Compressor diaphragm rupture early warning method and system based on time series data

Abstract

The present application relates to the technical field of compressor diaphragm detection, and particularly relates to a compressor diaphragm rupture early warning method and system based on time series data, which comprises the following steps: obtaining multi-dimensional signals in the operation process of the compressor and performing band-pass filtering and resampling processing to obtain a standardized compression cycle waveform; extracting a characteristic frequency band signal, calculating an early damage degree positively correlated with the kurtosis deviation degree of the signal; combining the early damage degree with the recent history mean value and the waveform shape deviation degree to construct a medium-term damage degree; further combining the medium-term damage degree with the abnormal score excess ratio to construct a late damage degree; inputting the damage degrees of each stage as weights to dynamically determine the adaptive window length; and finally calculating the abnormal score on the current vibration signal and outputting the early warning. The present application effectively eliminates false alarms caused by real working condition drift, and improves the capture sensitivity of extremely weak early rupture precursors.

Description

Technical Field

[0001] This invention relates to the field of compressor diaphragm detection technology, and in particular to a method and system for early warning of compressor diaphragm rupture based on time-series data. Background Technology

[0002] Diaphragm compressors, due to their leak-free operation, corrosion resistance, and suitability for high-pressure clean media, are used in high-safety applications such as hydrogen storage and transportation, petrochemical distillation, and semiconductor process gas delivery. Their core sealing component is a metal diaphragm, which, under the combined effects of long-term alternating pressure loads and corrosive media, can develop and propagate fatigue cracks, ultimately leading to sudden rupture. Diaphragm rupture will cause large-scale leakage of process media, easily resulting in major safety accidents such as combustion, explosion, and poisoning, along with unplanned downtime losses. Therefore, it is necessary to utilize multi-dimensional time-series sensor data during compressor operation to establish a precise early warning mechanism for diaphragm rupture, ensuring the safe and stable operation of the compressor.

[0003] Chinese patent application CN115561575A discloses a method for identifying electrical anomalies in offshore wind farms based on multidimensional matrix profiles. Taking common fault states and disturbance phenomena in offshore wind farms as the subject of the invention, the effectiveness of the algorithm is demonstrated through theoretical analysis, simulation calculation and actual data verification.

[0004] In existing technologies, the core idea of ​​the matrix profilometry algorithm is to construct a global profile vector by calculating the normalized Euclidean distance between a time-series subsequence and its nearest neighbor, thereby locating abnormal time-series subsequences. This is suitable for real-time early warning of continuously acquired time-series data in industrial settings. However, during the operation of a diaphragm compressor, the vibration and chamber pressure time-series signals exhibit strong quasi-periodicity, and the length of this period continuously and dynamically drifts with load adjustments and speed fluctuations. That is, even under normal operating conditions, the signal period varies across different operating periods. The matrix profilometry algorithm relies on a fixed-length window for subsequence similarity matching. When operating conditions drift, the period misalignment between the normal operating segment and the historical reference template will lead to artificially high distance values, causing false alarms. Furthermore, the slight variations in amplitude and frequency caused by actual early diaphragm damage are easily drowned out by the background noise generated by operating condition drift, resulting in missed detection of diaphragm rupture damage. Summary of the Invention

[0005] To address the problem that existing compressor diaphragm rupture early warning methods, which employ fixed-window matrix profile algorithms, are prone to false alarms and missed detections under varying operating conditions, this invention provides a compressor diaphragm rupture early warning method and system based on time-series data.

[0006] In a first aspect, the present invention provides a compressor diaphragm rupture early warning method based on time-series data, employing the following technical solution: A compressor diaphragm rupture early warning method based on time-series data includes the following steps: S1: Acquire vibration acceleration signals, gas chamber pressure signals, and oil chamber pressure signals during compressor operation; perform bandpass filtering on the vibration acceleration signals to obtain characteristic frequency band signals; and perform resampling processing on the vibration acceleration signals to obtain standardized compression cycle waveforms; S2: Segment the characteristic frequency band signals according to the analysis window length, and calculate the early damage degree of each analysis window. The early damage degree is positively correlated with the kurtosis deviation of the characteristic frequency band signals within the analysis window; S3: Calculate the waveform shape deviation of the current compression cycle, and construct the intermediate damage degree by combining it with the recent historical average of the early damage degree. The intermediate damage degree is positively correlated with both the waveform shape deviation and the recent historical average of the early damage degree; S4: Calculate the abnormal score over-ratio, and construct the late damage degree by combining it with the historical accumulation degree of the intermediate damage degree; S5: Dynamically determine the adaptive window length using the early damage degree, intermediate damage degree, and late damage degree as weight inputs; S6: Calculate the abnormal score on the vibration signal of the current compression cycle using the adaptive window length, and output an early warning signal based on the abnormal score.

