Fault diagnosis intelligent analysis method and system for industrial equipment and storage medium

Abstract

The invention provides a fault diagnosis intelligent analysis method and system for industrial equipment, and a storage medium. The method comprises the following steps: segmenting monitoring time sequence data into at least one sub-sequence based on statistical characteristics and working conditions of the monitoring time sequence data; determining the length of an embedded window based on the local characteristics of the subsequences, constructing a trajectory matrix by using the length of the embedded window, and performing singular value decomposition on the trajectory matrix to obtain a primitive matrix and a corresponding singular value; forming candidate groups for the primitive matrixes of different preset combinations, calculating a separability metric value of each candidate group, and determining a component group based on the separability metric values and the singular values; performing diagonal averaging reconstruction on the selected component group to obtain signal components, and calculating preset characteristic parameters of each signal component; and according to the preset feature parameters, target signal components related to the preset fault mode are screened out, feature vectors representing the fault state are extracted from the target signal components, the feature vectors are input into a pre-configured fault diagnosis model, and fault diagnosis information is output.

Description

Technical Field

[0001] This application belongs to the field of fault diagnosis, and particularly relates to an intelligent analysis method, system, and storage medium for fault diagnosis of industrial equipment. Background Art

[0002] During the long-term service of industrial equipment, various rotating machinery, reciprocating machinery, and complex equipment systems will inevitably generate various faults due to factors such as fatigue, wear, aging, or external impacts. If these faults are not diagnosed and processed in a timely and accurate manner, it will lead to a decline in equipment performance and an increase in energy consumption at best, and at worst, it will trigger catastrophic shutdown accidents, causing huge economic losses and even casualties. Traditional equipment maintenance strategies mostly rely on regular inspections or after-fact repairs. The former may lead to over-maintenance or under-maintenance, while the latter often comes with high fault repair costs and production interruption losses. Predictive maintenance based on condition monitoring analyzes various monitoring data generated during equipment operation, such as vibration, current, temperature, pressure, acoustic signals, etc., to evaluate the health status of the equipment in real time, and can diagnose and give early warnings at the early stage of fault occurrence.

[0003] Singular spectrum analysis is a non-parametric time series analysis method that does not require the establishment of an accurate physical model. It can adaptively decompose a one-dimensional time series into several additive signal components with clear physical meanings, and is particularly suitable for processing non-linear and non-stationary industrial vibration signals. Its basic principles include constructing a trajectory matrix, performing singular value decomposition (SVD) on the trajectory matrix, grouping the singular values and the corresponding primitive matrices, and reconstructing each component in the original signal through diagonal averaging. However, if the embedding window length is too small, it may not be able to effectively identify long-period components; if the window length is too large, it may smooth out short-time impact features or introduce too much noise, and the grouping lacks an objective optimization criterion, which easily leads to the mixing of useful signals and noise, affecting the accurate extraction of subsequent fault features. How to selectively screen out the components most relevant to specific fault modes from the numerous signal components decomposed by SSA is of great significance for improving its application efficiency in industrial equipment fault diagnosis. Summary of the Invention

[0004] Aiming at the technical problem of inaccurate identification caused by inaccurate windows and grouping in existing singular spectrum analysis, this application proposes an intelligent analysis method for fault diagnosis of industrial equipment, including:

[0005] Obtain the monitored time series data of the industrial equipment to be diagnosed, and divide the monitored time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitored time series data;

[0006] Determine the embedding window length based on the local characteristics of the subsequence, construct a trajectory matrix using the embedding window length, perform singular value decomposition on the trajectory matrix to obtain the primitive matrix and the corresponding singular values; form candidate groups for different preset combinations of the primitive matrix, calculate the separability metric values of each candidate group, and determine the component groups based on the separability metric values and the singular values; perform diagonal averaging reconstruction on the selected component groups to obtain signal components, and calculate the preset characteristic parameters of each signal component;

[0007] According to the preset characteristic parameters, screen out the target signal components related to the preset fault mode, extract the feature vectors characterizing the fault state from the target signal components, and input the feature vectors into a pre-configured fault diagnosis model to output fault diagnosis information.

[0008] Optionally, the step of splitting the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data includes:

[0009] Calculate the statistical characteristics of the monitoring time series data, including mean, variance, skewness, kurtosis, autocorrelation coefficient, and / or power spectral density;

[0010] According to the preset working condition rules, identify the working condition change points in the monitoring time series data, including equipment start / stop, load change, and speed change;

[0011] According to the statistical characteristics and the working condition change points, split the monitoring time series data into at least one subsequence.

