recalculated until the indirect influence value is within the normal parameter range, and the final parameter adjustment amount is obtained; According to the emergency level in the collaborative control demand information, the final parameter adjustment amount of each control object is assigned a control priority; According to the time window requirement in the collaborative control demand information, the final parameter adjustment amount of each control object is assigned a control execution time period; According to the control priority and the control execution time period, a control execution sequence is generated, and the identification, the final parameter adjustment amount, the control execution time period and the control execution sequence of each control object are integrated to generate a single-object control instruction. Arranging all single-object regulation instructions in a regulation execution order, generating multi-device collaborative operation regulation instructions, and adding regulation batch identification and timeliness identification to the multi-device collaborative operation regulation instructions, the timeliness identification being used to indicate an effective execution time period of the multi-device collaborative operation regulation instructions; Extracting key regulation parameters in the multi-device collaborative operation regulation instructions, and generating a regulation parameter abstract, the regulation parameter abstract including a core adjustment amount and a key execution time period.

7. The machine learning based SHELTER equipment operating condition monitoring method of claim 6, wherein, The construction process of the collaborative regulation model includes: Collecting a sample collaborative regulation data set, the sample collaborative regulation data set including sample collaborative regulation demand information, a sample mutual feedback correlation feature set, sample multi-device collaborative operation regulation instructions, and a corresponding sample synchronous operation feedback signal set; Data arrangement is performed on the sample collaborative regulation data set, the sample collaborative regulation demand information and the corresponding sample mutual feedback correlation feature set are taken as input samples of the collaborative regulation model, and the sample multi-device collaborative operation regulation instructions are taken as output samples of the collaborative regulation model; The sample collaborative regulation demand information in the input samples of the collaborative regulation model is subjected to structured processing, and a sample emergency level, a sample regulation object list, a sample regulation target, and a sample time window requirement are extracted to generate structured input features; The sample mutual feedback correlation feature set in the input samples of the collaborative regulation model is subjected to feature screening, a sample regulation correlation feature subset corresponding to the sample regulation object list is extracted as correlation input features, and the structured input features and the correlation input features are integrated to form the input samples of the collaborative regulation model, and the parameter adjustment amount, the regulation priority, and the regulation execution time period in the sample multi-device collaborative operation regulation instructions are taken as key features of the output samples of the collaborative regulation model; The complete input samples of the collaborative regulation model and the output samples of the collaborative regulation model are divided into a model training set, a model validation set, and a model test set according to a preset ratio; An initial structure of the collaborative regulation model is constructed, the initial structure including a demand analysis layer, a correlation feature analysis layer, a parameter adjustment calculation layer, a priority allocation layer, and a time period allocation layer; Initial parameters of each layer of the initial structure are set, including an initial value of a coupling coefficient of the parameter adjustment calculation layer, an initial value of a weight of the priority allocation layer, and initial parameters of an algorithm of the time period allocation layer; The model training set is input into the initial structure, an error between the output of the collaborative regulation model and the output samples of the model training set is calculated through a gradient descent algorithm, and the parameters of each layer are adjusted, in the training process, after a preset training batch is completed, the model validation set is used to evaluate the matching degree between the regulation instructions generated by the collaborative regulation model and the output samples of the validation set through a matching degree evaluation algorithm, the matching degree change trend is recorded, if the matching degree does not improve for a continuous preset batch, the loss function weight of the collaborative regulation model is adjusted, and the training is performed again, until the matching degree reaches a preset validation matching threshold. The trained collaborative regulation model is tested using a model test set, and the compliance rate of the collaborative nature and the compliance rate of the regulation effect of the regulation instructions generated by the collaborative regulation model on the test set are calculated. If the compliance rate of the collaborative nature is lower than a preset collaborative threshold or the compliance rate of the regulation effect is lower than a preset effect threshold, the network structure of the collaborative regulation model is adjusted, including increasing the operation dimension of the parameter adjustment calculation layer and optimizing the calculation logic of the priority allocation layer, and retraining and testing are performed; The test-passed collaborative regulation model is logically optimized, and an abnormal situation processing branch is added, so that the collaborative regulation model can generate reasonable regulation instructions when the regulation demand information is incomplete or the mutual feedback correlation feature set is abnormal. The optimized collaborative regulation model and the training parameters, verification results, and test results in the collaborative regulation model construction process are stored.