[0007] The technical advantages of this invention are as follows: Addressing the problem that existing anomaly detection algorithms, relying on fixed-length analysis windows, are prone to serious false alarms and missed early, subtle damage when facing compressor operating condition drift, this invention introduces a damage evolution verification mechanism based on temporal causal correlation. By sequentially constructing damage indices for early impact, mid-term waveform deviation, and late-term persistent anomalies, and dynamically allocating weights at each stage to adaptively calculate the optimal observation window length, the system can accurately match the optimal feature scale for different stages from crack initiation to critical rupture under complex and variable operating conditions. This overcomes the dilution effect of a single fixed window on multi-scale anomaly features, eliminates false alarms caused by normal load switching, and improves the sensitivity to capturing early, extremely subtle diaphragm rupture precursors, thereby enhancing real-time early warning capabilities.

[0008] Preferably, the step of bandpass filtering the vibration acceleration signal is as follows: identifying the inherent resonance peak of the diaphragm structure in a frequency range that is higher than a first preset multiple of the rotation fundamental frequency; regarding the peak whose amplitude exceeds a second preset multiple of the average background noise as the inherent resonance peak of the diaphragm, taking the frequency range covering the inherent resonance peak of the diaphragm as the characteristic frequency band and setting the cutoff frequency of the bandpass filter.

[0009] The technical effect of this invention is as follows: by adaptively identifying the inherent resonance spectrum peaks of the diaphragm structure in a specific high-frequency range much higher than the rotation fundamental frequency, the characteristic frequency band is dynamically defined and the filtering threshold is set. Compared with the traditional filtering method using a fixed frequency range, this effectively isolates the huge low-frequency mechanical main vibration energy and environmental electromagnetic interference during the normal operation of the compressor. Thus, the transient high-frequency impact signal caused by early micro-crack friction is purified in an environment with extremely high signal-to-noise ratio, ensuring that diaphragms of different geometric sizes or materials can obtain high-fidelity early damage physical feature input under different speed conditions.

[0010] Preferably, the method for obtaining the standardized compression cycle waveform includes: locating the starting point of each compression cycle by the local maxima of the air chamber pressure signal, continuously detecting the time interval between adjacent maxima and taking the average value to obtain the nominal cycle duration, and then calculating the nominal cycle sampling point number; applying linear resampling to each compression cycle to the nominal cycle sampling point number, subtracting the mean of the resampled segment from each segment and dividing by the standard deviation of the segment to obtain the standardized compression cycle waveform.

[0011] The technical advantages of this invention are as follows: By utilizing the inherent peak characteristics of the air chamber pressure signal in each compression stroke, the true starting point of the compression cycle is accurately anchored. By dynamically calculating the nominal cycle duration and linear interpolation resampling, the measured signal with variable cycle is strictly aligned to a unified phase grid. This effectively overcomes the serious timing stretching and misalignment problems caused by load adjustment and speed fluctuations in industrial settings, ensuring that each sampling point has an absolute and consistent physical phase correspondence during subsequent waveform comparison and analysis. This also prevents false waveform deviations caused by operating cycle drift from the source of data preprocessing.

[0012] Preferably, the method for calculating the early damage degree of each analysis window is as follows: calculate the ratio of the kurtosis of the characteristic frequency band signal within the analysis window to the kurtosis of the healthy baseline; subtract 1 from the ratio to obtain the difference, and select the maximum value between the difference and 0 as the early damage degree.

[0013] The technical effect of this invention is as follows: the impact intensity is quantified by the ratio of the measured kurtosis of the characteristic frequency band signal within the analysis window to the kurtosis of the healthy baseline, and the pure deviation increment is extracted by subtracting the benchmark and truncating the negative value. This makes the early damage index not only able to keenly reflect the sudden non-Gaussian impact pulse brought about by the closing friction of the micro-crack surface, but also able to adaptively offset the systematic baseline deviation caused by sensor installation differences or the original background noise of the equipment.

[0014] Preferably, the method for calculating the waveform shape deviation of the current compression cycle is as follows: calculate the sum of the deviation values ​​of each sampling point between the standardized compression cycle waveform of the current compression cycle and the preset reference template; divide the sum by the total energy of the reference template to obtain the waveform shape deviation.

[0015] The technical effect of this invention is as follows: by accurately calculating the sum of the deviations of each corresponding sampling point between the current normalized waveform and the smoothly updated historical reference template, and normalizing it with the total energy of the template, it can macroscopically and extremely stably quantify the systematic shape distortion in the compression and expansion stroke caused by the degradation of local stiffness of the diaphragm. Compared with simple time-domain peak or root mean square comparison, this deviation calculation method effectively eliminates the interference of overall amplitude fluctuations and accurately focuses on the waveform profile variation level that characterizes the essence of mid-term physical damage.

