Medical equipment fault detection method and system based on machine learning

Abstract

The invention provides a medical equipment fault detection method and system based on machine learning, and relates to the technical field of fault detection. The method comprises the steps of monitoring motor start-stop division operation cycles, synchronously collecting and processing vibration, frequency difference and voltage signals, calculating characteristic difference to generate a dynamic sequence, extracting a non-convergence trend and phase deviation, fusing voltage and frequency fluctuation characteristics, constructing a threshold rule to judge abnormity, and outputting a light fault early warning signal. According to the invention, through precise monitoring of motor start-stop nodes, period division, collection of vibration, frequency and voltage signals, combination of time alignment and noise filtering, improvement of signal quality, cross-period calculation of characteristic difference, and extraction of non-convergence trend and phase offset, the ability of capturing tiny anomalies is enhanced; according to the method, multi-dimensional indexes such as voltage fluctuation and frequency deviation are fused, a threshold judgment mechanism in a continuous window is constructed, stable identification and early warning of equipment light faults are achieved, and the detection accuracy and response timeliness are improved.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a method and system for fault detection of medical equipment based on machine learning. Background Technology

[0002] The field of fault detection technology encompasses the means and methods for monitoring, identifying, and determining whether the operating status of various devices or systems is abnormal. The core of this technology lies in the real-time acquisition and analysis of multiple key parameters during equipment operation to discover potential fault symptoms, determine the type and location of faults, and assist in maintenance decisions. Data acquisition methods in this field include sensor signal acquisition, operation log recording, and environmental parameter acquisition. Analysis methods include rule-based judgment mechanisms and data-driven modeling. As the complexity of equipment operating status increases, fault detection has gradually evolved from traditional static detection to an intelligent analysis process relying on large-scale data mining and modeling, covering multiple sub-directions, including industrial equipment diagnostics, transportation vehicle health management, and medical instrument monitoring, and is gradually becoming more automated, real-time, and systematic.

[0003] The machine learning-based medical equipment fault detection method and system refers to a systematic approach that uses statistical learning theory and training model construction methods to classify and model structured data signals such as voltage, current, vibration frequency, and temperature collected during the operation of medical equipment, and then uses the trained discriminant model to identify the state of real-time data. This patent focuses on the technical issue of medical equipment fault diagnosis, covering steps such as standardizing and preprocessing historical operating data, training a multi-class classification model using supervised learning algorithms, and using this model to perform input mapping and state label judgment on newly collected data. In specific implementation, this method constructs a training sample set based on equipment type features and fault condition labels, uses decision trees or support vector machines to learn the discriminant boundary, continuously collects real-time signals during equipment operation, and uses the model to determine whether the current state belongs to a fault category label.

[0004] Existing fault detection technologies based on traditional modeling methods generally rely on the overall modeling and training process of large amounts of historical structured data. When faced with unstable fluctuations in actual operation, they lack fine-grained segmentation of state changes within a cycle, making it difficult to quickly respond to transient behaviors of equipment operation. In actual data acquisition, different types of sensor signals suffer from acquisition delays, timestamp errors, and data drift, failing to achieve precise alignment at the cycle level, resulting in unstable feature representation and affecting the real-time performance of anomaly identification. Most methods use static feature input, lacking a mechanism to track the feature evolution process, making it difficult to detect subtle but continuous trend changes over multiple cycles, causing some minor anomalies to be ignored due to insignificant feature differences. In the state discrimination logic, existing methods focus on judging static classification labels, ignoring the correlation between signals and temporal logic, resulting in insufficient sensitivity to anomalies in the early stages, often failing to trigger early warnings when faults occur. For example, when voltage is slightly unstable or vibration frequency drifts slowly, static models fail to meet preset thresholds, making timely intervention difficult, delaying equipment maintenance, and increasing risk exposure time. The aforementioned issues result in insufficient robustness of the existing solutions in actual operation, affecting the accuracy of fault prediction and the efficiency of clinical support. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a machine learning-based method and system for detecting faults in medical devices.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning-based method for fault detection in medical devices includes the following steps: S1: Monitor the start and stop nodes of medical equipment drive, divide the structural motion into multiple operating cycles, synchronously collect vibration signals, frequency peaks and voltage signals in each cycle, perform time stamp alignment and noise filtering on multi-source signals, and generate a multi-channel synchronous signal set; S2: Call the frequency peak difference, vibration peak interval and voltage standard deviation of each cycle in the multi-channel synchronization signal set, calculate the average difference of features in adjacent cycles, and construct a dynamic feature sequence; 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, and generate non-convergence trend marker and phase shift marker. S4: By merging the non-convergent trend marker and phase offset marker with the voltage standard deviation in the dynamic feature sequence according to the time window, the Pearson correlation coefficient of voltage fluctuation and frequency fluctuation and the cumulative offset of vibration interval offset are extracted to generate a multi-dimensional fusion feature table. S5: Based on the multi-dimensional fusion feature table, construct threshold triggering rules, determine the status of feature data within the time window, and generate a minor fault warning signal when the continuous window abnormal results reach the equipment operation statistical control limit.

[0007] Optionally, the multi-channel synchronization signal set includes vibration signals, frequency peaks, and voltage signals; the dynamic feature sequence includes frequency peak difference, vibration peak interval, and voltage standard deviation; the non-convergence trend marker and phase offset marker include the non-convergence of frequency fluctuations and the phase offset of vibration intervals; the multi-dimensional fusion feature table includes the Pearson correlation coefficient between voltage fluctuations and frequency fluctuations and the cumulative offset of vibration interval offset; and the equipment minor fault early warning signal includes anomaly judgment results that meet statistical process control limits.

