A machine learning-based medical device failure detection method and system

Abstract

The application provides a kind of medical equipment fault detection method and system based on machine learning, it is related to fault detection technical field.The method comprises: monitoring motor start-stop division operation cycle, vibration, frequency difference, voltage signal is synchronously collected and processed, feature difference is calculated to generate dynamic sequence, non-convergent trend and phase shift are extracted, voltage and frequency fluctuation characteristics are fused, threshold rule is determined to judge anomaly, and light fault early warning signal is output.In the application, by accurately monitoring the motor start-stop node, the cycle is divided and the vibration, frequency and voltage signals are collected, combined with time alignment and noise filtering, the signal quality is improved, the feature difference is calculated across the cycle, the non-convergent trend and phase shift are extracted, the ability to capture subtle abnormalities is enhanced, the multi-dimensional indicators such as voltage fluctuation and frequency offset are fused, the threshold judgment mechanism in the continuous window is constructed, the stable identification and early warning of light fault of equipment are realized, and the detection accuracy and response time are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault detection, and particularly relates to a medical equipment fault detection method and system based on machine learning. BACKGROUND

[0002] The technical field of fault detection includes means and methods for monitoring, identifying and determining whether the running state of various devices or systems is abnormal. The core content of this technical field is to discover potential fault signs, determine fault types and locations, and assist maintenance decisions by collecting and analyzing various key parameters in the running process of the device. The data collection methods involved in this field include sensor signal collection, operation log recording and environmental parameter acquisition, and the analysis methods include rule-based judgment mechanism and data-based modeling method. With the increasing complexity of device running state, fault detection has gradually developed from traditional static detection to intelligent analysis process relying on large-scale data mining and modeling, covering multiple sub-directions, including industrial equipment diagnosis, transportation tool health management and medical instrument monitoring, and gradually tending to automation, real-time and systematization.

[0003] Among them, the medical equipment fault detection method and system based on machine learning refers to the use of statistical learning theory and training model construction method to classify and model the structured data signals such as voltage, current, vibration frequency and temperature collected in the running process of medical equipment, and to use the trained discriminant model to identify the state of real-time data. The patent subject covers the steps of standardizing the historical running data, using supervised learning algorithm to complete the training process of multi-class classification model, and using the model to input the newly collected data and judge the state label. In specific implementation, the method constructs a training sample set based on device type characteristics and fault working condition labels, uses decision tree or support vector machine to complete discriminant boundary learning, and continuously collects real-time signals in the running process of the device, and judges whether the current state belongs to the fault class label through the model.

[0004] The existing technology based on the traditional modeling method in the fault detection process generally relies on the overall modeling and training process of a large amount of historical structured data. When facing the non-stable fluctuation state in actual operation, it is difficult to quickly respond to the transient behavior of the device operation due to the lack of fine division of the state change in the period. In the actual collection process, different types of sensing signals have collection delay, timestamp error and data drift, and the period granularity cannot be accurately aligned, resulting in unstable feature expression and affecting the real-time performance of abnormal identification. Most methods use static feature input form, lack of tracking mechanism for feature evolution process, and it is difficult to find subtle but continuous trend changes in multiple periods, causing some minor abnormalities to be ignored due to insignificant feature differences. In the state discrimination logic, the existing method focuses on the judgment of static classification labels and ignores the correlation and time sequence logic between signals, resulting in insufficient sensitivity to early-stage abnormalities, and often failing to trigger early warning when the fault occurs. For example, when the voltage is slightly unstable or the vibration frequency slowly drifts, the static model cannot intervene in time due to not meeting the preset threshold, delaying the device maintenance rhythm and increasing the risk exposure time. The above problems result in insufficient robustness of the existing scheme in actual operation, affecting the accuracy of fault prediction and the efficiency of clinical guarantee. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and to provide a medical device fault detection method and system based on machine learning.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] A medical device fault detection method based on machine learning, comprising the following steps:

[0008] S1: Monitor the start-stop nodes of the medical device drive, divide the structure movement into multiple running periods, synchronously collect the vibration signal, frequency peak value and voltage signal in each period, and perform timestamp alignment and noise filtering on the multi-source signal to generate a multi-channel synchronous signal set;

[0009] S2: Calculate the feature average difference in adjacent periods by calling the frequency peak value difference, vibration peak interval and voltage standard deviation of each period in the multi-channel synchronous signal set, and construct a dynamic feature sequence;

[0010] S3: Calculate the non-convergence of frequency fluctuation and the phase shift of vibration interval by calling the frequency peak value difference and vibration peak interval in the dynamic feature sequence, and generate non-convergence trend markers and phase shift markers;

[0011] S4: merge the non-convergent trend mark and phase offset mark with the voltage standard deviation in the time window, extract the Pearson correlation coefficient of the voltage fluctuation and the frequency fluctuation, the cumulative offset of the vibration interval offset, and generate a multi-dimensional fusion feature table;

[0012] S5: construct a threshold trigger rule based on the multi-dimensional fusion feature table, judge the feature data state in the time window, and generate a device light fault early warning signal when the abnormal result of the continuous window reaches the equipment operation statistical control limit.

[0013] Optionally, the multi-channel synchronous signal set includes vibration signals, frequency peaks, and voltage signals, the dynamic feature sequence includes frequency peak differences, vibration peak interval, and voltage standard deviation, the non-convergent trend mark and phase offset mark include frequency fluctuation non-convergence and vibration interval phase offset, and the multi-dimensional fusion feature table includes the Pearson correlation coefficient of the voltage fluctuation and the frequency fluctuation and the cumulative offset of the vibration interval offset. The device light fault early warning signal includes an abnormal judgment result that meets the statistical process control limit.

[0014] Optionally, the specific steps of S1 are as follows:

[0015] S101: Monitor the motor drive start-stop node of the medical device, record the rising edge and falling edge of the control voltage signal, divide the running period of the device structure movement, assign the period number and start and end time stamp, and obtain the period time interval value;

[0016] S102: Based on the period time interval value, synchronously collect vibration signals, frequency peaks, and voltage signals in multiple periods, uniformly label the sensing channels, align the time stamps and establish data mapping, and obtain a period synchronous signal set;

[0017] S103: According to the period synchronous signal set, judge the amplitude increment of the vibration and voltage signals, filter out abnormal values according to the peak noise amplitude threshold, reconstruct the signal curve, and generate a multi-channel synchronous signal set.