[0016] Preferably, the method for constructing the intermediate damage degree by combining the recent historical average of the early damage degree includes: calculating the average of the early damage degree of all analysis windows within multiple compression cycles as the recent historical average; and multiplying the waveform shape deviation by the recent historical average to obtain the intermediate damage degree.

[0017] The technical effect of this invention is as follows: By establishing a multiplicative correlation constraint between the waveform shape deviation and the recent historical average of the early damage degree, and utilizing the rigid temporal causal law that the physical evolution of diaphragm rupture inevitably first produces micro-crack impact and then leads to overall stiffness degradation and deformation, the system can automatically suppress and return the intermediate damage index to zero when faced with occasional waveform distortion caused by the instantaneous switching of normal unit load, due to the lack of the precondition for early damage. This fundamentally eliminates false abnormal alarms caused by such normal operating condition disturbances and enhances the rigor of online diagnostic logic.

[0018] Preferably, the method for constructing late damage is as follows: calculate the ratio of the abnormal score of the previous compression cycle to the late activation baseline threshold, subtract 1 from the ratio and take the maximum value of the ratio and 0 to obtain the abnormal score overload ratio; calculate the average of the mid-term damage of multiple compression cycles as the historical accumulation degree; multiply the abnormal score overload ratio by the historical accumulation degree to obtain the late damage degree.

[0019] The technical effect of this invention is that when judging late-stage severe damage, the score excess ratio, which represents the current abnormal severity, is further multiplied and fused with the degree of historical accumulation of mid-stage damage, which represents the depth of early-stage defect accumulation. This extends the closed-loop of the complete life cycle time constraint chain from early to mid-stage evolution to late-stage deterioration to the final decision-making stage, effectively avoiding the misjudgment of the diaphragm being on the verge of rupture due to the continuous baseline shift caused by long-term continuous start-stop or severe load oscillation.

[0020] Preferably, the method for calculating the anomaly score is as follows: calculate the normalized Euclidean distance between the current subsequence and each non-adjacent historical subsequence in the historical buffer; select the minimum value among all calculated normalized Euclidean distances as the anomaly score for the current compression cycle.

[0021] The technical effect of this invention is as follows: by using the minimum normalized Euclidean distance between a subsequence and a historical non-adjacent subsequence in the matrix contour algorithm to define the anomaly score, the early warning system can achieve efficient online self-matching and quantitative evaluation of unknown and gradually changing anomaly patterns by relying on real-time data streams without the need to pre-build a large and complex failure mode feature library. While strictly eliminating the interference of self-matching of its own neighboring data, it endows the edge computing device with real-time time-series distance measurement capabilities with low latency and low computing power overhead.

[0022] Preferably, the specific steps for outputting early warning signals based on anomaly scores are as follows: During the health baseline stage, based on the mean and standard deviation of anomaly scores over multiple compression periods, a three-level early warning threshold is defined by an increasing multiple of the standard deviation; the current anomaly score is compared sequentially with the three-level early warning threshold, and based on the comparison results, a first-level early warning signal to shorten the monitoring period, a second-level early warning signal to generate a planned maintenance instruction, or a third-level early warning signal to execute an emergency shutdown are output respectively.

[0023] The technical effect of this invention is as follows: Based on the statistical characteristics of the equipment health baseline stage, a three-level incremental early warning threshold is dynamically divided and bound to a tiered equipment intervention command. This not only realizes the refined risk characterization of the entire range from weak precursors to critical failure, but also achieves closed-loop linkage with the on-site industrial automation control system. This provides on-site maintenance personnel with direct and executable digital decision support for formulating scientific maintenance schedules or emergency shutdown operations based on different levels of urgency.

[0024] Secondly, this invention provides a compressor diaphragm rupture early warning system based on time-series data, employing the following technical solution: A compressor diaphragm rupture early warning system based on time-series data includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the compressor diaphragm rupture early warning method based on time-series data as described above.

[0025] The present invention has the following technical effects: To address the problem that existing anomaly detection methods are prone to false alarms and missed detections due to fixed feature windows during equipment operating condition drift, this invention proposes a multi-dimensional signal early warning architecture system based on temporal causal chains and adaptive window scheduling. By integrating high-frequency impact to keenly perceive early crack initiation and waveform shape deviation to measure mid-term stiffness degradation, and strictly coupling these with historical evolution trajectories multiplicatively, a deep reconstruction and characterization of the physical laws governing the stepwise evolution of diaphragm damage is achieved. This method can not only autonomously avoid false alarms caused by normal operating condition switching, but also dynamically match the optimal feature observation scale for each decay stage, providing a full life-cycle intelligent safety defense mechanism and refined operation and maintenance decision support for high-risk industrial compressors under complex field conditions. Attached Figure Description

[0026] Figure 1 This is a flowchart of the compressor diaphragm rupture early warning method based on time-series data according to the present invention.

