A machine learning based centrifuge failure prediction system

Abstract

The present application relates to the technical field of fault prediction, in particular to a centrifuge fault prediction system based on machine learning, which comprises a data channel synchronization module, a multi-dimensional feature extraction module, a state evolution index construction module, a trend aggregation trajectory identification module and a fault section identification module.In the present application, the multi-channel signal segmentation processing mechanism based on periodic state synchronously aligns different types of data, combines the uniform sampling structure based on time scale to construct the state evolution sequence, establishes the state characteristic trajectory through the multi-dimensional parameter set, constructs the trend change index relying on the root mean square of the difference between adjacent states, identifies the aggregated section using the continuous trend mutation and counts the trajectory deviation frequency, realizes the dynamic migration of the trajectory boundary and the intelligent identification of the fault section, avoids the boundary failure problem caused by the static preset condition, and effectively enhances the continuous prediction stability of the long-period equipment and the adaptive range under non-standard working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault prediction, in particular to a centrifuge fault prediction system based on machine learning. BACKGROUND

[0002] The technical field of fault prediction mainly involves real-time monitoring and data analysis of equipment operating state to identify potential fault signs and predict the time and location of their possible occurrence, including equipment state parameter collection, feature extraction, historical operation data analysis and life prediction modeling, and is widely used in health management of key equipment in manufacturing, energy, transportation and other industries. The traditional centrifuge fault prediction system refers to obtaining parameters such as vibration signals, current fluctuations and speed changes of the centrifuge during operation, combining time series-based anomaly detection methods to compare and analyze the data to determine whether the equipment has potential anomalies, usually using acceleration sensors to collect vibration information of the equipment bearings or housings, converting the original vibration signals into frequency domain features through Fourier transform, and comparing the amplitude change trend in the preset frequency band to identify abnormal states or wear conditions. At the same time, an experience threshold determination mechanism is used to determine whether the equipment has reached the alarm standard to prompt possible fault risks.

[0003] In the existing centrifuge fault prediction process, the frequency domain transformation of a single vibration signal is mainly relied on as the basis for fault identification, which is prone to misjudgment in running scenarios with multiple source state disturbances or insufficient signal stability, especially in complex states such as thermal deformation, current disturbance or speed fluctuation of the centrifuge after long-time operation. Fourier transform cannot accurately reveal the time evolution trend, resulting in deviation of the frequency band amplitude change from the actual abnormal position. At the same time, the static experience threshold is fixed and cannot be dynamically adjusted with the state change, which is prone to boundary lag or early triggering problems in the initial stage of parameter mutation, making it difficult to capture the trajectory change under the gradual state deviation and affecting the timeliness and continuity of equipment fault warning. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to provide a centrifuge fault prediction system based on machine learning.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: a centrifuge fault prediction system based on machine learning comprises:

[0006] A data channel synchronization module acquires continuous signal data of the starting point of the centrifuge main shaft cycle, divides the vibration analysis cycle into segments, and synchronizes the time scale of all signal segments to generate a unified segmented signal index matrix;

[0007] The multi-dimensional feature extraction module calculates a vibration signal peak position time difference based on the unified segmented signal index matrix, counts a current signal mutation point quantity proportion, obtains a rotating speed signal continuous fluctuation amplitude mean value, calculates a temperature signal temperature rise rate index, and sequentially constructs a shape formation cycle state evolution vector sequence.

[0008] The state evolution index construction module calculates a root mean square of each dimension difference value as a trend change index based on the cycle state evolution vector sequence, records a trend mutation point position, and generates a trend mutation state change index set.

[0009] The trend aggregation trajectory identification module performs cosine similarity analysis and labels an aggregation section based on the trend mutation state change index set, performs trajectory offset comparison and counts a number of out-of-bound times to obtain a trajectory offset identification sequence.

[0010] The fault section identification module calculates a difference degree to judge a boundary update amplitude and performs boundary replacement based on the trajectory offset identification sequence, labels a fault area, and generates a centrifuge fault prediction result.

[0011] As a further scheme of the application, the unified segmented signal index matrix includes a cycle synchronization number, a time scale alignment identifier, a signal channel mapping relationship, a segmented time sequence index, and a sampling interval record; the cycle state evolution vector sequence includes a vibration position difference value, a current mutation proportion factor, a rotating speed fluctuation mean amplitude index, a temperature rise rate change amount, and a time sequence correlation code; the trend mutation state change index set includes a trend response amplitude index, a cycle fluctuation mean square difference value, an abnormal evolution position index, and a vector difference expression degree; the trajectory offset identification sequence includes a similarity aggregation label, a direction consistency count, a trajectory change interval envelope value, and an out-of-bound cycle position record; and the centrifuge fault prediction result includes a predicted state section number, a fault evolution path label, a trajectory center difference measure, and a fault level discrimination result.

[0012] As a further scheme of the application, the data channel synchronization module includes:

[0013] The signal acquisition sub-module acquires continuous signal data of a vibration acceleration sensor, an axial rotating speed recorder, a current monitoring channel, and a temperature sensing device, acquires a vibration analysis cycle based on a current main shaft cycle starting point as a reference benchmark, and acquires all signal data segments to generate a multi-source synchronous signal segment set.

[0014] The time synchronization submodule reads timestamp data inside each signal segment according to the multi-source synchronization signal segment set, calculates time deviation amounts of each signal segment from a main shaft cycle starting point, adjusts time axis data of each signal segment in combination with a main shaft cycle starting point time reference, compares whether the deviation amounts exceed a time synchronization threshold, and if so, adds or subtracts adjusts the signal time axis until the time axis of each signal segment meets a synchronization standard, and obtains a unified time alignment signal segment set;

[0015] The signal index generation submodule records starting and ending time points, signal lengths, sampling rates and channel serial numbers of each signal segment according to the unified time alignment signal segment set, integrates the time points and channel numbers of each signal segment in combination with a main shaft cycle starting point reference, establishes a signal segment and main shaft cycle reference point mapping index, and generates a unified segmented signal index matrix.

[0016] As a further scheme of the application, the multi-dimensional feature extraction module comprises:

[0017] The vibration feature calculation submodule obtains vibration acceleration signal data segments based on the unified segmented signal index matrix, selects all vibration data points of each signal segment, detects vibration amplitudes point by point, records time positions of all peak points, calculates time difference values between each pair of adjacent peak points, and sums up to obtain vibration signal peak time difference amounts;

[0018] The current and rotating speed feature extraction submodule reads current signal data segments according to the vibration signal peak time difference amounts, performs point-by-point amplitude difference calculation on all sampling points, judges whether the difference data exceeds a set current mutation threshold, counts the number of mutation points, synchronously calculates a current signal mutation point number proportion, and performs amplitude change amount absolute value operation on rotating speed data of continuous sampling points, calculates a fluctuation amplitude mean value, and obtains a current mutation proportion and a rotating speed fluctuation mean value;

[0019] The temperature feature analysis submodule reads temperature signal data segments according to the current mutation proportion and the rotating speed fluctuation mean value, performs point-by-point difference on temperature values of all sampling points, calculates a unit time temperature rise change rate in combination with a time interval, averages to obtain an average temperature rise rate index, sorts each feature result in chronological order in combination with the vibration signal peak time difference amount, the current mutation proportion and the rotating speed fluctuation mean value, groups according to the time sequence, and establishes a cycle state evolution vector sequence.