[0018] Optionally, the specific steps of S2 are as follows:

[0019] S201: Obtain the frequency peak difference, vibration peak interval, and period voltage standard deviation in each period in the multi-channel synchronous signal set, calculate the maximum frequency point difference value, vibration peak time difference average, and voltage signal standard deviation, respectively, and generate a period characteristic value set;

[0020] S202: Call the corresponding characteristic values of the adjacent two periods in the period characteristic value set, calculate the difference of the same characteristic in the continuous period and take the average, and generate a channel characteristic difference sequence;

[0021] S203: Calculate the time variation intensity based on the time sequence difference of the features in the channel feature difference sequence, and construct a channel sequence to generate a dynamic feature sequence.

[0022] Optionally, the specific steps of S3 are:

[0023] S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence by a fixed window, extract the corresponding relationship of the frequency peak value and vibration interval in the window, compare the frequency peak difference and vibration interval value, and obtain the frequency difference trend degree;

[0024] S302: According to the frequency difference trend degree, extract the start and end values of the window frequency peak value sequence, calculate the linear regression slope, compare the slope difference value with the adjacent window, extract the window position exceeding the frequency slope threshold, and obtain the frequency non-convergent fluctuation value;

[0025] S303: Based on the frequency non-convergent fluctuation value, extract the start and end difference value of the corresponding vibration peak interval, judge the deviation from the vibration interval reference value, filter the interval whose deviation exceeds the reference coefficient, and combine the frequency non-convergent window position to obtain the non-convergent trend marker and phase shift marker.

[0026] Optionally, the specific steps of S4 are:

[0027] S401: Based on the non-convergent trend marker and phase shift marker and the voltage standard deviation sequence, align and pair the voltage standard deviation and the two types of marker values by time window, establish a marker strength value table, and then combine it into a marker voltage joint sequence;

[0028] S402: Call the voltage standard deviation and frequency fluctuation value in the marker voltage joint sequence, calculate the fluctuation amount by pairing according to the time label, calculate the Pearson correlation coefficient by time window, and obtain the fluctuation correlation measure value;

[0029] S403: Based on the time label of the fluctuation correlation measure value, extract the position of the phase shift marker in the marker voltage joint sequence, calculate the cumulative offset value of the vibration interval, and combine it with the fluctuation correlation measure value to generate a multi-dimensional fusion feature table.

[0030] Optionally, the specific steps of S5 are:

[0031] S501: Obtain the feature value of each time window in the multi-dimensional fusion feature table, judge the normal or abnormal state of each window according to the set judgment threshold, record the feature state of each time window, and generate a feature state interval value;

[0032] S502: Based on the feature state interval value statistics frequency of abnormal state in continuous time window, the continuity and abnormal distribution are constructed to build a judgment list, the potential failure window meeting the lower limit of device running stability is marked, and a continuous trigger index value is generated;

[0033] S503: According to the comparison between the continuous trigger index value and the device statistical process control limit, the window exceeding the control limit is screened, the light failure trigger mark is generated according to the feature abnormal type, and the device light failure early warning signal is obtained.

[0034] A medical device fault detection system based on machine learning, comprising:

[0035] The signal acquisition module monitors the motor drive start-stop node in the medical device, divides the structure movement into multiple running periods, collects the vibration signal, frequency peak value and voltage signal in each period, aligns the signals in the period according to the time stamp and executes noise filtering, and generates a multi-channel synchronous signal set;

[0036] The feature construction module calls the frequency peak difference, vibration peak interval and voltage standard deviation of the running period in the multi-channel synchronous signal set, calculates the change amplitude according to the average difference of the feature values between adjacent periods, and generates a dynamic feature sequence;

[0037] The trend determination module divides the frequency peak difference and vibration peak interval in the dynamic feature sequence into windows, calculates the linear regression slope of frequency change and the adjacent offset value of vibration interval, respectively judges the non-convergence of frequency change and the phase offset of vibration rhythm, and generates a trend offset mark set;

[0038] The data fusion module merges the data in the time window dimension based on the trend offset mark set and the voltage standard deviation in the dynamic feature sequence, extracts the Pearson correlation coefficient between the voltage standard deviation and the frequency fluctuation in each window, and calculates the cumulative difference value of the phase offset, and correspondingly fuses the two with the time sequence index respectively, to generate a multi-dimensional fusion feature table;

[0039] The fault warning module calls the feature values of the time window in the multi-dimensional fusion feature table, judges the interval with the set fault judgment threshold, screens the interval range meeting the abnormal state condition and exceeding the statistical process control limit value in the continuous time window, and generates a device light failure early warning signal.

[0040] Compared with the prior art, the advantages and positive effects of the present application are:

[0041] In the present application, by accurately monitoring the motor start-stop node, dividing the cycle and collecting vibration, frequency and voltage signals, combining time alignment and noise filtering, improving signal quality, calculating feature differences across cycles, extracting non-convergent trends and phase shifts, enhancing the ability to capture subtle abnormalities, integrating multi-dimensional indicators such as voltage fluctuations and frequency shifts, and constructing a threshold judgment mechanism within a continuous window, the method realizes stable identification and early warning of device light faults, and improves detection accuracy and response timeliness. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 It is a step flowchart of the present application.

[0043] Figure 2 It is a S1 refinement flowchart of the present application.

[0044] Figure 3 It is a S2 refinement flowchart of the present application.

[0045] Figure 4 It is a S3 refinement flowchart of the present application.

[0046] Figure 5 It is a S4 refinement flowchart of the present application.

[0047] Figure 6 It is a S5 refinement flowchart of the present application.