[0027] Figure 2 This is a schematic diagram of the change curve of the damage degree of the present invention during the membrane damage evolution process.

[0028] Figure 3 This is a schematic diagram of the evolution curve of the effective window length of the present invention within a complete damage cycle. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention discloses a compressor diaphragm rupture early warning method based on time-series data, referring to... Figure 1 This includes the following steps: Step S1: Data acquisition and preprocessing.

[0031] A piezoelectric accelerometer is installed near the diaphragm cavity on the compressor housing to collect real-time vibration acceleration signals. Gas pressure sensors are installed at the inlet and outlet of the gas chamber to collect real-time gas chamber pressure signals. An industrial pressure transmitter is installed in the hydraulic oil chamber to collect real-time oil chamber pressure signals. .

[0032] During the equipment commissioning phase, a calibration vibration signal was first acquired at an initial sampling frequency of 100kHz. Calculate the power spectral density above the rotating fundamental frequency. Identify the intrinsic resonance spectral peaks of the diaphragm structure in a frequency range more than 10 times higher, among which , The rated speed is expressed in rpm. The compressor's rotary excitation energy is concentrated in the range of 1 to 8 times the fundamental frequency. The diaphragm's natural frequency is determined by its geometric dimensions and the material's elastic modulus, and is typically more than 10 times higher than the fundamental frequency. Using this multiple as a lower limit, the diaphragm resonance region can be reliably separated from the rotary excitation region. Within this frequency range, spectral peaks with amplitudes exceeding three times the average background noise are considered diaphragm natural resonance peaks. The frequency range covering these peaks is the characteristic frequency band, with its upper bound denoted as . The lower boundary is denoted as This calibration process is performed once during equipment debugging, and then at the sampling frequency during normal operation. Continuous data acquisition. Feature frequency bands. Used for setting the cutoff frequency of the bandpass filter in step S2.

[0033] Butterworth bandpass filters were applied to the vibration acceleration signal, air chamber pressure signal, and oil chamber pressure signal to filter out DC bias and high-frequency electrical interference. The air chamber pressure signal reaches its peak value during each compression cycle when the diaphragm compresses the gas to its highest pressure; this peak value is significantly higher than other local fluctuations within the compression stroke. A time interval constraint between adjacent maxima was set during detection, i.e., the interval is not less than 0.8 times the nominal cycle length, to filter out intermediate fluctuation points within the compression stroke. The local maxima are used to locate the starting point of each compression cycle, and continuous detection is performed. The nominal period length is obtained by averaging the time intervals between adjacent maxima. This leads to the nominal number of periodic sampling points. ,in This is the rounding function.

[0034] The vibration signal is segmented according to the compression period, and the actual number of sampling points in each segment is determined by the interval between adjacent maxima. Linear resampling is applied to each segment, meaning the actual number of sampling points is less than [a certain value]. Interpolation point supplementation, more than Samples are drawn proportionally, and the output is consistent in both cases. Each segment is resampled to exactly [number] sampling points. The sampling point, the first Each sampling point corresponds to the compression cycle and the compression travel progress. Phase position at that point After resampling, subtracting the mean of each segment and dividing by the standard deviation of that segment yields the standardized compressed periodic waveform. , This is a compressed periodic index.

[0035] Take the equipment health operation phase continuously Each compression cycle constitutes a healthy baseline phase, used for calculating baseline parameters in subsequent steps. Provide sufficient statistical samples to ensure the stability of baseline parameters under typical operating conditions.

[0036] Step S2: Construct early damage level.

[0037] Under long-term alternating stress, fatigue cracks initiate in the high-stress concentration zone of the diaphragm. During the compression stroke, the crack surface experiences contact friction under cyclic loads, generating transient impact pulses of extremely short duration. The frequency of these pulses is much higher than the compressor's fundamental rotation frequency, and the energy is concentrated in the characteristic frequency band. It is clearly separated from the low-frequency main vibration component in the frequency domain.

[0038] Kurtosis is a statistic used to measure the impact of a signal. The theoretical kurtosis value for a normal signal is 3. Kurtosis increases when an impact component is present, and it is widely used in the field of mechanical impact detection. When calculating the kurtosis of broadband vibration signals, the energy of the low-frequency dominant vibration is much higher than the energy of the high-frequency impact pulse, and the overall kurtosis has a very limited response to early impacts.

[0039] Taking measured data from the crack initiation stage of a certain type of diaphragm as an example: the kurtosis of the characteristic frequency band signal increased from a baseline value of 3 to 9, while the kurtosis of the broadband signal only increased from 3.1 to 3.5 during the same period, a difference of about 5 times. In the context of fluctuating operating conditions, such a small change in broadband kurtosis is easily masked. By restricting kurtosis calculation to the characteristic frequency band and using the average kurtosis value of the healthy baseline stage as a reference, the absolute kurtosis value is converted into a deviation ratio relative to the healthy state, maintaining a stable perception of early impact characteristics under varying operating conditions.