[0020] As a further scheme of the application, the state evolution index construction module comprises:

[0021] The difference extraction submodule sequentially obtains three adjacent vector groups based on the sequence of the periodic state evolution vectors, extracts the numerical values of each group of three vectors in each same dimension, respectively performs a subtraction operation on the numerical values between two adjacent periods in the same dimension, and repeats the operation to all dimensions to obtain a sequence of periodic state difference values after collation;

[0022] The trend calculation submodule obtains a root mean square trend change value in a corresponding dimension by performing a square operation and an average root operation on the difference values of each dimension according to the sequence of the periodic state difference values, and obtains a trend change strength by performing an operation on the data before and after the current period to obtain a trend change strength sequence;

[0023] The mutation judgment submodule sets a fault diagnosis threshold based on the trend change strength sequence, performs a threshold comparison operation on the trend strength values of three consecutive periods, and judges whether the three-period sequence is greater than the fault diagnosis threshold one by one, and if the condition is met, records the current period number as a trend mutation point position, extracts and collates the state evolution data corresponding to all mutation points, and establishes a trend mutation state change index set.

[0024] As a further scheme of the present application, the trend aggregation trajectory recognition module comprises:

[0025] The direction consistency recognition submodule obtains the state vectors of two periods before and after each mutation point position based on the trend mutation state change index set, respectively extracts the vector data of each dimension of the periods, calculates the cosine of the angle, judges whether the cosine value is greater than or equal to a direction consistency reference threshold, and if the condition is met, marks the dimension as a direction consistent dimension, otherwise marks the dimension as a direction inconsistent dimension, and obtains a sequence of direction consistent dimension quantity values;

[0026] The aggregation segment determination submodule obtains the number of direction consistent dimensions of each period according to the sequence of the direction consistent dimension quantity values, judges whether the number of direction consistent dimensions of each period is greater than or equal to a principal component contribution rate threshold, and if the condition is met, marks the position of the corresponding period as an aggregation period, identifies a continuous period number range and sets it as an aggregation segment, compares the minimum and maximum values of each dimension state value, obtains the upper and lower boundary values of the state envelope trajectory in the segment, and obtains a state envelope trajectory interval;

[0027] The trajectory offset statistical submodule judges the state vector of each subsequent period in sequence according to the state envelope trajectory interval, compares it with the maximum and minimum values of the envelope interval of the corresponding dimension, judges whether it exceeds the interval boundary range, and if the dimension state value is greater than the maximum value or less than the minimum value, marks the corresponding dimension as an out-of-bound dimension, counts the number of out-of-bound dimensions of each period and calculates the out-of-bound proportion, and if it is greater than a trajectory offset judgment reference value, marks the period state as having a trajectory offset, and generates a trajectory offset identification sequence.

[0028] As a further scheme of the present application, the fault section identification module comprises:

[0029] The out-of-bound cumulative determination submodule traverses all cycles based on the trajectory offset identification sequence, reads each cycle value in the offset identification value sequence, and records the last cycle of the section as an out-of-bound determination trigger point for recording if there are more than three consecutive cycles, counts all cycle numbers that satisfy more than three times of continuous out-of-bound, and obtains a trigger cycle sequence.

[0030] The boundary update decision submodule reads a state vector of five cycles before each trigger cycle according to the trigger cycle sequence, calculates a five-cycle mean value for each dimension, obtains upper and lower boundaries in an aggregated section state envelope trajectory, calculates a dimension difference strength, records an update marker number, and after the cumulative number reaches three times, takes the current mean value interval as an updated boundary to form a corrected state boundary interval.

[0031] The fault section labeling submodule executes tracking determination on the cycles that still continuously offset after updating according to the corrected state boundary interval and the trajectory offset identification sequence, detects whether continuous offset is still in an out-of-bound state after three boundary updates, and if so, marks the corresponding cycle as a fault section cycle, outputs a fault cycle number, and obtains a centrifuge fault prediction result.

[0032] Compared with the prior art, the present application has the following advantages and positive effects:

[0033] In the present application, a multi-channel signal segmentation processing mechanism based on cycle states synchronously aligns different types of data, constructs a state evolution sequence based on a unified sampling structure with a time scale, establishes a state feature trajectory by extracting a parameter set including time displacement difference, current mutation strength, fluctuation amplitude, and temperature rise change, constructs a trend change index based on the root mean square of the difference between adjacent states, identifies an aggregated section using continuous trend mutation and counts trajectory offset frequency, realizes dynamic migration of trajectory boundaries and intelligent identification of fault sections, avoids boundary failure problems caused by static preset conditions, constructs real-time discrimination logic for state evolution, solves the problem of insufficient model adaptation capability under varying operation cycles, and effectively enhances the continuous prediction stability of long-cycle equipment and the adaptation range under non-standard working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The system flowchart of the present application is shown in the figure.

[0035] Figure 2 The data channel synchronization module flowchart of the present application is shown in the figure.

[0036] Figure 3 The multi-dimensional feature extraction module flowchart of the present application is shown in the figure.

[0037] Figure 4 A flow chart of the state evolution index construction module of the present application is shown in FIG. 1;

[0038] Figure 5 A flow chart of the trend aggregation trajectory identification module of the present application is shown in FIG. 2;

[0039] Figure 6 A flow chart of the fault section identification module of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is 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.

[0041] 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 used to facilitate the description of the present application and simplify 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, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0042] Referring to Figure 1 A centrifuge fault prediction system based on machine learning includes:

[0043] The data channel synchronization module acquires continuous signal data of the vibration acceleration sensor, the axial rotation speed recorder, the current monitoring channel and the temperature sensing device deployed in the centrifuge device after the start point of the current main shaft cycle, divides the vibration analysis cycle (30 seconds) into segments with the start point of the cycle time as the reference, synchronizes the time scale of all signal segments (executes the precise time protocol), and generates a unified segmented signal index matrix;

[0044] The multi-dimensional feature extraction module calculates the time difference of the vibration signal peak position based on the unified segmented signal index matrix, counts the proportion of the number of current signal mutation points, acquires the mean value of the continuous fluctuation amplitude of the rotation speed signal, calculates the temperature signal temperature rise rate index (in accordance with the temperature rise test standard of high-voltage switchgear), and constructs a state time sequence in order to form a cycle state evolution vector sequence;

[0045] The state evolution index construction module sequentially obtains the difference values of each dimension between three groups of adjacent vectors based on the periodic state evolution vector sequence, calculates the root mean square of each dimension difference value as a trend change index, records the trend mutation point position when the trend change index of three consecutive periods exceeds the fault diagnosis threshold, and generates a trend mutation state change index set;

[0046] The trend aggregation trajectory recognition module performs cosine similarity analysis (threshold set ≥ 0.85) on the state vectors of the two periods before and after each mutation point position based on the trend mutation state change index set, judges whether the change directions of each dimension are consistent, and if the number of consistent dimensions exceeds the principal component contribution rate threshold (set as 70% of the total dimensions), marks it as an aggregation section, calculates the corresponding change amplitude interval to form a state envelope trajectory, and compares the trajectory offset of each subsequent period state vector, counts the number of out-of-bound times, and obtains a trajectory offset identification sequence.