[0012] Optionally, the step of determining the embedding window length based on the local characteristics of the subsequence includes:

[0013] Calculate the local characteristics of the subsequence, where the local characteristics include the preliminary estimation of the data fluctuation amplitude and the suspected periodic components;

[0014] Preliminarily determine the embedding window according to the data fluctuation amplitude, and adjust the preliminarily determined embedding window using the suspected periodic components to obtain the embedding window length L, where the value range of L is [2, N / 2], and N is the length of the subsequence.

[0015] Optionally, the step of calculating the separability metric values of each candidate group includes:

[0016] Calculate the separability metric index between the signal components in each candidate group, where the separability metric index is the cross-correlation coefficient, the frequency band overlap degree, or the mutual information.

[0017] Optionally, the step of determining the component groups based on the separability metric values and the singular values is specifically:

[0018] For each candidate group, calculate the sum of the squares of the singular values to obtain the total energy of the candidate group;

[0019] The weighted sum of the total energy and the separability metric value is calculated to obtain the score of the candidate group, and at least one candidate group with the largest score is used as the component group.

[0020] Optionally, the selected component group is reconstructed by diagonal averaging to obtain signal components, and preset characteristic parameters of each signal component are calculated, including:

[0021] The selected component group is reconstructed by diagonal averaging to obtain signal components;

[0022] Calculate the preset characteristic parameters of each signal component, and the preset characteristic parameter is the energy proportion or the entropy value; among them, the calculation method of the energy proportion is: the energy of the signal component divided by the total energy of all signal components.

[0023] Optionally, screening out the target signal components related to the preset fault mode includes:

[0024] According to the characteristics of the preset fault mode, set the screening conditions for the target signal components;

[0025] Calculate the preset characteristic parameters of each signal component and compare them with the screening conditions;

[0026] Screen out the signal components that meet the screening conditions as the target signal components.

[0027] Optionally, extracting the feature vector characterizing the fault state from the target signal components includes:

[0028] Extract time-domain features from the target signal components, and the time-domain features include mean value, variance, skewness, kurtosis, and autocorrelation coefficient;

[0029] Extract frequency-domain features from the target signal components, and the frequency-domain features include power spectral density, band energy, and band center frequency;

[0030] Combine the time-domain features and the frequency-domain features to form a feature vector, where the dimension of the feature vector is the sum of the dimensions of the time-domain features and the frequency-domain features.

[0031] This application also proposes a fault diagnosis intelligent analysis system for industrial equipment, including:

[0032] An acquisition unit for obtaining the monitoring time series data of the industrial equipment to be diagnosed, and dividing the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data;

[0033] A feature parameter determination unit, which is used to determine the embedding window length based on the local characteristics of the subsequence, construct a trajectory matrix by using the embedding window length, perform singular value decomposition on the trajectory matrix to obtain a primitive matrix and corresponding singular values; form candidate groups for different preset combinations of primitive matrices, calculate the separability metric values of each candidate group, and determine the component groups based on the separability metric values and singular values; perform diagonal averaging reconstruction on the selected component groups to obtain signal components, and calculate the preset feature parameters of each signal component;

[0034] An analysis unit, which is used to screen out target signal components related to preset fault modes according to the preset feature parameters, extract feature vectors representing the fault state from the target signal components, and input the feature vectors into a pre-configured fault diagnosis model to output fault diagnosis information.

[0035] Optionally, the splitting of the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data includes:

[0036] Calculating the statistical characteristics of the monitoring time series data, including mean, variance, skewness, kurtosis, autocorrelation coefficient and / or power spectral density;

[0037] Identifying the working condition change points in the monitoring time series data according to preset working condition rules, including equipment start / stop, load change, and speed change;

[0038] Splitting the monitoring time series data into at least one subsequence according to the statistical characteristics and working condition change points.

[0039] Optionally, the determining of the embedding window length based on the local characteristics of the subsequence includes:

[0040] Calculating the local characteristics of the subsequence, where the local characteristics include a preliminary estimate of the data fluctuation amplitude and suspected periodic components;

[0041] Preliminarily determining an embedding window according to the data fluctuation amplitude, and adjusting the preliminarily determined embedding window by using the suspected periodic components to obtain the embedding window length L, where the value range of L is [2, N / 2], and N is the length of the subsequence.

[0042] Optionally, the calculating of the separability metric values of each candidate group includes:

[0043] Calculating the separability metric index between signal components in each candidate group, and the separability metric index is the cross-correlation coefficient, frequency band overlap degree or mutual information.

[0044] Optionally, the determining of the component groups based on the separability metric values and singular values is specifically:

[0045] For each candidate group, calculate the sum of the squares of the singular values to obtain the total energy of the candidate group;

[0046] Perform a weighted sum of the total energy and the separability metric value to obtain the score of the candidate group, and use at least one candidate group with the largest score as the component group.