[0040] The matrix contour algorithm uses subsequence similarity as its core computational unit, and the window length determines the time scale the algorithm perceives. The duration of the early impact pulse corresponds to a sub-period scale. When running with a fixed window covering the entire period, the proportion of the impact pulse within the window is extremely low, and its contribution to the subsequence distance is diluted by the remaining normal segments, leading to a decrease in the algorithm's sensitivity in the early damage stage. Therefore, it is necessary to construct an index that can quantify the early impact intensity at a sub-period scale.

[0041] right Apply cutoff frequency is and Bandpass filter to extract characteristic frequency band signals .Will By analysis window length Divide into segments, with each segment not overlapping, where: , This is the rounding function. Corresponding to a quarter-compression cycle duration, while ensuring the statistical stability of kurtosis calculation, the analysis time granularity is limited to a sub-periodic scale, matching the typical duration of early shock pulses. During the healthy baseline phase, for... The healthy baseline kurtosis is obtained by averaging the characteristic band kurtosis of all analysis windows within a compression cycle. .

[0042] Let the first The kurtosis of the characteristic frequency band signal within each analysis window is Define early damage level :

[0043] in, For the first Early damage degree of characteristic frequency band signals within an analysis window. and Both are dimensionless kurtosis values. The ratio of the two reflects the multiple of the current kurtosis relative to the healthy baseline. Subtracting 1 gives the proportion exceeding the baseline. Truncation of negative values ​​ensures that the index is non-negative.

[0044] Actual measurements during the healthy operation phase The value is close to the theoretical kurtosis value of 3 for a normal signal, but it may deviate due to factors such as sensor installation method and structural background noise. Using the measured mean value instead of the theoretical value 3 as a benchmark can eliminate the systematic deviation of specific equipment and make the index adaptable to different equipment.

[0045] When the membrane is healthy, exist Nearby fluctuations, Approaching zero. After crack initiation, the impact pulse causes the kurtosis of the characteristic frequency band to continuously increase. Exceeding , The increase is positively correlated with the degree of crack propagation.

[0046] Step S3: Construct the intermediate damage level.

[0047] During the mid-stage of diaphragm damage, crack propagation leads to local stiffness degradation. This change in local stiffness alters the diaphragm's bending deformation characteristics at each stage of its stroke, consequently changing the oil cavity pressure distribution and causing a systematic shift in the shape of the vibration waveform during each compression cycle. The compression cycle waveform of a healthy diaphragm exhibits high repeatability; after standardization, the waveforms of each compression cycle are highly consistent at the sampling point level. However, after stiffness degradation, the peak position, rise slope, and overall profile of the waveform undergo systematic changes, and the degree of these changes monotonically increases with the depth of damage.

[0048] When detecting waveform shape deviation, there are unique interference situations, including: when the compressor load changes, the change in the operating point also causes a slight adjustment in the waveform shape. The amplitude normalization in step S1 eliminates the overall amplitude difference, but the slight change at the shape level still causes the waveform deviation to increase temporarily, resulting in false activation of intermediate damage.

[0049] The evolution of diaphragm damage follows a specific temporal pattern: intermediate stiffness degradation is predicated on early crack initiation. When there is no continuous accumulation of early damage, temporary deviations in waveform shape are more likely due to load switching than intermediate damage. Using the recent historical average of the early damage degree in step S2 as a multiplicative factor for the intermediate index, when the accumulation of early damage is close to zero, even if the waveform deviation is temporarily high, the intermediate index is suppressed. By introducing temporal constraints on damage evolution, we can distinguish between false activation caused by load switching and intermediate damage.

[0050] Step S1 applies linear interpolation resampling to each compression cycle until exactly The sampling point, the first Each sampling point corresponds to a travel route The phase position at the location. The compression cycle in both the healthy baseline phase and the current monitoring phase represents one complete compression-expansion cycle. After the same resampling process, the phase positions of the two phases are... Each sampling point is strictly consistent in phase, and there is a clear physical correspondence between the points.

[0051] Based on the health baseline stage Standardized compression cycle waveform for each compression cycle For the location of each sampling point The average of 100 compression cycles is used to obtain the reference template. , After the healthy baseline phase, for each new compression cycle acquired, the earliest compression cycle is removed in a sliding manner, and the most recent compression cycle is re-acquired. The reference template is updated by averaging the values ​​at each compression cycle. This allows it to adjust smoothly as it drifts under slow operating conditions, while The larger window size makes the contamination effect of a small number of occasional abnormal cycles on the reference template negligible.