[0047] The fault section recognition module calculates the difference between the mean value of the five period state vectors before each out-of-bound point and the center point of the aggregation trajectory based on the trajectory offset identification sequence, judges the boundary update amplitude and performs boundary replacement, and when the update accumulates to three times and is still in the out-of-bound state, marks it as a fault area, and generates a centrifuge fault prediction result.

[0048] The unified segmented signal index matrix includes period synchronization number, time scale alignment identifier, signal channel mapping relationship, segmented time sequence index, and sampling interval record; the periodic state evolution vector sequence includes vibration position difference value, current mutation proportion factor, speed fluctuation average amplitude index, temperature rise rate change, and time sequence correlation code; the trend mutation state change index set includes trend response amplitude index, period fluctuation mean square error value, abnormal evolution position index, and vector difference expression degree; the trajectory offset identification sequence includes similarity aggregation label, direction consistency count, trajectory change interval envelope value, and out-of-bound period position record; the centrifuge fault prediction result includes prediction state section number, fault evolution path label, trajectory center difference measure, and fault level discrimination result.

[0049] Please refer to Figure 2 , the data channel synchronization module includes:

[0050] The signal acquisition sub-module acquires continuous signal data of the vibration acceleration sensor, the axial speed recorder, the current monitoring channel and the temperature sensing device, divides the vibration analysis period based on the current main shaft period starting point as the reference benchmark, acquires all signal data segments, and generates a multi-source synchronous signal segment set;

[0051] The continuous signal data of the vibration acceleration sensor, the axial rotation speed recorder, the current monitoring channel and the temperature sensing device are acquired. Firstly, the vibration acceleration channel is detected. A three-axis vibration acceleration sensor of CA-YD102 is selected, the sampling frequency is set to 1024 Hz, the real-time data is acquired through the PLC acquisition module, 1024 data points are collected per second, the sampling time is set to 30 seconds, the total number of collected points is 1024*30 = 30720 points, and then the data is written into the industrial database through the local buffer cache. Subsequently, the axial rotation speed recorder is monitored. A ZX-RPM300 series recorder is selected, the sampling frequency is 256 Hz, and the total sampling points in 30 seconds are 256*30 = 7680 points. The data is saved in CSV format. Then, the signal of the current monitoring channel is acquired. A current transmitter with a rated accuracy of ±0.5% FS is selected, the sampling frequency is set to 100 Hz, and the number of collected points in 30 seconds is 100*30 = 3000 points. The CAN bus protocol is used for data uploading. Finally, the temperature sensing device is collected. A Pt100 thermal resistance temperature sensor is used, the sampling frequency is 10 Hz, and the number of collected points in 30 seconds is 10*30 = 300 points. The Modbus RTU communication protocol is used for real-time uploading to the control system server. After the acquisition of the signal data of the above four channels is completed, the software module of the data receiving end is used to distinguish and store all the signal data according to the channel type, mark the start time, sampling frequency and channel number of each signal segment, integrate them into a unified data format, form a multi-channel and multi-time sequence structure, and combine the current main shaft cycle start point time mark to number and group all the signal segments according to the start and end time period. A multi-source synchronous signal segment set is obtained. The specific sampling parameters are shown in Table 1.

[0052] Table 1 Signal acquisition parameter table:

[0053]

[0054] As shown in Table 1, the specific sampling parameters of each signal channel are listed.

[0055] The time synchronization submodule reads the timestamp data in each signal segment according to the multi-source synchronous signal segment set, calculates the time deviation of each signal segment from the main shaft cycle start point, adjusts the time axis data of each signal segment in combination with the main shaft cycle start point time reference, compares whether the deviation exceeds the time synchronization threshold, and adjusts the signal time axis if it exceeds. Until the time axis of each signal segment meets the synchronization standard, a unified time alignment signal segment set is obtained.

[0056] According to the multi-source synchronization signal segment set, the timestamp data recorded in each signal segment is called. First, the starting timestamp data of the vibration acceleration channel is extracted, the starting time is 0.0025 seconds, and the time deviation from the main shaft cycle starting point time 0 seconds is 2.5 milliseconds. Then, the time deviation of the axial speed signal segment is detected, the starting time is-0.003 seconds, and the deviation is-3.0 milliseconds. Next, the time deviation of the current monitoring channel is calculated, the starting time is 0.0018 seconds, and the deviation is 1.8 milliseconds. Finally, the temperature sensing signal segment is detected, the starting time is-0.0022 seconds, and the deviation is-2.2 milliseconds. The synchronization threshold is set to 5 milliseconds. Whether the time deviation of each signal segment is greater than or less than the threshold is judged. According to the comparison of the absolute value of the deviation and the 5-millisecond threshold, all signal segment deviations meet the threshold condition. The signal segment with a deviation greater than zero is subjected to a time axis left shift operation, and the moving time is the corresponding deviation amount. For example, the starting time of the vibration acceleration signal segment is adjusted from 0.0025 seconds to 0 seconds, and the deviation is less than zero. The signal segment is subjected to a time axis right shift operation. The starting time of the axial speed signal segment is adjusted from-0.003 seconds to 0 seconds, the current monitoring signal segment is adjusted from 0.0018 seconds to 0 seconds, and the temperature sensing signal segment is adjusted from-0.0022 seconds to 0 seconds. After adjustment, the starting time of the four signal segments is rechecked to confirm that all signal segment starting times are aligned with the main shaft cycle starting point 0 seconds. A unified time alignment signal segment set is obtained, and the related deviation adjustment data is shown in Table 2.

[0057] Table 2 Time synchronization adjustment table:

[0058]

[0059] As shown in Table 2, the time deviation detection and adjustment record of each signal segment is shown.