[0008] Optionally, the specific steps of S1 are as follows: S101: Monitor the start and stop nodes of the motor drive of medical equipment, record the rising and falling edges of the control voltage signal, divide the operating cycle of the equipment structure movement, assign cycle number and start and end timestamp, and obtain the cycle time interval value. S102: Based on the periodic time interval value, simultaneously collect vibration signals, frequency peaks and voltage signals within multiple periods, unify sensor channel labels, align timestamps and establish data mapping to obtain a periodic synchronization signal set; S103: Based on the periodic synchronization signal set, determine the amplitude increment of vibration and voltage signals, filter out abnormal values ​​according to the peak noise amplitude threshold, reconstruct the signal curve, and generate a multi-channel synchronization signal set.

[0009] Optionally, the specific steps of S2 are as follows: S201: Obtain the frequency peak difference, vibration peak interval and period voltage standard deviation of each cycle in the multi-channel synchronization signal set, calculate the maximum frequency point difference, the mean vibration peak time difference and the voltage signal standard deviation respectively, and generate a set of period feature values; S202: Call the corresponding feature values ​​of two adjacent periods in the set of periodic feature values, calculate the difference of the same feature of the channel in consecutive periods and take the average to generate a channel feature difference sequence; S203: Based on the temporal order differences of features in the channel feature difference sequence, calculate the intensity of time change and summarize to construct a channel sequence, generating a dynamic feature sequence.

[0010] Optionally, the specific steps of S3 are as follows: S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence according to a fixed window, extract the correspondence between frequency peak and vibration interval within the window, compare the frequency peak difference and vibration interval value, and obtain the frequency difference change trend. S302: Based on the frequency difference change trend, extract the start and end values ​​of the window frequency peak sequence, calculate the linear regression slope, compare it with the slope difference of adjacent windows, extract the window position that exceeds 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, determine the deviation from the vibration interval reference value, filter the intervals that deviate beyond the reference coefficient, and combine the frequency non-convergent window position to obtain the non-convergent trend mark and phase offset mark.

[0011] Optionally, the specific steps of S4 are as follows: S401: Based on the non-convergent trend marker, phase offset marker, and voltage standard deviation sequence, the voltage standard deviation and the two types of marker values ​​are aligned and paired according to the time window, and a marker intensity value table is established and then merged into a marker voltage joint sequence. S402: Call the voltage standard deviation and frequency fluctuation value in the marked 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 metric value; S403: Based on the time tag of the fluctuation correlation metric, extract the position of the phase offset tag in the joint sequence of the marked voltage, calculate the cumulative offset value of the vibration interval, and merge it with the fluctuation correlation metric to generate a multi-dimensional fusion feature table.

[0012] Optionally, the specific steps of S5 are as follows: S501: Obtain the feature value of each time window in the multidimensional fusion feature table, judge the normal or abnormal state of the feature of each window according to the set judgment threshold, record the feature state of each time window, and generate feature state interval values. S502: Based on the characteristic state interval values, count the frequency of abnormal states in a continuous time window, construct a judgment list according to continuity and abnormal distribution, mark potential fault windows that meet the lower limit of equipment operation stability, and generate continuous trigger index values. S503: Based on the comparison between the continuous trigger index value and the equipment statistical process control limit, the window that exceeds the control limit is filtered out, and a minor fault trigger mark is generated according to the characteristic abnormality type to obtain the equipment minor fault early warning signal.

[0013] A machine learning-based medical device fault detection system includes: The signal acquisition module monitors the start and stop nodes of the motor drive in the medical equipment, divides the structural motion into multiple operating cycles, collects the vibration signal, frequency peak and voltage signal in each cycle, aligns the signals in the cycle according to the timestamp and performs noise filtering, and generates a multi-channel synchronous signal set. The feature construction module calls the frequency peak difference, vibration peak interval and voltage standard deviation of the multi-channel synchronous signal concentrated operating cycle, calculates the change amplitude according to the average difference of feature values ​​between adjacent cycles, and generates a dynamic feature sequence. The trend determination module performs window segmentation based on the frequency peak difference and vibration peak interval in the dynamic feature sequence, calculates the linear regression slope of frequency change and the adjacent offset value of vibration interval, determines the non-convergence of frequency change and the phase shift of vibration rhythm respectively, and generates a trend offset label set. The data fusion module merges data according to the time window dimension based on the trend offset marker set and the voltage standard deviation in the dynamic feature sequence, extracts the Pearson correlation coefficient between the voltage standard deviation and frequency fluctuation in each window, calculates the cumulative difference of phase offset, and fuses the two with the time series index respectively 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, filters the interval range within the continuous time window that meets the abnormal state conditions and exceeds the statistical process control limit, and generates a minor equipment fault early warning signal.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by accurately monitoring the motor start-stop nodes, dividing the period and collecting vibration, frequency and voltage signals, and combining time alignment and noise filtering, the signal quality is improved. The feature differences are calculated across periods, and non-convergent trends and phase shifts are extracted to enhance the ability to capture subtle anomalies. By integrating multi-dimensional indicators such as voltage fluctuations and frequency shifts, a threshold judgment mechanism within a continuous window is constructed to achieve stable identification and early warning of minor equipment faults, thereby improving detection accuracy and response time. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed flowchart of S1 of the present invention; Figure 3 This is a detailed flowchart of the S2 process of the present invention; Figure 4 This is a detailed flowchart of the S3 process of the present invention; Figure 5 This is a detailed flowchart of the S4 process of the present invention; Figure 6 This is a detailed flowchart of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] Please see Figure 1 This invention provides a machine learning-based method for detecting faults in medical devices, comprising the following steps: S1: Monitor the start and stop nodes of the motor drive of medical equipment, divide the structural motion into multiple operating cycles, synchronously collect vibration signals, frequency peaks and voltage signals in each cycle, perform timestamp alignment and noise filtering on the multi-source signals collected in the same cycle, and generate a multi-channel synchronous signal set. (Frequency peak is a known parameter that represents the frequency signal within a single period and is used to quantify the frequency fluctuation range.) S2: Call the frequency peak difference, vibration inter-peak interval and voltage standard deviation of each cycle in the multi-channel synchronization signal set, and construct a dynamic feature sequence by calculating the average difference of features in adjacent cycles; (The vibration peak interval is the time difference between adjacent peaks in the vibration signal, used to characterize the periodicity of vibration; the voltage standard deviation is a statistical measure of the voltage signal fluctuation amplitude.) S3: Call the frequency peak difference and vibration interpeak 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 marker and phase shift marker. (The slope of linear regression is the standard statistic for measuring the direction of data change in trend analysis, and it is calculated using the least squares method.) S4: By merging the voltage standard deviation in the dynamic feature sequence with the non-convergent trend marker and phase offset marker according to the time window, the Pearson correlation coefficient between voltage fluctuation and frequency fluctuation and the cumulative offset of vibration interval offset are extracted to generate a multi-dimensional fusion feature table. (The Pearson correlation coefficient is a standard statistic that measures the linear correlation between two sets of data; the cumulative offset is the sum of phase offset markers, used to quantify the offset trend.) S5: Based on the multi-dimensional fusion feature table, a threshold triggering rule is constructed to determine the feature data status within each time window. When the abnormal judgment result of the continuous window meets the statistical process control limit of the equipment operation, a minor fault warning signal of the equipment is generated. (The threshold triggering rule is a classification logic based on a preset threshold; the statistical process control limit is a process stability judgment standard commonly used in the industrial field, which is calculated through historical data.)

