A multivariate vibration anomaly detection method and system based on data quality gating

Abstract

The present application relates to the technical field of industrial equipment state monitoring, in particular to a multivariate vibration anomaly detection method and system based on data quality gating. The present application obtains multiple vibration variables collected synchronously and performs time series coding, determines data unreliability in combination with the local quality degradation degree of each variable, the unexplained degree obtained by the deviation of the input mapping and the effective explanation degree obtained by the contribution of the output mapping; as the data unreliability increases, the representation sent by the variable to other variables is attenuated and the unattenuated direct path of the variable is reserved, accordingly the vibration variable is reconstructed, the complementary distribution of the reconstruction deviation is allocated as the equipment vibration anomaly score and the data quality anomaly score, and then the two are compared with the normal verification threshold respectively and the corresponding state is output. The present application can limit the cross-variable interference of unreliable channels, reserve the abnormal evidence of the variable, and distinguish the equipment vibration anomaly and the data quality anomaly which can exist simultaneously, and is suitable for equipment state monitoring with multiple synchronous vibration measuring points.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, specifically to a multivariable vibration anomaly detection method and system based on data quality gating. Background Technology

[0002] Industrial equipment condition monitoring often involves deploying multiple vibration acquisition channels at bearing housings, enclosures, or support locations. The same mechanical excitation, along different transmission paths, generates vibration responses with varying amplitudes, phases, or times at different measurement points. In operating environments lacking fault labels, existing multivariate vibration anomaly detection typically standardizes and slides windows the synchronous vibration sequence. Window quality values ​​are generated based on local variations, sliding standard deviations, or high-frequency residuals. Temporal convolutional networks are used to extract the characteristics of each variable, and information between variables is fused using an attention mechanism. The decoder then reconstructs the input window, generating anomaly scores based on reconstruction errors, potential spatial distances, or statistical biases. This identifies detection windows that deviate from normal operating patterns and suppresses the influence of low-quality channels on the fusion results.

[0003] When a sensor becomes loose, has poor contact, experiences cable jitter, sampling jitter, or is subject to localized electromagnetic interference, causing abrupt changes, increased high-frequency components, or increased fluctuations in a single acquisition channel, its local morphology may resemble the common response formed by multiple measurement points due to mechanical faults. Existing methods rely on single-channel quality indicators to directly determine gating weights and scale the channel as a whole or fuse potential representations. This approach cannot identify whether local anomalies are supported by the common responses of other measurement points, nor can it determine whether the channel's representation is still useful for reconstructing other channels. Overly strong gating weakens the true equipment anomaly components corroborated by other measurement points and loses evidence of the original variable; insufficient gating allows disturbed representations to propagate to other channels and increase multiple reconstruction biases. This results in two interrelated technical problems: the difficulty in distinguishing between true mechanical anomalies and acquisition degradation, and the difficulty in simultaneously addressing the limitations on the external representations of suspicious channels and the retention of evidence of anomalies in the original variable. Therefore, a multivariate vibration anomaly detection method and system based on data quality gating is needed to solve the aforementioned problems of the difficulty in distinguishing between true mechanical anomalies and acquisition degradation, and the difficulty in simultaneously addressing the limitations on the external representations of suspicious channels and the retention of evidence of anomalies in the original variable. Summary of the Invention

[0004] In view of the above-mentioned problems in the prior art, the present invention provides the following technical solution: In a first aspect, the present invention provides a multivariable vibration anomaly detection method based on data quality gating, comprising: S1. Acquire multiple vibration variables that are synchronously collected, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position; S2. For each vibration variable, determine the data unreliability based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. S3. As the unreliability of the data increases, the variable-wise representations sent to the remaining vibration variables are attenuated, and the unattenuated direct path of the variable-wise representations is retained to obtain the representations to be reconstructed. S4. Reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. Allocate the reconstruction deviation according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. S5. Determine the equipment vibration abnormality threshold and data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, output the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, output the data quality abnormality status.

[0005] Optionally, the plurality of vibration variables may be at least three vibration variables acquired by mutually independent acquisition channels.

[0006] Optionally, the data unreliability is determined according to... Calculate, where, The unreliability of the data, and The terms represent the degree of local quality degradation, the degree of unexplained quality, and the degree of effective explanation, respectively, and all three are taken from 0 to 1. The data unreliability increases with the degree of local quality degradation and the degree of unexplainedness, and decreases with the degree of effective explanation.

[0007] Optionally, determining the degree of local quality degradation includes: The variation of adjacent sampled values, local dispersion, and high-frequency residual energy of the vibration variable are calculated and combined to obtain local mass statistics. The local quality statistics are normalized according to the normal training distribution to obtain the degree of local quality degradation.

[0008] Optionally, determining the degree of unexplainedness includes: Within the normal time offset range, the remaining vibration variables that have an effective mapping relationship with the vibration variable to be evaluated are used as inputs to map the vibration variable to be evaluated, thereby obtaining a mapping characterization; The degree of unexplainedness is obtained by comparing the variable-wise representation of the vibration variable to be evaluated with the mapping representation.