[0060] The signal index generation submodule records the start and end time points, signal length, sampling rate and channel serial number of each signal segment according to the unified time alignment signal segment set, integrates the time points and channel numbers of each signal segment, establishes a signal segment and main shaft cycle reference point mapping index, and generates a unified segmented signal index matrix;

[0061] According to the unified time alignment signal segment set, the start and end time points, data length, sampling rate and channel serial number of four types of signal segments of vibration acceleration, axial speed, current monitoring and temperature sensing are extracted. Firstly, the vibration acceleration signal segment is numbered, the channel number is 1, the corresponding time period is 0-30000 milliseconds, the data length is 30720 points, the index serial number is set as 101, then the axial speed signal segment is numbered as 2, the corresponding time period is also 0-30000 milliseconds, the data length is 7680 points, the index serial number is 102, the current monitoring signal segment is numbered as 3, the time period is 0-30000 milliseconds, the data length is 3000 points, the index serial number is 103, the temperature sensing signal segment is numbered as 4, the time period is 0-30000 milliseconds, the data length is 300 points, the index serial number is 104, by reading the channel number and time period information of each signal segment, the signal segments are sorted according to the ascending order of channel number, the mapping relationship is established according to the signal segment number sequence, the signal segment number, corresponding time period, channel number and index serial number are integrated, the signal segment index data set of four rows and one column is formed, and finally the unified segmented signal index matrix is generated, as shown in Table 3.

[0062] Table 3 signal index mapping table:

[0063]

[0064] As shown in Table 3, it is the detailed mapping data of signal segment and corresponding index serial number.

[0065] Please refer to Figure 3 , the multi-dimensional feature extraction module comprises:

[0066] The vibration feature calculation sub-module obtains the vibration acceleration signal data segment based on the unified segmented signal index matrix, selects all vibration data points of each signal segment, detects the vibration amplitude point by point, records the time position of all peak points, calculates the time difference between each pair of adjacent peak points, and obtains the vibration signal peak time difference.

[0067] The uniform segmented signal index matrix is obtained, all data segments of the vibration acceleration signal channel are extracted in turn, point-by-point vibration amplitude detection is performed on each data segment, the vibration signal data with a sampling frequency of 1024 Hz is first read, 1024 data points collected per second are distributed to the time axis, peak value detection is performed using a threshold value of 0.8 g, whether each sampling point meets the peak value condition is judged, when the vibration amplitude of a point exceeds 0.8 g, the point is identified as a peak value point, for example, the vibration amplitude of the 125th point is 1.2 g, the vibration amplitude of the 280th point is 1.1 g, and the vibration amplitude of the 430th point is 1.3 g, the timestamp data of these peak value points is extracted, the time difference of all adjacent peak value pairs is calculated, for example, the 125th point corresponds to a time of 0.122 seconds, the 280th point corresponds to a time of 0.273 seconds, and the time difference between the two points is 0.151 seconds, the time difference of all subsequent peak value pairs is sequentially solved, the number of peak value pairs is counted, it is assumed that 150 peak value pairs are detected in a 30-second data segment, the total time difference cumulative value is 1.86 seconds, the total time difference is divided by the number of peak value pairs by using an arithmetic average method, that is, 1.86 / 150, the average peak value time difference is 0.0124 seconds, which is converted to 12.4 milliseconds, and finally the vibration signal peak value time difference is generated, as shown in Table 4.

[0068] Table 4: Vibration peak value time difference calculation table

[0069]

[0070] As shown in Table 4, the data of the time difference calculation process of the vibration signal peak value pair in the embodiment is shown.

[0071] The current and speed characteristic extraction submodule reads the current signal data segment according to the vibration signal peak value time difference, performs point-by-point amplitude difference calculation on all sampling points, judges whether the difference data exceeds the set current mutation threshold value, counts the number of mutation points, synchronously calculates the proportion of the number of current signal mutation points, and performs amplitude change amount absolute value operation on the speed data of the continuous sampling points, calculates the fluctuation amplitude mean value, and obtains the current mutation proportion and the speed fluctuation mean value;

[0072] According to the vibration signal peak time difference, the current signal data segment is extracted, the current difference of all adjacent sampling points is calculated, the current sampling frequency is set to 100 Hz, the sampling time interval is 0.01 seconds, the current change of each pair of adjacent sampling points is calculated, the threshold ±2A is used to judge the mutation point, for example, the current of the 100th point is 8.0A, the current of the 101th point is 10.8A, the difference value is 2.8A, which is judged as a mutation point, and the number of mutation points is accumulated, for example, 150 mutation points are detected in 3000 sampling points, the proportion of mutation points is 150 / 3000=0.05, that is, 5%, then the speed signal data segment is called, the speed signal with a sampling frequency of 256 Hz is read, the speed change between adjacent sampling points is processed by absolute value, for example, the speed of the 500th point is 1490 rpm, the speed of the 501th point is 1495 rpm, the single point change is 5 rpm, and the 7680 point data is calculated pair by pair, the total fluctuation change is 24576 rpm after the sum of all changes, and the fluctuation amplitude mean value is 24576 / 7680=3.2 rpm, the current mutation proportion and the speed fluctuation mean value are finally obtained, as shown in Table 5.

[0073] Table 5: Current and speed characteristic calculation table

[0074]

[0075] As shown in Table 5, the calculation process data of current mutation proportion and speed fluctuation mean value.

[0076] The temperature feature analysis submodule reads the temperature signal data segment according to the current mutation proportion and the speed fluctuation mean value, calculates the temperature difference of all sampling points, calculates the temperature rise rate per unit time combined with the time interval, and obtains the average temperature rise rate index. According to the vibration signal peak time difference, the current mutation proportion and the speed fluctuation mean value, the characteristic results are sorted in time sequence, combined in time sequence, and a periodic state evolution vector sequence is established.

[0077] According to the current sudden change proportion and the average speed fluctuation, the temperature signal data segment is extracted, the temperature sampling data with a sampling frequency of 10Hz is read, the sampling time interval is 0.1s, the temperature difference calculation is performed on each pair of adjacent sampling points, for example, the temperature of the 50th point is 25.0℃, the temperature of the 51st point is 25.3℃, the temperature difference is 0.3℃, the temperature rise rate is 0.3 / 0.1=3.0℃ / s, the temperature rise rate calculation is performed on the whole 299 pairs of temperature points in turn, the total temperature rise rate cumulative sum is 450℃ / s, then divided by 299 pairs of points, the average temperature rise rate is 450 / 299=1.51℃ / s, then combined with the vibration signal peak time difference 12.4ms, the current sudden change proportion and the average speed fluctuation 0.05 and 3.2rpm, and the temperature rise rate 1.51℃ / s, the characteristic values are arranged in time sequence in turn, and the four-dimensional vector sequence is obtained, which represents the state evolution process under the corresponding period, and finally the period state evolution vector sequence is established, as shown in Table 6.

[0078] Table 6: Period state evolution vector table

[0079]

[0080] As shown in Table 6, the specific characteristic quantity summary of the period state evolution vector sequence is shown.