[0048] Figure 7 It is a system module diagram of the present application. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0050] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation of the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0051] Please refer to Figure 1 The embodiment of the present application provides a medical equipment fault detection method based on machine learning, comprising the following steps:

[0052] S1: Monitor the start-stop nodes of the motor drive of the medical equipment, divide the structural movement into multiple operation cycles, synchronously collect the vibration signals, frequency peaks, and voltage signals in each cycle, time-stamp align and noise filter the multiple-source signals collected in the same cycle, and generate a multi-channel synchronous signal set;

[0053] (Frequency peak is a known parameter representing the frequency signal in a single cycle, used to quantify the frequency fluctuation range);

[0054] S2: Call the frequency peak difference, vibration peak interval, and voltage standard deviation of each cycle in the multi-channel synchronous signal set, calculate the average difference of the features in adjacent cycles, and construct a dynamic feature sequence;

[0055] (Vibration peak interval is the time difference between adjacent peaks in the vibration signal, used to represent the periodicity of vibration; Voltage standard deviation is a statistical quantity of the fluctuation amplitude of the voltage signal);

[0056] S3: Call the frequency peak difference and vibration peak interval in the dynamic feature sequence, calculate the non-convergence of frequency fluctuation and the phase shift of vibration interval through window segmentation and linear regression slope, and generate non-convergence trend markers and phase shift markers;

[0057] (The linear regression slope is a standard statistical quantity for measuring the direction of data change in trend analysis, calculated by the least squares method);

[0058] S4: Merge the non-convergence trend markers and phase shift markers with the voltage standard deviation in the dynamic feature sequence according to the time window, extract the Pearson correlation coefficient of voltage fluctuation and frequency fluctuation, and the cumulative offset of vibration interval offset, and generate a multi-dimensional fusion feature table;

[0059] (The Pearson correlation coefficient is a standard statistical quantity for measuring the linear correlation between two groups of data; The cumulative offset is the cumulative sum of the phase shift markers, used to quantify the offset trend);

[0060] S5: Based on the multi-dimensional fusion feature table, construct a threshold triggering rule to judge the feature data state in each time window. When the abnormal judgment result of consecutive windows meets the statistical process control limit value of equipment operation, generate a device light fault warning signal;

[0061] (The threshold triggering rule is a classification logic based on a pre-set threshold; The statistical process control limit is a commonly used process stability judgment standard in the industry, calculated by historical data).

[0062] The multi-channel synchronous signal set includes a vibration signal, a frequency peak value, and a voltage signal, the dynamic characteristic sequence includes a frequency peak value difference, a vibration peak interval, and a voltage standard deviation, the non-convergent trend marker and the phase offset marker include non-convergence of frequency fluctuation and phase offset of vibration interval, and the multi-dimensional fusion feature table includes a Pearson correlation coefficient of voltage fluctuation and frequency fluctuation and a cumulative offset amount of vibration interval offset, and the device light fault early warning signal includes an abnormal judgment result meeting a statistical process control limit.

[0063] Please refer to Figure 2 , and specific steps of S1 are as follows:

[0064] S101: Monitor a motor drive start-stop node of a medical device, record rising and falling edges of a control voltage signal, divide a running period of device structure movement, assign a period number and start and end time stamps, and obtain a period time interval value;

[0065] In the process of monitoring the motor drive start-stop node of the medical device, it is first determined that the motor control mode used is a voltage drive type, a selected data acquisition module is connected to an output port of a controller, and a sampling frequency is set to 10,000 times per second, so that the change of the voltage signal can be captured in real time. A voltage threshold value of the control signal is set, for example, 2.5 volts. If the voltage jumps from below the threshold to the high level, the change time point is recorded as the start time stamp through edge detection. The falling edge recognition rule is set, that is, the moment when the voltage falls from high to low is regarded as the stop time stamp. The two time points are written into the memory buffer area through the embedded interrupt mechanism. If the duration between the two time points exceeds a certain set value, for example, 0.2 seconds, it is confirmed as a complete motor running period. The period number is numbered in time sequence, forming a period identifier. For example, the start and end times of the first period are 0.5 seconds to 2.1 seconds, that is, period one, and the time interval is 1.06 seconds. This is written into a structured array as a period running record. Through continuous monitoring, the start time, end time and duration of multiple periods are obtained. The period number is used for index identification, the time stamp is used as a boundary attribute, and the duration is used as a supplementary parameter. All period records are classified into a database through a unified data structure. In the record structure, the period number is defined as the primary key field, and the time interval attribute field and the time difference field are defined. When a new period record is generated, a write operation is dynamically executed to save the data completely.

[0066] S102: Based on the period time interval value, a vibration signal, a frequency peak value and a voltage signal in multiple periods are synchronously collected, a sensing channel label is unified, time stamps are aligned and data mapping is established, and a period synchronous signal set is obtained;

[0067] According to each calibrated cycle time interval, the data acquisition operation of the multi-channel sensor is started synchronously, first, each sensor is assigned a clear label, for example, the first channel is the vibration signal, the second channel is the frequency signal, and the third channel is the control voltage signal, and the sampling frequency is set uniformly, for example, ten kilohertz, to ensure that the number of data points collected in each cycle is consistent, the number of required sampling points in the cycle is obtained by multiplying the time interval length by the sampling frequency, for example, the cycle length is one second and five hundred milliseconds, and the corresponding sampling point number is fifteen hundred points, in the data acquisition process, each sampling point of each channel is marked with an absolute time stamp, and the starting time of the cycle is added sequentially to build a time-based data axis, through the acquisition device, for example, the embedded main control module, the data stream is uniformly processed, each data point is assigned time and channel label, and the time alignment between channels is completed, the acquisition result is structured and organized as a multi-dimensional array, arranged according to cycle number, channel identification and data point sequence number, through defining mapping rules, any time point and its corresponding channel data can be quickly retrieved, for example, under the condition of cycle number three, starting time five seconds and cycle duration one second and two hundred milliseconds, the data point time axis of the cycle is established at a frequency of ten thousand times per second, to ensure that the data of each channel corresponds one by one at the same time point, and finally the synchronous data of each channel in the whole cycle is written into the data table, each record contains time, cycle number, channel identification and corresponding data value.