[0047] Optionally, perform diagonal averaging reconstruction on the selected component group to obtain signal components, and calculate preset characteristic parameters of each signal component, including:

[0048] Perform diagonal averaging reconstruction on the selected component group to obtain signal components;

[0049] Calculate preset characteristic parameters of each signal component, where the preset characteristic parameter is the energy proportion or the entropy value; among them, the calculation method of the energy proportion is: the energy of the signal component divided by the total energy of all signal components.

[0050] Optionally, screening out target signal components related to a preset fault mode includes:

[0051] According to the characteristics of the preset fault mode, set the screening conditions for the target signal components;

[0052] Calculate preset characteristic parameters of each signal component and compare them with the screening conditions;

[0053] Screen out the signal components that meet the screening conditions as the target signal components.

[0054] Optionally, extracting a feature vector representing the fault state from the target signal components includes:

[0055] Extract time-domain features from the target signal components, where the time-domain features include the mean value, variance, skewness, kurtosis, and autocorrelation coefficient;

[0056] Extract frequency-domain features from the target signal components, where the frequency-domain features include the power spectral density, band energy, and band center frequency;

[0057] Combine the time-domain features and the frequency-domain features to form a feature vector, where the dimension of the feature vector is the sum of the dimensions of the time-domain features and the frequency-domain features.

[0058] This application also proposes a storage medium, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the fault diagnosis intelligent analysis method in any of the above possible implementation manners.

[0059] By segmenting the monitored time series data based on statistical characteristics, preset working conditions, etc., and adaptively determining the embedding window length for the local characteristics of each subsequence, the present application can better adapt to the variability of signal characteristics of industrial equipment in different operating stages or different fault states, and avoid the limitations brought by global fixed parameters. Moreover, by introducing separability metrics and singular values to optimize the grouping process of the primitive matrix, the combination method that can achieve the optimal discrimination degree between the reconstructed signal components is selected, which improves the purity and accuracy of separating the useful signal components from background noise and other interferences. In addition, after reconstructing the signal components, according to the preset characteristic parameters such as the energy, periodicity, complexity, etc. of each component, and combining the prior knowledge of the preset fault modes, the target signal components most relevant to potential faults are selectively screened out, making the subsequent feature extraction and diagnostic analysis more focused and efficient, and enhancing the detection ability for specific fault modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flowchart of the first embodiment;

[0061] Figure 2 is a schematic diagram of the embedding window length;

[0062] Figure 3 is a schematic diagram of the sequence data;

[0063] Figure 4 is a schematic diagram of the singular values of the sequence data;

[0064] Figure 5 is a schematic diagram of the reconstructed components of the sequence data. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0066] In the description of the present application, the terms "first", "second" and their corresponding term numbers in the specification, claims and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, which is only a way of distinguishing objects with the same attributes when describing the embodiments of the present application. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device including a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these processes, methods, products or devices.

[0067] In addition, in the description of the present application, unless otherwise specified, the meaning of "a plurality" is two or more. The term "and / or" or the character " / " in the present application is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B, or A / B, which can represent: A exists alone, A and B exist simultaneously, and B exists alone.

[0068] Specific embodiments, a fault diagnosis intelligent analysis method for industrial equipment, as Figure 1 shown, includes:

[0069] S1, obtaining the monitored time series data of the industrial equipment to be diagnosed, and segmenting the monitored time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitored time series data;

[0070] The monitored time series data is sourced from various sensors installed at key parts of the industrial equipment, such as vibration sensors, temperature sensors, acoustic sensors, current sensors or pressure sensors, etc. The sensors can collect in real time or regularly the physical quantities reflecting the operating state of the equipment and convert them into digital time series signals. The data obtained can be single-channel signals or multi-channel synchronous signals from multiple sensors. Parameters such as the sampling frequency and recording duration of the data are reasonably set according to the characteristics of the equipment to be diagnosed and the typical frequency range of potential fault modes to ensure that sufficient fault information is captured. For example, for the bearing fault of high-speed rotating machinery, the data collected by high-frequency vibration sensors is required; for some slowly developing wear processes, the vibration or wear amount data over a longer time span is concerned.

[0071] The operating state of industrial equipment is not always constant. There are different operating conditions such as startup, shutdown, speed change, and load change. The analysis of statistical characteristics includes but is not limited to significant changes in indicators such as the mean, variance, energy, kurtosis, and information entropy of data segments. For example, Bayesian online change point detection, CUSUM algorithm, or hypothesis testing methods based on sliding windows are used to identify the time points when the statistical properties in the data stream change suddenly, so as to perform segmentation. Segmentation based on preset condition rules depends on the operating conditions obtained from the equipment control system such as the SCADA system, or thresholds related to certain measurable parameters such as rotational speed and load current set through expert knowledge. For example, when it is monitored that the rotational speed signal exceeds a certain preset value and remains stable for a period of time, it is considered that the equipment enters a certain specific operating condition, and the data in the corresponding time period is divided into a subsequence. In one embodiment, unsupervised clustering algorithms such as K-Means clustering based on the energy spectrum characteristics of the short-time Fourier transform are used to perform automatic pattern recognition and segmentation on continuous data segments.

