Multidimensional signal fault detection method under low signal-to-noise ratio and computer device

Abstract

The application discloses a multi-dimensional signal fault detection method and computer device under low signal-to-noise ratio, noise types and intensity are quantified through noise imaging, and a target de-noising operator is determined, which improves fault feature separability and reduces noise dominant risk; through cross-channel consistency features, common evidence of multi-dimensional signals is used for enhancing detection, so that weak faults invisible in a single channel can be stably identified after multi-dimensional fusion, thereby improving the accuracy of fault detection under low signal-to-noise ratio.

Description

Technical Field

[0001] This application relates to the field of industrial equipment testing technology, and in particular to a multi-dimensional signal fault detection method and computer device under low signal-to-noise ratio conditions. Background Technology

[0002] In industrial equipment condition monitoring, the complex on-site working conditions and installation conditions lead to the prevalence of low signal-to-noise ratio (Low SNR) phenomena, making fault impacts easily overwhelmed by strong background noise, structural resonance, and non-stationary components caused by variable load and speed.

[0003] Currently, acquired signals can be filtered using fixed filtering parameters (such as bandpass filtering / envelope demodulation) to determine fixed features (such as peak-to-peak value, kurtosis, spectral peaks, and envelope spectral peaks), and fault detection can be performed based on these fixed features. However, at low signal-to-noise ratios, fixed filtering parameters and fixed features are sensitive to noise, and the feature variance is easily dominated by noise, resulting in low accuracy of fault detection. Summary of the Invention

[0004] This application provides a multi-dimensional signal fault detection method and computer device under low signal-to-noise ratio conditions. The method quantifies the noise type and intensity by noise profiling to improve the accuracy of fault detection in low signal-to-noise ratio scenarios.

[0005] In a first aspect, a method for multi-dimensional signal fault detection under low signal-to-noise ratio is provided, comprising: obtaining the operating signal of a target device within a target time window through multiple acquisition channels, wherein the operating signal corresponds to multiple dimensions, and each dimension corresponds to one or more acquisition channels; determining the noise data of the operating signal, and determining the noise profile of the operating signal based on the noise data; determining a target denoising operator from a set of denoising operators based on the noise profile, and using the target denoising operator to denoise the operating signal to obtain a denoised signal; determining a fault score based on the target features of the denoised signal, wherein the target features include at least a consistency enhancement feature, and the consistency enhancement feature characterizes the degree of consistency of each denoised signal under multiple acquisition channels; and determining the existence of a fault and outputting alarm information when the fault score reaches the target threshold and the target alarm condition is met.

[0006] In some embodiments, determining the target denoising operator from the set of denoising operators based on the noise profile includes: denoising the running signal using multiple denoising operators in the set of denoising operators to obtain a denoised signal to be tested; determining a fidelity score for each of the denoised signals to be tested based on a target fidelity index, wherein the fidelity score characterizes the degree to which the denoised signal to be tested retains non-noise signals, and the target fidelity index includes at least one of the following: band energy, impulse sparsity, and cross-channel consistency; and determining the target denoising operator from the set of denoising operators based on the fidelity score and each of the denoised signals to be tested.

[0007] In some embodiments, determining the target denoising operator from the set of denoising operators based on the fidelity score and each of the denoised signals under test includes: determining the signal-to-noise ratio gain of each of the denoised signals under test; determining a comprehensive score of each of the denoised signals under test based on the signal-to-noise ratio gain and the fidelity score; and determining the target denoising operator from the set of denoising operators based on the comprehensive score.

[0008] In some embodiments, determining the fault score based on the target features of the denoised signal includes: robustly normalizing the target features to determine normalized features; determining multiple sub-scores based on the normalized features, the sub-scores including at least two of the following: impulse sub-score, modulation sub-score, energy drift sub-score, and consistency anomaly sub-score, wherein the impulse sub-score is determined by envelope spectral peaks, spectral kurtosis, and impulse consistency; the modulation sub-score is determined by the ratio of sideband energy to order domain sideband; the energy drift sub-score is determined by the offset of target frequency band energy relative to the baseline; and the consistency anomaly sub-score is determined by coherence degradation or impulse consistency anomaly; and weightedly fusing the sub-scores to determine the fault score.

[0009] In some embodiments, determining the noise data of the operating signal includes: determining the intra-channel noise signal of the operating signal within a single channel and the inter-channel noise signal of the operating signal among multiple channels, wherein the intra-channel noise signal includes at least one of the following: noise variance estimation, noise floor power spectral density, narrowband interference, impulse noise figure, and frequency band signal-to-noise ratio estimation of the target frequency band; the inter-channel noise signal includes at least one of the following: inter-channel coherence ratio and inter-channel common-mode ratio; and determining the noise data based on the intra-channel noise signal and the inter-channel noise signal.

[0010] In some embodiments, the process of determining the target threshold includes: determining a target operating condition box from multiple operating condition boxes according to the operating condition corresponding to the operating signal, wherein different operating condition boxes correspond to operating data under different operating conditions in the target time window; determining the data volume of the target operating data corresponding to the target operating condition box; and determining the target threshold according to the data volume and the controllable false alarm rate.

[0011] In some embodiments, determining the target threshold based on the data volume and the controllable false alarm rate includes: determining a target quantile based on the controllable false alarm rate when the data volume does not reach the target data volume; determining the target threshold based on the target quantile of the target running data; and determining the target threshold using extreme value theory when the data volume reaches the target data volume, based on the controllable false alarm rate and the target running data.