[0068] S103: According to the cycle synchronization signal set, the amplitude increment of the vibration and voltage signal is judged, the abnormal value is screened out according to the peak noise amplitude threshold, the signal curve is reconstructed, and the multi-channel synchronous signal set is generated;

[0069] In the processing link of the multi-channel synchronous signal set, the amplitude variation of the vibration signal and the voltage signal in each channel needs to be judged. The amplitude variation of each sampling point is defined as the value of the current point minus the value of the previous point. If the difference exceeds the set noise peak threshold, it is considered that the current data has a mutation. According to the historical data statistics of the average amplitude and the standard deviation of the vibration channel, for example, the average value of the vibration signal is zero point zero five gravity acceleration, and the standard deviation is zero point zero one. The noise threshold is set to be the average value plus three times the standard deviation, that is, zero point zero eight gravity acceleration. All vibration sampling points in the period are traversed, the amplitude difference between adjacent two points is calculated, and compared with the threshold. If the difference value at a certain time is higher than the threshold, it is considered that the point is an abnormal value. The average value of a plurality of normal data points before and after the abnormal value is used to replace the abnormal value. For example, if the amplitude of a point is zero point one three, and the amplitude of the previous point is zero point zero seven, the difference is zero point zero six. If the value exceeds the threshold, the average value of the previous five normal data points is used to replace the current abnormal value. After the vibration channel is repaired, the voltage channel is reconstructed with the same strategy. The repaired vibration and voltage data are kept synchronous in the time dimension to form an updated multi-channel synchronous signal record. The record structure retains the cycle number, sampling time, channel number and repaired data value, and is stored in a special data structure for subsequent use.

[0070] Please refer to Figure 3 The specific steps of S2 are:

[0071] S201: Obtain the frequency peak difference, vibration peak interval and cycle voltage standard deviation in each cycle of the multi-channel synchronous signal set, calculate the maximum frequency point difference value, vibration peak time difference average and voltage signal standard deviation respectively, and generate a cycle characteristic value set;

[0072] In processing the frequency peak difference, vibration peak interval and voltage standard deviation of each cycle in the multi-channel synchronous signal set, the original signal data is first divided according to the set cycle length, for example, if the set cycle length is 100 milliseconds and the sampling frequency is 10 kHz, then each cycle signal contains 1000 data points. For each cycle signal, the frequency spectrum information is extracted by fast Fourier transform, and the frequency component with the maximum amplitude is identified as the frequency peak point. Then the numerical difference of the peak frequency point in different channels is compared to obtain the numerical value of the frequency peak difference. Subsequently, for the vibration signal in the same cycle, the position of the vibration peak point is determined by extracting the local maximum value, and the amplitude exceeding twice the average value can be used as the peak effective criterion. Based on this, the positions of all peak points are extracted, and the time interval between two adjacent peak values is calculated, and the average value of these intervals is taken as the average interval of the vibration peaks in the cycle. For example, if the peak values in a cycle appear at 5 milliseconds, 25 milliseconds, 45 milliseconds, 65 milliseconds and 85 milliseconds, the intervals are 20 milliseconds, and the average is 20 milliseconds. For the voltage signal, the mean value and the variance of the deviation from the mean value are calculated based on the voltage amplitude of each sampling point, and then the square root of the variance is taken as the standard deviation of the voltage signal in the cycle. If the voltage value sequence is 2.1, 2.3, 2.0, 2.4 and 2.2, the mean value is 2.2, the deviation square is 0.01, 0.01, 0.04, 0.04 and 0, the average value is 0.02, and the square root is about 0.14. After extracting the above three characteristics of each cycle in each channel, the characteristics are sorted according to the type of characteristics to form a cycle characteristic value set, which is used for subsequent difference analysis.

[0073] S202: The corresponding characteristic values of the adjacent two cycles in the cycle characteristic value set are called, the difference value of the same characteristic in the continuous cycle is calculated and averaged, and the channel characteristic difference sequence is generated;

[0074] On the basis of the set of periodic characteristic values, corresponding characteristic values in adjacent periods are extracted, a two-dimensional matrix structure of the channel and the period dimension is established, and for each type of characteristic value corresponding to each channel, the numerical difference between adjacent periods is calculated in turn according to the period order. For example, if the voltage standard deviation of a certain channel in five consecutive periods is 0.15, 0.17, 0.20, 0.18 and 0.19, then the difference between adjacent periods is calculated as 0.02, 0.03, 0.02 and 0.01, and then the difference values are averaged to obtain the average change of the voltage standard deviation in consecutive periods, and the average value is 0.02. In order to avoid the influence of abnormal fluctuations on the analysis accuracy, the upper limit of the difference value can be set as a filtering condition, such as by statistical analysis of a large number of historical period samples, more than 90% of the data change range is within 0.05, then the upper limit can be set to 0.1, if a pair of difference values exceeds this range, then the pair of data is excluded from the average calculation. Each type of feature (frequency peak difference, vibration peak interval, voltage standard deviation) is processed in the same way, and finally a sequence of average period differences of three types of features is formed for each channel, that is, a channel feature difference sequence.

[0075] S203: Based on the time sequence difference of the features in the channel feature difference sequence, calculate the time change intensity and aggregate to build a channel sequence to generate a dynamic feature sequence;

[0076] First, create a sequential structure containing time dimension for each channel corresponding to three types of feature difference values, for example, the voltage standard deviation difference value is arranged in time sequence as 0.02, 0.01, 0.03, 0.02, 0.01 in 5 consecutive periods, corresponding to time points t1 to t5 respectively, then calculate the amplitude of numerical change between two adjacent values in time sequence, and obtain a new difference change sequence by taking the absolute value of the difference between adjacent values, if the difference value decreases from 0.02 to 0.01, the change value is 0.01, if it increases from 0.01 to 0.03, the change value is 0.02, and so on, a sequence representing the change intensity of the difference is obtained, after this step, a local analysis window is introduced to continuously aggregate the difference intensity sequence, when the window length is selected as 3, the average operation is performed on every three consecutive change values, for example, the first three change values are 0.01, 0.02, 0.01, the corresponding window aggregation value is 0.013, then slide one position to the right and repeat the operation to obtain the complete sliding average change intensity sequence, then set a reference standard for identifying the strength of fluctuation, the standard can be extracted from a large number of historical feature difference change intensities under normal working conditions, for example, if the change intensity of more than 95% of the samples is less than 0.02, set 0.02 as the fluctuation recognition threshold of this type of feature, compare each window result one by one, if the average value in the window is greater than 0.02, it is determined as a high dynamic change section, if it is less than or equal to 0.02, it is divided into a stable section, perform the above operation for each type of feature, and record the marking results as flag information in the time sequence, for example, the state marks of the voltage feature corresponding to each window may be stable, high dynamic, stable, etc., perform the same processing on the other two types of features such as vibration peak interval and frequency peak difference to obtain the corresponding dynamic state marking sequence, then combine the state of each type of feature at the same time point to form a composite dynamic state point, for example, at a certain time, the voltage is stable, the vibration is stable, and the frequency is high dynamic, the combined state is [stable, stable, high dynamic], concatenate all the combined states in time sequence to form a complete channel dynamic feature sequence, which provides a data basis for subsequent state recognition or behavior modeling.