[0081] Please refer to Figure 4 , the state evolution index construction module comprises:

[0082] The difference extraction submodule sequentially obtains three adjacent vector groups based on the period state evolution vector sequence, extracts the values of each group of three vectors in each same dimension, and performs subtraction operation on the values between two adjacent periods in the same dimension, and repeats the operation until all dimensions are obtained. The period state difference value sequence table is obtained by arranging;

[0083] Based on the periodic state evolution vector sequence, the state vector samples of the adjacent three periods are extracted in turn, the numerical values of each group of vectors in four dimensions (vibration signal, current signal, speed signal, temperature signal) are extracted, and the difference operation is performed on the same dimension values between adjacent periods, for example, between the 1st period and the 2nd period, the vibration dimension value is 12.4ms and 11.7ms, the difference is 0.7ms, the current mutation proportion is 0.08 and 0.05, the difference is 0.03, the speed fluctuation average is 3.2rpm and 7.2rpm, the difference is 4rpm, the temperature rise rate is 1.5℃ / s and 1.2℃ / s, the difference is 0.3℃ / s, the same operation is performed on all dimensions, and the difference value data is recorded, and the period number is traversed to ensure that the difference is performed on adjacent period data each time, for example, after repeating the above operation on the 2nd period and the 3rd period, the vibration difference value is 1.1ms, the current difference value is 0.02, the speed difference value is 5rpm, and the temperature rise difference value is 0.4℃ / s, after processing the 3rd period and the 4th period, the vibration difference value is 0.5ms, the current difference value is 0.06, the speed difference value is 3rpm, and the temperature rise difference value is 0.5℃ / s, on this basis, a multi-dimensional difference value data set is formed, the difference values of each period are numbered and established into a structured table, and the results are shown in Table 7 to obtain a period state difference value sequence table.

[0084] Table 7: Period state difference value example table

[0085]

[0086] As shown in Table 7, the four-dimensional feature difference values between adjacent periods are recorded for subsequent trend strength index calculation and calling.

[0087] The trend calculation submodule performs square operation and average root operation on each dimension difference item according to the period state difference value sequence table, obtains the root mean square trend change value under the corresponding dimension, combines the data before and after the current period, and uses the formula:

[0088] ;

[0089] The operation obtains the trend change strength to obtain a trend change strength sequence, wherein, indicates the trend change strength, the unit is the same as the difference value unit of change, indicates the difference value of the current period state vector and the previous period state vector in the dimension, indicates the average value of all dimension difference values, indicates the trend change gradient of the dimension, indicates the trend change gradient of the dimension state change amplitude value, the unit is the same as same, denotes the cycle state change trend offset value, the unit is consistent with the denominator;

[0090] According to the cycle state difference value sequence, first, the difference value set under each dimension is counted, such as vibration difference value [0.7, 1.1, 0.5], current difference value [0.03, 0.02, 0.06], rotating speed difference value [4, 5, 3], temperature rise difference value [0.3, 0.4, 0.5], the average value of each dimension difference sequence is calculated respectively, the vibration average difference value is 0.8, the current average difference value is 0.037, the current change is small and does not need to be standardized, the rotating speed average value is 4.0 rpm, and the temperature rise average value is 0.4℃ / s, then taking the vibration dimension as an example, the trend gradient index sequence is constructed, which is defined as the continuous cycle difference value change rate, which can be expressed as , wherein the cycle sampling interval is 1 unit cycle, then the G sequence is [0.4, -0.6] and the like, and the trend intensity value is calculated according to the formula:

[0091] Taking a numerical example, suppose:

[0092] The vibration dimension difference value sequence is ;

[0093] The average value is ;

[0094] The gradient value sequence is ;

[0095] The state change amplitude sequence is ;

[0096] The trend offset is ;

[0097] In turn:

[0098] The numerator is:

[0099] ;

[0100] The denominator is:

[0101] ;

[0102] The trend change intensity value is:

[0103] ;

[0104] The result shows that the trend change intensity is-0.0346, and the trend change intensity sequence is finally obtained.

[0105] ​The trend change intensity is a comprehensive measurement index for measuring the change trend of the periodic state vector in continuous time series, which reflects the coupling degree of the fluctuation amplitude and the change rate of the system in multiple characteristic dimensions relative to the average state. The index not only considers the difference amplitude of each dimension between the current period and the adjacent period, but also introduces the trend gradient as a dynamic parameter, thereby realizing the quantitative evaluation of whether the change is continuously intensified. At the same time, by integrating the multi-dimensional state amplitude fluctuation and the periodic offset degree to construct a normalized benchmark, it can maintain dimensional consistency and comparability at different scales and units. Therefore, the trend change intensity not only reflects the severity of state change, but also reflects the coordination and mutation between the changes of each dimension, which is the key basis for subsequent judgment of the trend mutation point position.

[0106] The formula molecule part is obtained by taking the absolute value of the difference between the current period difference value of each dimension and the average difference value of all dimensions and multiplying it by the trend change gradient of the corresponding dimension , which realizes the dual weighting of the difference degree and the change rate. The absolute value operation is used to remove the directional influence, so that the formula focuses on the change amount rather than the positive or negative fluctuation. The multiplication operation reflects the dynamic amplification or attenuation of the difference value under the trend change, and the summation operation uniformly accumulates the trend difference effects of all dimensions to form the global trend driving total amount. The denominator part is composed of the square sum of the state change amplitudes of all dimensions , which is an energy type index, and then the root operation is used to restore the original change amount dimension, and the absolute value of the periodic trend offset degree is added to construct the trend disturbance total resistance term. The square sum reflects the current overall change scale of the system, the root operation makes it comparable with the numerator term, and the absolute value term is a disturbance factor introduced by the irregular fluctuation across periods, which constitutes a dynamic adjustment factor for the denominator to the trend intensity. Therefore, the overall formula forms a ratio between the "trend incentive amount (numerator)" and the "system resistance amount (denominator)" in structure, which has the functions of direction-independent, amplitude-weighted, and disturbance-controllable multi-dimensional trend intensity judgment.

[0107] The mutation judgment submodule is based on the trend change intensity sequence, sets a fault diagnosis threshold, and performs a threshold comparison operation on the trend intensity values of three consecutive periods. It judges whether each three-period sequence is greater than the fault diagnosis threshold. If the condition is met, the current period number is recorded as the trend mutation point position, and the state evolution data corresponding to all mutation points are extracted and organized to establish a trend mutation state change index set.