[0009] Optionally, determining the effective explanatory power includes: The vibration variables to be evaluated are retained in the mapping, and the first mapping deviation weighted sum of the remaining vibration variables is calculated; In the mapping, the vibration variables to be evaluated are masked, and the second mapping deviation weighted sum of the remaining vibration variables is calculated; When the weighted sum of the second mapping deviations is greater than the weighted sum of the first mapping deviations, the effective explanatory power is obtained by subtracting the weighted sum of the first mapping deviations from the weighted sum of the second mapping deviations. When the weighted sum of the second mapping deviations is less than or equal to the weighted sum of the first mapping deviations, the effective explanatory power is determined to be 0.

[0010] Optionally, the attenuation of the variable-wise characterization transmitted to the remaining vibration variables as the data unreliability increases includes: make Let be the variable index of the vibration variable, and denote the data unreliability as . ; according to Calculate the gating coefficient. This is the gate control safety factor, where, ; Using the gating coefficient as a multiplier, the variable-wise representation to be transmitted is multiplied and then transmitted to the remaining vibration variables.

[0011] Optionally, the step of complementaryly allocating the reconstruction bias according to the data unreliability includes: make For the variable index of the vibration variable, and The first The data unreliability and reconstruction bias of the aforementioned vibration variables These are the non-negative variable weights that sum to 1; according to and The equipment vibration anomaly score and the data quality anomaly score are calculated separately.

[0012] Optionally, the determination of the equipment vibration anomaly threshold adopts the quantile or robust upper bound of the normal verification sample distribution of the equipment vibration anomaly score, and the determination of the data quality anomaly threshold adopts the quantile or robust upper bound of the normal verification sample distribution of the data quality anomaly score.

[0013] Secondly, the present invention provides a multivariable vibration anomaly detection system based on data quality gating, for implementing the aforementioned method, comprising: The window and encoding module are used to acquire multiple vibration variables that are collected synchronously, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position. The quality assessment module is used to determine the data unreliability for each vibration variable based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the incoming mapping deviation, and the degree of effective explanation obtained from the outgoing mapping contribution. The intervariable transfer control module is used to attenuate the variable representations sent to the remaining vibration variables as the data unreliability increases, and retain the unattenuated direct path of the variable representations to obtain the representations to be reconstructed. The reconstruction scoring module is used to reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. The reconstruction deviation is then distributed according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. The status determination module is used to determine the equipment vibration abnormality threshold and the data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, the module outputs the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, the module outputs the data quality abnormality status.

[0014] The technical solution provided by this invention has the following beneficial effects: 1. Under the condition that multiple vibration measurement points have a stable and normal coupling relationship and the main acquisition chains are independent of each other, the data unreliability is formed by combining the degree of local quality degradation, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. This allows channels with only local morphological anomalies and lacking bidirectional intervariate support to obtain higher unreliability, thereby reducing the situation where the common mechanical response of multiple measurement points is weakened as acquisition degradation, and improving the distinction between evidence of real equipment vibration anomalies and isolated acquisition degradation.

[0015] 2. The gating coefficient corresponding to the unreliability of the data is applied only to the representation of the source variable sent to other variables, while the unfaded direct path of the variable's own representation is preserved. This restricts the abnormal representation of the disturbed channel from propagating to other channels, while the abnormal information of the variable itself can still enter its representation to be reconstructed. This balances reducing the impact of the disturbed representation on the reconstruction results of other channels and preserving the abnormal evidence of the variable itself.

[0016] 3. Using the same data unreliability to allocate the reconstruction bias of each variable with complementary weights, the equipment vibration anomaly score and the data quality anomaly score jointly retain the abnormal evidence before allocation, and are independently compared with the normal verification threshold, thereby avoiding the direct loss of the bias reduced by gating. This allows the equipment monitoring end and the data acquisition and maintenance end to obtain the corresponding abnormal state respectively, and allows the retention of both states when equipment response and data acquisition degradation coexist. Attached Figure Description

[0017] Figure 1 A schematic diagram illustrating the steps of a multivariable vibration anomaly detection method based on data quality gating provided by the present invention; Figure 2 This invention provides an overall flowchart of a multivariable vibration anomaly detection method based on data quality gating. Figure 3 This invention provides a schematic diagram of the multi-vibration measuring point independent acquisition chain structure. Figure 4 This invention provides a schematic diagram illustrating the relationship between data unreliability formation and dual-path multiplexing. Figure 5 This is a schematic diagram of the source variable gating transmission and the direct path of the variable provided by the present invention; Figure 6 A schematic diagram of a multivariable vibration anomaly detection system based on data quality gating provided by the present invention; Figure 7 The diagram shows the variation of gating and allocation coefficients driven by data unreliability, as provided in this invention. Detailed Implementation

[0018] 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] This embodiment takes multi-point vibration monitoring of gear transmission equipment as an example. At least three vibration acquisition components are installed at the bearing housing, housing, or support location of the equipment, and each acquisition component outputs vibration variables under the same sampling time base. The processing system can be deployed in an on-site industrial computer, edge computing device, or server, and receives vibration sampling values ​​with channel identifiers, sampling time, sampling rate, and equipment operating condition identifiers through a synchronous vibration acquisition interface.