[0077] Please refer to Figure 4 The specific steps of S3 are:

[0078] S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence by a fixed window, extract the corresponding relationship of frequency peak value and vibration interval in the window, compare the frequency peak difference and vibration interval value, and obtain the frequency difference change trend degree;

[0079] After the dynamic characteristic sequence is acquired, first, the acceleration or displacement signal is subjected to fast Fourier transform (FFT), and in the frequency domain, effective frequency peaks exceeding a set threshold (such as the mean value plus twice the standard deviation) are extracted, and the frequency difference between adjacent peaks is calculated; then in the time domain, the time points at which the vibration amplitude peaks corresponding to the frequency peaks appear are extracted, and the time interval between each pair of adjacent vibration peaks is calculated. Each group of frequency difference and corresponding time interval forms a pair, forming a structured characteristic pair sequence. Then the entire sequence is divided into fixed length sliding windows, all the pairs contained in each window are retained, and the ratio of the frequency difference to the time interval is calculated for each pair, obtaining the ratio set in the window; by aggregating the set (such as calculating the average), the synchronization change trend between the frequency characteristics and the vibration behavior in the time period is measured. By comparing the change amplitude of the average ratio in the adjacent two windows, the stability of the trend is judged, and if the change amplitude exceeds a set threshold (such as 10%), it is considered as a trend mutation area. The method can further depict the change trajectory and distribution state between the frequency difference and the vibration response through the trend evolution of multiple windows, providing quantitative support for dynamic characteristic recognition and state analysis.

[0080] S302: According to the frequency difference change trend degree, the start and end values of the window frequency peak sequence are extracted, the linear regression slope is calculated, the slope difference value of the adjacent window is compared, the window position exceeding the frequency slope threshold is extracted, and the frequency non-convergent fluctuation value is obtained;

[0081] Based on the frequency peak sequence, linear regression analysis is performed on the start and end points of the peaks in each fixed window, the slope of the frequency change trend is calculated, and the slope value sequence of a group of consecutive windows is obtained after the frequency sequence is divided. By comparing the slope difference between adjacent windows, it is judged whether it exceeds the preset mutation judgment threshold (the threshold can be determined by statistics on the slope distribution of normal samples, such as setting the maximum slope value of 90% of the samples as the upper limit), once the slope change amplitude of a window and its adjacent window exceeds the threshold (such as the current window slope is 0.6, and the last one is 0.2, the difference is 0.4, which exceeds the threshold of 0.3), it is determined that the frequency trend in the window has mutated, and the time segment is marked as a frequency fluctuation abnormal area, and is defined as a frequency non-convergent fluctuation value, which means that the frequency peak has a sudden trend change in the time sequence, and no longer tends to be a significant fluctuation characteristic, which is used to identify the key time position of frequency anomaly, and support further analysis and anomaly positioning.

[0082] S303: Based on the frequency non-convergent fluctuation value, the start and end difference value of the corresponding vibration peak interval is extracted, the deviation from the vibration interval reference value is judged, the interval whose deviation exceeds the reference coefficient is screened, and the non-convergent trend mark and phase shift mark are obtained in combination with the frequency non-convergent window position;

[0083] On the basis of the positioned frequency fluctuation window, the occurrence interval of the vibration amplitude peak value in the corresponding time period is further analyzed, the time difference value of all adjacent vibration peaks in the window is extracted, and the interval change amount at the start and end of the window is counted; The change amount is compared with the pre-set reference interval value, which is usually derived from the data analysis results under stable working conditions, for example, the average value of 20 milliseconds between vibration peaks is taken as the reference value, and the allowed fluctuation range is set, such as ±20% of the reference value, that is, the upper and lower limits are 24 milliseconds and 16.67 milliseconds respectively; If the interval change in the window exceeds the threshold value, it is determined that there is a trend deviation, and it is marked as a non-convergent trend section; Then compare the trend section with the position of the frequency non-convergent window, if the overlapping time length accounts for more than 30% of the total length of the window, it is confirmed that the area is a non-convergent trend marker; On this basis, the phase angle of each vibration peak value relative to the ideal cycle position is further calculated, if the phase deviation of a peak value appearing in the reference cycle exceeds the set threshold value (such as 90 degrees), it is marked as a phase deviation point, for example, the reference cycle is 20 milliseconds and the peak value appears at the position of 27 milliseconds, the corresponding deviation angle is about 126 degrees, which exceeds the range, that is, it is considered as a phase anomaly; Finally, the above non-convergent trend marker and phase deviation marker are used as an important basis for identifying the non-steady-state characteristics of the system.