[0108] According to the trend change intensity sequence, a decision operation is performed on the continuous period T value sequence, the trend mutation threshold is defined as 3.5, the threshold setting is based on the 95% confidence upper limit of the state change intensity in the system operation, and in large-scale operation condition monitoring, after 1000 groups of different period state data are counted, it is observed that the average amplitude of the difference value in all dimensions is not more than 0.9, the gradient value is not more than ±0.6, and the period trend offset degree is maintained between 0.3 and 0.6, and in this combination of conditions, the T value is distributed in the range of 0.5 to 3.2, so the upper interval is expanded by 0.1 margin as the mutation recognition critical point is set to 3.5, which is most sensitive to fluctuation intensity, especially with significant upward trend with the increase of vibration difference and temperature rise fluctuation, and has a quantifiable judgment effect, and the decision condition is that the trend intensity values of three consecutive periods are all greater than the threshold value, for example, a sequence T = [3.2, 3.7, 4.0, 3.6, 3.3, 3.8], whether the sequence segment meets the condition is detected in a sliding window manner, the T values in the 2nd to 4th period interval are [3.7, 4.0, 3.6], all of which are greater than the threshold value 3.5, which meets the mutation condition, so the period 4 is recorded as the trend mutation point position, the period numbers of all periods meeting the condition are recorded, the number of period mutation points is counted, and the corresponding original state feature vector data is extracted, that is, the original state sequence of period 4 is vibration time difference 12.8 ms, current mutation proportion 0.085, speed fluctuation 3.9 rpm, and temperature rise rate 1.65℃ / s, the multi-dimensional state vector data corresponding to the period is taken as the mutation state flag sample, all mutation point samples are recorded repeatedly, and finally a trend mutation state change index set is established.

[0109] Please refer to Figure 5 , the trend aggregation trajectory recognition module comprises:

[0110] The direction consistency recognition submodule is based on the trend mutation state change index set, obtains the state vectors of the two periods before and after each mutation point position, respectively extracts the vector data of each dimension of the period, calculates the cosine value, judges whether the cosine value is greater than or equal to the direction consistency reference threshold, and if the condition is met, marks it as a direction consistent dimension, otherwise marks it as a direction inconsistent dimension, and obtains a sequence of the number of direction consistent dimensions.

[0111] Based on the trend mutation state change index set, the period numbers of all mutation points are extracted, and the state vectors of the two periods before and after each mutation period are obtained, and the mutation point period is set as , the four-dimensional state values of the periods and are extracted respectively as follows: , the vibration dimension: , the current dimension: , the speed dimension: The state value pair corresponding to each dimension is respectively constituted into a vector With , the cosine similarity calculation method is used for direction judgment, and the dot product calculation is the multiplication of each dimension component and the summation, for example, the dot product of the current dimension is , the module length is , the module length product is , the cosine value is , the direction consistency threshold is set to , the value is obtained by collecting 256 mutation cycle points in a day and the corresponding state vectors before and after the cycle, calculating the cosine value distribution of each dimension, the results show that more than 99% of the cosine value of the normal operation cycle is concentrated in the interval of 0.88 to 0.99, and the cosine value of most disturbance dimensions in all state mutation cycles is less than 0.80, therefore, 0.85 is selected as the dividing point to exclude normal weak fluctuations and cover the common mutation dimension direction reversal, and the threshold value will decrease slightly with the increase of the range of vibration dimension and speed dimension, there is empirical relationship, wherein indicates the average value of the multi-dimensional state variance, whether the cosine value is greater than the threshold value is judged, if the condition is met, the direction of the dimension is consistent, and the above process is repeated in each dimension to obtain the direction consistency of each dimension in cycle 3 as follows:

[0112] Table 8 Direction consistency judgment table of cycle state vector:

[0113]

[0114] Repeat the above process for each cycle to count the number of direction consistent dimensions, and the number of direction consistent dimensions in cycle 3 is , and the number of direction consistent dimensions is obtained.

[0115] The aggregation section judgment submodule obtains the number of direction consistent dimensions of each cycle according to the number of direction consistent dimensions, judges whether the number of direction consistent dimensions of each cycle is greater than or equal to the principal component contribution rate threshold, if the condition is met, marks the position of the corresponding cycle as an aggregation cycle, identifies the continuous cycle number range and sets it as an aggregation section, compares the minimum and maximum values of each dimension state value, obtains the upper and lower boundary values of the state in the section, and obtains the state envelope trajectory interval;

[0116] According to the number of direction consistent dimensions , the number of direction consistent dimensions of each cycle is obtained, the total number of dimensions is set to , the principal component contribution rate threshold is set to , the aggregation judgment threshold is calculated as , and the upper integer is obtained as , if The period is marked as the aggregation period, for example, the period 4 is 6, and the period 5 is 7, forming a continuous section , all the period state vectors of the section are extracted, and the upper and lower boundary extraction operations are performed on each dimension, and the vibration dimension value is , the current dimension is , the rotation speed dimension is , and the temperature rise dimension is , and the envelope interval is constructed as follows:

[0117] Table 9: Aggregated section state envelope trajectory interval table

[0118]

[0119] The state envelope trajectory interval is finally obtained , wherein is the dimension number, is the lower boundary, is the upper boundary.

[0120] The trajectory offset statistical submodule sequentially judges the state vector of each subsequent period according to the state envelope trajectory interval, compares it with the maximum value and the minimum value of the envelope interval of the corresponding dimension, judges whether it exceeds the interval boundary range, if the dimension state value is greater than the maximum value or less than the minimum value, the corresponding dimension is marked as an out-of-bound dimension, the number of out-of-bound dimensions per period is counted and the out-of-bound proportion is calculated, if it is greater than the trajectory offset judgment reference value, the period state is marked as trajectory offset, and the trajectory offset identification sequence is generated;

[0121] According to the state envelope trajectory interval , the state vector of the period is judged, if the state value of a certain dimension , it is considered as out-of-bound, in the period 6, the state vector is: vibration , current , rotation speed , and temperature rise , the out-of-bound dimensions are vibration and current; the number of out-of-bound dimensions is 2, and the proportion is , no trajectory offset is triggered; in the period 7, the state is: vibration , current , rotation speed , and temperature rise , all the four dimensions are out-of-bound, the number of out-of-bound is , the proportion is , the trajectory offset judgment threshold is met, the period 7 is recorded as the trajectory offset period, the period offset situation is summarized after the period 8 is sequentially judged, and the trajectory offset identification sequence is generated , wherein represents the period offset, means no offset.

[0122] Please refer to Figure 6 , the fault section identification module comprises:

[0123] The out-of-bound cumulative determination submodule traverses all periods based on the trajectory offset identification sequence, reads each period value in the offset identification value sequence, and if there are more than three consecutive periods, records the last period of the section as an out-of-bound determination trigger point, counts all period numbers that satisfy more than three consecutive out-of-bound times, and obtains a trigger period sequence.