[0020] The processing system reads the currently valid model version before starting a processing window. A model version must be bound to at least the sampling rate, window length, window step, variable identity table, normalization parameters, temporal coding parameters, valid variable pairs, allowed time offset range, intervariate mapping parameters, local quality statistics parameters, error scale, gating floor coefficient, variable weights, and two types of normal validation thresholds. New versions are only released after integrity and normal validation have passed; windows that have already started processing continue to use the version locked at startup, and subsequent windows only switch to the new version.

[0021] The variable identity table uses the sensor serial number, acquisition channel identifier, installation location code, and variable index as a joint key. Upon receiving a window, the processing system matches the joint key item by item and verifies the sampling rate, sample count, timestamp continuity, and operating condition range. If the joint key is inconsistent, the same identity is duplicated, a necessary channel is missing, the synchronization residual exceeds the allowable time offset range, or the configuration verification fails, the window is set to an invalid input or invalid configuration final state, the reason code and the original window are saved, and no further abnormal scores are generated.

[0022] During the model building phase, data from the same device, with the same set of variable identities, and within normal operating range are used. Normal data, isolated by time period, is divided into normal training windows and normal validation windows. The normal training window is used to fit channel-wise standardized parameters, the temporal encoder, the intervariate mapper, and the reconstruction decoder, and to establish normal scales for local quality statistics, mapping bias, and reconstruction bias. The normal validation window does not participate in parameter updates; it is only used to check the model version and obtain decision thresholds. During training, window reconstruction bias and intervariate mapping bias are used as loss terms, calculated batch-wise forward and updated backward, stopping when the validation loss no longer decreases after several consecutive rounds. Finally, a snapshot of the parameters with the lowest validation loss is retained.

[0023] like Figure 1 As shown, this embodiment provides a multivariable vibration anomaly detection method based on data quality gating, including: S1. Acquire multiple vibration variables that are synchronously collected, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position; S2. For each vibration variable, determine the data unreliability based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. S3. As the unreliability of the data increases, the variable-wise representations sent to the remaining vibration variables are attenuated, and the unattenuated direct path of the variable-wise representations is retained to obtain the representations to be reconstructed. S4. Reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. Allocate the reconstruction deviation according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. S5. Determine the equipment vibration abnormality threshold and data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, output the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, output the data quality abnormality status.

[0024] like Figure 2 As shown, the synchronous multivariate vibration window is time-coded to form a variable-wise representation. This variable-wise representation is used to evaluate and determine data unreliability based on local quality and bidirectional intervariate interpretation evidence. The variable-wise representation, as the controllable representation, along with the data unreliability, enters intervariate transfer control to form a reconstructable representation. The reconstructable representation is then used to reconstruct vibration variables to obtain a reconstruction bias. Under data unreliability control, this reconstruction bias is complementaryly allocated to form equipment vibration anomaly scores and data quality anomaly scores.

[0025] Figure 2 The terminal compares the equipment vibration anomaly score with the equipment vibration anomaly threshold, and the data quality anomaly score with the data quality anomaly threshold, to obtain the equipment vibration anomaly state and the data quality anomaly state, respectively. The two comparisons are independent of each other. Regardless of the comparison result of the equipment branch, the comparison of the data quality branch continues to be performed. Therefore, the two states can be true individually or simultaneously.

[0026] For S1, specifically, the processing system reads multiple vibration variables with the same window start time and sample index range from the synchronization cache. Each variable is first subtracted from the normal training mean stored in its model version, and then divided by the corresponding normal training scale. The processing system then divides the variable into standardized synchronization windows with clear variable number, time position number, and window identifier according to a fixed window length and window step size.

[0027] In one example configuration, the sampling rate is 25.6 kHz, each window contains 2048 sampling points, adjacent windows overlap by 50%, and the device rotation speed is within the same set of operating parameters around 1500 r / min. The residual synchronization error after timestamp alignment is less than 0.25 sampling intervals, and each valid variable pair is allowed to align within the normal time offset range of three consecutive sampling points. The above values ​​are for illustrative purposes only and do not limit other sampling parameters to match the device's frequency band and rotation speed.

[0028] During temporal encoding, each variable is passed through a one-dimensional temporal convolutional layer with different dilation rates. Each convolutional layer extracts short-term shocks and longer-term periodic variations along the time position, maps them to the same dimension after feature dimension mapping, and adds the positional encoding to the convolutional features at the corresponding time position. The processing system thus obtains a pervariate representation that corresponds one-to-one with the input variables and retains the time position, and passes each pervariate representation, along with variable identity, window identifier, and model version, to subsequent processing.

[0029] Furthermore, the plurality of vibration variables are at least three vibration variables obtained by mutually independent acquisition channels.

[0030] Specifically, such as Figure 3 As shown, the mechanical vibration response of the gear transmission equipment is transmitted to monitoring position A. The vibration signal generated at monitoring position A is converted into a sampling signal by the vibration acquisition component of channel A. This sampling signal is then conditioned by the analog front end of channel A, converted into a digital vibration sampling value by analog-to-digital conversion of channel A, and finally sent to the synchronous vibration acquisition interface as the vibration sampling value of channel A via the data transmission chain of channel A.