[0012] In some embodiments, the alarm information includes the confidence level of the fault score and explanatory information, wherein the explanatory information includes at least one of the following: information on the target feature that contributes the most to the fault score, information on the target test chamber, and summary information of the noise profile; the confidence level is determined by at least one of the data quality of the operating signal, the degree to which the denoised signal preserves non-noise signals, channel delay stability, or coherent baseline.

[0013] In some embodiments, the method further includes: if the number of times the fault score continuously reaches the target threshold exceeds a target number, suspending the data in the target operating data corresponding to the fault score from participating in the determination of the target threshold.

[0014] In a second aspect, a computer device is provided, including a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program to implement the multidimensional signal fault detection method under low signal-to-noise ratio as described in the first aspect.

[0015] By applying the above technical solutions, the operating signals of the target device within a target time window are obtained based on multiple acquisition channels. These operating signals correspond to multiple dimensions, with each dimension corresponding to one or more acquisition channels. Noise data of the operating signals is determined, and a noise profile is established based on this data. A target denoising operator is selected from a set of denoising operators based on the noise profile, and this operator is used to denoise the operating signals, obtaining denoised signals. A fault score is determined based on the target features of the denoised signals. These target features include at least consistency enhancement features, which characterize the consistency of each denoised signal across multiple acquisition channels. If the fault score reaches a target threshold and meets the target alarm conditions, a fault is identified, and an alarm is output. This approach quantifies noise type and intensity through noise profiling and determines the target denoising operator, improving fault feature separability and reducing the risk of noise dominance. By using cross-channel consistency features to leverage common evidence from multi-dimensional signals for enhanced detection, weak faults invisible in a single channel can be stably identified after multi-dimensional fusion, thereby improving the accuracy of fault detection in low signal-to-noise ratio scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a multidimensional signal fault detection method under low signal-to-noise ratio according to an embodiment of this application; Figure 2 This is a flowchart of the target denoising operator in an embodiment of this application; Figure 3 This is a flowchart illustrating the determination of fault scores in an embodiment of this application; Figure 4 The process for determining the target threshold in the embodiments of this application Figure 1 ; Figure 5 The process for determining the target threshold in the embodiments of this application Figure 2 ; Figure 6 This is a structural block diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0018] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0019] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0020] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0021] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0022] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0023] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0024] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0025] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0026] This application provides a method for multidimensional signal fault detection under low signal-to-noise ratio conditions. By quantifying the noise type and intensity through noise profiling and determining the target denoising operator, the method improves the separability of fault features and reduces the risk of noise dominance. By using cross-channel consistency features to enhance detection by combining common evidence from multidimensional signals, weak faults that are not visible in a single channel can be stably identified after multidimensional fusion, thereby improving the accuracy of fault detection in low signal-to-noise ratio scenarios.

[0027] like Figure 1 As shown, this multidimensional signal fault detection method under low signal-to-noise ratio includes the following steps: Step S101: Obtain the operating signal of the target device within the target time window through multiple acquisition channels. The operating signal corresponds to multiple dimensions, and each dimension corresponds to one or more acquisition channels.

[0028] In this embodiment, the execution entity of the fault detection method can be an electronic device acting as an edge device or a server. The target device includes one or more of the following: motor, bearing, gearbox, pump, fan, robot reducer, etc. The multi-dimensional operating signals can include various signals such as vibration (multi-axis), current, voltage, rotational speed, temperature, acoustic emission, pressure, and torque. The operating signal within the target time window is a data segment within a target duration. The target time window slides according to a target step size; for example, the window length of the target time window can be 1s-10s, and the step size can be 0.2s-1s. In some cases, the signal-to-noise ratio of this operating signal is lower than the target signal-to-noise ratio, making it a low signal-to-noise ratio signal.

[0029] The operating signals of the target device within a target time window are obtained through multiple acquisition channels. Each dimension corresponds to one or more acquisition channels. For example, if the multiple acquisition channels include a first acquisition channel, a second acquisition channel, a third acquisition channel, and a fourth acquisition channel, current can be acquired through the first acquisition channel, voltage through the second acquisition channel, temperature through the third acquisition channel, and torque through the fourth acquisition channel. If the multiple acquisition channels include three acquisition channels, triaxial vibration signals belonging to the same dimension can be acquired through the three acquisition channels, with each acquisition channel acquiring the vibration signal of one axis.

[0030] Step S102: Determine the noise data of the operating signal, and determine the noise profile of the operating signal based on the noise data.

[0031] In this embodiment, noise data is determined by identifying noise in the operating signal, and a noise profile of the operating signal is determined based on the noise data. This noise profile can characterize the noise features of the noise data.

[0032] Step S103: Determine the target denoising operator from the set of denoising operators based on the noise profile, and use the target denoising operator to denoise the running signal to obtain a denoised signal.

[0033] In this embodiment, the denoising operator set may include multiple denoising operators, each with a parameter space. For example, each denoising operator may include: Adaptive wavelet shrinkage, parameters: wavelet basis, number of decomposition levels, threshold coefficient; Wiener / Spectral Subtraction Enhancement, parameters: noise PSD (power spectral density), smoothing factor; Robust bandpass + notch filter, parameters: passband, notch filter frequency; SSA / Low-rank decomposition denoising, parameters: embedding dimension, rank; Kalman smoothing / state space.