[0052] set up , The value exceeds the duration of typical operating condition switching, i.e., 5 to 10 compression cycles, so that the historical average can distinguish between short-term operating condition disturbances and continuous damage accumulation.

[0053] Let the first The standardized compression cycle waveform for each compression cycle is as follows: The corresponding reference template is , define the first The waveform shape deviation of each compression cycle is :

[0054] The numerator is the sum of the deviations between the current compression cycle and the reference template, and the denominator is the total energy of the reference template. Both are the sum of dimensionless standardized signals, reflecting the degree of deviation of the current waveform from the overall shape of the reference template.

[0055] Each compression cycle corresponds to four non-overlapping analysis windows, each with a length of [missing information]. .set up For the first The early damage value of the analysis window, the first The global index of the last analysis window in each compression cycle is Then it is near Within each compression cycle The early indicator mean of each analysis window is :

[0056] Reflecting on recent The degree of early damage accumulation within each compression cycle, and Alignment of the calculation cycle range.

[0057] Definition of the first The mid-term damage degree of each compression cycle is :

[0058] When the membrane is healthy, Approaching zero Approaching zero. After mid-term stiffness degradation, the waveform systematically deviates from the reference template. Increase; at the same time, the accumulation of cracks in the early stage causes Increase, the product of the two makes Significantly increased. Load switching caused... If there is a temporary spike in blood pressure, and there has been no prior accumulation of damage, Approaching zero, making It is suppressed.

[0059] Step S4: Construct late-stage damage.

[0060] In the late stage of diaphragm damage, the crack extends to near the critical size, the structural integrity is severely reduced, and the abnormal score output by the matrix profile algorithm remains high for several consecutive compression cycles and increases monotonically as the damage deepens.

[0061] Judging late-stage damage solely based on the current moment's abnormal score presents the following challenges: during compressor start-up and shutdown or multiple consecutive load changes, the abnormal score also remains consistently high across several compression cycles, making it difficult to distinguish from the score performance of late-stage damage. The fundamental difference between late-stage damage and continuous operating disturbances in terms of scoring timing lies in the fact that the degree of elevation in scores during each compression cycle under continuous operating disturbances is not fixed, while the late-stage damage score monotonically increases with increasing damage and maintains a stable amplitude. Therefore, relying solely on statistical analysis of the scoring timing still carries the risk of misjudgment in real-world scenarios.

[0062] The temporal constraints in step S3 also hold true in the late stage: late-stage structural failure is predicated on mid-term stiffness degradation. When the mid-term damage accumulation approaches zero, persistently high anomaly scores are more likely due to continuous operating disturbances than late-stage diaphragm damage. By incorporating the recent historical average of the mid-term damage degree from step S3 into the calculation of late-stage indicators, and suppressing the activation of late-stage indicators when the mid-term damage accumulation approaches zero, the temporal constraints from early to mid-term and then to late stage are extended to the final stage, forming a complete phased causal verification chain.

[0063] Anomaly score in the previous compression cycle of the late-stage index dependency matrix profile algorithm Step S6 writes the data to the history buffer at the end of each compression cycle. During the healthy baseline phase, [the data is then processed]. The average of the anomaly scores for each compression cycle was taken. and standard deviation Late activation baseline threshold defined according to the Raida criterion :

[0064] The distribution of abnormal scores in the healthy stage is based on Centered on, The upper bound of the corresponding health score distribution is reasonable. Scores exceeding this value correspond to a real abnormal state with a high probability. Calculating score excess based on this benchmark has a statistical basis.

[0065] Late activation baseline threshold Based on this, define the abnormal scoring excess ratio. :

[0066] and Both values ​​are distance values ​​output from the matrix contour, with the same dimensions. The ratio between the two is dimensionless. Subtracting 1 yields the relative amount by which the anomaly score exceeds the baseline. ensure Non-negative.

[0067] With near The average damage level during each compression cycle is used as the historical accumulation level:

[0068] in, For the first Mid-term damage value for each compression cycle. For the first The most recent compression cycle The average value of the mid-term damage over a compression cycle reflects the near-term damage. The degree of accumulation of intermediate damage within a compression cycle.

[0069] Definition of the first Late-stage damage per compression cycle :

[0070] When the membrane is healthy, exist the following, Zero, Zero. Late-stage damage makes Continue to exceed , The value remains greater than zero; meanwhile, mid-term damage accumulation causes... Increase, the product of the two makes Significantly increased. Continuous operating condition disturbances lead to... If the level remains consistently high, and there has been no prior accumulation of mid-term damage, Approaching zero It is suppressed.

[0071] Step S5: Calculate the adaptive window.