[0031] The mechanical vibration response of the gear transmission equipment is also transmitted to monitoring positions B and C. The vibration signal at monitoring position B sequentially passes through the B-channel vibration acquisition unit, the B-channel analog front-end, the B-channel analog-to-digital converter, and the B-channel data transmission chain to form the B-channel vibration sample value. Similarly, the vibration signal at monitoring position C sequentially passes through the C-channel vibration acquisition unit, the C-channel analog front-end, the C-channel analog-to-digital converter, and the C-channel data transmission chain to form the C-channel vibration sample value. Both channels also sequentially undergo the conversion of acquired signals, conditioned signals, and digital vibration sample values.

[0032] The synchronous vibration acquisition interface aligns the vibration sample values ​​of channel A, channel B, and channel C using a common sampling clock or a calibrated timestamp, and then transmits the synchronously acquired multiple vibration variables to the processing system. Each of the three channels has an independent vibration acquisition component, analog front-end, analog-to-digital converter, and data transmission chain. Local cable jitter or poor contact in any channel will not be directly written to other channels along the acquisition hardware.

[0033] Channel independence is verified through a link list before model release and saved along with the variable identity table. If multiple channels share the same analog front-end, analog-to-digital converter, or transmission chain, they are marked as a common module group. Variables within a common module group are not used as the sole source of explanation for each other unless an independent reference channel is set up and common interference is filtered out first.

[0034] For S2, specifically, the quality assessment process reads the standardized vibration variables and variable-by-variable representations from the same window, and calculates the degree of local quality degradation, the degree of unexplained ingress, and the degree of effective explanatory ingress one by one according to the variable index. All three types of results are converted to 0 to 1 and saved along with their input verification status; if any result is missing or out of range, the processing system records the calculation anomaly and does not replace it with a temporary default value.

[0035] In some implementations, the data unreliability is determined according to... Calculate, where, The unreliability of the data, , and The terms represent the degree of local quality degradation, the degree of unexplained quality, and the degree of effective explanation, respectively, and all three are taken from 0 to 1. The data unreliability increases with the degree of local quality degradation and the degree of unexplainedness, and decreases with the degree of effective explanation.

[0036] Specifically, such as Figure 4 As shown, for the first One variable, the degree of local quality degradation. and degree of unexplained ingress Data unreliability Forming a co-directional contribution, resulting in a higher degree of effective explanation. through Reverse contribution to data unreliability When local quality deteriorates and there is a lack of two-way intervariate explanation, the data unreliability increases; when it can be explained by other variables or can effectively explain other variables, the corresponding unreliability is suppressed.

[0037] Data unreliability This data is read by the subsequent two stages. The gating branch multiplies the unreliability compensation of the data by the difference between 1 and the safety factor, and then adds it back to the safety factor to obtain the gating coefficient. Gating coefficient Control the outward transmission strength of source variables and apply it to cross-variable representation transmission. Assign branch routing data unreliability. Forming equipment allocation weights And data quality allocation weight The two weights are used to allocate the reconstruction bias respectively. This forms the abnormal vibration score of the equipment. and data quality anomaly score .

[0038] For example, when the local quality degradation of a variable within a certain window is 0.90, the unexplained degree is 0.85, and the effective explained degree is 0.05, the aforementioned data unreliability calculation relationship yields 0.727. Within another window, when the three degrees are 0.80, 0.40, and 0.80 respectively, the result is 0.064. This example is only for verifying the direction of the calculation and does not represent actual measured performance.

[0039] Furthermore, the determination of the degree of local quality degradation includes: The variation of adjacent sampled values, local dispersion, and high-frequency residual energy of the vibration variable are calculated and combined to obtain local mass statistics. The local quality statistics are normalized according to the normal training distribution to obtain the degree of local quality degradation.

[0040] Specifically, for a standardized vibration window, the processing system first calculates the absolute value of the difference between every two adjacent sampled values ​​and takes the median of all absolute values ​​as the change in adjacent sampled values; then it calculates the absolute deviation of the window sampled values ​​relative to the median and takes the median of the absolute deviation as the local dispersion; subsequently, it filters out the low-frequency part using the high-pass cutoff frequency saved in the model version, and then accumulates the squares of the remaining high-frequency residuals point by point and divides them by the number of sampling points to obtain the high-frequency residual energy.

[0041] The processing system multiplies the change in adjacent sampled values, the degree of local dispersion, and the high-frequency residual energy by three non-negative combination coefficients stored in the model version, and then adds the three products together to obtain the local quality statistics. Each combination coefficient contains the unit conversion scale of the corresponding index, so that the added terms have the same dimensions; the combination coefficients are fixed by the normal training window and do not change with the test window.

[0042] The processing system then reads the median and pre-selected high quantile values ​​of the local quality statistics within the normal training window. It outputs 0 when the measured statistic is not higher than the median, 1 when it is not lower than the high quantile, and 1 when it falls between the two. For values ​​in between, it performs linear interpolation based on the statistic's position within the median to high quantile range to obtain the degree of local quality degradation. This maintains the processing order of first combining three indicators to form local quality statistics, and then normalizing these statistics according to their normal training distribution.

[0043] Furthermore, determining the degree of unexplainedness includes: Within the normal time offset range, the remaining vibration variables that have an effective mapping relationship with the vibration variable to be evaluated are used as inputs to map the vibration variable to be evaluated, thereby obtaining a mapping characterization; The degree of unexplainedness is obtained by comparing the variable-wise representation of the vibration variable to be evaluated with the mapping representation.