[0034] Different noise profiles require different denoising operators for their runtime data. One or more target denoising operators can be determined from the set of denoising operators based on the noise profile, and then used to denoise the runtime signal to obtain a denoised signal. For example, if the proportion of narrowband interference lines represented by the noise profile reaches a first proportion, a robust bandpass filter with notch filtering is preferentially used as the target denoising operator. If the noise profile represents approximately broadband noise and the SNR_hat_k (the estimated signal-to-noise ratio of the kth predefined frequency band (structural resonance band, envelope band, meshing frequency band, etc.)) is lower than the target estimate, adaptive wavelet shrinkage and Wiener / spectral attenuation enhancement are preferentially used as the target denoising operators. If the noise profile characterizes the common-mode ratio to the second proportion, the target denoising operator needs to perform projection elimination on the common-mode components before proceeding with denoising. For example, if it is confirmed that the first principal component is mainly dominated by common-mode noise (e.g., by jointly judging through coherence baseline and common-mode ratio), PCA (Principal Component Analysis) is performed on the channel matrix to remove the first principal component or weaken it proportionally to avoid accidentally deleting fault signals. If the impulse noise index is higher than the target index and the impulse needs to be retained, the target denoising operator needs to avoid excessive smoothing and prioritize the wavelet threshold strategy (reducing the threshold coefficient) to retain high-frequency details.

[0035] Step S104: Determine the fault score based on the target features of the denoised signal. The target features include at least a consistency enhancement feature, which characterizes the degree of consistency of each denoised signal across multiple acquisition channels.

[0036] In this embodiment, fault scoring is performed based on the target features of the denoised signal to determine the fault score. The target features include at least a consistency enhancement feature, which characterizes the degree of consistency of each denoised signal across multiple acquisition channels. In some embodiments of this application, the consistency enhancement feature includes at least one of the following: the mean and variance of cross-spectrum / coherence in the target frequency band, the impact arrival consistency feature, and the multidimensional energy directionality feature. The impact arrival consistency feature characterizes the calculation of the cross-channel event overlap rate for the detected transient event time set {τ_i^c}. The multidimensional energy directionality feature can characterize triaxial vibration. For triaxial vibration signals, the principal direction energy ratio (i.e., the energy proportion of the first principal component after PCA of the triaxial signal matrix) can be used.

[0037] Step S105: If the fault score reaches the target threshold and the target alarm condition is met, a fault is determined to exist, and an alarm message is output.

[0038] In this embodiment, after determining the fault score, the fault detection result is determined based on the comparison between the fault score and the target threshold. For example, if the fault score reaches the target threshold and meets the target alarm condition, then a fault is determined to exist, and an alarm message is output. Otherwise, it is determined that no fault exists.

[0039] In some embodiments of this application, before determining the noise data of the running signal, the process may further include: preprocessing the running signal, the preprocessing including at least one of the following: synchronization processing, calibration and normalization processing, quality gating processing, anti-aliasing and detrending processing.

[0040] In this embodiment, the synchronization process may include: if there is a fixed delay or clock drift between multiple acquisition channels of the same target device, the inter-channel delay Δτ_c is estimated and corrected using a cross-correlation method.

[0041] Calibration and normalization processes may include: converting analog quantities into physical quantities, and applying gain and / or bias corrections to the data from each acquisition channel.

[0042] Quality gating processing may include: determining the data quality of the running signal, where data quality includes at least one of missing rate, saturation rate, clipping count, and timestamp jitter; if the data quality does not meet the target conditions, preventing data in the target time window from participating in the determination of the target threshold and issuing a low-confidence alarm signal; for example, the target condition corresponding to the saturation rate may be that the saturation rate does not exceed the threshold; if the saturation rate exceeds the threshold (e.g., 1%), it is determined that the saturation rate does not meet the target conditions, the data in the target time window is not participating in the determination of the target threshold, and a "low-confidence output" is issued for alarm purposes.

[0043] Anti-aliasing and de-stressing processing may include: configuring anti-aliasing filters according to the sampling rate; and performing high-pass / de-stressing on low-frequency drift (e.g., 0.5–5 Hz, depending on the specific target device).

[0044] By preprocessing the operating signals, the accuracy of the operating signals is further improved, thereby improving the accuracy of fault detection.

[0045] The multi-dimensional signal fault detection method under low signal-to-noise ratio (SNR) conditions in this application obtains the operating signal of the target device within a target time window through multiple acquisition channels. The operating signal corresponds to multiple dimensions, and each dimension corresponds to one or more acquisition channels. Noise data of the operating signal is determined, and a noise profile of the operating signal is determined based on the noise data. A target denoising operator is determined from a set of denoising operators based on the noise profile, and the operating signal is denoised using the target denoising operator to obtain a denoised signal. A fault score is determined based on the target features of the denoised signal. The target features include at least consistency enhancement features, which characterize the consistency of each denoised signal across multiple acquisition channels. When the fault score reaches a target threshold and meets the target alarm conditions, a fault is determined, and an alarm message is output. This method quantifies the noise type and intensity through the noise profile and determines the target denoising operator, improving the separability of fault features and reducing the risk of noise dominance. By using cross-channel consistency features to utilize common evidence from multi-dimensional signals to enhance detection, weak faults invisible in a single channel can be stably identified after multi-dimensional fusion, thereby improving the accuracy of fault detection in low SNR scenarios.