[0084] Please refer to Figure 5 , the specific steps of S4 are:

[0085] S401: Based on the non-convergent trend marker and the phase deviation marker and the voltage standard deviation sequence, align and pair the voltage standard deviation and the two types of marker values according to the time window, establish a marker strength value table, and then combine it into a marker voltage joint sequence;

[0086] The non-converging trend marker needs to be extracted according to the continuous change of the standard deviation in the voltage signal. In operation, the voltage data is segmented by sliding at a fixed time interval, for example, every 10 seconds as a window, the change amplitude of the voltage standard deviation is compared window by window, if the standard deviation change in the last three windows is greater than the set threshold, the current time point is marked as a non-converging trend, the threshold can be set to 0.02 volts of standard deviation amplitude, the phase shift marker is judged by real-time monitoring of the phase value of the voltage signal by the phasor measurement device, when the phase difference between a certain time point and its previous time point is more than 5 degrees, it is considered that the phase shift occurs, the voltage standard deviation sequence is obtained by statistics of the discrete degree of voltage data in each time period, after extraction, the time labels of the three types of data are uniformly processed, and are aligned by seconds, if there are non-converging trend markers and phase shift markers in a certain second, and there is a voltage standard deviation value, the three are paired and combined into a composite record, the marker values are given different weights, for example, the non-converging trend marker is given a weight of 0.6, and the phase shift marker is given a weight of 0.4, a comprehensive marker intensity value is obtained by weighted calculation, and a voltage standard deviation at the same time point is combined to form a marker voltage joint record, the process is repeated to generate continuous time sequence data, in the sequence, each record contains a comprehensive marker intensity value and a voltage standard deviation value, for example, a combination of 0.6 and 0.12 volts at a certain time, and a complete joint sequence is continuously constructed.

[0087] S402: Call the voltage standard deviation and frequency fluctuation value in the marker voltage joint sequence, calculate the fluctuation amount by pairing according to the time label, calculate the Pearson correlation coefficient according to the time window, and obtain the fluctuation correlation degree value;

[0088] When extracting the voltage standard deviation value and the frequency fluctuation value, the data of the same time label needs to be read from the joint sequence and the frequency record respectively, the voltage standard deviation has been obtained in the joint sequence, and the frequency fluctuation needs to be judged by the difference between the frequency values of adjacent time points. For example, if the current frequency is 50.04 Hz and the frequency of the previous second is 50.01 Hz, the frequency fluctuation value is 0.03 Hz. The frequency fluctuation sequence is calculated second by second in this way, and is paired with the voltage standard deviation to form paired data. Then, a fixed time period is divided, for example, every 30 seconds as a window, the overall consistency of all paired data in the window is counted, the linear correlation is judged by using statistical analysis method, and a value representing the correlation degree is obtained. The closer the correlation degree is to 1, the stronger the relationship is, and the closer the correlation degree is to 0, the weaker the relationship is. The correlation value and its time label are combined to form a new fluctuation correlation degree sequence. The data can show its performance in different time windows.

[0089] S403: Based on the time label of the wave correlation measure value, the position of the phase shift marker in the labeled voltage joint sequence is extracted, the cumulative offset value of the vibration interval is calculated, and is combined with the wave correlation measure value to generate a multi-dimensional fusion feature table;

[0090] First, the time information is extracted from the wave correlation measure sequence, for example, the time points of 10 seconds, 40 seconds, 70 seconds, etc. are identified, indicating that there is a certain degree of linear correlation between the voltage standard deviation and the frequency fluctuation in these time periods. Then, records matching the above time points are screened out from the joint sequence, and it is checked whether the records contain phase shift markers. If there is a marker, the current time point is marked as an effective vibration point, and the corresponding phase angle value is recorded. Then, all effective vibration points are arranged in chronological order, and the interval length and phase angle change between adjacent time points are calculated one by one. The interval length is the value of the next time point minus the previous time point, for example, the interval between 40 seconds and 70 seconds is 30 seconds. If the corresponding phase angles are 123° and 115° respectively, the change is 8°. The phase shift increment of this segment is recorded in absolute value. All vibration point pairs are processed in this way to obtain multiple vibration intervals and their corresponding angle change values. Then, these change values are accumulated to obtain the cumulative offset value corresponding to each time period.

[0091] In identifying the phase shift, according to whether the difference between the phase angles of adjacent time points exceeds a set threshold value, for example, set to 5°, if the current time is the 50th second, the previous time is the 49th second, the phase angles are 117° and 111° respectively, the difference is 6°, it is considered that the phase shift occurs, and the current time point is recorded as the effective vibration point; in the subsequent processing process, if the phase angle change value between two vibration points exceeds another set of discrimination values, for example, 10°, the segment is marked as an abnormal offset segment, and the offset value still participates in the subsequent cumulative calculation; such cumulative offset value will be paired with the original fluctuation correlation value according to the time label one by one, forming a multi-dimensional fusion record, each record contains the correlation value between the time point, the voltage standard deviation and the frequency fluctuation, whether there is a phase shift in the time period, the corresponding cumulative offset angle, and whether it belongs to an abnormal offset segment, for example, the record of the 70th second, the fluctuation correlation is 0.93, there is a phase shift, the cumulative offset angle is 12°, and it is marked as an abnormal offset segment, which will be included in the fusion feature table; In addition, the correlation value is derived from the sliding window calculation in the previous process, and the window length is set to 30 seconds, and a group of voltage standard deviation values and frequency fluctuation values are sampled every second, for example, in a certain window, the voltage standard deviation is 0.06, 0.07, 0.08, and the frequency fluctuation is 0.01, 0.02, 0.03, by comparing the overall change direction between the two columns of values, the correlation degree is judged, if the consistency of the value trend is high, it is determined that the correlation is strong, and the strong interval can be set to 0.85 to 1.00, and the weak interval is 0.00 to 0.30, if the correlation value of the current window is 0.93, it is classified into the strong correlation range, and the corresponding cumulative offset value is recorded; The final output fusion feature table continuously covers the entire analysis period, forming a structured data sequence, which is used for subsequent feature modeling or abnormality discrimination process.

[0092] Please refer to Figure 6 , the specific steps of S5 are:

[0093] S501: Obtain the feature value of each time window in the multi-dimensional fusion feature table, judge the normal or abnormal state of each window according to the set judgment threshold, record the feature state of each time window, and generate a feature state interval value;

[0094] The feature values of each time window in the multi-dimensional fusion feature table can be obtained by first collecting the real-time running state of the equipment through multiple sensing channels, such as collecting the temperature (in Celsius), vibration (in millimeters per second), current (in amperes), and other multi-dimensional signals of the fan, and performing noise reduction, normalization, and time alignment processing on the obtained data to fuse the features of each channel under a unified time reference. Each ten minutes is divided into a time window to divide the feature set, the mean value of the temperature can be extracted, the standard deviation of the vibration is extracted, the peak value of the current is extracted, and the like. After the feature extraction is completed, state recognition is performed according to the pre-set judgment threshold, for example, the temperature is set to 85 degrees, the vibration threshold is set to 6, and the current peak value threshold is set to 60. If the temperature is greater than 85, it is marked as abnormal, if the vibration value is higher than 6, it is also marked as abnormal, and if the current exceeds 60, it is considered as abnormal, otherwise it is normal. In each time window, the above features are judged, for example, in a certain ten-minute window, the temperature is 83, the vibration is 7, and the current is 55. The temperature and current are normal, and the vibration is abnormal. The state corresponding to this window can be recorded as "normal, abnormal, normal". In this way, the feature judgment results of each window are recorded in sequence to form a state sequence in time order, and the state of each feature in each window forms a feature state interval value.