[0019] The multi-channel synchronization signal set includes vibration signals, frequency peaks, and voltage signals. The dynamic feature sequence includes frequency peak difference, vibration peak interval, and voltage standard deviation. The non-convergence trend marker and phase offset marker include the non-convergence of frequency fluctuations and the phase offset of vibration intervals. The multi-dimensional fusion feature table includes the Pearson correlation coefficient between voltage fluctuations and frequency fluctuations and the cumulative offset of vibration interval offset. The equipment minor fault early warning signal includes the abnormal judgment results that meet the statistical process control limits.

[0020] Please see Figure 2 The specific steps of S1 are as follows: S101: Monitor the start and stop nodes of the motor drive of medical equipment, record the rising and falling edges of the control voltage signal, divide the operating cycle of the equipment structure movement, assign cycle number and start and end timestamp, and obtain the cycle time interval value. In monitoring the start and stop points of the motor drive in medical equipment, it is first necessary to determine that the motor control method used is voltage-driven. A selected data acquisition module is connected to the controller's output port, and the sampling frequency is set to 10,000 times per second to ensure real-time capture of voltage signal changes. A voltage threshold value for the control signal is established, for example, 2.5 volts. If the voltage jumps from below this threshold to a high level, the change time is recorded as a start timestamp using edge detection. Simultaneously, a falling edge recognition rule is set, where the instant the voltage drops from high to low is considered the stop timestamp. These two timestamps are written to a memory buffer area via an embedded interrupt mechanism. If the duration between the two timestamps exceeds a certain set value, such as 0.2 seconds... If a cycle is identified, it is considered a complete motor operation cycle. Each cycle is numbered according to the time sequence to form a cycle identifier. For example, the start and end times of the first cycle are 0.5 seconds to 2.1 seconds, which is cycle one, and its time interval is 1 second and 60 milliseconds. This is written into a structured array as a cycle operation record item. The start time, end time and duration of multiple cycles are obtained through continuous monitoring. The cycle number is used for index identification, the timestamp is used as a boundary attribute, and the duration is used as a supplementary parameter. All cycle records are included in the database through a unified data structure. In this record structure, the cycle number is used as the primary key field, and time interval attribute fields and time difference fields are defined. When a new cycle record is generated, the write operation is dynamically performed to save the data completely.

[0021] S102: Based on the periodic time interval value, synchronously collect vibration signals, frequency peaks and voltage signals within multiple periods, unify sensor channel labels, align timestamps and establish data mapping to obtain a set of periodic synchronization signals; Based on each calibrated time interval, data acquisition from multiple sensors is initiated synchronously. First, each sensor is assigned a specific label; for example, the first channel is for vibration signals, the second for frequency signals, and the third for control voltage signals. A unified sampling frequency is set, such as 10 kHz, to ensure a consistent number of data points collected within each cycle. The required number of sampling points within a cycle is calculated by multiplying the time interval length by the sampling frequency. For example, a cycle length of 1.50 seconds corresponds to 15,000 sampling points. During data acquisition, an absolute timestamp is assigned to each sampling point in each channel. The sampling interval is then accumulated sequentially based on the start time of the cycle, constructing a time-based data axis. The acquisition device, such as the embedded main control module, processes the data stream uniformly, assigns time and channel labels to each data point, and then completes the time alignment between channels. The acquisition results are structured into a multi-dimensional array, arranged according to the period number, channel identifier, and data point sequence number. By defining mapping rules, the data of any time point and its corresponding channel can be quickly retrieved. For example, under the conditions of period number three, start time five seconds, and period duration of one second and two hundred milliseconds, the data point time axis of this period is established at a frequency of 10,000 times per second to ensure that the data of each channel corresponds one-to-one at the same time point. Finally, the synchronous data of each channel in the entire period is written into the data table, and each record contains time, period number, channel identifier, and corresponding data value.