[0046] In some embodiments of this application, the step of determining the target denoising operator from the set of denoising operators based on the noise profile is as follows: Figure 2 As shown, it includes the following steps: Step S1031: Use multiple denoising operators in the denoising operator set to denoise the running signal respectively to obtain the denoised signal to be tested.

[0047] In this embodiment, the multiple denoising operators can be all denoising operators in the denoising operator set, or only a portion of the denoising operators (e.g., excluding denoising operators that clearly do not match the noise profile). The running signal is denoised using the multiple denoising operators in the denoising operator set to obtain the denoised signal to be tested.

[0048] Step S1032: Determine the fidelity score of each of the denoised signals to be tested based on the target fidelity index. The fidelity score characterizes the degree to which the denoised signal to be tested retains non-noise signals. The target fidelity index includes at least one of the following: band energy, impulse sparsity, and cross-channel consistency.

[0049] In this embodiment, the frequency band energy can be the energy of the target frequency band, which may include one or more of the following: structural resonance band, envelope band, and band near the meshing frequency. The score can be determined by whether the target frequency band energy ratio remains within the range [r_min, r_max] (e.g., 0.6–1.6). The score can also be determined by whether the variation in sparsity indicators (such as L1 / L2 ratio, kurtosis) does not exceed a threshold. Finally, the score can be determined by whether the cross-channel coherence boost Δcoh is positive and not excessively concentrated at a single frequency point.

[0050] Each denoised signal under test is scored based on the target fidelity index to determine the fidelity score. The fidelity score characterizes the degree to which the denoised signal under test retains non-noise signals in the running signal.

[0051] Step S1033: Determine the target denoising operator from the set of denoising operators based on the fidelity score and each of the denoised signals to be tested.

[0052] In this embodiment, the appropriate denoising operator is determined based on the fidelity score and each denoised signal to be tested to determine whether it is suitable for denoising the running data, and the denoising operator suitable for denoising the running data is determined as the target denoising operator.

[0053] By considering the fidelity score corresponding to the denoising operator, non-noise signals (such as fault signals) can be avoided during the denoising process, thus preserving key fault information even at low SNR (Signal-to-Noise Ratio) and improving the accuracy of fault detection.

[0054] In some embodiments of this application, determining the target denoising operator from the set of denoising operators based on the fidelity score and each of the denoised signals to be tested includes: Determine the signal-to-noise ratio gain of each of the denoised signals to be tested; Based on the signal-to-noise ratio gain and fidelity score of each of the denoised signals under test, a comprehensive score for each of the denoised signals under test is determined. The target denoising operator is determined from the set of denoising operators based on the comprehensive scores.

[0055] In this embodiment, the signal-to-noise ratio (SNR) gain of each denoised signal to be tested is first determined. Then, for each denoised signal to be tested, its comprehensive score is determined based on its SNR gain and fidelity score. Finally, by comparing the comprehensive scores, the target denoising operator is determined from the set of denoising operators based on the comprehensive scores, including the following cases: Case 1: Determine the highest comprehensive score. There is only one highest comprehensive score. The denoising operator corresponding to the highest comprehensive score is determined as the target denoising operator. Scenario 2: Determine the highest comprehensive score. There are multiple highest comprehensive scores. Select one of the denoising operators corresponding to the highest comprehensive score as the target denoising operator. Scenario 3: Determine the top-k comprehensive scores with the highest scores, and identify the denoising operators corresponding to the top-k comprehensive scores as multiple target denoising operators. Each target denoising operator can be denoised in series or in parallel when denoising the running signal, and then the denoised signal is output as the best result.

[0056] Case 4: Select multiple parameter candidate combinations. Each parameter candidate combination includes n denoising operators (n>1). For each parameter candidate combination, determine the comprehensive score of each denoising operator and sum them with weights to determine the total score of the parameter candidate combination. The denoising operators in the parameter candidate combination with the largest total score are determined as the target denoising operators. When denoising the running signal in the subsequent process, the target denoising operators can be denoised in series or in parallel and then the denoised signal is output selectively.

[0057] In this way, the comprehensive score is determined by the signal-to-noise ratio gain and fidelity score of the denoised signal under test, so as to achieve more efficient and accurate determination of the target denoising operator.

[0058] In some embodiments of this application, the comprehensive score is determined by formula (1):

[0059] Where R represents the overall score, and α represents the weight. Indicates the signal-to-noise ratio gain. This represents the fidelity score. In some embodiments of this application, α = 0.4~0.7; the number of denoising operator / parameter candidate combinations corresponding to the target time window is 5–30.

[0060] In some embodiments of this application, as an alternative, a denoised signal is first obtained using a preset default denoising operator, and then the denoised signal is comprehensively scored based on formula (1). If the comprehensive score is less than the target comprehensive score, another denoising operator is used to denoise and score again in a preset order until the comprehensive score reaches the target or all denoising operators in the candidate set are traversed. This avoids removing non-noise signals (such as fault signals) during the denoising process, thus retaining key fault information even at low SNR and improving the accuracy of fault detection.

[0061] In some embodiments of this application, the step of determining the fault score based on the target features of the denoised signal is as follows: Figure 3 As shown, it includes the following steps: Step S1041: Robustly standardize the target features to determine the standardized features.