[0044] Specifically, during the model training phase, the band coherence within the normal window and the change in prediction bias before and after masking the source variable are statistically analyzed for each directed variable pair. Variable pairs whose band coherence continuously reaches the minimum stable value saved in the model version, or whose prediction bias steadily decreases after the source variable is involved, are written into the effective variable pair table; each record simultaneously saves the source variable identity, target variable identity, allowed time offset range, and mapping weights.

[0045] At runtime, with the first The system selects records from the effective variable pair table whose target identity is equal to the variable identity but different from the source identity. Each source variable moves its time position one by one within the allowed time offset range, and outputs candidate representations through a locked intervariate mapper. The system selects the time position with the smallest robust mapping deviation and adds them up according to the mapping weights in the effective variable pair table to obtain the mapping representation of the variable to be evaluated.

[0046] The processing system subtracts the variable-wise representation and the mapped representation of the variable to be evaluated item by item according to the same time position and feature dimension, takes the absolute value of the difference and calculates a weighted average to obtain the ingress mapping bias. Then, it performs piecewise interpolation and amplitude limiting on the median and high quantile values ​​of this bias in the normal training window, which are the same as the local quality statistics, to obtain the unexplained degree from 0 to 1; the larger the bias, the greater the unexplained degree.

[0047] Furthermore, the determination of the effective explanatory power includes: The vibration variables to be evaluated are retained in the mapping, and the first mapping deviation weighted sum of the remaining vibration variables is calculated; In the mapping, the vibration variables to be evaluated are masked, and the second mapping deviation weighted sum of the remaining vibration variables is calculated; When the weighted sum of the second mapping deviations is greater than the weighted sum of the first mapping deviations, the effective explanatory power is obtained by subtracting the weighted sum of the first mapping deviations from the weighted sum of the second mapping deviations. When the weighted sum of the second mapping deviations is less than or equal to the weighted sum of the first mapping deviations, the effective explanatory power is determined to be 0.

[0048] Specifically, the processing system performs forward processing twice on the variable to be evaluated, using the same window snapshot, the same set of target variables, the same time offset range, and the same mapping parameters. In the first pass, the variable to be evaluated is retained as input to the other variables, and the robust mapping biases of the other target variables are obtained one by one. These biases are then summed according to the target weights in the table based on the effective variables, forming a first weighted sum of mapping biases. In the second pass, only the external representation of the variable to be evaluated is set to 0, while other inputs remain unchanged. These are then summed again using the same criteria, forming a second weighted sum of mapping biases.

[0049] The processing system subtracts the weighted sum of the first mapping bias from the weighted sum of the second mapping bias. When the difference is greater than 0, the robust scale of the positive errors in the normal training window is read, the difference is divided by the scale and narrowed to 0 to 1 to obtain the effective explanation level; when the difference is less than or equal to 0, 0 is output directly. The model parameters are not updated, the effective variable pairs are not changed, and the intermediate results of the current window are not fed back to the input between two forward processing steps, thus maintaining a consistent benchmark.

[0050] When a variable to be evaluated lacks at least one valid alternative variable, or when there are fewer than three valid channels, the window does not meet the operating conditions for bidirectional interpretation. The processing system sets the window to an invalid input final state, records the missing variable pairs or channel identities, and retains the calculated variable-wise representations for troubleshooting, but does not output a unique conclusion of device anomaly or data quality anomaly. In cases where multiple acquisition channels simultaneously exhibit highly correlated common-mode interference, the independent reference channel first performs common component filtering; if filtering cannot be completed, it also enters the invalid input final state.

[0051] For S3, specifically, the intervariate transitive control process reads the variable-wise representation and data unreliability of each source variable, and converts the unreliability into a gating coefficient specific to the source variable. Gating is only set on the path from the source variable to other variables; the direct representation of the variable entering its own reconstruction branch is not multiplied by the gating coefficient to avoid the direct deletion of local anomalous evidence.

[0052] Furthermore, the variable-wise characterization that attenuates the transmission of data to the remaining vibration variables as the data unreliability increases includes: make Let be the variable index of the vibration variable, and denote the data unreliability as . ; according to Calculate the gating coefficient. This is the gate control safety factor, where, ; Using the gating coefficient as a multiplier, the variable-wise representation to be transmitted is multiplied and then transmitted to the remaining vibration variables.

[0053] Specifically, such as Figure 5 As shown, the unreliability of the source variable data Determine the gating coefficients of the source variables Source variable representation by variable Entering the representation of this variable to be reconstructed along the undecayed direct path Meanwhile, the source variables are represented variable by variable. The source variable values ​​are represented by value mapping. Source variable gating coefficient Representation of source variable values Multiply to form the gated representation of the source variable values. .

[0054] Gated source variable value representation Send to other variables along the gating's outward transmission path, where and characterize it variable-by-variable with other variables. The undecayed direct paths converge on the target variable side, yielding the representations of other variables to be reconstructed. During aggregation, the processing system adds the representations of other variables themselves to the gated post-value representations of each source variable according to the locked cross-variable association weights; the association weights only describe the normal variable relationships and do not replace the gating coefficients of the source variables.