[0072] S2. Determine the embedding window length based on the local characteristics of the subsequence, construct a trajectory matrix using the embedding window length, perform singular value decomposition on the trajectory matrix to obtain the primitive matrix and the corresponding singular values; form candidate groups for different preset combinations of the primitive matrix, calculate the separability metric values of each candidate group, and determine the component groups based on the separability metric values and singular values; perform diagonal averaging reconstruction on the selected component groups to obtain signal components, and calculate the preset characteristic parameters of each signal component.

[0073] In one embodiment, the embedding window length L is determined by analyzing the autocorrelation function of the subsequence, and finding the position of its first significant zero crossing or peak, which are often related to the main period or characteristic scale of the signal; or calculating the average mutual information of the subsequence. Map the one-dimensional time subsequence to a multi-dimensional trajectory matrix through windowing. Figure 2 Shows a schematic diagram of the data sequence and the window. Specifically, let the subsequence be , ,..., . After determining the embedding window length L, the i-th column of the trajectory matrix X is , and the dimension of the trajectory matrix is L×K, where K = N - L + 1.

[0074] Singular value decomposition SVD decomposes the trajectory matrix into , where are the singular values arranged in descending order, and are the corresponding left and right singular vectors respectively, is the primitive matrix or basic matrix, and d is the rank of the trajectory matrix. The magnitude of each singular value represents the importance of the corresponding primitive matrix in the original signal.Figure 3 is the original signal, Figure 4 is the singular value.

[0075] The original signal may contain various components of different types and scales, such as long-term trends, periodic fluctuations of different frequencies, noise, etc. After SVD decomposition, the energy and structural information of these components will be dispersed into different eigenmatrices. By trying different combinations of eigenmatrices, it is possible to better reconstruct a signal structure with practical significance and help separate the aliased components. There are various ways to determine the candidate groups, including but not limited to that the signal characteristics under certain working conditions are reflected by a specific subset of eigenmatrices, and the combination can be preset according to this information; or a random combination method can be used to form the candidate groups. In another embodiment, a preliminary division is performed based on the inflection points of the singular value spectrum, or several adjacent eigenmatrices are combined. In one embodiment, the separability metric value uses the similarity between different reconstructed components, Figure 5 shows different reconstructed components. A similarity value close to 1 indicates a high correlation between the two, which are grouped together, and a value close to 0 indicates no correlation, belonging to different groups. In another embodiment, distance correlation or a kernel-based method is used. When determining the component groups from multiple candidate groups, a method combining the separability metric value and the singular value is used. It is also possible to determine the component groups only according to the separability metric value, and select the group with the optimal metric value, so that the components within the group are highly correlated, while the components between the groups are as independent or distinguishable as possible.

[0076] Sum the eigenmatrices within the selected group, and then through a diagonal averaging process, convert the reconstructed matrix back to a one-dimensional time series signal component, and calculate the characteristic parameters of each reconstructed signal component. In one embodiment, the characteristic parameters are energy parameters, periodic parameters, and / or complexity parameters, etc. Among them, the energy parameter is the root mean square value, variance, etc. of the signal, and the periodic parameter extracts the main frequency, harmonic components, spectral peak amplitude, etc. by performing a fast Fourier transform on the signal component; the complexity parameter is permutation entropy, sample entropy, etc.

[0077] S3. According to the preset characteristic parameters, screen out the target signal components related to the preset fault mode, extract the feature vectors characterizing the fault state from the target signal components, and input the feature vectors into a pre-configured fault diagnosis model to output fault diagnosis information.

[0078] The screening methods include, but are not limited to, the feature descriptions of specific fault modes in the expert knowledge base. For example, if it is known that a certain fault will generate impact vibration within a specific frequency range, the screening condition is set as follows: the energy of the signal components is concentrated within this frequency range, and its kurtosis value or complexity index is significantly higher than that of the normal signal. In another embodiment, a rule-based inference system is used for screening, where the rule base contains the associations between different fault modes and the characteristic parameters of signal components; or a threshold method is used for screening. For example, the energy ratio of the target signal components exceeds a certain threshold, and at the same time its main frequency falls near the preset fault characteristic frequency. After screening out the target signal components, a set of numerical features characterizing the fault state of the device are extracted to form a feature vector. The feature vector is input into a fault diagnosis model to obtain fault diagnosis information, and the fault diagnosis model is a support vector machine, a neural network, etc.