[0124] Based on the offset state identification value in the trajectory offset identification sequence, the offset flag corresponding to each period in the sequence is read If four consecutive periods satisfy the condition of being 1, the last period is recorded as an out-of-bound determination trigger point. First, the offset flag value of the full sequence is obtained. In the example, the offset flag values of periods 2 to 5 are all 1, which satisfies the condition of four consecutive out-of-bound periods. Therefore, period 5 is marked as a trigger period. Continue to judge the values of periods 6 to 9 in the sample data. Among them, periods 7 to 9 are consecutive three periods of offset, which do not satisfy the recording condition, and the judgment is terminated. In actual operation, a sliding window operation needs to be performed on the offset identification array with a length of , to determine whether each subarray with a length of 4 is all 1. Through a loop, the window segment is extracted and a logical judgment operation is performed Each time it is satisfied, the is added to the trigger period set. Table 10 gives a set of sample data.

[0125] Table 10: Trajectory offset identification sequence sample table:

[0126]

[0127] As shown in Table 10, the system identifies only one window that satisfies the condition, and the corresponding trigger period number is "period 5". In a real scenario, this operation can process a state sequence with a sampling period of 0.5 seconds, indicating that the system is always in an abnormal offset state within 2 seconds. Finally, the trigger period sequence is obtained.

[0128] The boundary update decision submodule reads the state vector of the five periods before each trigger period according to the trigger period sequence, calculates the five-period mean for each dimension, and obtains the upper and lower boundaries in the aggregated section state envelope trajectory interval using the formula:

[0129] ;

[0130] The operation obtains the dimension difference intensity, records the update mark times, and after the cumulative times reach three times, takes the current mean interval as an update boundary to form a correction state boundary interval, wherein, representing the period The normalized difference intensity on the dimension representing the period The state mean on the dimension in the previous 5 periods, The state value on the dimension in the first period, The lower boundary and the upper boundary of the state envelope on the dimension of the aggregation section, representing the fluctuation span of the dimension state trajectory, used for normalization processing;

[0131] According to the trigger period sequence obtained in the foregoing, for each trigger period , a state trajectory tracing operation is performed to obtain the state vector from period to , five state values are extracted for each dimension and the mean value is calculated, the preset state envelope trajectory boundary is called again, the boundary center point and the interval width are obtained, then the deviation amount of the mean value from the center point is compared, and the historical deviation sum under the dimension is added, and the formula is used for calculation.

[0132] The example data is shown in Table 11.

[0133] Table 11: Example table of state values in the previous five periods (temperature rise dimension) of period 5:

[0134]

[0135] Suppose the temperature rise envelope boundary of the aggregation section is , then , the center point is , the mean value deviation is (1.346-1.33=0.016), and the historical fluctuation sum is:

[0136] ;

[0137] Substituting into the formula gives:

[0138] ;

[0139] ​The comparison update threshold is set to 0.15, the update condition is met, so the recording period 5 is triggered for one update in the temperature rise dimension, and if the cumulative trigger is more than three times in the subsequent period, the state envelope boundary is updated, and finally the corrected state boundary interval is generated.

[0140] The dimension difference intensity is used to measure the overall deviation of a certain dimension in the current period relative to the historical behavior and boundary reference, which essentially reflects whether the dimension shows a trend of divergence or abnormal fluctuation in a certain time window. This index considers two aspects of information: one is the distance between the dimension state mean value and the center position of the aggregated section in the current rolling period, representing the overall deviation trend; the other is the fluctuation degree of the historical period state value relative to the mean value, reflecting the local stability and disturbance activity. By weighting and combining the two parts and normalizing them within the aggregated section state envelope width, the dimension difference intensity can directly represent whether the dimension deviates from the original stable aggregation state and has a trend of transition. Therefore, the larger the value, the more likely the dimension trajectory is in a critical state of deviation or instability, and the higher the risk of dynamic disturbance.

[0141] In the formula, represents the period The absolute deviation of the mean value of the first five periods from the envelope center is used to describe the overall trajectory deviation trend, and the deviation of the mean value of the last five periods from the envelope center is used to describe the local fluctuation trend. The fluctuation amplitude of the historical state value relative to the mean value is used to reflect the trajectory fluctuation activity, and the sum of the two parts forms a complete periodic deviation index. The denominator represents the state fluctuation range of the aggregated section in this dimension, which is used as a normalization factor to scale and standardize the overall deviation amplitude. Therefore, the fractional form is used to scale the numerator deviation item, ensuring comparability in different dimensions and different dimensions. The pre-coefficient is used to average the five-period historical deviation, so that the overall formula considers both the absolute deviation trend and the local fluctuation intensity, and uniformly adjusts the scale, building an index system that reflects both the trajectory deviation direction and the local disturbance sensitivity.

[0142] The fault section labeling submodule, according to the corrected state boundary interval, combines the trajectory deviation identification sequence, and performs tracking judgment on the periods that continue to deviate after updating. It detects whether the continuous deviation is still in the out-of-boundary state after completing three boundary updates. If it meets the condition, it will mark the corresponding period as a fault section period, output the fault period number, and obtain the centrifuge fault prediction result.

[0143] Based on the obtained corrected state boundary interval, the period state value is detected again to see if it is still out of boundary after updating. The state of each period is judged, and whether or The case where The updated boundary value, for example, the state value of cycle 6 in the temperature rise dimension is , has exceeded the updated interval and is marked as out of range. If this out of range condition occurs after three boundary updates and still satisfies the out of range, the cycle will be determined as a failure cycle, the cycle number is recorded and all such cycles are summarized, and finally the centrifuge failure prediction result is obtained.

[0144] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can use the disclosed technical content to make changes or modifications as 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, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. A machine learning based centrifuge failure prediction system, characterized by, The system comprises: The data channel synchronization module acquires continuous signal data of the current main shaft cycle starting point of the centrifuge, divides the vibration analysis cycle segment, and synchronizes the time scale of all signal segments to generate a unified segmented signal index matrix; The multi-dimensional feature extraction module calculates the vibration signal peak position time difference based on the unified segmented signal index matrix, counts the proportion of current signal mutation points, acquires the mean value of the continuous fluctuation amplitude of the speed signal, calculates the temperature signal temperature rise rate index, and sequentially constructs a cycle state evolution vector sequence; The state evolution index construction module calculates the root mean square of each dimension difference as a trend change index based on the cycle state evolution vector sequence, records the trend mutation point position, and generates a trend mutation state change index set; The trend aggregation trajectory identification module performs cosine similarity analysis and labels the aggregation segment based on the trend mutation state change index set, compares the trajectory offset and counts the number of out-of-bound times to obtain a trajectory offset identification sequence; The fault segment identification module calculates the difference degree to judge the boundary update amplitude and performs boundary replacement based on the trajectory offset identification sequence, marks the fault area, and generates a centrifuge fault prediction result.