[0072] The three damage indicators in steps S2 to S4 correspond to the activation levels of the matrix profilometry algorithm at the optimal observation scales for the early, middle, and late damage stages, respectively. The original matrix profilometry algorithm operates with a fixed window length. The early impact pulse, middle waveform deformation, and late-stage persistent deviation correspond to three optimal scales: sub-period, single compression period, and multiple compression period, respectively. The single fixed window dilutes the features at each stage to varying degrees. This step uses the three damage indicators as weight inputs to dynamically determine the window length of the matrix profilometry algorithm, causing the window to converge towards the corresponding optimal scale at each damage stage.

[0073] Let the lengths of the three candidate windows be respectively , , All are expressed in units of the number of sampling points and have the same dimensions. The analysis window length is consistent with that in step S2; Corresponding to a single compression cycle; Three compression periods were used to cover the cumulative effect of amplifying late-stage successive anomalies in sufficient historical subsequences.

[0074] , , All measurements at each stage are dimensionless deviations from the healthy baseline, with the physical meaning unified as the relative deviation of the damage degree at the corresponding stage, and are then normalized. At that time, directly ordered The algorithm operates with a fixed window and a single compression period, making it equivalent to the original matrix contour algorithm. When the sum of the three factors is greater than zero, the weights for each scale are:

[0075]

[0076]

[0077] The sum of the three weights is 1, and the adaptive window length is calculated using a weighted average:

[0078] in, For the first The actual adaptive window length used by the matrix contouring algorithm during each compression cycle; , , These are respectively a quarter compression cycle, a single compression cycle, and three compression cycles.

[0079] To suppress frequent window changes caused by short-term fluctuations in indicators, when At that time, the matrix contour history buffer is cleared, and the reference baseline is re-established with the current window length to prevent historical subsequences across the window length from interfering with the distance calculation.

[0080] Step S6: Anomaly score calculation and graded early warning.

[0081] The adaptive window length obtained in step S5 As a window parameter for the matrix profile algorithm, online matrix profile calculation is performed on the vibration signal of the current compression cycle. The normalized Euclidean distance between the current subsequence and each non-adjacent historical subsequence in the historical buffer is calculated, and the minimum value is taken as the anomaly score of the current compression cycle. and will Write to the history buffer for calculation in step S4 during the next compression cycle. Use when needed.

[0082] The current subsequence is a truncation with the end time of the current compression period as its right endpoint and a length of [missing information]. The time sequence segment; the historical subsequence is a historical buffer containing historical moments. The time segments are time segments with the same window length and right endpoints; non-adjacent segments require a time interval between their right endpoints greater than [a certain value]. To ensure no overlap in time, self-matching interference is eliminated.

[0083] At the health baseline stage, based on Mean of anomaly score for each compression cycle with standard deviation Set three warning thresholds:

[0084]

[0085]

[0086] in, , , Three warning boundaries are defined using increasing standard deviation multiples, covering the entire risk range from minor anomalies to critical rupture. The three-level warning threshold and the late activation baseline threshold are also used. All data originate from the same health baseline statistics, and the parameter system is unified.

[0087] Current anomaly score Compare with the three threshold levels sequentially: In a timely manner, a Level 3 warning is issued, and an emergency shutdown is initiated. Output a level-two early warning and generate planned maintenance instructions; It can issue a Level 1 early warning in real time, shortening the monitoring cycle; The system continues to operate at the current frequency.

[0088] The fixed window of the original matrix contour algorithm is thus replaced by an adaptive window that perceives the damage stage: early damage stage. activation, Towards Convergence: The algorithm operates on a sub-period scale, enabling the impact pulse to achieve a higher signal-to-noise ratio within a shorter window; Mid-stage damage phase. activation, Towards Convergence: The algorithm runs in a single compression cycle, fully capturing waveform morphology deviations; Late damage stage. activation, Towards Convergence: The algorithm runs in multiple compression cycles to amplify the cumulative effect of continuous deviation, achieving full-process early warning coverage from early crack initiation to late critical rupture.

[0089] The technical effects of this invention can also be illustrated in conjunction with the accompanying drawings. Figure 2 This is a schematic diagram showing the changes of the three damage levels during the membrane damage evolution process. It rises first after the cracks initiate. exist It will activate after accumulating to a certain level. Activated after continuous accumulation of damage in the intermediate stage, the three respond in order from early to intermediate to late stage, verifying the effectiveness of temporal causal constraints in the actual damage evolution process.

[0090] Figure 3 This is a schematic diagram of the evolution curve of the effective window length during the complete damage cycle, in the healthy phase. Maintained at a single compression cycle scale At the site, the early damage stage extends to the sub-periodic scale. Convergence, regression during the intermediate damage phase The late-stage damage extends to multiple cycle scales. The window automatically adjusts to the corresponding optimal observation scale as the damage stage progresses, verifying the correctness of the adaptive mechanism.