[0062] In this embodiment, the target features, in addition to consistency enhancement features, may also include at least one of the following: time-domain features, frequency-domain features, and time-frequency / modulation-domain features. The time-domain features may include one or more of RMS (Root Mean Square), MAD (Median Absolute Deviation), IQR (Interquartile Range), peak factor, kurtosis (robust truncated version), and impulse index. In low SNR scenarios, robust statistics based on median absolute deviation (MAD) / interquartile range (IQR) are used instead of pure variance in the time-domain features to reduce the contamination of fault features by noise spikes. The frequency-domain features may include one or more of the following: target frequency band energy, spectral peak ratio, spectral kurtosis peak value and position, and narrowband line remnant (used to identify cases where electromagnetic interference is not completely suppressed, reducing confidence).

[0063] Time-frequency domain / modulation domain features may include: envelope spectrum, amplitude / sideband energy at fault characteristic frequencies (such as BPFO / BPFI / FTF / BSF, gear meshing and sideband) after Hilbert envelope; "order tracking" or "speed normalized frequency axis" is used when the speed changes: the frequency is mapped to the order using rpm, and features are extracted in the order domain to improve non-stationary robustness; short-time Fourier / continuous wavelet energy ridges are used to extract energy migration and modulation depth.

[0064] The target features are robustly standardized using formula (2):

[0065] in, The standardized feature represents the j-th target feature. Represents the j-th target feature. This represents the median of the j-th target feature. Let Q3-Q1 represent the interquartile range of the j-th target feature. This indicates the minimum value to prevent the denominator from being zero.

[0066] Step S1042: Determine multiple sub-fractions based on the normalization features. The sub-fractions include at least two of the following: impulse sub-fraction, modulation sub-fraction, energy drift sub-fraction, and consistency anomaly sub-fraction. The impulse sub-fraction is determined by the envelope spectral peak, spectral kurtosis, and impulse consistency. The modulation sub-fraction is determined by the ratio of sideband energy to order domain sideband. The energy drift sub-fraction is determined by the offset of the target frequency band energy relative to the baseline. The consistency anomaly sub-fraction is determined by coherence degradation or impulse consistency anomaly.

[0067] In this embodiment, multiple sub-scores are determined to achieve scoring from multiple different perspectives.

[0068] Step S1043: Weighted fusion of each sub-score is performed to determine the fault score.

[0069] In this embodiment, the fault score can be determined using formula (3):

[0070] in, This represents the fault score at time t. This represents the weight of the k-th sub-score. This represents the k-th sub-fraction. It can be fixed (in the early stages of the project) or adapted according to the noise profile (e.g., increasing the weight of consistency anomaly sub-scores in the case of low signal-to-noise ratio).

[0071] By robustly standardizing the target features and determining the fault score through multiple sub-scores of the standardized features, the accuracy of the fault score is further improved.

[0072] In some embodiments of this application, determining the noise data of the operating signal includes: The method determines the intra-channel noise signal of the operating signal within a single channel and the inter-channel noise signal of the operating signal across multiple channels. The intra-channel noise signal includes at least one of the following: noise variance estimation, noise floor power spectral density, narrowband interference, impulse noise figure, and frequency band signal-to-noise ratio estimation of the target frequency band. The inter-channel noise signal includes at least one of the following: inter-channel coherence ratio and inter-channel common-mode ratio. The noise data is determined based on the noise signal within the channel and the noise signal between the channels.

[0073] In this embodiment, noise variance estimation may include estimation in the high-frequency tail or silent band, or using wavelet detail coefficients MAD: sigma_hat = median(|d1|) / 0.6745. PSD_c(f) can be obtained using the Welch method, and the noise floor power spectral density is obtained by smoothing with the moving median. The smoothing window width is typically much smaller than the width bw of the narrowband interference line (e.g., the moving window corresponding to bw = 0.5~5Hz is approximately 1 / 5~1 / 3 of bw). Peaks (peak values ​​exceeding the neighborhood median K_line, such as 8–15) are found in the PSD, and narrowband interference is recorded for notch filtering / suppression. The impulse noise index may include kurtosis / kurtosis, spectral kurtosis peak value, and Teager energy statistics.

[0074] The signal-to-noise ratio of the target frequency band is estimated using formula (4):

[0075] in, This represents the bandwidth signal-to-noise ratio estimate of the target frequency band. Indicates the target frequency band Internal signal power, This represents the noise power of the target frequency band. This indicates a minimum value to prevent the denominator from being zero. The noise floor power spectral density (noise_floor_psd(f)) can be used to calculate the minimum value in the target frequency band. Numerical integration is performed inside to obtain ,Right now The original signal PSD is in Integral minus P noise Get P signal The estimate.

[0076] The inter-channel coherence ratio and inter-channel common-mode ratio can be determined by calculating the baseline and common-mode ratio of the inter-channel coherence coefficient. The inter-channel coherence ratio and inter-channel common-mode ratio are used to identify "common noise sources" (electromagnetic interference, loose installation resonance, etc.).

[0077] This method determines noise data based on noise signals within and between channels, and by considering both noise within and between channels, it improves the accuracy of noise data.

[0078] In some embodiments of this application, such as Figure 4 As shown, the process of determining the target threshold includes the following steps: Step S201: Based on the operating conditions corresponding to the operating signal, determine the target operating condition box from multiple operating condition boxes. Different operating condition boxes correspond to operating data under different operating conditions in the target time window.