2. The machine learning based centrifuge failure prediction system of claim 1, wherein, The unified segmented signal index matrix includes cycle synchronization number, time scale alignment identifier, signal channel mapping relationship, segmented time sequence index, and sampling interval record; the cycle state evolution vector sequence includes vibration position difference value, current mutation proportion factor, speed fluctuation mean amplitude index, temperature rise rate change, and time sequence correlation code; the trend mutation state change index set includes trend response amplitude index, cycle fluctuation mean square difference value, abnormal evolution position index, and vector difference expression degree; the trajectory offset identification sequence includes similarity aggregation label, direction consistency count, trajectory change interval envelope value, and out-of-bound cycle position record; and the centrifuge fault prediction result includes prediction state segment number, fault evolution path label, trajectory center difference measure, and fault level discrimination result.

3. The machine learning based centrifuge failure prediction system of claim 1, wherein, The data channel synchronization module comprises: The signal acquisition submodule acquires continuous signal data of the vibration acceleration sensor, the axial speed recorder, the current monitoring channel, and the temperature sensing device, divides the vibration analysis cycle based on the current main shaft cycle starting point as the reference benchmark, and acquires all signal data segments to generate a multi-source synchronized signal segment set; The time synchronization submodule reads the timestamp data inside each signal segment according to the multi-source synchronized signal segment set, calculates the time deviation of each signal segment from the main shaft cycle starting point, adjusts the time axis data of each signal segment in combination with the main shaft cycle starting point time benchmark, compares whether the deviation exceeds the time synchronization threshold, and if it does, adjusts the signal time axis by adding or subtracting until the time axis of each signal segment meets the synchronization standard to obtain a unified time-aligned signal segment set; The signal index generation submodule records the start and end time points, signal length, sampling rate and channel serial number of each signal segment according to the uniform time alignment signal segment set, integrates the time points and channel numbers of each signal segment in combination with the main shaft cycle start point reference benchmark, establishes a signal segment and main shaft cycle reference point mapping index, and generates a uniform segmented signal index matrix.

4. The machine learning based centrifuge failure prediction system of claim 1, wherein, The multi-dimensional feature extraction module comprises: The vibration feature calculation submodule obtains a vibration acceleration signal data segment based on the uniform segmented signal index matrix, selects all vibration data points of each signal segment, detects vibration amplitudes point by point, records the time positions of all peak points, calculates the time difference between each pair of adjacent peak points, and sums up to obtain a vibration signal peak time difference amount; The current and rotating speed feature extraction submodule reads a current signal data segment according to the vibration signal peak time difference amount, performs point-by-point amplitude difference calculation on all sampling points, judges whether the difference data exceeds a set current mutation threshold, counts the number of mutation points, synchronously calculates the proportion of the number of current signal mutation points, and performs amplitude change amount absolute value operation on the rotating speed data of continuous sampling points, calculates the fluctuation amplitude mean value, and obtains the current mutation proportion and the rotating speed fluctuation mean value; The temperature feature analysis submodule reads a temperature signal data segment according to the current mutation proportion and the rotating speed fluctuation mean value, performs point-by-point difference on the temperature values of all sampling points, calculates the unit time temperature rise rate in combination with the time interval, averages to obtain an average temperature rise rate index, sorts the feature results in chronological order in combination with the vibration signal peak time difference amount, the current mutation proportion and the rotating speed fluctuation mean value, combines in chronological order, and establishes a cycle state evolution vector sequence.

5. The machine learning based centrifuge failure prediction system of claim 1, wherein, The state evolution index construction module comprises: The difference extraction submodule sequentially obtains adjacent three groups of vectors based on the cycle state evolution vector sequence, extracts the values of each group of three vectors in each same dimension, respectively performs subtraction operation on the values between two adjacent cycles in the same dimension, and repeats the operation to all dimensions to obtain a cycle state difference value list; The trend calculation submodule performs square operation and average root operation on each dimension difference item according to the cycle state difference value list, obtains the root mean square trend change value in the corresponding dimension, calculates the trend change strength in combination with the data before and after the current cycle, obtains a trend change strength sequence; The mutation judgment submodule sets a fault diagnosis threshold based on the trend change strength sequence, performs threshold comparison operation on the trend strength values of three continuous cycles, judges whether the three-cycle sequence is greater than the fault diagnosis threshold one by one, and if the condition is met, records the current cycle number as a trend mutation point position, extracts and arranges the state evolution data corresponding to all mutation points, and establishes a trend mutation state change index set.

6. The machine learning based centrifuge failure prediction system of claim 1, wherein, The trend aggregation trajectory recognition module comprises: The direction consistency identification submodule obtains state vectors of two periods before and after each mutation point position based on the set of trend mutation state change indicators, respectively extracts each dimension vector data of the periods, calculates a cosine value, judges whether the cosine value is greater than or equal to a direction consistency reference threshold, and if the condition is met, marks the dimension as a direction consistent dimension, otherwise marks the dimension as a direction inconsistent dimension, and obtains a sequence of direction consistent dimension quantity values; The aggregation section determination submodule obtains a direction consistent dimension value of each period according to the sequence of direction consistent dimension quantity values, judges whether the number of direction consistent dimensions of each period is greater than or equal to a principal component contribution rate threshold, and if the condition is met, marks the position of the corresponding period as an aggregation period, identifies a continuous period number range and sets it as an aggregation section, compares the minimum and maximum values of each dimension state value, obtains state upper and lower boundary values in the section, and obtains a state envelope trajectory interval; The trajectory offset statistical submodule judges the state vector of each subsequent period in sequence according to the state envelope trajectory interval, compares the maximum and minimum values of the envelope interval of the corresponding dimension, judges whether it is out of the interval boundary range, and if the dimension state value is greater than the maximum value or less than the minimum value, marks the corresponding dimension as an out-of-bound dimension, counts the number of out-of-bound dimensions of each period and calculates the out-of-bound proportion, and if it is greater than a trajectory offset judgment reference value, marks that the period state has a trajectory offset, and generates a trajectory offset identification sequence.

7. The machine learning based centrifuge failure prediction system of claim 1, wherein, The fault section identification module comprises: The out-of-bound cumulative determination submodule traverses and judges all periods based on the trajectory offset identification sequence, reads each period value in the offset identification value sequence, records the last period of the section as an out-of-bound judgment trigger point if there are more than three consecutive periods, counts all period numbers that satisfy the condition of more than three consecutive out-of-bound times, and obtains a trigger period sequence; The boundary update decision submodule reads the state vectors of the five periods before each trigger period according to the trigger period sequence, calculates the five-period mean value of each dimension, obtains the upper and lower boundaries in the state envelope trajectory interval of the aggregation section, calculates the dimension difference intensity, records the number of update marks, and after the cumulative number reaches three times, takes the current mean value interval as the updated boundary to form a corrected state boundary interval; The fault section labeling submodule performs tracking judgment on the periods that still continuously offset after updating according to the corrected state boundary interval and the trajectory offset identification sequence, detects whether the continuous offset is still in the out-of-bound state after completing three boundary updates, and if the condition is met, marks the corresponding period as a fault section period, outputs the fault period number, and obtains a centrifuge fault prediction result.

Images