[0022] S103: Based on the periodic synchronization signal set, determine the amplitude increment of vibration and voltage signals, filter out outliers according to the peak noise amplitude threshold, reconstruct the signal curve, and generate a multi-channel synchronization signal set; In the processing of multi-channel synchronous signal sets, it is necessary to determine the amplitude changes of vibration and voltage signals in each channel. The amplitude change at each sampling point is defined as the current value minus the previous value. If this difference exceeds a set noise peak threshold, it is considered that there is a sudden change in the current data. Based on historical data, the average amplitude and standard deviation of the vibration channels are statistically analyzed. For example, if the average vibration signal is 0.05 gravitational acceleration and the standard deviation is 0.01, the noise threshold is set to the average value plus three times the standard deviation, i.e., 0.08 gravitational acceleration. All vibration sampling points within the cycle are iterated, the amplitude difference between adjacent points is calculated, and compared with the threshold. If the difference at a certain moment... If the value is higher than the threshold, the point is considered an outlier. The outlier is replaced by the average of several normal data points before and after it. For example, if the amplitude of a point is 0.13 and the previous point is 0.07, the difference is 0.06. If the value exceeds the threshold, the current outlier is replaced by the average of five normal data points before and after it. After the vibration channel is repaired, the voltage channel is reconstructed using the same strategy. The repaired vibration and voltage data are synchronized in time, forming an updated multi-channel synchronization signal record. The record structure retains the period number, sampling time, channel number, and repaired data value, and stores them in a dedicated data structure for subsequent use.

[0023] Please see Figure 3 The specific steps of S2 are as follows: S201: Obtain the frequency peak difference, vibration peak interval and period voltage standard deviation of each cycle in the multi-channel synchronization signal set, calculate the maximum frequency point difference, the mean vibration peak time difference and the voltage signal standard deviation respectively, and generate a set of period feature values; When processing the frequency peak difference, vibration peak interval, and voltage standard deviation for each cycle of a multi-channel synchronous signal set, the original signal data is first divided according to a set cycle length. For example, if the cycle length is set to 100 milliseconds and the sampling frequency is 10 kHz, then each cycle signal contains 1000 data points. For each cycle signal, the spectral information is extracted using Fast Fourier Transform (FFT) to identify the frequency component with the largest amplitude, which is the frequency peak point. Then, the numerical differences of this peak frequency point in different channels are compared to obtain the frequency peak difference. Subsequently, for the vibration signal within the same cycle, the location of the vibration peak point is determined by extracting local maxima. When the amplitude exceeds twice the mean, it can be used as a valid peak criterion. Based on this, the locations of all peak points are extracted, and the time interval between two adjacent peaks is calculated. Then, the average of these intervals is taken as the average vibration peak interval within that cycle. Taking the peaks occurring at 5 milliseconds, 25 milliseconds, 45 milliseconds, 65 milliseconds, and 85 milliseconds within a certain channel cycle as an example, the intervals are 20 milliseconds, and the average is 20 milliseconds. For voltage signals, the mean and variance of the voltage amplitude at each sampling point are calculated. The square root of the variance is then taken as the standard deviation of the voltage signal within that period. If the voltage value sequence is 2.1, 2.3, 2.0, 2.4, 2.2, then the voltage mean is 2.2, the squared deviations are 0.01, 0.01, 0.04, 0.04, and 0 respectively, the average is 0.02, and the square root is approximately 0.14. After extracting these three features for each period of each channel, they are organized according to feature type to form a set of periodic feature values ​​for subsequent difference analysis.

[0024] S202: Call the corresponding feature values ​​of two adjacent periods in the periodic feature value set, calculate the difference of the same feature of the channel in consecutive periods and take the average to generate the channel feature difference sequence; Based on the set of periodic feature values, corresponding feature values ​​within adjacent periods are extracted, establishing a two-dimensional matrix structure of channel and period dimensions. For each type of feature value corresponding to each channel, the numerical differences between adjacent periods are calculated sequentially 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 differences between adjacent periods are calculated to be 0.02, 0.03, 0.02, and 0.01. These differences are then averaged to obtain the average change in voltage standard deviation over consecutive periods, with an average value of 0.02. To avoid the impact of abnormal fluctuations on the accuracy of the analysis, an upper limit for the difference can be set as a filtering condition. For example, if more than 90% of the data in a large number of historical periodic samples have a change range within 0.05, then the upper limit can be set to 0.1. If a pair of differences exceeds this range, then that pair of data is excluded from the averaging calculation. Each type of feature (frequency peak difference, vibration peak interval, voltage standard deviation) is processed in the same way, and finally a mean sequence of periodic differences of the three types of features is formed for each channel, that is, the channel feature difference sequence.

[0025] S203: Based on the temporal order differences of features in the channel feature difference sequence, calculate the intensity of temporal change and summarize them to construct the channel sequence, generating a dynamic feature sequence; First, a sequential structure containing a time dimension needs to be created for the three types of feature difference values ​​corresponding to each channel. For example, the voltage standard deviation difference values ​​are arranged in chronological order over five consecutive periods as 0.02, 0.01, 0.03, 0.02, and 0.01, corresponding to time points t1 to t5 respectively. Then, the magnitude of the numerical change is calculated for each pair of adjacent difference values ​​in chronological order. A new difference change sequence is obtained 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 the value increases from 0.01 to 0.03, the change value is 0.02. This process continues to obtain a sequence representing the intensity of the difference. After this step, a local analysis window is introduced to continuously aggregate the difference intensity sequence. When the window length is selected as 3, every three consecutive change values ​​are averaged. For example, if the first three change values ​​are 0.01, 0.02, and 0.01, the corresponding aggregated value within the window is 0.013. This process is then repeated one position to obtain the complete moving average change intensity sequence. A reference standard is then set to identify the strength of the fluctuation. In weak cases, this standard can extract statistical results from the intensity of characteristic differences under a large number of historical normal operating conditions. If the intensity of change in more than 95% of the samples is less than 0.02, then 0.02 is set as the fluctuation recognition threshold for this type of feature. The results of each window are compared one by one. If the average value within the window is greater than 0.02, it is determined to be a high dynamic change segment. If it is less than or equal to 0.02, it is classified as a stable segment. The above operation is performed for each type of feature, and the labeling results are recorded as the marker information in the time series. For example, the state label corresponding to each window of the voltage feature may be stable, high dynamic, stable, etc. Then, the other two types of features, such as the vibration peak interval and the frequency peak difference, are processed in the same way to obtain the corresponding dynamic state label sequences. Then, the states of each type of feature at the same time point are combined to form a composite dynamic state point. For example, if the voltage is stable, the vibration is stable, and the frequency is high dynamic at a certain moment, then the combined state is [stable, stable, high dynamic]. All combined states are spliced ​​in chronological order to form a complete channel dynamic feature sequence. This sequence provides a data foundation for subsequent state recognition or behavior modeling.