[0079] In this embodiment, the operating conditions can be pre-stratified based on various operating condition signals to determine multiple operating condition boxes. These operating condition signals include, for example, speed, load, temperature, signal-to-noise ratio estimation, and noise floor. In the absence of operating condition signals, multiple operating condition boxes can also be determined using signals such as the dominant frequency peak position and total power. Optionally, multiple operating condition boxes can be determined by different intervals; for example, speed and load can be divided by interval, and noise floor can be divided by quantile interval. Multiple operating condition boxes can also be determined by clustering, for example, by performing online K-means (K=5~30) on speed, load, and noise floor. Each operating condition box maintains its independent feature distribution and threshold model to avoid distortion of the target threshold caused by mixing different operating conditions.

[0080] The target operating condition box is determined from multiple operating condition boxes based on the operating condition corresponding to the operating signal.

[0081] Step S202: Determine the amount of target operating data corresponding to the target operating condition box.

[0082] The target operating data corresponding to the target operating condition box is within the target time window, and its data volume can vary with the sampling frequency and sampling duration.

[0083] Step S203: Determine the target threshold based on the data volume and the controllable false alarm rate.

[0084] In this embodiment, the controllable false alarm rate is, for example, ≤1 time per channel per day or ≤0.1 times per hour. The target threshold is determined based on the data volume and the controllable false alarm rate, and the target threshold can be updated online.

[0085] By establishing a threshold model based on operating conditions and noise levels, the threshold can be automatically adjusted according to changes in noise level and operating conditions, significantly reducing false alarms and missed alarms.

[0086] In some embodiments of this application, the step of determining the target threshold based on the data volume and the controllable false alarm rate is as follows: Figure 5 As shown, it includes the following steps: Step S2031: If the data volume does not reach the target data volume, determine the target quantile based on the controllable false alarm rate; determine the target threshold based on the target quantile of the target running data.

[0087] In this embodiment, if the data volume does not reach the target data volume, the target quantile is determined based on the controllable false alarm rate; the target threshold is determined based on the target quantile of the target running data. The target threshold can be determined using formula (5):

[0088] Where TH represents the target threshold and FAR represents the controllable false alarm rate. TH represents the target quantile, and Y represents the target operational data. For example, if the controllable false alarm rate is 5%, then TH = The target threshold is the 95th percentile of the target running data.

[0089] Step S2032: When the data volume reaches the target data volume, the target threshold is determined using extreme value theory based on the controllable false alarm rate and the target operating data.

[0090] If the data volume reaches the target data volume, the target threshold is determined using extreme value theory (EVT(POT)). An initial threshold u (e.g., the 95th percentile) is selected, and a GPD (Generalized Pareto Distribution) is fitted to the exceedance amount Yu to calculate the target threshold TH_evt under a controllable false alarm rate. Simultaneously, the threshold confidence interval is output (given by the parameter estimation variance). Here, EVT adopts the POT-GPD scheme, where the number N of exceedances above the initial threshold u is greater than or equal to a lower limit, which could be, for example, 2000.

[0091] By using different methods to determine the target threshold based on the different amounts of target operational data, the threshold can be automatically adjusted according to changes in operating conditions, reliably reducing false alarms and missed alarms.

[0092] In some embodiments of this application, the alarm information includes the confidence level of the fault score and explanatory information. The explanatory information includes at least one of the following: information on the target feature that contributes the most to the fault score, information on the target test chamber, and summary information of the noise profile. The confidence level is determined by at least one of the following: the data quality of the operating signal, the degree to which the denoised signal retains non-noise signals, channel delay stability, or coherent baseline.

[0093] In this embodiment, the target alarm condition may include one or more combinations of the following: the number of target time windows in which the fault score continuously exceeds the target threshold reaches a target number (e.g., 2-5); the confidence level reaches a target confidence level; and the difference between the fault score and the target threshold reaches a target difference. If the fault score reaches the target threshold and the target alarm condition is met, a fault is determined to exist, and alarm information is output. The alarm information may include at least one of the following: information on the target feature that contributes the most to the fault score, information on the target operating condition box, and summary information of the noise profile. The information on the target feature that contributes the most to the fault score may include a feature identifier and the corresponding sub-score. The information on the target operating condition box may include the identifier of the target operating condition box.

[0094] By determining the confidence level of the fault score and including confidence level and explanatory information in the alarm information, users can intuitively determine the fault status of the target device, thus improving the user experience.

[0095] In some embodiments of this application, the target threshold is based on a Schmitt trigger / hysteresis threshold, including a first threshold TH_high corresponding to an upward trend and a second threshold TH_low corresponding to a downward trend. An alarm is triggered when the fault score rises from below TH_low to above TH_high, and the alarm is deactivated when it falls from above TH_high to below TH_low, thereby suppressing alarm jitter near the threshold. For example, TH_low = 0.9 * TH_high.

[0096] In some embodiments of this application, the multidimensional signal fault detection method under low signal-to-noise ratio further includes: If the number of times the fault score continuously reaches the target threshold exceeds the target number, the data in the target operating data corresponding to the fault score will be suspended from participating in the determination of the target threshold.

[0097] In this embodiment, if the number of times the fault score reaches the target threshold consecutively exceeds the target number, it indicates a real fault. Therefore, the data corresponding to the fault score in the target operating data is suspended from participating in the determination of the target threshold, preventing the target threshold from being abnormally inflated and improving its stability. In some embodiments of this application, once the fault score recovers to below the target threshold and remains below the target duration, the corresponding data can be allowed to participate in the determination of the target threshold again.