[0079] In this embodiment, the monitoring data is segmented according to statistical characteristics and working conditions. By determining the embedding window length, constructing a trajectory matrix, performing singular value decomposition, component grouping and reconstruction, signal components are extracted and their characteristic parameters are calculated. The target signal components related to the preset faults are screened, fault feature vectors are extracted, and finally the results are output by inputting them into the diagnosis model. By segmenting the data to adapt to the non-stationarity of the device operation, using signal decomposition to extract information related to faults and obtain features, the pertinence and accuracy of the diagnosis are improved.

[0080] The monitoring data of industrial equipment often has significant differences under different working conditions such as start-stop, different loads, and different speeds. If the entire time series is analyzed without discrimination, the data characteristics under different working conditions will interfere with each other, which may mask the weak features of early faults or cause normal working condition fluctuations to be misjudged as faults. Through segmentation, subsequences with similar characteristics can be separated, enabling subsequent feature extraction and fault diagnosis models to more accurately model and analyze the data behavior under specific working conditions. In an alternative embodiment, segmenting the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data includes:

[0081] Calculating the statistical characteristics of the monitoring time series data, including mean, variance, skewness, kurtosis, autocorrelation coefficient, and / or power spectral density;

[0082] Identifying the working condition change points in the monitoring time series data according to the preset working condition rules, including equipment start-stop, load change, and speed change;

[0083] Segmenting the monitoring time series data into at least one subsequence according to the statistical characteristics and working condition change points.

[0084] Specifically, for each data segment within a window, calculate its key statistical characteristics, such as mean, variance / standard deviation, skewness, kurtosis, autocorrelation coefficient, power spectral density, etc. Moreover, systems such as SCADA, DCS, and PLC record the timestamps of operations such as equipment start / stop instructions, load setpoint changes, and speed setpoint changes, and these timestamps directly correspond to the operating condition change points. Obtain the time points when the equipment starts up and begins to shut down, the start and end times of switching from one stable level to another, the time when the speed adjustment starts and reaches the new stable speed, etc. The change in the operating condition means a fundamental change in the equipment operation mechanism. Take the identified operating condition change points as forced segmentation points. Then, within the data segments initially segmented by the operating condition change points, for example, apply a change point detection algorithm to the time series of statistical characteristics such as mean and variance. If a significant and continuous mutation in the statistical characteristics is detected, even if there is no externally recorded operating condition change, a segmentation point should be added here. When the variance calculated by the sliding window increases by more than a certain preset ratio within a short period of time, it is taken as a segmentation point.

[0085] The embedding window length is the key to constructing the trajectory matrix, directly affecting the quality and interpretability of the decomposition result. If the window length is too short, it may not be able to fully capture the features with longer periods or slower evolution in the signal, such as the weak signals of some early faults or the low-frequency vibration modes of equipment operation, resulting in information loss; conversely, if the window length is too long, it may smooth out the important details in the signal, or mix the signal components from different physical sources together, reducing the resolution of the components and the accuracy of subsequent analysis. In an optional embodiment, determining the embedding window length based on the local characteristics of the subsequence includes:

[0086] Calculate the local characteristics of the subsequence, where the local characteristics include a preliminary estimate of the data fluctuation amplitude and suspected periodic components;

[0087] Preliminarily determine the embedding window according to the data fluctuation amplitude, and adjust the initially determined embedding window using the suspected periodic components to obtain the embedding window length L, where the value range of L is [2, N / 2], and N is the length of the subsequence.

[0088] To evaluate the data fluctuation amplitude, a sliding window technique is adopted to calculate statistical indicators within a local region on the subsequence, such as variance, standard deviation, peak-to-peak value, or energy. For example, a relatively small detection window is set and slid on the subsequence to calculate the standard deviation sequence of the data within each small window, so as to understand the distribution of signal energy over time and the degree of intensity of change. For the preliminary estimation of suspected periodic components, preferably, the autocorrelation function (ACF) of the subsequence is calculated. Whether there are significant peaks in the ACF other than the zero-delay peak, and the delay times corresponding to these peaks may indicate the potential periodic length. In yet another embodiment, the fast Fourier transform of the subsequence is calculated to obtain its power spectral density (PSD), and frequency peaks with significant energy concentration are searched for on the power spectrum. The reciprocals of these frequency peaks also correspond to potential periods.