[0095] S502: Based on the feature state interval value, the frequency of abnormal state in the continuous time window is counted, a judgment list is constructed according to the continuity and abnormal distribution, and a potential fault window that meets the lower limit of the equipment running stability is marked, and a continuous trigger index value is generated;

[0096] On the basis of the obtained feature state interval value, a fixed number of continuous time windows can be set for sliding statistics, for example, every five windows are a group, the number of abnormal states in which is counted, and each type of feature is counted separately, for example, the number of temperature abnormalities is three, the number of vibration abnormalities is two, and the number of current abnormalities is one. Therefore, the frequency of temperature abnormalities is high in this window group. If the pre-set lower limit of abnormal judgment is three times, that is, any feature is abnormal for not less than three times in five windows, it can be determined as a potential fault window. After each sliding, the window is pushed forward by one, and the next group of statistics is continued, forming the abnormal statistical results corresponding to the continuous multiple window groups. For some window groups, the frequency of abnormalities continuously reaches the set lower limit, that is, the seventh window is marked as an abnormal state in three consecutive groups of statistics, and the continuous trigger number is three. The number is recorded as a continuous trigger index value, and the corresponding window position index is marked to form a judgment list. For example, the seventh, ninth, and tenth windows are marked as potential abnormalities for multiple times, and a continuous trigger index list can be established in sequence as a basis for subsequent light fault judgment.

[0097] S503: According to the comparison of the continuous trigger index value and the equipment statistical process control limit, the windows exceeding the control limit are screened, and the light fault trigger marks are generated according to the characteristic abnormal types to obtain the equipment light fault early warning signal;

[0098] On the basis of the established continuous trigger index value, the control range is statistically calculated according to the historical acquisition data during the equipment operation, such as the average level of all continuous trigger values is one point five, and the fluctuation degree is zero point seven. By setting a reasonable upper limit value, the control threshold value is obtained, for example, by defining the maximum acceptable range by adding three times the fluctuation value to the average level, that is, if the continuous trigger index value exceeds the upper limit value of about three point six, it can be determined that the abnormality is concentrated and the situation is judged, that is, the light fault early warning window is marked. After screening these windows exceeding the limit, according to the feature state record in each window, it can be traced back to which type of feature or which combination of features appears abnormal in each window exceeding the limit, for example, the temperature and vibration state in the ninth window are both abnormal, which is marked as "high temperature aggravating vibration", and if only the vibration is abnormal, it is marked as "high intensity vibration". Each mark needs to be based on the specific combination of abnormal features in the window to form a light fault early warning signal list one by one, which is corresponding to the window index for subsequent correlation analysis and record tracking.

[0099] Please refer to Figure 7 , the medical equipment fault detection system based on machine learning comprises:

[0100] The signal acquisition module monitors the motor drive start-stop nodes in the medical equipment, divides the structure movement into multiple running periods, acquires the vibration signal, frequency peak value and voltage signal in each period, aligns the signals in each period according to the time stamp and performs noise filtering to generate a set of multi-channel synchronous signals;

[0101] The feature construction module calls the frequency peak value difference, vibration peak interval and voltage standard deviation of the running period in the set of multi-channel synchronous signals, calculates the change amplitude according to the average difference of the feature values between adjacent periods, and generates a dynamic feature sequence;

[0102] The trend determination module divides the frequency peak value difference and the vibration peak interval in the dynamic feature sequence into windows, calculates the linear regression slope of the frequency change and the adjacent offset value of the vibration interval, respectively judges the non-convergence of the frequency change and the phase offset of the vibration rhythm, and generates a set of trend offset marks;

[0103] The data fusion module merges the data according to the time window dimension based on the trend offset mark set and the voltage standard deviation in the dynamic feature sequence, extracts the Pearson correlation coefficient between the voltage standard deviation and the frequency fluctuation in each window, and calculates the cumulative difference value of the phase offset, and corresponds to the time sequence index respectively to generate a multi-dimensional fusion feature table;

[0104] The fault early warning module calls the feature value of the time window in the multi-dimensional fusion feature table, performs interval judgment with a set fault judgment threshold, screens the interval range that meets the abnormal state condition and exceeds the statistical process control limit value in the continuous time window, and generates a device light fault early warning signal.

[0105] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any person skilled in the art can use the disclosed technical content to make changes or modifications into equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution content of the present application, and according to the technical essence of the present application, still belong to the protection scope of the technical solution of the present application.