[0055] Gate control minimum guarantee coefficient In the model version, a fixed value greater than 0 and less than 1 is used to ensure that highly unreliable source variables retain limited external context. If any gated input is missing, out of bounds, or multiplication fails, the processing system stops further scoring for that window, writes the failed final state and the identity of the erroneous variable, and does not refill the gated coefficients from the previous window.

[0056] For S4, specifically, the reconstruction scoring process reads the representation to be reconstructed for each variable. The decoder first performs one-dimensional convolution and feature dimension mapping along the time position, and outputs a reconstruction window with the same number of variables, number of sampling points, and time order as the normalized input window; if the shape or time signature is inconsistent, the processing system records the processing failure and stops scoring.

[0057] For each variable, the processing system subtracts the standardized input value from the corresponding reconstructed value point by point. After applying a robust cutoff boundary saved in the model version to the difference, the absolute value is taken, and then the results within a window are averaged to obtain the reconstruction bias of that variable. The robust cutoff boundary is obtained from the high quantile of the normal training bias, used to limit the infinite amplification of window bias by a single extreme sampling point, without changing the variable's identity and temporal correspondence.

[0058] Further, the reconstruction bias is complementaryly allocated according to the data unreliability, including: make For the variable index of the vibration variable, and The first The data unreliability and reconstruction bias of the aforementioned vibration variables The weights of variables that are non-negative and sum to 1; according to and The equipment vibration anomaly score and the data quality anomaly score are calculated separately.

[0059] Specifically, variable weights are fixed by the stability of each variable and the coverage of effective variable pairs in the normal training window, and are normalized to a sum of 1 when the model version is released; when no difference weights are set, each effective variable uses the same weight. For each variable, the processing system uses the complement of data unreliability as the equipment allocation coefficient and the data unreliability itself as the data quality allocation coefficient; the two coefficients are multiplied by the reconstruction bias and the variable weight of the variable, respectively, and then summed across variables to form the equipment vibration anomaly score and the data quality anomaly score in the same window.

[0060] The two types of allocations use the same reconstruction bias and have complementary weights. Therefore, when the unreliability of a variable increases, the share of its reconstruction bias entering the equipment vibration anomaly score decreases synchronously, while the share entering the data quality anomaly score increases synchronously; the reduced equipment-side evidence is not discarded, but transferred to the data quality side for collection and maintenance.

[0061] For S5, specifically, the state determination process reads the equipment vibration anomaly score and data quality anomaly score under the same window and model version, and reads the corresponding normal verification threshold. If the version identifiers of the score and threshold are different, the processing system sets the window to the configuration invalid final state and does not perform cross-version comparison.

[0062] Furthermore, the determination of the equipment vibration anomaly threshold adopts the quantile or robust upper bound of the normal verification sample distribution of the equipment vibration anomaly score, and the determination of the data quality anomaly threshold adopts the quantile or robust upper bound of the normal verification sample distribution of the data quality anomaly score.

[0063] Specifically, before model release, the processing system uses locked parameters to process normal validation windows, which are temporally isolated from the normal training windows, one by one, forming two types of normal validation score sequences. Each sequence is sorted in ascending numerical order, and the highest quantile value satisfying the predetermined upper limit of normal false alarms is selected as the corresponding threshold; alternatively, the median and median absolute deviation of the sequence are first obtained, then the median absolute deviation is multiplied by a fixed factor and added to the median, and the resulting robust upper bound is used as the corresponding threshold. The threshold and its acquisition method are written into the same model version; the test window and fault labels do not participate in threshold updates.

[0064] During operation, the processing system always performs two independent threshold comparisons. When the equipment vibration anomaly score exceeds the corresponding threshold, the equipment status is marked as abnormal; otherwise, it is marked as undetected anomaly. When the data quality anomaly score exceeds the corresponding threshold, the data quality status is marked as abnormal; otherwise, it is marked as undetected anomaly. This results in four combined states: undetected on both the equipment and quality sides, equipment-only anomaly, quality-only anomaly, and both sides simultaneously anomaly.

[0065] When only the device side is abnormal, the window identifier, reconstruction deviation of relevant variables, and device status are sent to the device monitoring terminal. When only the quality side is abnormal, the window identifier, variable identity with high data unreliability, and quality status are sent to the data acquisition and maintenance terminal. When both sides are abnormal, two types of records are sent separately, without one type of status overriding the other. If no abnormality is detected on either side, the normal completion record is saved, and no abnormality alarm is sent.

[0066] Each window uses its start time, variable identity set, and model version as the idempotent processing key, and sequentially goes through the states of pending alignment, encoded, evaluated, gated, scored, and decided. The final states are normal completion, invalid input, invalid configuration, and processing failure. During each state transition, the input validation summary, the validation value of the intermediate object, and the next pending execution state are simultaneously saved.

[0067] After the processing process restarts, it first searches for the same idempotent processing key. If a final state record already exists, the saved result is returned directly without resending; if a non-final state record exists and the original window, locked model version, and intermediate object verification values ​​are complete, it continues from the first incomplete state; if the conditions are incomplete, it transitions to a processing failure final state. The failure window releases its temporary buffer, and the next window restarts with a new idempotent processing key; the failure value of the previous window is not carried over to the next window.