[0089] In the initial stage, a candidate range of the embedding window is initially set according to the general situation of the data fluctuation amplitude. For example, if the analysis shows that the subsequence contains a large number of high-frequency and rapidly changing components, a relatively small initial window length candidate value is preferably selected to capture these dynamics more precisely; conversely, if the subsequence shows gentle fluctuations and is mainly composed of low-frequency components, a larger initial window length candidate value is selected. Then, the initially set embedding window is adjusted using the estimated suspected periodic components. If one or more significant suspected periods are identified from the ACF or PSD, the embedding window length is set to an integer multiple of these periods, or at least ensure that L can cover the most dominant or longest periodic length among them. For example, L is equal to a certain longest period or twice the longest period, etc.; wherein, whether the significant suspected period is significant is determined according to the energy of the suspected period, for example, the energy is greater than the average value. Thus, the periodic components can be fully presented within the embedding window and effectively separated in the subsequent decomposition. The final embedding window length must also meet the constraint condition, that is, its value range is between [2, N / 2], where N is the length of the currently processed subsequence. In yet another embodiment, the adjustment method is to determine the set composed of integer multiples of the suspected period, and adjust the initial window length candidate value to the value closest to the value in the set. For example, if the suspected period is 2 and the set is [2, 4, 8], and the initial window is 3.6, then it jumps to 4.

[0090] To ensure that each signal component finally used for fault diagnosis has a high fault indication significance and minimized information redundancy. In an alternative embodiment, calculating the separability metric values of each candidate group includes:

[0091] Calculating the separability metric indicators between signal components in each candidate group, and the separability metric indicator is the cross-correlation coefficient, the frequency band overlap degree, or the mutual information.

[0092] After performing singular value decomposition (SVD) on the trajectory matrix of the subsequence, a series of primitive matrices sorted by the magnitudes of their corresponding singular values are obtained. Each primitive matrix consists of a singular value, the corresponding left singular vector, and right singular vector. These primitive matrices represent the energy and structure of the signal in different directions. Based on the magnitudes of the singular values, several primitive matrices with relatively large and close singular values are grouped together, and they jointly constitute a main signal component; or a primitive matrix with a significantly prominent single singular value is regarded as a component alone. Another approach is to attempt to group the first k (k takes different values) primitive matrices into one group, and the remaining ones into another group or multiple groups. It is also possible to aggregate primitive matrices with similar characteristics according to a preliminary judgment on the types of components that the signal may contain, such as trend terms, periodic terms of specific frequencies, noise, etc. Each candidate group represents a potential decomposition method for the original signal, and each sub-group reconstructs a specific signal component through diagonal averaging.

[0093] For each previously generated candidate group, quantitatively evaluate the independence between different signal components within it. Specifically, by calculating the Pearson cross-correlation coefficient between them in the time domain, the closer this value is to 0, the weaker the linear correlation between these two components in time, that is, the more separated they are; or calculate the power spectral density (PSD) of each signal component, and then, for any two components, analyze the overlapping degree of their power spectra on the frequency axis. The smaller the overlapping degree, the more distinct the components are in the frequency domain; mutual information can also be used. The smaller the mutual information value, the lower the statistical dependence between the two components, that is, the less helpful the knowledge of one component is in predicting the other component, and they are more separated.

[0094] Relying solely on separability metrics may select some components that are statistically independent but have little physical meaning or extremely low energy contribution. The singular value directly reflects the energy contribution or importance of each signal component in the original signal. By combining the two, select the component groups that can be clearly separated in signal structure and indeed carry the main information of the original signal, thereby improving the effectiveness of subsequent fault feature extraction and the accuracy of diagnostic results. In an optional embodiment, determining the component group based on the separability metric value and the singular value is specifically as follows:

[0095] For each candidate group, calculate the sum of the squares of the singular values to obtain the total energy of the candidate group;

[0096] Perform a weighted sum of the total energy and the separability metric value to obtain the score of the candidate group, and take at least one candidate group with the maximum score as the component group.

[0097] For each candidate group, calculate its total energy. This energy is obtained by summing the squares of the singular values corresponding to the primitive matrices within the group. To eliminate the influence of dimensions and make them comparable, normalization is performed, for example, calculating its proportion in the total energy of all singular values. It should be noted that some operations that may be involved in this application need to be normalized. Those skilled in the art know when normalization should be performed. For conventional operations such as normalization, they will not be elaborated, but the absence of description does not mean that there are no conventional operations such as normalization.

[0098] For each candidate group, a separability metric value has been pre-calculated, which is used to evaluate the degree of independence or distinguishability between the signal components within the group. For example, it is based on indicators such as the average of cross-correlation coefficients, frequency band non-overlap degree, or mutual information. The larger the value, the better the separability. To comprehensively consider the energy contribution and separability, a comprehensive score is calculated for each candidate group. This is achieved by performing a weighted sum of the normalized energy proportion and the separability metric value. Sort the scores from high to low. Finally, select one or more candidate groups with the highest scores as the finally determined component groups.

[0099] In an optional embodiment, the selected component group is subjected to diagonal averaging reconstruction to obtain signal components, and preset characteristic parameters of each signal component are calculated, including:

[0100] The selected component group is subjected to diagonal averaging reconstruction to obtain signal components;

[0101] Calculate the preset characteristic parameters of each signal component. The preset characteristic parameters are the energy proportion or entropy value; among them, the calculation method of the energy proportion is: the energy of the signal component divided by the total energy of all signal components.