Claims

1. A machine learning based medical device failure detection method, characterized by, The method comprises the following steps: S1: monitoring the start-stop node of the medical equipment drive, dividing the structural motion into multiple running cycles, synchronously collecting the vibration signal, frequency peak value and voltage signal in each cycle, time stamping and aligning the multi-source signals, and generating a multi-channel synchronous signal set; S2: calling the frequency peak value difference, vibration peak interval and voltage standard deviation of each cycle in the multi-channel synchronous signal set, calculating the feature average difference in adjacent cycles, and constructing a dynamic feature sequence; S3: calling the frequency peak value difference and vibration peak interval in the dynamic feature sequence, calculating the non-convergence of frequency fluctuation and the phase offset of vibration interval, and generating a non-convergence trend mark and a phase offset mark; S4: merging the non-convergence trend mark and the phase offset mark with the voltage standard deviation in the dynamic feature sequence according to a time window, extracting the Pearson correlation coefficient of voltage fluctuation and frequency fluctuation, the cumulative offset of vibration interval offset, and generating a multi-dimensional fusion feature table; S5: constructing a threshold trigger rule based on the multi-dimensional fusion feature table, judging the feature data state in the time window, and generating a device minor fault warning signal when the abnormal results of continuous windows reach the statistical control limit of device operation. 2.The machine learning based medical device failure detection method of claim 1, wherein, The multi-channel synchronous signal set comprises vibration signals, frequency peak values and voltage signals, the dynamic feature sequence comprises frequency peak value difference, vibration peak interval and voltage standard deviation, the non-convergence trend mark and the phase offset mark comprise the non-convergence of frequency fluctuation and the phase offset of vibration interval, and the multi-dimensional fusion feature table comprises the Pearson correlation coefficient of voltage fluctuation and frequency fluctuation and the cumulative offset of vibration interval offset. 3.The machine learning based medical device fault detection method of claim 1, wherein, The specific steps of S1 are: S101: monitoring the start-stop node of the medical equipment motor drive, recording the rising edge and falling edge of the control voltage signal, dividing the running cycle of the device structure motion, assigning cycle number and start and end time stamp, and obtaining cycle time interval value; S102: based on the cycle time interval value, synchronously collecting vibration signals, frequency peak values and voltage signals in multiple cycles, uniformly labeling the sensing channels, aligning the time stamps and establishing data mapping, and obtaining a cycle synchronous signal set; S103: according to the cycle synchronous signal set, judging the amplitude increment of vibration and voltage signals, filtering out abnormal values according to the peak noise amplitude threshold, reconstructing the signal curve, and generating a multi-channel synchronous signal set.

4. The machine learning based medical device fault detection method of claim 3, wherein, The specific steps of S2 are: S201: obtaining the frequency peak value difference, vibration peak interval and cycle voltage standard deviation of each cycle in the multi-channel synchronous signal set, respectively calculating the maximum frequency point difference value, vibration peak time difference average and voltage signal standard deviation, and generating a cycle characteristic value set; S202: calling the corresponding characteristic values of the adjacent two cycles in the cycle characteristic value set, calculating the difference of the same characteristic in the channel in the continuous cycles and taking the average, and generating a channel feature difference sequence; S203: based on the time sequence difference of the features in the channel feature difference sequence, calculating the time change intensity and summarizing to construct a channel sequence, and generating a dynamic feature sequence.

5. The machine learning based medical device fault detection method of claim 4, wherein, The specific steps of S3 are: S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence by a fixed window, extract the corresponding relationship between the frequency peak and the vibration interval in the window, compare the frequency peak difference and the vibration interval value, and obtain the frequency difference trend degree; S302: According to the frequency difference trend degree, extract the start and end values of the window frequency peak sequence, calculate the linear regression slope, compare the slope difference value with the adjacent window slope, extract the window position exceeding the frequency slope threshold, and obtain the frequency non-convergent fluctuation value; S303: Based on the frequency non-convergent fluctuation value, extract the start and end difference value of the corresponding vibration peak interval, judge the deviation from the vibration interval reference value, filter the intervals whose deviation exceeds the reference coefficient, and combine the frequency non-convergent window position to obtain the non-convergent trend mark and phase shift mark.

6. The machine learning based medical device fault detection method of claim 5, wherein, The specific steps of S4 are: S401: Based on the non-convergent trend mark and the phase shift mark and the voltage standard deviation sequence, align and pair the voltage standard deviation and the two types of mark values by time window, establish a mark strength value table, and then combine it into a mark voltage joint sequence; S402: Call the voltage standard deviation and frequency fluctuation value in the mark voltage joint sequence, pair and calculate the fluctuation amount according to the time label, calculate the Pearson correlation coefficient according to the time window, and obtain the fluctuation correlation measure value; S403: Based on the time label of the fluctuation correlation measure value, extract the position of the phase shift mark in the mark voltage joint sequence, calculate the cumulative offset value of the vibration interval, and combine it with the fluctuation correlation measure value to generate a multi-dimensional fusion feature table.

7. The machine learning based medical device fault detection method of claim 6, wherein, The specific steps of S5 are: S501: Obtain the feature value of each time window in the multi-dimensional fusion feature table, judge the normal or abnormal state of each window according to the set judgment threshold, record the feature state of each time window, and generate a feature state interval value; S502: Based on the feature state interval value, count the frequency of abnormal state in the continuous time window, construct a judgment list according to the continuity and abnormal distribution, mark the potential fault window that meets the lower limit of the device running stability, and generate a continuous trigger index value; S503: According to the comparison between the continuous trigger index value and the device statistical process control limit, filter the windows exceeding the control limit, and generate a light fault trigger mark according to the feature abnormal type to obtain a device light fault early warning signal.

8. A machine learning based medical device failure detection system, characterized in that, The machine learning-based medical device fault detection method according to any one of claims 1-7, the system comprising: The signal acquisition module monitors the motor drive start-stop nodes in the medical device, divides the structure movement into multiple running periods, collects the vibration signal, frequency peak and voltage signal in each period, aligns the signals in the period by time stamp and performs noise filtering to generate a multi-channel synchronous signal set; The feature construction module calls the frequency peak difference, vibration peak interval and voltage standard deviation of the running period in the multi-channel synchronous signal set, calculates the change amplitude according to the average difference of the feature values between adjacent periods, and generates a dynamic feature sequence; The trend determination module performs window segmentation according to the difference between frequency peak values and the vibration peak interval in the dynamic characteristic sequence, calculates the linear regression slope of frequency variation and the adjacent offset value of vibration interval, respectively determines the non-convergence of frequency variation and the phase offset of vibration rhythm, and generates a trend offset label set; The data fusion module performs data merging in the time window dimension based on the trend offset label set and the voltage standard deviation in the dynamic characteristic sequence, extracts the Pearson correlation coefficient between the voltage standard deviation and the frequency fluctuation in each window, calculates the cumulative difference value of the phase offset, and correspondingly fuses the time sequence index to generate a multi-dimensional fusion feature table; The fault early warning module calls the feature values of the time window in the multi-dimensional fusion feature table, performs interval judgment with the set fault judgment threshold, screens the interval range that meets the abnormal state condition and exceeds the statistical process control limit value in the continuous time window, and generates a device light fault early warning signal.

Images