[0068] The results of the operation and manual inspection can be added to the offline sample library, but the unreliability of the data being modified in the window cannot be reported, such as the gating coefficient, allocation coefficient, or threshold. Only after offline retraining, normal validation, and version integrity verification have all passed can the relevant changes be used as a new model version for subsequent windows, thus ensuring that the processing chain of a single window remains deterministic and reproducible.

[0069] In some embodiments, the system is used to implement the aforementioned method, and the system includes: The window and encoding module are used to acquire multiple vibration variables that are collected synchronously, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position. The quality assessment module is used to determine the data unreliability for each vibration variable based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. The intervariable transfer control module is used to attenuate the variable representation sent to the remaining vibration variables as the data unreliability increases, and retain the unattenuated direct path of the variable representation to obtain the representation to be reconstructed. The reconstruction scoring module is used to reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. The reconstruction deviation is then distributed according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. The status determination module is used to determine the equipment vibration abnormality threshold and the data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, the module outputs the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, the module outputs the data quality abnormality status.

[0070] Specifically, such as Figure 6 As shown, the window and encoding module executes S1, receiving multiple vibration variables synchronously acquired from the synchronous vibration acquisition interface; the window and encoding module then transmits the variable-wise representation to the quality assessment module. The quality assessment module executes S2 and transmits the data unreliability to the intervariate transfer control module. The intervariate transfer control module executes S3 and transmits the representation to be reconstructed to the reconstruction scoring module. The reconstruction scoring module executes S4 and transmits the equipment vibration anomaly score and data quality anomaly score to the state determination module. The state determination module executes S5.

[0071] The model parameters and normal baseline memory provide model version parameters to the window and encoding module and the quality assessment module according to version, and provide normal verification thresholds to the status determination module. The status determination module outputs abnormal equipment vibration status to the equipment monitoring terminal and abnormal data quality status to the acquisition and maintenance terminal; the two output interfaces use the same window identifier, but save and confirm the reception status separately.

[0072] The aforementioned modules can be stored as program modules in memory, loaded and executed by the processor; alternatively, they can be deployed separately in synchronous acquisition devices, edge computing devices, and servers. In distributed deployment, each module uses a message passing intermediary object containing window identifier, variable identity, model version, object type, and checksum. The receiving module first matches these five fields before consuming the message. Messages with inconsistent fields are rejected and an invalid configuration record is created, preventing cross-contamination of objects from different windows or versions.

[0073] When the system starts, the model parameters and normal benchmark memory read the list and file verification values ​​of the published versions, checking each required parameter to ensure it exists and can be parsed. If the verification passes, the version is set to available, and the window and encoding module are allowed to receive new windows; if the verification fails, the system remains in an invalid configuration state, outputs the reason code for missing fields or files that failed verification, and starts the detection without using some parameters.

[0074] From an overall technical perspective, the fundamental problem this solution addresses is that both real-world equipment malfunctions and localized data acquisition degradation can cause sudden vibration changes, increased fluctuations, or an increase in high-frequency components. Local statistics of the variable alone cannot reliably distinguish between the two. Therefore, the processing system first retains the temporal location representation of each variable. Then, it uses whether the variable can be explained by other valid variables and whether it can reduce the mapping bias of other variables as two opposing cross-variable pieces of evidence, which, together with localized quality degradation, converge to form data unreliability. This intermediate value only increases significantly when localized morphological degradation, unexplainable ingress, and a lack of effective contribution from the egress all occur simultaneously.

[0075] The same data unreliability is then reused in two stages: in the information propagation stage, only the transmission intensity of the suspicious source variable to other variables is reduced, while its unattenuated direct path is preserved; in the evidence attribution stage, the same reconstruction bias is complementaryly distributed between the device side and the data quality side. The former process reduces the contamination of the normal channel by isolated disturbed channels, while the latter process ensures that the bias that has been downweighted by the device side still enters the acquisition and maintenance chain; the two types of thresholds are then compared independently, allowing simultaneous output when device anomalies and acquisition degradation coexist. Identity matching, version locking, validity verification, abnormal final state, and idempotent recovery jointly ensure that the above causal chain does not crosstalk between different windows and can be reproduced.

[0076] like Figure 7 As shown, the horizontal axis represents the unreliability of the data. The vertical axis represents the coefficients, and the three curves represent the gating coefficients. Equipment allocation coefficient and data quality allocation coefficient .when As data unreliability increases from 0 to 1, the gating coefficient decreases continuously from 1 to 0.2 according to the aforementioned gating relationship, the equipment allocation coefficient decreases continuously from 1 to 0, and the data quality allocation coefficient increases continuously from 0 to 1; the two allocation coefficients are complementary at any point. Example point correspond This is used to verify the trend shown in the figure and the calculations in the aforementioned examples, and does not represent the actual measured performance curve.

[0077] In other implementations, univariate temporal encoding can be achieved using a one-dimensional convolutional network, recurrent network, or temporal attention network that preserves the temporal position; cross-variable mapping can be achieved using multi-head attention with source variable masking, graph messaging, or multivariate regression. Regardless of the implementation used, the same set of target variables and error caliber should be used to perform two comparisons of the preserved source variable and the masked source variable, and the gating should only apply to the representation sent by the source variable to other variables.