[0102] A selected component group contains one or more primitive matrices. If a component group contains multiple primitive matrices, these primitive matrices are first summed to form a new matrix, and a one-dimensional time series is restored by averaging the elements on the anti-diagonal of this new matrix. Calculate the preset characteristic parameters of each reconstructed signal component, such as the energy proportion or entropy value. After calculating these preset characteristic parameters, each signal component is represented by a set of values, for example, its energy proportion, a certain entropy value, etc. These values constitute a part of the feature vector input to the fault diagnosis model subsequently.

[0103] After the monitoring signals of industrial equipment are decomposed, multiple signal components may be generated, and these components may correspond to the normal operating state of the equipment, different types of potential faults, background noise, or other environmental impacts. Not all components have the same value for the diagnosis of specific faults. By pre-defining the range or pattern of characteristic parameters related to known fault modes and applying them to the characteristic parameters of each calculated signal component, target-oriented component selection is achieved, ensuring that only those components most likely to carry information about the target fault are further processed, thus avoiding wasting computing resources on a large amount of irrelevant data and reducing the risk of normal fluctuations or non-target faults confusing the diagnostic results. In an alternative embodiment, screening out the target signal components related to the preset fault mode includes:

[0104] According to the characteristics of the preset fault mode, set the screening conditions for the target signal components;

[0105] Calculate the preset characteristic parameters of each signal component and compare them with the screening conditions;

[0106] Screen out the signal components that meet the screening conditions as the target signal components.

[0107] Determine each type or category of preset fault mode based on the prior knowledge of a specific industrial equipment and its common fault modes. A certain fault may cause a significant increase in the signal energy within a specific frequency range, then the energy ratio or the energy of a specific frequency band becomes a key characteristic parameter. Another fault may cause a decrease in the regularity and an increase in the complexity of the signal, then the sample entropy or permutation entropy may be more sensitive.

[0108] The screening conditions are based on threshold comparisons. For example, the energy ratio of a certain signal component is greater than , or its sample entropy is less than 0.1. It can also be based on pattern matching. For example, the power spectrum of a certain signal component shows an expected peak pattern in a specific frequency band, or it is based on the combined logic of multiple characteristic parameters. For example, high energy ratio AND high energy in a specific frequency band.

[0109] Calculate the preset characteristic parameters of each signal component obtained in the previous step, and compare these calculated characteristic parameter values with the screening conditions set in the previous step one by one, and finally screen out the signal components that meet all the specified conditions as the target signal components. After decomposing and reconstructing the original subsequence and calculating the preset characteristic parameters such as energy proportion and entropy value of each signal component, substitute these actually calculated characteristic parameter values into the screening conditions set for the specific fault mode for judgment. For each signal component, check whether its calculated characteristic parameters meet the screening conditions of the preset fault mode A. If so, the component is marked as a target signal component related to fault mode A. The same process will be carried out for preset fault modes B, C, etc. A signal component may meet the screening conditions of multiple fault modes at the same time, or it may not meet the conditions of any preset fault mode. Output a set of one or more target signal components, which are signal parts that are considered to be highly correlated with one or more preset fault modes of current concern. If no component meets the screening conditions, the preset fault characteristics are not shown in the current data.

[0110] The original target signal component itself is still time series data. Directly inputting it into the diagnostic model may cause problems such as high dimensionality, information redundancy, and insensitivity to small changes. By extracting time domain and frequency domain features such as mean, variance, power spectrum density, and frequency band energy, the essential information related to the fault state in the signal can be condensed into a few eigenvalues, which not only greatly reduces the data dimension but also improves the ability to distinguish. In an optional embodiment, the feature vector representing the fault state is extracted from the target signal component, including:

[0111] Extracting time domain features from target signal components, wherein the time domain features include mean, variance, skewness, kurtosis, and autocorrelation coefficient;

[0112] Extract frequency domain features from the target signal component, wherein the frequency domain features include power spectrum density, frequency band energy, and frequency band center frequency;

[0113] The time domain features and the frequency domain features are combined to form a feature vector, where the dimension of the feature vector is the sum of the dimensions of the time domain features and the frequency domain features.

[0114] After calculating the time domain features and frequency domain features, they are arranged and combined in a predetermined order to obtain a feature vector, which is provided as input to subsequent fault diagnosis models such as support vector machines, neural networks, decision trees, etc. for classification.

[0115] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some feature data can be ignored or not executed. Another point is that the displayed or discussed coupling, direct coupling, or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical, or other forms.

[0116] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0117] In addition, in each embodiment of the present application, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0118] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical disks and other various media that can store program codes.