[0091] This invention also discloses a compressor diaphragm rupture early warning system based on time-series data, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the compressor diaphragm rupture early warning method based on time-series data according to this invention.

[0092] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0093] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A compressor diaphragm rupture early warning method based on time-series data, characterized in that, Includes the following steps: S1: Acquire the vibration acceleration signal, gas chamber pressure signal and oil chamber pressure signal during the operation of the compressor, perform bandpass filtering on the vibration acceleration signal to obtain the characteristic frequency band signal, and perform resampling processing on the vibration acceleration signal to obtain the standardized compression cycle waveform; S2: The characteristic frequency band signal is segmented according to the length of the analysis window, and the early damage degree of each analysis window is calculated. The early damage degree is positively correlated with the degree of kurtosis deviation of the characteristic frequency band signal within the analysis window. The method for calculating the early damage degree of each analysis window is to calculate the ratio of the kurtosis of the characteristic frequency band signal within the analysis window to the kurtosis of the healthy baseline. Subtract 1 from the ratio to obtain the difference, and select the maximum value between the difference and 0 as the early damage degree; S3: Calculate the waveform shape deviation of the current compression cycle, and construct the intermediate damage degree by combining it with the recent historical average of the early damage degree. The intermediate damage degree is positively correlated with both the waveform shape deviation and the recent historical average of the early damage degree; The method for calculating the waveform shape deviation of the current compression cycle is to calculate the sum of the deviation values ​​of each sampling point between the standardized compression cycle waveform of the current compression cycle and the preset reference template; Dividing the sum by the total energy of the reference template yields the waveform shape deviation. The method for constructing intermediate damage by combining the recent historical average of early damage degree includes: calculating the average of early damage degree across all analysis windows within multiple compression cycles as the recent historical average; multiplying the waveform shape deviation by the recent historical average to obtain the intermediate damage degree; S4: calculating the abnormal score excess ratio and constructing late damage degree by combining the historical accumulation degree of intermediate damage degree; the method for constructing late damage degree is as follows: calculating the ratio of the abnormal score of the previous compression cycle to the preset late activation baseline threshold, subtracting 1 from this ratio and taking the maximum value of 0 to obtain the abnormal score excess ratio; averaging the intermediate damage degree across multiple compression cycles as the historical accumulation degree; multiplying the abnormal score excess ratio by the historical accumulation degree to obtain the late damage degree; S5: dynamically determining the adaptive window length using early damage degree, intermediate damage degree, and late damage degree as weight inputs; S6: calculating the abnormal score on the vibration signal of the current compression cycle using the adaptive window length and outputting a warning signal based on the abnormal score.

2. The compressor diaphragm rupture early warning method based on time-series data according to claim 1, characterized in that, The steps for bandpass filtering the vibration acceleration signal are as follows: identify the inherent resonance spectral peaks of the diaphragm structure in the frequency range that is higher than the first preset multiple of the rotation fundamental frequency; regard the spectral peaks whose amplitude exceeds the second preset multiple of the average background noise as the inherent resonance spectral peaks of the diaphragm; take the frequency range covering the inherent resonance spectral peaks of the diaphragm as the characteristic frequency band and set the cutoff frequency of the bandpass filter.

3. The compressor diaphragm rupture early warning method based on time-series data according to claim 1, characterized in that, The method for obtaining the standardized compression cycle waveform includes: using the local maxima of the air chamber pressure signal to locate the starting point of each compression cycle, continuously detecting the time interval between adjacent maxima and taking the average value to obtain the nominal cycle duration, and then calculating the nominal cycle sampling point number; applying linear resampling to each compression cycle to the nominal cycle sampling point number, subtracting the mean of the resampled segment from each segment and dividing by the standard deviation of the segment to obtain the standardized compression cycle waveform.

4. The compressor diaphragm rupture early warning method based on time-series data according to claim 1, characterized in that, The method for calculating anomaly scores is as follows: calculate the normalized Euclidean distance between the current subsequence and each non-adjacent historical subsequence in the historical buffer; The minimum value among all calculated normalized Euclidean distances is selected as the anomaly score for the current compression cycle.

5. The compressor diaphragm rupture early warning method based on time-series data according to claim 1, characterized in that, The specific steps for outputting early warning signals based on anomaly scores are as follows: During the health baseline phase, based on the mean and standard deviation of anomaly scores over multiple compression periods, a three-level early warning threshold is defined by an increasing multiple of the standard deviation; the current anomaly score is compared sequentially with the three-level early warning threshold, and based on the comparison results, a first-level early warning signal to shorten the monitoring cycle, a second-level early warning signal to generate a planned maintenance instruction, or a third-level early warning signal to execute an emergency shutdown are output respectively.

6. A compressor diaphragm rupture early warning system based on time-series data, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the compressor diaphragm rupture early warning method based on time-series data according to any one of claims 1-5.

Images