[0026] Please see Figure 4 The specific steps of S3 are as follows: S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence according to a fixed window, extract the correspondence between the frequency peak and vibration interval within the window, compare the frequency peak difference and vibration interval values, and obtain the frequency difference change trend. After acquiring the dynamic feature sequence, the acceleration or displacement signal is first subjected to a Fast Fourier Transform (FFT) to extract effective frequency peaks exceeding a set threshold (e.g., mean plus twice the standard deviation) in the frequency domain, and the frequency difference between adjacent peaks is calculated. Then, in the time domain, the occurrence times of the vibration amplitude peaks corresponding to these frequency peaks are extracted, and the time interval between each pair of adjacent vibration peaks is calculated. Each set of frequency differences and corresponding time intervals constitutes a pair, forming a structured feature pair sequence. Next, the entire sequence is divided into fixed-length sliding windows, retaining all pairs within each window, and the ratio of frequency difference to time interval is calculated separately for each pair, resulting in a set of ratios within the window. This set is then aggregated (e.g., by calculating the average value) to measure the synchronous change trend between frequency characteristics and vibration behavior within that time period. By comparing the change amplitude of the average ratio within two adjacent windows, the stability of the trend is determined; if the change amplitude exceeds a set threshold (e.g., 10%), it is considered a trend abrupt change zone. This method can further depict the trajectory and distribution of the frequency difference and vibration response through the trend evolution of multiple windows, providing quantitative support for dynamic feature identification and state analysis.

[0027] S302: Based on the trend of frequency difference change, extract the start and end values ​​of the window frequency peak sequence, calculate the linear regression slope, compare it with the slope difference of adjacent windows, extract the window position that exceeds the frequency slope threshold, and obtain the frequency non-convergent fluctuation value. Based on the frequency peak sequence, a sliding window approach is used to perform linear regression analysis on the start and end points of the peak within each fixed window to calculate the slope of the frequency change trend. The entire frequency sequence is then segmented to obtain a series of slope values ​​for consecutive windows. By comparing the slope difference between adjacent windows, it is determined whether it exceeds a preset abrupt change threshold (this threshold can be determined statistically by analyzing the slope distribution of normal samples, such as using the maximum slope value of 90% of the samples as the upper limit). Once the slope change of a certain window exceeds this threshold (e.g., if the current window slope is 0.6 and the previous one was 0.2, the difference is 0.4, exceeding the threshold of 0.3), it can be determined that a sudden change in the frequency trend has occurred within that window. This time segment is marked as a frequency fluctuation abnormal region and defined as a frequency non-convergent fluctuation value. Its meaning is: the frequency peak shows a sudden trend change in the time series and no longer tends to be stable, which is used to identify the key time position of frequency anomalies and support subsequent in-depth analysis and anomaly localization.

[0028] S303: Based on the frequency non-convergent fluctuation value, extract the start and end difference value of the corresponding vibration peak interval, determine the deviation from the vibration interval benchmark value, filter the intervals that deviate beyond the benchmark coefficient, and combine the frequency non-convergent window position to obtain the non-convergent trend mark and phase offset mark. Based on the identified frequency fluctuation window, further analysis is performed on the intervals between peak vibration amplitudes within the corresponding time period. The time difference between all adjacent vibration peaks within the window is extracted, and the interval change at the beginning and end of the window is statistically analyzed. This change is compared with a pre-set benchmark interval value, which is typically derived from data analysis results under stable operating conditions. For example, if the average value between vibration peaks is 20 milliseconds, this is used as the benchmark value, and an allowable fluctuation range is set, such as ±20% of the benchmark value, i.e., upper and lower limits of 24 milliseconds and 16.67 milliseconds, respectively. If the interval change within the window exceeds this threshold, it is determined that there is a trend shift and marked as a non-convergent trend segment. The trend segment is then compared with the non-convergent window position. If the overlap time length accounts for more than 30% of the total window length, the region is confirmed as a non-convergent trend marker. Based on this, the phase angle of each vibration peak relative to the ideal period position is further calculated. If the phase shift of a peak within the reference period exceeds a set threshold (such as 90 degrees), it is marked as a phase shift point. For example, if the reference period is 20 milliseconds and the peak appears at 27 milliseconds, the corresponding shift angle is about 126 degrees, which is outside the range and is considered a phase anomaly. Finally, the above non-convergent trend markers and phase shift markers are used as important bases for identifying the non-steady-state characteristics of the system.