Claims

1. An intelligent analysis method for fault diagnosis of industrial equipment, characterized in that, Including: Obtain the monitoring time series data of the industrial equipment to be diagnosed, and divide the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data; Determine the embedding window length based on the local characteristics of the subsequence, construct a trajectory matrix using the embedding window length, perform singular value decomposition on the trajectory matrix to obtain the primitive matrix and corresponding singular values; Form candidate groups for different preset combinations of primitive matrices, calculate the separability metric values of each candidate group, determine the component groups based on the separability metric values and singular values; perform diagonal averaging reconstruction on the selected component groups to obtain signal components, and calculate the preset characteristic parameters of each signal component; According to the preset characteristic parameters, screen out the target signal components related to the preset fault mode, extract the feature vectors representing the fault state from the target signal components, and input the feature vectors into the pre-configured fault diagnosis model to output fault diagnosis information.

2. The method according to claim 1, wherein The dividing the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data includes: Calculate the statistical characteristics of the monitoring time series data, including mean, variance, skewness, kurtosis, autocorrelation coefficient, and / or power spectral density; According to the preset working condition rules, identify the working condition change points in the monitoring time series data, including equipment start / stop, load change, and speed change; According to the statistical characteristics and working condition change points, divide the monitoring time series data into at least one subsequence.

3. The method according to claim 1, characterized in that The determining the embedding window length based on the local characteristics of the subsequence includes: Calculate the local characteristics of the subsequence, and the local characteristics include the preliminary estimation of the data fluctuation amplitude and suspected periodic components; Preliminarily determine the embedding window according to the data fluctuation amplitude, and adjust the preliminarily determined embedding window using the suspected periodic components to obtain the embedding window length L, where the value range of L is [2, N / 2], and N is the length of the subsequence.

4. The method according to claim 1, wherein The calculating the separability metric values of each candidate group includes: Calculate the separability metric index between the signal components in each candidate group, and the separability metric index is the cross-correlation coefficient, band overlap degree, or mutual information.

5. The method according to claim 1, wherein The determining the component groups based on the separability metric values and singular values is specifically: For each candidate group, calculate the sum of the squares of the singular values to obtain the total energy of the candidate group; Perform weighted summation on the total energy and the separability metric value to obtain the score of the candidate group, and take at least one candidate group with the maximum score as the component group.

6. The method according to claim 1, characterized in that, The performing diagonal averaging reconstruction on the selected component groups to obtain signal components and calculating the preset characteristic parameters of each signal component includes: Perform diagonal averaging reconstruction on the selected component groups to obtain signal components; Calculate the preset characteristic parameters of each signal component, and the preset characteristic parameters are the energy ratio or entropy value; among them, the calculation method of the energy ratio is: the energy of the signal component divided by the total energy of all signal components.

7. The method according to claim 1, characterized in that The screening out the target signal components related to the preset fault mode includes: Set the screening conditions for the target signal components according to the characteristics of the preset fault mode; Calculate the preset characteristic parameters of each signal component and compare them with the screening conditions; Filter out the signal components that meet the filtering conditions as the target signal components.

8. The method according to claim 1, wherein Extracting the feature vector characterizing the fault state from the target signal components includes: Extract time-domain features from the target signal components, where the time-domain features include mean, variance, skewness, kurtosis, and autocorrelation coefficient; Extract frequency-domain features from the target signal components, where the frequency-domain features include power spectral density, band energy, and band center frequency; Combine the time-domain features and the frequency-domain features to form a feature vector, where the dimension of the feature vector is the sum of the dimensions of the time-domain features and the frequency-domain features.

9. An intelligent analysis system for fault diagnosis of industrial equipment, characterized in that, It includes: An acquisition unit for obtaining the monitoring time series data of the industrial equipment to be diagnosed, and segmenting the monitoring time series data into at least one subsequence based on the statistical characteristics and working conditions of the monitoring time series data; A feature parameter determination unit for determining the embedding window length based on the local characteristics of the subsequence, constructing a trajectory matrix using the embedding window length, performing singular value decomposition on the trajectory matrix to obtain the primitive matrix and the corresponding singular values; Form candidate groups for different preset combinations of primitive matrices, calculate the separability metric values of each candidate group, determine the component groups based on the separability metric values and the singular values; perform diagonal averaging reconstruction on the selected component groups to obtain signal components, and calculate the preset feature parameters of each signal component; An analysis unit for screening out the target signal components related to the preset fault mode according to the preset feature parameters, extracting the feature vector characterizing the fault state from the target signal components, inputting the feature vector into a pre-configured fault diagnosis model, and outputting the fault diagnosis information.

10. A storage medium, characterized in that, At least one program code is stored in the storage medium, and the at least one program code is loaded and executed by a processor to implement the fault diagnosis intelligent analysis method according to any one of claims 1 to 8.

Images