8. The machine learning based SHEL method of claim 1, wherein, The synchronous operation feedback signal set is input into the mutual feedback correlation processing link, and the extraction rule of the mutual feedback correlation feature set is updated, including: Analyzing the synchronous operation feedback signal set, extracting the deviation data of the short-term operation state evolution sequence prediction value of the operation state deduction model and the actual operation state value in the synchronous operation feedback signal set, generating a prediction deviation feature set, the prediction deviation feature set includes component parameter prediction deviation values and overall operation collaboration degree prediction deviation values; Performing feature decomposition processing on the prediction deviation feature set, extracting the fluctuation period feature and the change rate feature of the component parameter prediction deviation value, and the time sequence change trend feature of the overall operation collaboration degree prediction deviation value; Correlation analysis is performed on the fluctuation period feature and the change rate feature of the component parameter prediction deviation value and the real-time parameter signal in the synchronous operation feedback signal set, and the deviation sensitivity factor of each operation component parameter is calculated, the deviation sensitivity factor is used to quantify the contribution degree of parameter fluctuation to prediction deviation; Correlation analysis is performed on the time sequence change trend feature of the overall operation collaboration degree prediction deviation value and the inter-device interaction signal and the shelter environment comprehensive influence signal in the synchronous operation feedback signal set, and the collaborative deviation conduction path feature is extracted, the collaborative deviation conduction path feature includes the delay feature of the inter-device interaction signal and the indirect influence feature of the environmental parameter on the collaboration degree; Based on the deviation sensitivity factor and the collaborative deviation conduction path feature, a feature correlation strength evaluation matrix is constructed, which is used to describe the correlation closeness between different types of signal features; According to the feature correlation strength evaluation matrix, the extraction rule of the mutual feedback correlation feature set is adjusted, including: For the extraction rule of the component parameter coupling feature, the coupling coefficient calculation dimension between the operation component parameters whose deviation sensitivity factors exceed the preset threshold is increased; For the extraction rule of the inter-device signal conduction feature, the weight allocation of the inter-device interaction signal delay feature in the collaborative deviation conduction path feature is strengthened; For the extraction rule of the environmental parameter cross feature, a dynamic evaluation mechanism of the indirect influence feature of the environmental parameter on the collaboration degree is introduced; The extraction rule of the adjusted mutual feedback correlation feature set and the collection time period information of the synchronous operation feedback signal set are stored in association to form a feature extraction rule library that is dynamically updated over time; The updated mutual feedback correlation feature set is verified using historical synchronous operation data sets, the prediction deviation change rate before and after the feature extraction rule adjustment is calculated, if the prediction deviation change rate does not reach the preset optimization threshold, the feature correlation strength evaluation matrix construction and the extraction rule adjustment are re-performed until the preset optimization threshold is met. 9.A machine learning based smart shelter equipment operation state monitoring system, characterized in that, Comprise: A processor; A machine readable storage medium for storing machine executable instructions of the processor; Wherein the processor is configured to execute the machine executable instructions to perform the machine learning based intelligent shelter equipment operation state monitoring method of any one of claims 1 to 8.

10. A computer program product, characterised in that, The computer program product comprises machine executable instructions stored in a computer readable storage medium, and a processor of a computer device reads the machine executable instructions from the computer readable storage medium, and the processor executes the machine executable instructions, so that the computer device executes the machine learning based intelligent shelter equipment operation state monitoring method of any one of claims 1 to 8.

Images