[0029] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the non-convergent trend markers, phase shift markers, and voltage standard deviation sequences, the voltage standard deviations and the two types of marker values ​​are aligned and paired according to the time window, and a marker intensity value table is established and then merged into a marker-voltage joint sequence. Non-convergent trend markers are extracted based on the continuous changes in the standard deviation of the voltage signal. During operation, the voltage data is divided into segments at fixed time intervals, such as 10 seconds per window. The change in the standard deviation is compared window by window. If the change in standard deviation exceeds a set threshold for three consecutive windows, the current time point is marked as non-convergent. The threshold can be set to a standard deviation amplitude of 0.02 volts. Phase shift markers are determined by real-time monitoring of the voltage signal's phase value using phasor measurement equipment. A phase shift is considered to have occurred when the phase difference between a given time point and its previous time point exceeds 5 degrees. The voltage standard deviation sequence is obtained by statistically analyzing the dispersion of the voltage data within each time period. After extraction, the three types of data are analyzed... The data is processed uniformly according to time stamps and aligned at the second level. If a non-convergent trend marker, a phase shift marker, and a voltage standard deviation exist within a certain second, the three are paired and combined into a composite record. Different weights are assigned to the marker values, for example, a non-convergent trend marker is weighted at 0.6 and a phase shift marker at 0.4. A comprehensive marker strength value is obtained through weighted calculation, which, together with the voltage standard deviation at the same time point, forms a marker-voltage joint record. This process is repeated to generate continuous time series data. In this series, each record contains a comprehensive marker strength value and a voltage standard deviation value, for example, a combination of 0.6 and 0.12 volts at a certain moment, continuously building up to form a complete joint sequence.

[0030] S402: Call the voltage standard deviation and frequency fluctuation value in the marked 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 metric value; When extracting voltage standard deviation and frequency fluctuation values, data with the same time label need to be read from the constructed joint sequence and frequency record, respectively. The voltage standard deviation has already been obtained from the joint sequence, while the frequency fluctuation needs to be judged by the difference in frequency values ​​between adjacent time points. For example, if the frequency at the current time is 50.04 Hz and the previous second was 50.01 Hz, then the frequency fluctuation value is 0.03 Hz. The frequency fluctuation sequence is calculated second by second in this way, and paired with the voltage standard deviation to form paired data. Then, the data is divided into fixed time periods, such as 30 seconds per window, and the overall consistency of the changes of all paired data within the window is statistically analyzed. The linear correlation is judged using statistical analysis methods to obtain a value representing the degree of correlation. The closer the correlation is to 1, the stronger the relationship between the two, and the closer it is to 0, the weaker the relationship. This correlation value is combined with its time label to form a new fluctuation correlation measurement sequence. The data can be displayed segment by segment in different time windows.

[0031] S403: Based on the time stamp of the wave correlation metric, extract the position of the phase offset marker in the joint sequence of marked voltages, calculate the cumulative offset value of the vibration interval, and merge it with the wave correlation metric to generate a multi-dimensional fusion feature table; First, time information is extracted from the fluctuation correlation measurement sequence, such as identifying time points like the 10th second, 40th second, and 70th second, indicating a certain degree of linear correlation between voltage standard deviation and frequency fluctuation within these time periods. Next, records matching the above time points are selected from the joint sequence, and it is checked whether these records contain phase offset markers. If a marker exists, the current time point is marked as a valid vibration point, and the corresponding phase angle value is recorded. Then, all valid 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 the 40th second and the 70th second is 30 seconds. If the corresponding phase angles are 123° and 115° respectively, the change is 8°. The phase offset increment of this segment is recorded as an absolute value. All vibration point pairs are processed in this way to obtain multiple vibration intervals and their corresponding angle change values. Then, these changes are accumulated to obtain the cumulative offset value corresponding to each time period. When identifying phase shifts, the difference in phase angle between adjacent time points is considered to exceed a set threshold, such as 5°. If the current time is 50 seconds and the previous time was 49 seconds, with phase angles of 117° and 111° respectively, the difference is 6°, indicating a phase shift. The current time point is then recorded as a valid vibration point. In subsequent processing, if the change in phase angle between two vibration points exceeds another set threshold, such as 10°, that segment is marked as an abnormal shift segment, and its shift value still participates in subsequent cumulative calculations. These cumulative shift values ​​are paired with the original fluctuation correlation metrics according to time tags to form a multi-dimensional fusion record. Each record includes the time point, the correlation value between voltage standard deviation and frequency fluctuation, whether a phase shift exists in that time period, the corresponding cumulative shift angle, and whether it belongs to an abnormal shift segment. For example, in the record at 70 seconds, the fluctuation correlation is 0.93, indicating a phase shift. A cumulative offset angle of 12° is marked as an abnormal offset segment, and this record will be fully included in the fusion feature table. Furthermore, the correlation value is derived from the sliding window calculation in the previous process. The window length is set to 30 seconds, and a set of voltage standard deviation and frequency fluctuation values ​​are sampled every second. For example, within a certain window, the voltage standard deviation is 0.06, 0.07, and 0.08, and the frequency fluctuation is 0.01, 0.02, and 0.03, respectively. By comparing the overall direction of change between the two columns of values, their correlation is determined. If the numerical trends are highly consistent, it is determined to be a strong correlation. The strong range can be set between 0.85 and 1.00, and the weak range between 0.00 and 0.30. If the correlation value of the current window is 0.93, it is classified into the strong correlation range and recorded in combination with the corresponding cumulative offset value. The final output fusion feature table continuously covers the entire analysis cycle, forming a structured data sequence for subsequent feature modeling or anomaly detection processes.