[0078] In other implementations, normalization of the normal training distribution can be replaced by an empirical cumulative distribution instead of linear interpolation from the median to the high quantile. Reconstruction bias can be aggregated using absolute bias, squared bias with fixed truncation, or smoothed absolute bias. The threshold can be a quantile or a robust upper bound. The parameters of the alternative approach are all fixed by the normal training window or normal validation window, and the processing relationships between data unreliability, gating coefficients, complementary assignments, and independent state comparisons are maintained.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multivariable vibration anomaly detection method based on data quality gating, characterized in that, include: S1. Acquire multiple vibration variables that are synchronously collected, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position; S2. For each vibration variable, determine the data unreliability based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. S3. As the unreliability of the data increases, the variable-wise representations sent to the remaining vibration variables are attenuated, and the unattenuated direct path of the variable-wise representations is retained to obtain the representations to be reconstructed. S4. Reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. Allocate the reconstruction deviation according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. S5. Determine the equipment vibration abnormality threshold and data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, output the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, output the data quality abnormality status.

2. The method according to claim 1, characterized in that, The multiple vibration variables are at least three vibration variables obtained from mutually independent acquisition channels.

3. The method according to claim 1, characterized in that, The data unreliability is based on Calculate, where, The unreliability of the data, and The terms represent the degree of local quality degradation, the degree of unexplained quality, and the degree of effective explanation, respectively, and all three are taken from 0 to 1. The data unreliability increases with the degree of local quality degradation and the degree of unexplainedness, and decreases with the degree of effective explanation.

4. The method according to claim 3, characterized in that, The determination of the degree of local quality degradation includes: The variation of adjacent sampled values, local dispersion, and high-frequency residual energy of the vibration variable are calculated and combined to obtain local mass statistics. The local quality statistics are normalized according to the normal training distribution to obtain the degree of local quality degradation.

5. The method according to claim 3, characterized in that, The determination of the degree of unexplainedness includes: Within the normal time offset range, the remaining vibration variables that have an effective mapping relationship with the vibration variable to be evaluated are used as inputs to map the vibration variable to be evaluated, thereby obtaining a mapping characterization; The degree of unexplainedness is obtained by comparing the variable-wise representation of the vibration variable to be evaluated with the mapping representation.

6. The method according to claim 3, characterized in that, The determination of the effective explanatory power includes: The vibration variables to be evaluated are retained in the mapping, and the first mapping deviation weighted sum of the remaining vibration variables is calculated; In the mapping, the vibration variables to be evaluated are masked, and the second mapping deviation weighted sum of the remaining vibration variables is calculated; When the weighted sum of the second mapping deviations is greater than the weighted sum of the first mapping deviations, the effective explanatory power is obtained by subtracting the weighted sum of the first mapping deviations from the weighted sum of the second mapping deviations. When the weighted sum of the second mapping deviations is less than or equal to the weighted sum of the first mapping deviations, the effective explanatory power is determined to be 0.

7. The method according to claim 1, characterized in that, The variable-wise characterization that attenuates the transmission of data to the remaining vibration variables as the data unreliability increases includes: make Let be the variable index of the vibration variable, and denote the data unreliability as . ; according to Calculate the gating coefficient, where, This is the minimum threshold for gate control. ; Using the gating coefficient as a multiplier, the variable-wise representation to be transmitted is multiplied and then transmitted to the remaining vibration variables.

8. The method according to claim 1, characterized in that, The method of complementaryly allocating the reconstruction bias according to the data unreliability includes: make For the variable index of the vibration variable, and The first The data unreliability and reconstruction bias of the aforementioned vibration variables These are the non-negative variable weights that sum to 1; according to and The equipment vibration anomaly score and the data quality anomaly score are calculated separately.

9. The method according to claim 1, characterized in that, The threshold for abnormal equipment vibration is determined by the quantile or robust upper bound of the normal verification sample distribution of the abnormal equipment vibration score, and the threshold for abnormal data quality is determined by the quantile or robust upper bound of the normal verification sample distribution of the abnormal data quality score.

10. A multivariable vibration anomaly detection system based on data quality gating, characterized in that, The system is used to implement the method of any one of claims 1 to 9, and the system comprises: The window and encoding module are used to acquire multiple vibration variables that are collected synchronously, perform time-series encoding on the vibration variables, and obtain a variable-wise representation of the retained time position. The quality assessment module is used to determine the data unreliability for each vibration variable based on the degree of local quality degradation of the vibration variable, the degree of unexplainedness obtained from the input mapping deviation, and the degree of effective explanation obtained from the output mapping contribution. The intervariable transfer control module is used to attenuate the variable representation sent to the remaining vibration variables as the data unreliability increases, and retain the unattenuated direct path of the variable representation to obtain the representation to be reconstructed. The reconstruction scoring module is used to reconstruct the vibration variable based on the character to be reconstructed and obtain the reconstruction deviation. The reconstruction deviation is then distributed according to the data unreliability to obtain the equipment vibration anomaly score that decreases as the data unreliability increases and the data quality anomaly score that increases accordingly. The status determination module is used to determine the equipment vibration abnormality threshold and the data quality abnormality threshold based on the normal verification window. When the equipment vibration abnormality score exceeds the equipment vibration abnormality threshold, the module outputs the equipment vibration abnormality status. When the data quality abnormality score exceeds the data quality abnormality threshold, the module outputs the data quality abnormality status.

Images