[0098] In some embodiments of this application, the update rate of the target threshold is no greater than the target change rate (e.g., 5% / day), thereby preventing target threshold fluctuations and improving stability.

[0099] In some embodiments of this application, the data structure of the target time window can be: Segment S_t = { asset_id, sensor_group_id channels: {x_t^(1)...x_t^(C)}, / / C-dimensional channels fs: sampling_rate, t_start, t_end, rpm(optional), load(optional), temperature(optional), quality: {missing_rate, saturation_rate, clip_count, timestamp_jitter} } Recommended window length: T_win = 1s~10s (1–2s for rotating machinery; 5–10s for low speed), step size T_hop = 0.2s~1s.

[0100] Wherein, asset_id: unique identifier for the asset; sensor_group_id: unique identifier for the sensor group; t_start, t_end: start and end timestamps of the data segment; fs: sampling rate (unit: Hz); channels: C-dimensional channel data, in the format {x_t^(1)...x_t^(C)}; operating status (optional): rpm (speed), load (load), temperature (temperature); Quality: Data quality; missing_rate: Missing rate; saturation_rate: Saturation rate; clip_count: Clipping count; timestamp_jitter: Timestamp jitter.

[0101] The structure of a noise profile can be: NoiseProfile NP_t = { per_channel: [ {sigma_hat, noise_floor_psd(f), narrowband_lines: [(f0,bw,amp)],impulsive_noise_index} ], cross_channel: {coherence_baseline, common_mode_ratio}, stationarity: {spectral_flux, adf_pvalue(optional)}, snr_band: {band_k: SNR_hat_k} } Wherein, per channel: single-channel noise parameters; sigma hat: noise standard deviation estimate (power is its square); noise floor psd(f): noise floor power spectral density (varying with frequency); narrowband lines: narrowband interference lines (including frequency f0, bandwidth bw, amplitude amp); impulsive noise index: impulse noise index.

[0102] Cross channel: noise parameters across channels; Coherence baseline: coherence baseline between channels; Common mode ratio: common mode noise ratio; Stationarity: noise stationarity index; Spectral flux: spectral flux (measures changes in power spectral density); ADF pvalue (optional): p-value for autoregressive stationarity test (optional); SNRband: signal-to-noise ratio band; Band k: k-th band; SNR hat k: estimated signal-to-noise ratio for this band.

[0103] The structure of features and fractions can be: FeatureVector F_t = { features: {f1..fM}, meta: {selected_bands, denoise_method_id, params}, confidence: {conf_signal, conf_sync, conf_denoise} } Score Y_t = { anomaly_score, / / Overall fault score (higher score indicates more abnormality) sub_scores: {impact, modulation, energy_shift, cross_channel_inconsistency}, confidence condition_bin_id } Where FeatureVector F_t is the feature vector; Score Y_t is the score.

[0104] Features: Contains feature vector elements f1 to fM, which are the original feature values ​​of the signal or data; Meta: Metadata, recording selected bands (feature frequency bands), denoise method ID (denoising operator identifier), and params (parameter configuration); Confidence: Confidence index, including conf signal (signal confidence), conf sync (synchronization confidence), and conf denoise (denoising confidence).

[0105] Anomaly score: Overall fault score; a higher value indicates a more severe anomaly. Sub-scores: Sub-scores including impact, modulation, energy shift, and cross-channel inconsistency. Confidence: Score confidence level, reflecting the reliability of the results. Conditionbin ID: Condition bin identifier.

[0106] In some embodiments of this application, the sampling rate can be: Vibration: 6.4 kHz to 51.2 kHz (i.e., 6400 to 51200 Hz).

[0107] Current: 1 kHz to 20 kHz (i.e., 1000 to 20000 Hz, depending on the equipment / fault frequency band).

[0108] In some embodiments of this application, narrowband line determination is as follows: peak value exceeds the neighborhood median K_line by 8 to 15 times, and bandwidth bw is 0.5 to 5 Hz (or adaptive according to resolution). Controllable false alarm rate FAR_target: can be set according to device level (e.g., 10). - ³~10 -5 Each window).

[0109] EVT (Extreme Value Theory) switching conditions: The number of normal samples N per test chamber is ≥2000 and the distribution stability index is met (e.g., quantile drift <10% / week).

[0110] By applying the above technical solutions, the following technical effects can be achieved: By quantifying the noise type and intensity through noise profiling and driving the selection of denoising operators and parameter adaptation, the feature failure problem caused by fixed filtering / fixed parameters under low SNR is solved.

[0111] By using a closed-loop constraint of "denoising fidelity scoring" to constrain the denoising process, fault impact / modulation components are avoided from being erased as noise, thus preserving key fault information even at low SNR.

[0112] By using cross-channel consistency features (coherence / cross-spectrum / impact consistency) to enhance the detection of common evidence from multi-dimensional signals, weak faults that are not visible in a single channel can be stably identified after multi-dimensional fusion.

[0113] By establishing a threshold model based on operating conditions and noise levels, and using quantiles / EVT to generate a controllable false alarm rate threshold, the false alarm rate can be stably converged to a certain tolerance range within the controllable false alarm rate. This enables the threshold to be automatically adjusted according to changes in noise level and operating conditions, significantly reducing false alarms and missed alarms.