[0032] Please see Figure 6 The specific steps of S5 are as follows: S501: Obtain the feature value of each time window in the multidimensional fusion feature table, judge the normal or abnormal state of the feature of each window according to the set judgment threshold, record the feature state of each time window, and generate the feature state interval value. To obtain the feature values ​​of each time window in the multidimensional fusion feature table, the operating status of the equipment can be collected in real time through multiple sensor channels. For example, multidimensional signals such as temperature (in degrees Celsius), vibration (in millimeters per second), and current (in amperes) of the fan can be collected. The acquired data is then processed for noise reduction, normalization, and time alignment to fuse the features of each channel under a unified time reference. The feature set is divided into time windows of 10 minutes each. The mean of temperature can be extracted, the standard deviation of vibration can be extracted, and the peak value of current can be extracted. After feature extraction, the status is identified according to a pre-set judgment threshold. For example, the judgment value for temperature is set to 85. The temperature is set to 85 degrees Celsius, the vibration threshold is set to 6, and the peak current threshold is set to 60. If the temperature is greater than 85 degrees Celsius, it is marked as abnormal. If the vibration value is higher than 6, it is also marked as abnormal. If the current exceeds 60, it is considered abnormal. Otherwise, it is normal. The above characteristics are judged in each time window. For example, if the temperature is 83 degrees Celsius, the vibration is 7, and the current is 55 in a 10-minute window, the temperature and current are normal, but the vibration is abnormal. The state corresponding to this window can be recorded as "normal, abnormal, normal". This process is repeated to record the characteristic judgment results of each window in sequence, forming a state sequence arranged in time order. The state of each characteristic under each window constitutes the characteristic state interval value.

[0033] S502: Based on the characteristic state interval values, count the frequency of abnormal states in a continuous time window, construct a judgment list according to continuity and abnormal distribution, mark potential fault windows that meet the lower limit of equipment operation stability, and generate continuous trigger index values; Based on the obtained characteristic state interval values, a fixed number of continuous time windows can be set for sliding statistics. For example, each group consists of five windows, and the number of occurrences of abnormal states is counted. Each type of feature is counted separately. For example, if the number of abnormal temperature occurrences is three, the number of abnormal vibration occurrences is two, and the number of abnormal current occurrences is one, then the frequency of abnormal temperature occurrences is relatively high in this window group. If the lower limit for abnormal judgment is set to three in advance, that is, if any feature is abnormal at least three times within five windows, it can be judged as a potential fault window. After each sliding, one window is advanced and the next group of statistics is performed to form the abnormal statistical results corresponding to multiple consecutive groups of windows. If the frequency of abnormality in certain window groups continuously reaches the set lower limit, it can be marked as a potential fault window. For example, if the seventh window is marked as an abnormal state in three consecutive statistical groups, then the number of consecutive triggers is three. This number is recorded as a continuous trigger index value, and a judgment list is formed by marking the corresponding window position index. For example, if the seventh, ninth, and tenth windows are marked as potential abnormalities multiple times, a continuous trigger index list can be established in sequence as the basis for subsequent minor fault judgment.

[0034] S503: Based on the comparison between the continuous trigger index value and the equipment statistical process control limit, the window that exceeds the control limit is filtered out, and a minor fault trigger mark is generated according to the characteristic abnormality type to obtain the equipment minor fault early warning signal; Based on the established continuous trigger index values, the control range is statistically controlled according to historical data collected during equipment operation. For example, if the average level of all continuous trigger values ​​is 1.5 and the fluctuation level is 0.7, a control threshold is obtained by setting a reasonable upper limit. For example, the maximum acceptable range is defined by adding three times the fluctuation value to the average level. That is, if the continuous trigger index value exceeds the upper limit of approximately 3.6, it can be judged as an abnormal concentrated outbreak, which is marked as a minor fault warning window. After filtering out these over-limit windows, based on the characteristic status records of each window in the previous period, it is possible to trace which specific characteristic or combination of characteristics is abnormal in each over-limit window. For example, if both temperature and vibration status are abnormal in the ninth window, it is marked as "high temperature aggravates vibration". If only vibration is abnormal, it is marked as "high intensity vibration". Each mark needs to be based on the specific combination of abnormal characteristics in the window to form a list of minor fault warning signals, and correspond one-to-one with the window index for subsequent correlation analysis and record tracking.

[0035] Please see Figure 7 A machine learning-based medical device fault detection system includes: The signal acquisition module monitors the start and stop nodes of the motor drive in the medical equipment, divides the structural motion into multiple operating cycles, collects the vibration signal, frequency peak and voltage signal in each cycle, aligns the signals in the cycle according to the timestamp and performs noise filtering, and generates a multi-channel synchronous signal set. The feature construction module calls the frequency peak difference, vibration inter-peak interval and voltage standard deviation of the multi-channel synchronous signal concentrated operation cycle, calculates the change amplitude according to the average difference of feature values ​​between adjacent cycles, and generates a dynamic feature sequence; The trend determination module divides the window based on the frequency peak difference and vibration peak interval in the dynamic feature sequence, calculates the linear regression slope of frequency change and the adjacent offset value of vibration interval, and determines the non-convergence of frequency change and the phase shift of vibration rhythm respectively, generating a trend offset label set. The data fusion module merges data according to the time window dimension based on the trend offset marker set and the voltage standard deviation in the dynamic feature sequence. It extracts the Pearson correlation coefficient between the voltage standard deviation and frequency fluctuation in each window and calculates the cumulative difference of phase offset. The two are then fused with the time series 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, filters the interval range within the continuous time window that meets the abnormal state conditions and exceeds the statistical process control limit, and generates a minor equipment fault early warning signal.

[0036] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A machine learning-based method for fault detection in medical equipment, characterized in that, Includes the following steps: S1: Monitor the start and stop nodes of medical equipment drive, divide the structural motion into multiple operating cycles, synchronously collect vibration signals, frequency peaks and voltage signals in each cycle, perform time stamp alignment and noise filtering on multi-source signals, and generate a multi-channel synchronous signal set; S2: Call the frequency peak difference, vibration peak interval and voltage standard deviation of each cycle in the multi-channel synchronization signal set, calculate the average difference of features in adjacent cycles, and construct a dynamic feature sequence; 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, and generate non-convergence trend marker and phase shift marker. S4: By merging the non-convergent trend marker and phase offset marker with the voltage standard deviation in the dynamic feature sequence according to the time window, the Pearson correlation coefficient of voltage fluctuation and frequency fluctuation and the cumulative offset of vibration interval offset are extracted to generate a multi-dimensional fusion feature table. S5: Based on the multi-dimensional fusion feature table, construct threshold triggering rules, determine the status of feature data within the time window, and generate a minor fault warning signal when the continuous window abnormal results reach the equipment operation statistical control limit.

2. The medical device fault detection method based on machine learning according to claim 1, characterized in that, The multi-channel synchronization signal set includes vibration signals, frequency peaks, and voltage signals. The dynamic feature sequence includes frequency peak difference, vibration peak interval, and voltage standard deviation. The non-convergence trend marker and phase offset marker include the non-convergence of frequency fluctuations and the phase offset of vibration intervals. The multi-dimensional fusion feature table includes the Pearson correlation coefficient between voltage fluctuations and frequency fluctuations and the cumulative offset of vibration interval offset. The equipment minor fault early warning signal includes the abnormal judgment result that meets the statistical process control limit.

3. The machine learning-based medical device fault detection method according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Monitor the start and stop nodes of the motor drive of medical equipment, record the rising and falling edges of the control voltage signal, divide the operating cycle of the equipment structure movement, assign cycle number and start and end timestamp, and obtain the cycle time interval value. S102: Based on the periodic time interval value, simultaneously collect vibration signals, frequency peaks and voltage signals within multiple periods, unify sensor channel labels, align timestamps and establish data mapping to obtain a periodic synchronization signal set; S103: Based on the periodic synchronization signal set, determine the amplitude increment of vibration and voltage signals, filter out abnormal values ​​according to the peak noise amplitude threshold, reconstruct the signal curve, and generate a multi-channel synchronization signal set.

4. The machine learning-based medical device fault detection method according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Obtain the frequency peak difference, vibration peak interval and period voltage standard deviation of each cycle in the multi-channel synchronization signal set, calculate the maximum frequency point difference, the mean vibration peak time difference and the voltage signal standard deviation respectively, and generate a set of period feature values; S202: Call the corresponding feature values ​​of two adjacent periods in the set of periodic feature values, calculate the difference of the same feature of the channel in consecutive periods and take the average to generate a channel feature difference sequence; S203: Based on the temporal order differences of features in the channel feature difference sequence, calculate the intensity of time change and summarize to construct a channel sequence, generating a dynamic feature sequence.

5. The machine learning-based medical device fault detection method according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Obtain the frequency peak difference and vibration peak interval in the dynamic feature sequence, divide the sequence according to a fixed window, extract the correspondence between frequency peak and vibration interval within the window, compare the frequency peak difference and vibration interval value, and obtain the frequency difference change trend. S302: Based on the frequency difference change trend, extract the start and end values ​​of the window frequency peak sequence, calculate the linear regression slope, compare it with the slope difference of adjacent windows, extract the window position that exceeds 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, determine the deviation from the vibration interval reference value, filter the intervals that deviate beyond the reference coefficient, and combine the frequency non-convergent window position to obtain the non-convergent trend mark and phase offset mark.

6. The machine learning-based medical device fault detection method according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Based on the non-convergent trend marker, phase offset marker, and voltage standard deviation sequence, the voltage standard deviation and the two types of marker values ​​are aligned and paired according to the time window, and a marker intensity value table is established and then merged into a marker voltage joint sequence. S402: Call the voltage standard deviation and frequency fluctuation value in the marked 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 metric value; S403: Based on the time tag of the fluctuation correlation metric, extract the position of the phase offset tag in the joint sequence of the marked voltage, calculate the cumulative offset value of the vibration interval, and merge it with the fluctuation correlation metric to generate a multi-dimensional fusion feature table.

7. The machine learning-based medical device fault detection method according to claim 6, characterized in that, The specific steps of S5 are as follows: S501: Obtain the feature value of each time window in the multidimensional fusion feature table, judge the normal or abnormal state of the feature of each window according to the set judgment threshold, record the feature state of each time window, and generate feature state interval values. S502: Based on the characteristic state interval values, count the frequency of abnormal states in a continuous time window, construct a judgment list according to continuity and abnormal distribution, mark potential fault windows that meet the lower limit of equipment operation stability, and generate continuous trigger index values. S503: Based on the comparison between the continuous trigger index value and the equipment statistical process control limit, the window that exceeds the control limit is filtered out, and a minor fault trigger mark is generated according to the characteristic abnormality type to obtain the equipment minor fault early warning signal.

8. A machine learning-based medical equipment fault detection system, characterized in that, The medical device fault detection method based on machine learning according to any one of claims 1-7, wherein the system comprises: The signal acquisition module monitors the start and stop nodes of the motor drive in the medical equipment, divides the structural motion into multiple operating cycles, collects the vibration signal, frequency peak and voltage signal in each cycle, aligns the signals in the cycle according to the timestamp and performs noise filtering, and generates a multi-channel synchronous signal set. The feature construction module calls the frequency peak difference, vibration peak interval and voltage standard deviation of the multi-channel synchronous signal concentrated operating cycle, calculates the change amplitude according to the average difference of feature values ​​between adjacent cycles, and generates a dynamic feature sequence. The trend determination module performs window segmentation based on the frequency peak difference and vibration peak interval in the dynamic feature sequence, calculates the linear regression slope of frequency change and the adjacent offset value of vibration interval, determines the non-convergence of frequency change and the phase shift of vibration rhythm respectively, and generates a trend offset label set. The data fusion module merges data according to the time window dimension based on the trend offset marker set and the voltage standard deviation in the dynamic feature sequence, extracts the Pearson correlation coefficient between the voltage standard deviation and frequency fluctuation in each window, calculates the cumulative difference of phase offset, and fuses the two with the time series index respectively 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, filters the interval range within the continuous time window that meets the abnormal state conditions and exceeds the statistical process control limit, and generates a minor equipment fault early warning signal.

Images