[0114] This application also proposes a computer device, such as... Figure 6 As shown, it includes a processor, a memory, and a computer program stored in the memory. The processor executes the computer program to implement the multi-dimensional signal fault detection method under low signal-to-noise ratio as described in any embodiment of this application.

[0115] The computer device in the embodiments of this application can be a terminal or other devices besides a terminal. For example, the computer device can be an industrial control computer, an edge computing gateway, an industrial server, an edge GPU server, a SCADA host computer, or a data acquisition card with processing capabilities, and can also be a server, a personal computer (PC), etc. The embodiments disclosed herein do not impose specific limitations.

[0116] The memory may include RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0117] The processors mentioned above can be general-purpose processors, including CPUs, NPs (Network Processors), etc.; they can also be DSPs (Digital Signal Processors), ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0118] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0119] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for multidimensional signal fault detection under low signal-to-noise ratio, characterized in that, include: The operating signals of the target device within the target time window are obtained through multiple acquisition channels. The operating signals correspond to multiple dimensions, and each dimension corresponds to one or more acquisition channels. Determine the noise data of the operating signal, and determine the noise profile of the operating signal based on the noise data; Based on the noise profile, a target denoising operator is determined from the set of denoising operators, and the target denoising operator is used to denoise the running signal to obtain a denoised signal. Fault scores are determined based on the target features of the denoised signals, wherein the target features include at least consistency enhancement features, and the consistency enhancement features characterize the degree of consistency of each of the denoised signals across multiple acquisition channels. If the fault score reaches the target threshold and the target alarm condition is met, a fault is determined to exist, and an alarm message is output.

2. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 1, characterized in that, The step of determining the target denoising operator from the set of denoising operators based on the noise profile includes: The running signal is denoised using multiple denoising operators in the set of denoising operators to obtain the denoised signal to be tested. The fidelity score of each of the denoised signals under test is determined based on the target fidelity index. The fidelity score characterizes the degree to which the denoised signal under test retains non-noise signals. The target fidelity index includes at least one of the following: band energy, impulse sparsity, and cross-channel consistency. The target denoising operator is determined from the set of denoising operators based on the fidelity score and each of the denoised signals to be tested.

3. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 2, characterized in that, The step of determining the target denoising operator from the set of denoising operators based on the fidelity score and each of the denoised signals to be tested includes: Determine the signal-to-noise ratio gain of each of the denoised signals to be tested; Based on the signal-to-noise ratio gain and fidelity score of each of the denoised signals under test, a comprehensive score for each of the denoised signals under test is determined. The target denoising operator is determined from the set of denoising operators based on the comprehensive scores.

4. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 1, characterized in that, Determining the fault score based on the target features of the denoised signal includes: The target features are robustly standardized to determine the standardized features; Multiple sub-scores are determined based on the standardized features, and the sub-scores include at least two of the following: impulse sub-score, modulation sub-score, energy drift sub-score, and consistency anomaly sub-score. The impulse sub-score is determined by envelope spectral peaks, spectral kurtosis, and impulse consistency. The modulation sub-score is determined by the ratio of sideband energy to order domain sidebands. The energy drift sub-score is determined by the offset of the target frequency band energy relative to the baseline. The consistency anomaly sub-score is determined by coherence degradation or impulse consistency anomaly. The fault score is determined by weighting and fusing the sub-scores.

5. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 1, characterized in that, The noise data used to determine the operating signal includes: The method determines the intra-channel noise signal of the operating signal within a single channel and the inter-channel noise signal of the operating signal across multiple channels. The intra-channel noise signal includes at least one of the following: noise variance estimation, noise floor power spectral density, narrowband interference, impulse noise figure, and frequency band signal-to-noise ratio estimation of the target frequency band. The inter-channel noise signal includes at least one of the following: inter-channel coherence ratio and inter-channel common-mode ratio. The noise data is determined based on the noise signal within the channel and the noise signal between the channels.

6. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 1, characterized in that, The process of determining the target threshold includes: Based on the operating conditions corresponding to the operating signal, a target operating condition box is determined from multiple operating condition boxes, and different operating condition boxes correspond to operating data under different operating conditions in the target time window; Determine the amount of target operating data corresponding to the target operating condition box; The target threshold is determined based on the amount of data and the controllable false alarm rate.

7. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 6, characterized in that, Determining the target threshold based on the data volume and the controllable false alarm rate includes: If the data volume does not reach the target data volume, the target quantile is determined based on the controllable false alarm rate; the target threshold is determined based on the target quantile of the target operating data. When the amount of data reaches the target amount of data, the target threshold is determined using extreme value theory based on the controllable false alarm rate and the target operating data.

8. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 6, characterized in that, The alarm information includes the confidence level and explanatory information of the fault score. The explanatory information includes at least one of the following: information on the target feature that contributes the most to the fault score, information on the target test chamber, and summary information of the noise profile. The confidence level is determined by at least one of the following: the data quality of the operating signal, the degree to which the denoised signal preserves non-noise signals, channel delay stability, or coherent baseline.

9. The multidimensional signal fault detection method under low signal-to-noise ratio as described in claim 6, characterized in that, Also includes: If the number of times the fault score continuously reaches the target threshold exceeds the target number, the data in the target operating data corresponding to the fault score will be suspended from participating in the determination of the target threshold.

10. A computer device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the multidimensional signal fault detection method under low signal-to-noise ratio as described in any one of claims 1-9.

Images